it works

grr heap usage!

.github/workflows/gofmt.yml

@@ -0,0 +1,34 @@
+name: gofmt
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    paths:
+      - '.github/workflows/gofmt.yml'
+      - '**.go'
+jobs:
+
+  gofmt:
+    name: Run gofmt
+    runs-on: ubuntu-latest
+    steps:
+
+    - uses: actions/checkout@v6
+
+    - uses: actions/setup-go@v6
+      with:
+        go-version: '1.25'
+        check-latest: true
+
+    - name: Install goimports
+      run: |
+        go install golang.org/x/tools/cmd/goimports@latest
+
+    - name: gofmt
+      run: |
+        if [ "$(find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -l)" ]
+        then
+          find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -d
+          exit 1
+        fi

.github/workflows/release.yml

@@ -10,7 +10,7 @@ jobs:
     name: Build Linux/BSD All
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
 
       - uses: actions/setup-go@v6
         with:
@@ -36,7 +36,7 @@ jobs:
       id-token: write
       contents: read
     steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
 
       - uses: actions/setup-go@v6
         with:
@@ -76,7 +76,7 @@ jobs:
       HAS_SIGNING_CREDS: ${{ secrets.AC_USERNAME != '' }}
     runs-on: macos-latest
     steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
 
       - uses: actions/setup-go@v6
         with:
@@ -134,7 +134,7 @@ jobs:
       # be overwritten
       - name: Checkout code
         if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: actions/checkout@v7
+        uses: actions/checkout@v6
 
       - name: Download artifacts
         if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
@@ -163,17 +163,14 @@ jobs:
           mkdir -p build/linux-{amd64,arm64}
           tar -zxvf artifacts/nebula-linux-amd64.tar.gz -C build/linux-amd64/
           tar -zxvf artifacts/nebula-linux-arm64.tar.gz -C build/linux-arm64/
-          docker buildx build . --push -f docker/Dockerfile --platform linux/amd64,linux/arm64 \
-            --build-arg VERSION="${GITHUB_REF#refs/tags/v}" \
-            --build-arg REVISION="${GITHUB_SHA}" \
-            --tag "${DOCKER_IMAGE_REPO}:${DOCKER_IMAGE_TAG}" --tag "${DOCKER_IMAGE_REPO}:${GITHUB_REF#refs/tags/v}"
+          docker buildx build . --push -f docker/Dockerfile --platform linux/amd64,linux/arm64 --tag "${DOCKER_IMAGE_REPO}:${DOCKER_IMAGE_TAG}" --tag "${DOCKER_IMAGE_REPO}:${GITHUB_REF#refs/tags/v}"
 
   release:
     name: Create and Upload Release
     needs: [build-linux, build-darwin, build-windows]
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v7
+      - uses: actions/checkout@v6
 
       - name: Download artifacts
         uses: actions/download-artifact@v8

.github/workflows/smoke-extra.yml

@@ -30,7 +30,7 @@ jobs:
       VAGRANT_DEFAULT_PROVIDER: libvirt
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:
@@ -62,7 +62,7 @@ jobs:
       VAGRANT_DEFAULT_PROVIDER: virtualbox
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:
@@ -88,7 +88,7 @@ jobs:
     runs-on: windows-latest
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:

.github/workflows/smoke.yml

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:
@@ -36,14 +36,6 @@ jobs:
       working-directory: ./.github/workflows/smoke
       run: ./smoke.sh
 
-    - name: setup docker image ipv6
-      working-directory: ./.github/workflows/smoke
-      run: SMOKE_OVERLAY_IPV6=1 ./build.sh
-
-    - name: run smoke ipv6
-      working-directory: ./.github/workflows/smoke
-      run: SMOKE_OVERLAY_IPV6=1 ./smoke.sh
-
     - name: setup relay docker image
       working-directory: ./.github/workflows/smoke
       run: ./build-relay.sh

.github/workflows/smoke/build.sh

@@ -5,19 +5,6 @@ set -e -x
 rm -rf ./build
 mkdir ./build
 
-if [ "$SMOKE_OVERLAY_IPV6" ]
-then
-    LIGHTHOUSE_NIP="fd00:4242:0:0:0:ffff:c0a8:6401"
-    HOST2_NIP="fd00:4242:0:0:0:ffff:c0a8:6402"
-    HOST3_NIP="fd00:4242:0:0:0:ffff:c0a8:6403"
-    HOST4_NIP="fd00:4242:0:0:0:ffff:c0a8:6404"
-else
-    LIGHTHOUSE_NIP="192.168.100.1"
-    HOST2_NIP="192.168.100.2"
-    HOST3_NIP="192.168.100.3"
-    HOST4_NIP="192.168.100.4"
-fi
-
 # Smoke containers run on a dedicated docker network whose subnet is allocated
 # at smoke time, not known at build time. Configs are written with TEST-NET-3
 # placeholder IPs (RFC 5737) and smoke.sh / smoke-vagrant.sh / smoke-relay.sh
@@ -44,24 +31,24 @@ LIGHTHOUSE_IP="203.0.113.2"
         ../genconfig.sh >lighthouse1.yml
 
     HOST="host2" \
-        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
         ../genconfig.sh >host2.yml
 
     HOST="host3" \
-        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
         INBOUND='[{"port": "any", "proto": "icmp", "group": "lighthouse"}]' \
         ../genconfig.sh >host3.yml
 
     HOST="host4" \
-        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
         OUTBOUND='[{"port": "any", "proto": "icmp", "group": "lighthouse"}]' \
         ../genconfig.sh >host4.yml
 
     ../../../../nebula-cert ca -curve "${CURVE:-25519}" -name "Smoke Test"
-    ../../../../nebula-cert sign -name "lighthouse1" -groups "lighthouse,lighthouse1" -ip "$LIGHTHOUSE_NIP/24"
-    ../../../../nebula-cert sign -name "host2" -groups "host,host2" -ip "$HOST2_NIP/24"
-    ../../../../nebula-cert sign -name "host3" -groups "host,host3" -ip "$HOST3_NIP/24"
-    ../../../../nebula-cert sign -name "host4" -groups "host,host4" -ip "$HOST4_NIP/24"
+    ../../../../nebula-cert sign -name "lighthouse1" -groups "lighthouse,lighthouse1" -ip "192.168.100.1/24"
+    ../../../../nebula-cert sign -name "host2" -groups "host,host2" -ip "192.168.100.2/24"
+    ../../../../nebula-cert sign -name "host3" -groups "host,host3" -ip "192.168.100.3/24"
+    ../../../../nebula-cert sign -name "host4" -groups "host,host4" -ip "192.168.100.4/24"
 )
 
 docker build -t "nebula:${NAME:-smoke}" .

.github/workflows/smoke/smoke.sh

@@ -47,19 +47,6 @@ HOST2_IP="$PREFIX.3"
 HOST3_IP="$PREFIX.4"
 HOST4_IP="$PREFIX.5"
 
-if [ "$SMOKE_OVERLAY_IPV6" ]
-then
-    LIGHTHOUSE_NIP="fd00:4242:0:0:0:ffff:c0a8:6401"
-    HOST2_NIP="fd00:4242:0:0:0:ffff:c0a8:6402"
-    HOST3_NIP="fd00:4242:0:0:0:ffff:c0a8:6403"
-    HOST4_NIP="fd00:4242:0:0:0:ffff:c0a8:6404"
-else
-    LIGHTHOUSE_NIP="192.168.100.1"
-    HOST2_NIP="192.168.100.2"
-    HOST3_NIP="192.168.100.3"
-    HOST4_NIP="192.168.100.4"
-fi
-
 # Sed the placeholder TEST-NET-3 IPs in the host configs to the real ones.
 # build/lighthouse1.yml has no IPs to rewrite so it's skipped.
 for f in build/host2.yml build/host3.yml build/host4.yml; do
@@ -93,28 +80,28 @@ docker exec host3 tcpdump -i eth0 -q -w - -U 2>logs/host3.outside.log >logs/host
 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 2000 &
-docker exec host3 ncat -nklv 2000 &
-docker exec host4 ncat -e '/usr/bin/echo helloagainfromhost4' -nkluv 4000 &
-docker exec host2 ncat -e '/usr/bin/echo host2' -nkluv 3000 &
-docker exec host3 ncat -e '/usr/bin/echo host3' -nkluv 3000 &
+docker exec host2 ncat -nklv 0.0.0.0 2000 &
+docker exec host3 ncat -nklv 0.0.0.0 2000 &
+docker exec host4 ncat -e '/usr/bin/echo helloagainfromhost4' -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 &
 
 set +x
 echo
 echo " *** Testing ping from lighthouse1"
 echo
 set -x
-docker exec lighthouse1 ping -c1 $HOST2_NIP
-docker exec lighthouse1 ping -c1 $HOST3_NIP
+docker exec lighthouse1 ping -c1 192.168.100.2
+docker exec lighthouse1 ping -c1 192.168.100.3
 
 set +x
 echo
 echo " *** Testing ping from host2"
 echo
 set -x
-docker exec host2 ping -c1 $LIGHTHOUSE_NIP
+docker exec host2 ping -c1 192.168.100.1
 # Should fail because not allowed by host3 inbound firewall
-! docker exec host2 ping -c1 $HOST3_NIP -w5 || exit 1
+! docker exec host2 ping -c1 192.168.100.3 -w5 || exit 1
 
 set +x
 echo
@@ -122,34 +109,34 @@ echo " *** Testing ncat from host2"
 echo
 set -x
 # Should fail because not allowed by host3 inbound firewall
-! docker exec host2 ncat -nzv -w5 $HOST3_NIP 2000 || exit 1
-! docker exec host2 ncat -nzuv -w5 $HOST3_NIP 3000 | grep -q host3 || exit 1
+! docker exec host2 ncat -nzv -w5 192.168.100.3 2000 || exit 1
+! docker exec host2 ncat -nzuv -w5 192.168.100.3 3000 | grep -q host3 || exit 1
 
 set +x
 echo
 echo " *** Testing ping from host3"
 echo
 set -x
-docker exec host3 ping -c1 $LIGHTHOUSE_NIP
-docker exec host3 ping -c1 $HOST2_NIP
+docker exec host3 ping -c1 192.168.100.1
+docker exec host3 ping -c1 192.168.100.2
 
 set +x
 echo
 echo " *** Testing ncat from host3"
 echo
 set -x
-docker exec host3 ncat -nzv -w5 $HOST2_NIP 2000
-docker exec host3 ncat -nzuv -w5 $HOST2_NIP 3000 | grep -q host2
+docker exec host3 ncat -nzv -w5 192.168.100.2 2000
+docker exec host3 ncat -nzuv -w5 192.168.100.2 3000 | grep -q host2
 
 set +x
 echo
 echo " *** Testing ping from host4"
 echo
 set -x
-docker exec host4 ping -c1 $LIGHTHOUSE_NIP
+docker exec host4 ping -c1 192.168.100.1
 # Should fail because not allowed by host4 outbound firewall
-! docker exec host4 ping -c1 $HOST2_NIP -w5 || exit 1
-! docker exec host4 ping -c1 $HOST3_NIP -w5 || exit 1
+! docker exec host4 ping -c1 192.168.100.2 -w5 || exit 1
+! docker exec host4 ping -c1 192.168.100.3 -w5 || exit 1
 
 set +x
 echo
@@ -157,10 +144,10 @@ echo " *** Testing ncat from host4"
 echo
 set -x
 # Should fail because not allowed by host4 outbound firewall
-! docker exec host4 ncat -nzv -w5 $HOST2_NIP 2000 || exit 1
-! docker exec host4 ncat -nzv -w5 $HOST3_NIP 2000 || exit 1
-! docker exec host4 ncat -nzuv -w5 $HOST2_NIP 3000 | grep -q host2 || exit 1
-! docker exec host4 ncat -nzuv -w5 $HOST3_NIP 3000 | grep -q host3 || exit 1
+! docker exec host4 ncat -nzv -w5 192.168.100.2 2000 || exit 1
+! docker exec host4 ncat -nzv -w5 192.168.100.3 2000 || exit 1
+! docker exec host4 ncat -nzuv -w5 192.168.100.2 3000 | grep -q host2 || exit 1
+! docker exec host4 ncat -nzuv -w5 192.168.100.3 3000 | grep -q host3 || exit 1
 
 set +x
 echo
@@ -172,7 +159,7 @@ set -x
 # cannot initiate UDP to host2. Once host2 initiates a flow to host4:4000,
 # conntrack must let host4's listener reply on that flow. If it doesn't,
 # the echo back from host4 never reaches host2.
-docker exec host2 sh -c "(/usr/bin/echo host2; sleep 2) | ncat -nuv $HOST4_NIP 4000" | grep -q helloagainfromhost4
+docker exec host2 sh -c "(/usr/bin/echo host2; sleep 2) | ncat -nuv 192.168.100.4 4000" | grep -q helloagainfromhost4
 
 docker exec host4 sh -c 'kill 1'
 docker exec host3 sh -c 'kill 1'

.github/workflows/test.yml

@@ -13,28 +13,20 @@ on:
       - 'go.sum'
 jobs:
 
-  static:
-    name: Static checks
+  test-linux:
+    name: Build all and test on ubuntu-linux
     runs-on: ubuntu-latest
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:
         go-version: '1.25'
         check-latest: true
 
-    - name: Install goimports
-      run: go install golang.org/x/tools/cmd/goimports@latest
-
-    - name: gofmt
-      run: |
-        if [ "$(find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -l)" ]
-        then
-          find . -iname '*.go' | grep -v '\.pb\.go$' | xargs goimports -d
-          exit 1
-        fi
+    - name: Build
+      run: make all
 
     - name: Vet
       run: make vet
@@ -44,41 +36,27 @@ jobs:
       with:
         version: v2.5
 
-  test:
-    name: Test ${{ matrix.name }}
-    runs-on: ${{ matrix.os }}
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-        - name: linux
-          os: ubuntu-latest
-          build-cmd: go build ./cmd/nebula ./cmd/nebula-cert
-          test-cmd: make test
-          e2e-cmd: make e2evv
-        - name: linux-boringcrypto
-          os: ubuntu-latest
-          build-cmd: make bin-boringcrypto
-          test-cmd: make test-boringcrypto
-          e2e-cmd: make e2e GOEXPERIMENT=boringcrypto CGO_ENABLED=1 TEST_ENV="TEST_LOGS=1" TEST_FLAGS="-v -ldflags -checklinkname=0"
-        - name: linux-pkcs11
-          os: ubuntu-latest
-          build-cmd: make bin-pkcs11
-          test-cmd: make test-pkcs11
-          e2e-cmd: ''
-        - name: macos
-          os: macos-latest
-          build-cmd: go build ./cmd/nebula ./cmd/nebula-cert
-          test-cmd: make test
-          e2e-cmd: make e2evv
-        - name: windows
-          os: windows-latest
-          build-cmd: go build ./cmd/nebula ./cmd/nebula-cert
-          test-cmd: make test
-          e2e-cmd: make e2evv
+    - name: Test
+      run: make test
+
+    - name: End 2 end
+      run: make e2evv
+
+    - name: Build test mobile
+      run: make build-test-mobile
+
+    - uses: actions/upload-artifact@v7
+      with:
+        name: e2e packet flow linux-latest
+        path: e2e/mermaid/linux-latest
+        if-no-files-found: warn
+
+  test-linux-boringcrypto:
+    name: Build and test on linux with boringcrypto
+    runs-on: ubuntu-latest
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:
@@ -86,65 +64,69 @@ jobs:
         check-latest: true
 
     - name: Build
-      run: ${{ matrix.build-cmd }}
-
-    - name: Cross-build darwin-amd64
-      if: matrix.name == 'macos'
-      run: GOARCH=amd64 go build -o /tmp/nebula-amd64 ./cmd/nebula && GOARCH=amd64 go build -o /tmp/nebula-cert-amd64 ./cmd/nebula-cert
+      run: make bin-boringcrypto
 
     - name: Test
-      run: ${{ matrix.test-cmd }}
+      run: make test-boringcrypto
 
     - name: End 2 end
-      if: matrix.e2e-cmd != ''
-      run: ${{ matrix.e2e-cmd }}
+      run: make e2e GOEXPERIMENT=boringcrypto CGO_ENABLED=1 TEST_ENV="TEST_LOGS=1" TEST_FLAGS="-v -ldflags -checklinkname=0"
 
-    - uses: actions/upload-artifact@v7
-      if: matrix.e2e-cmd != '' && always()
-      with:
-        name: e2e packet flow ${{ matrix.name }}
-        path: e2e/mermaid/
-        if-no-files-found: warn
-
-  cross-build:
-    name: Cross-build ${{ matrix.name }}
+  test-linux-pkcs11:
+    name: Build and test on linux with pkcs11
     runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-        - {name: linux-arm,   make-target: all-cross-linux-arm}
-        - {name: linux-mips,  make-target: all-cross-linux-mips}
-        - {name: linux-other, make-target: all-cross-linux-other}
-        - {name: freebsd,     make-target: all-freebsd}
-        - {name: openbsd,     make-target: all-openbsd}
-        - {name: netbsd,      make-target: all-netbsd}
-        - {name: windows,     make-target: all-cross-windows}
-        - {name: mobile,      make-target: build-test-mobile}
     steps:
 
-    - uses: actions/checkout@v7
+    - uses: actions/checkout@v6
 
     - uses: actions/setup-go@v6
       with:
         go-version: '1.25'
         check-latest: true
 
-    - name: Build ${{ matrix.name }}
-      run: make -j"$(nproc)" ${{ matrix.make-target }}
+    - name: Build
+      run: make bin-pkcs11
 
-  finish:
-    name: CI status
-    if: always()
-    needs: [static, test, cross-build]
-    runs-on: ubuntu-latest
+    - name: Test
+      run: make test-pkcs11
+
+  test:
+    name: Build and test on ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [windows-latest, macos-latest]
     steps:
 
-    - name: Fail if any upstream job failed
-      if: contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled')
-      run: |
-        echo "upstream results: ${{ toJSON(needs) }}"
-        exit 1
+    - uses: actions/checkout@v6
 
-    - name: All upstream jobs passed
-      run: echo "ok"
+    - uses: actions/setup-go@v6
+      with:
+        go-version: '1.25'
+        check-latest: true
+
+    - name: Build nebula
+      run: go build ./cmd/nebula
+
+    - name: Build nebula-cert
+      run: go build ./cmd/nebula-cert
+
+    - name: Vet
+      run: make vet
+
+    - name: golangci-lint
+      uses: golangci/golangci-lint-action@v9
+      with:
+        version: v2.5
+
+    - name: Test
+      run: make test
+
+    - name: End 2 end
+      run: make e2evv
+
+    - uses: actions/upload-artifact@v7
+      with:
+        name: e2e packet flow ${{ matrix.os }}
+        path: e2e/mermaid/${{ matrix.os }}
+        if-no-files-found: warn

Makefile

@@ -60,18 +60,6 @@ ALL = $(ALL_LINUX) \
 	windows-amd64 \
 	windows-arm64
 
-# Cross-build shards used by .github/workflows/test.yml — same as ALL_*
-# but with the arch that has a native CI runner removed, so the cross-build
-# job is not duplicating coverage the native test jobs already give.
-ALL_CROSS_LINUX = $(filter-out linux-amd64,$(ALL_LINUX))
-
-# ALL_CROSS_LINUX further split into family sub-shards so each can run on
-# its own CI runner in parallel. Union of the three must equal
-# ALL_CROSS_LINUX; adding a new linux arch goes into the matching family.
-ALL_CROSS_LINUX_ARM   = linux-arm-5 linux-arm-6 linux-arm-7 linux-arm64
-ALL_CROSS_LINUX_MIPS  = linux-mips linux-mipsle linux-mips64 linux-mips64le linux-mips-softfloat
-ALL_CROSS_LINUX_OTHER = linux-386 linux-ppc64le linux-riscv64 linux-loong64
-
 e2e:
 	$(TEST_ENV) go test -tags=e2e_testing -count=1 $(TEST_FLAGS) ./e2e
 
@@ -94,35 +82,6 @@ DOCKER_BIN = build/linux-amd64/nebula build/linux-amd64/nebula-cert
 
 all: $(ALL:%=build/%/nebula) $(ALL:%=build/%/nebula-cert)
 
-all-linux: $(ALL_LINUX:%=build/%/nebula) $(ALL_LINUX:%=build/%/nebula-cert)
-
-all-freebsd: $(ALL_FREEBSD:%=build/%/nebula) $(ALL_FREEBSD:%=build/%/nebula-cert)
-
-all-openbsd: $(ALL_OPENBSD:%=build/%/nebula) $(ALL_OPENBSD:%=build/%/nebula-cert)
-
-all-netbsd: $(ALL_NETBSD:%=build/%/nebula) $(ALL_NETBSD:%=build/%/nebula-cert)
-
-all-darwin: build/darwin-amd64/nebula build/darwin-amd64/nebula-cert build/darwin-arm64/nebula build/darwin-arm64/nebula-cert
-
-all-windows: build/windows-amd64/nebula.exe build/windows-amd64/nebula-cert.exe build/windows-arm64/nebula.exe build/windows-arm64/nebula-cert.exe
-
-# CI cross-build shards. darwin-arm64 is covered by the native macos-latest
-# job; windows-amd64 is covered by the native windows-latest job; both are
-# omitted here to avoid building them a second time. darwin-amd64 stays in
-# all-cross-darwin because intel mac is only a labeled/master-time native
-# job, so PRs still need cross-build coverage for it.
-all-cross-linux: $(ALL_CROSS_LINUX:%=build/%/nebula) $(ALL_CROSS_LINUX:%=build/%/nebula-cert)
-
-all-cross-linux-arm:   $(ALL_CROSS_LINUX_ARM:%=build/%/nebula)   $(ALL_CROSS_LINUX_ARM:%=build/%/nebula-cert)
-
-all-cross-linux-mips:  $(ALL_CROSS_LINUX_MIPS:%=build/%/nebula)  $(ALL_CROSS_LINUX_MIPS:%=build/%/nebula-cert)
-
-all-cross-linux-other: $(ALL_CROSS_LINUX_OTHER:%=build/%/nebula) $(ALL_CROSS_LINUX_OTHER:%=build/%/nebula-cert)
-
-all-cross-darwin: build/darwin-amd64/nebula build/darwin-amd64/nebula-cert
-
-all-cross-windows: build/windows-arm64/nebula.exe build/windows-arm64/nebula-cert.exe
-
 docker: docker/linux-$(shell go env GOARCH)
 
 release: $(ALL:%=build/nebula-%.tar.gz)
@@ -268,9 +227,6 @@ smoke-relay-docker: bin-docker
 	cd .github/workflows/smoke/ && ./build-relay.sh
 	cd .github/workflows/smoke/ && ./smoke-relay.sh
 
-smoke-docker-ipv6: export SMOKE_OVERLAY_IPV6 = 1
-smoke-docker-ipv6: smoke-docker
-
 smoke-docker-race: BUILD_ARGS = -race
 smoke-docker-race: CGO_ENABLED = 1
 smoke-docker-race: smoke-docker
@@ -280,5 +236,5 @@ smoke-vagrant/%: bin-docker build/%/nebula
 	cd .github/workflows/smoke/ && ./smoke-vagrant.sh $*
 
 .FORCE:
-.PHONY: all all-linux all-freebsd all-openbsd all-netbsd all-darwin all-windows all-cross-linux all-cross-linux-arm all-cross-linux-mips all-cross-linux-other all-cross-darwin all-cross-windows bench bench-cpu bench-cpu-long bin build-test-mobile e2e e2ev e2evv e2evvv e2evvvv proto release service smoke-docker smoke-docker-race test test-cov-html smoke-vagrant/%
+.PHONY: bench bench-cpu bench-cpu-long bin build-test-mobile e2e e2ev e2evv e2evvv e2evvvv proto release service smoke-docker smoke-docker-race test test-cov-html smoke-vagrant/%
 .DEFAULT_GOAL := bin

cert/helper_test.go

@@ -13,12 +13,6 @@ import (
 	"golang.org/x/crypto/ed25519"
 )
 
-// testCertNow is the reference "now" used to derive default before/after times
-// in NewTestCaCert and NewTestCert. Holding it fixed for the lifetime of the
-// test binary keeps CA and leaf defaults aligned at the same second, so a leaf
-// signed with default times can never expire after its CA on a rounding race.
-var testCertNow = time.Now().Round(time.Second)
-
 // NewTestCaCert will create a new ca certificate
 func NewTestCaCert(version Version, curve Curve, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (Certificate, []byte, []byte, []byte) {
 	var err error
@@ -40,10 +34,10 @@ func NewTestCaCert(version Version, curve Curve, before, after time.Time, networ
 	}
 
 	if before.IsZero() {
-		before = testCertNow.Add(time.Second * -60)
+		before = time.Now().Add(time.Second * -60).Round(time.Second)
 	}
 	if after.IsZero() {
-		after = testCertNow.Add(time.Second * 60)
+		after = time.Now().Add(time.Second * 60).Round(time.Second)
 	}
 
 	t := &TBSCertificate{
@@ -76,11 +70,11 @@ func NewTestCaCert(version Version, curve Curve, before, after time.Time, networ
 // Expiry times are defaulted if you do not pass them in
 func NewTestCert(v Version, curve Curve, ca Certificate, key []byte, name string, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (Certificate, []byte, []byte, []byte) {
 	if before.IsZero() {
-		before = testCertNow.Add(time.Second * -60)
+		before = time.Now().Add(time.Second * -60).Round(time.Second)
 	}
 
 	if after.IsZero() {
-		after = testCertNow.Add(time.Second * 60)
+		after = time.Now().Add(time.Second * 60).Round(time.Second)
 	}
 
 	if len(networks) == 0 {

cert_test/cert.go

@@ -14,12 +14,6 @@ import (
 	"golang.org/x/crypto/ed25519"
 )
 
-// testCertNow is the reference "now" used to derive default before/after times
-// in NewTestCaCert and NewTestCert. Holding it fixed for the lifetime of the
-// test binary keeps CA and leaf defaults aligned at the same second, so a leaf
-// signed with default times can never expire after its CA on a rounding race.
-var testCertNow = time.Now().Round(time.Second)
-
 // NewTestCaCert will create a new ca certificate
 func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (cert.Certificate, []byte, []byte, []byte) {
 	var err error
@@ -41,10 +35,10 @@ func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Ti
 	}
 
 	if before.IsZero() {
-		before = testCertNow.Add(time.Second * -60)
+		before = time.Now().Add(time.Second * -60).Round(time.Second)
 	}
 	if after.IsZero() {
-		after = testCertNow.Add(time.Second * 60)
+		after = time.Now().Add(time.Second * 60).Round(time.Second)
 	}
 
 	t := &cert.TBSCertificate{
@@ -77,11 +71,11 @@ func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Ti
 // Expiry times are defaulted if you do not pass them in
 func NewTestCert(v cert.Version, curve cert.Curve, ca cert.Certificate, key []byte, name string, before, after time.Time, networks, unsafeNetworks []netip.Prefix, groups []string) (cert.Certificate, []byte, []byte, []byte) {
 	if before.IsZero() {
-		before = testCertNow.Add(time.Second * -60)
+		before = time.Now().Add(time.Second * -60).Round(time.Second)
 	}
 
 	if after.IsZero() {
-		after = testCertNow.Add(time.Second * 60)
+		after = time.Now().Add(time.Second * 60).Round(time.Second)
 	}
 
 	var pub, priv []byte

cmd/nebula-cert/ca.go

@@ -97,19 +97,6 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		if err = mustFlagString("out-key", cf.outKeyPath); err != nil {
 			return err
 		}
-	} else {
-		// out-key is meaningless under PKCS#11 because the private key never
-		// leaves the HSM; reject it so we never silently accept or claim a
-		// stdout slot for it.
-		outKeySet := false
-		cf.set.Visit(func(f *flag.Flag) {
-			if f.Name == "out-key" {
-				outKeySet = true
-			}
-		})
-		if outKeySet {
-			return newHelpErrorf("cannot set -out-key with -pkcs11")
-		}
 	}
 	if err := mustFlagString("out-crt", cf.outCertPath); err != nil {
 		return err
@@ -184,21 +171,12 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		}
 	}
 
-	var claims ioClaims
-	if err := reserveOutputs(&claims,
-		"out-key", *cf.outKeyPath,
-		"out-crt", *cf.outCertPath,
-		"out-qr", *cf.outQRPath,
-	); err != nil {
-		return err
-	}
-
 	var passphrase []byte
 	if !isP11 && *cf.encryption {
 		passphrase = []byte(os.Getenv("NEBULA_CA_PASSPHRASE"))
 		if len(passphrase) == 0 {
 			for i := 0; i < 5; i++ {
-				errOut.Write([]byte("Enter passphrase: "))
+				out.Write([]byte("Enter passphrase: "))
 				passphrase, err = pr.ReadPassword()
 
 				if err == ErrNoTerminal {
@@ -283,16 +261,14 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		Curve:          curve,
 	}
 
-	if !isP11 && !isStdio(*cf.outKeyPath) {
+	if !isP11 {
 		if _, err := os.Stat(*cf.outKeyPath); err == nil {
 			return fmt.Errorf("refusing to overwrite existing CA key: %s", *cf.outKeyPath)
 		}
 	}
 
-	if !isStdio(*cf.outCertPath) {
-		if _, err := os.Stat(*cf.outCertPath); err == nil {
-			return fmt.Errorf("refusing to overwrite existing CA cert: %s", *cf.outCertPath)
-		}
+	if _, err := os.Stat(*cf.outCertPath); err == nil {
+		return fmt.Errorf("refusing to overwrite existing CA cert: %s", *cf.outCertPath)
 	}
 
 	var c cert.Certificate
@@ -318,7 +294,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 			b = cert.MarshalSigningPrivateKeyToPEM(curve, rawPriv)
 		}
 
-		err = writeOutput(*cf.outKeyPath, b, 0600, out)
+		err = os.WriteFile(*cf.outKeyPath, b, 0600)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
@@ -329,7 +305,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		return fmt.Errorf("error while marshalling certificate: %s", err)
 	}
 
-	err = writeOutput(*cf.outCertPath, b, 0600, out)
+	err = os.WriteFile(*cf.outCertPath, b, 0600)
 	if err != nil {
 		return fmt.Errorf("error while writing out-crt: %s", err)
 	}
@@ -340,7 +316,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}
 
-		err = writeOutput(*cf.outQRPath, b, 0600, out)
+		err = os.WriteFile(*cf.outQRPath, b, 0600)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -356,7 +332,6 @@ func caSummary() string {
 func caHelp(out io.Writer) {
 	cf := newCaFlags()
 	out.Write([]byte("Usage of " + os.Args[0] + " " + caSummary() + "\n"))
-	out.Write([]byte(stdioHelpText))
 	cf.set.SetOutput(out)
 	cf.set.PrintDefaults()
 }

cmd/nebula-cert/ca_test.go

@@ -27,7 +27,6 @@ func Test_caHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" ca <flags>: create a self signed certificate authority\n"+
-			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -argon-iterations uint\n"+
 			"    \tOptional: Argon2 iterations parameter used for encrypted private key passphrase (default 1)\n"+
 			"  -argon-memory uint\n"+
@@ -85,7 +84,7 @@ func Test_ca(t *testing.T) {
 		err:      nil,
 	}
 
-	pwPromptEB := "Enter passphrase: "
+	pwPromptOb := "Enter passphrase: "
 
 	// required args
 	assertHelpError(t, ca(
@@ -169,8 +168,8 @@ func Test_ca(t *testing.T) {
 	eb.Reset()
 	args = []string{"-version", "1", "-encrypt", "-name", "test", "-duration", "100m", "-groups", "1,2,3,4,5", "-out-crt", crtF.Name(), "-out-key", keyF.Name()}
 	require.NoError(t, ca(args, ob, eb, testpw))
-	assert.Empty(t, ob.String())
-	assert.Equal(t, pwPromptEB, eb.String())
+	assert.Equal(t, pwPromptOb, ob.String())
+	assert.Empty(t, eb.String())
 
 	// test encrypted key with passphrase environment variable
 	os.Remove(keyF.Name())
@@ -208,8 +207,8 @@ func Test_ca(t *testing.T) {
 	eb.Reset()
 	args = []string{"-version", "1", "-encrypt", "-name", "test", "-duration", "100m", "-groups", "1,2,3,4,5", "-out-crt", crtF.Name(), "-out-key", keyF.Name()}
 	require.Error(t, ca(args, ob, eb, errpw))
-	assert.Empty(t, ob.String())
-	assert.Equal(t, pwPromptEB, eb.String())
+	assert.Equal(t, pwPromptOb, ob.String())
+	assert.Empty(t, eb.String())
 
 	// test when user fails to enter a password
 	os.Remove(keyF.Name())
@@ -218,8 +217,8 @@ func Test_ca(t *testing.T) {
 	eb.Reset()
 	args = []string{"-version", "1", "-encrypt", "-name", "test", "-duration", "100m", "-groups", "1,2,3,4,5", "-out-crt", crtF.Name(), "-out-key", keyF.Name()}
 	require.EqualError(t, ca(args, ob, eb, nopw), "no passphrase specified, remove -encrypt flag to write out-key in plaintext")
-	assert.Empty(t, ob.String())
-	assert.Equal(t, strings.Repeat(pwPromptEB, 5), eb.String()) // prompts 5 times before giving up
+	assert.Equal(t, strings.Repeat(pwPromptOb, 5), ob.String()) // prompts 5 times before giving up
+	assert.Empty(t, eb.String())
 
 	// create valid cert/key for overwrite tests
 	os.Remove(keyF.Name())
@@ -248,67 +247,3 @@ func Test_ca(t *testing.T) {
 	os.Remove(keyF.Name())
 
 }
-
-func Test_ca_stdio(t *testing.T) {
-	nopw := &StubPasswordReader{}
-
-	keyF, err := os.CreateTemp("", "ca.key")
-	require.NoError(t, err)
-	os.Remove(keyF.Name())
-	defer os.Remove(keyF.Name())
-
-	crtF, err := os.CreateTemp("", "ca.crt")
-	require.NoError(t, err)
-	os.Remove(crtF.Name())
-	defer os.Remove(crtF.Name())
-
-	// out-crt on stdout, out-key on disk
-	ob := &bytes.Buffer{}
-	eb := &bytes.Buffer{}
-	require.NoError(t, ca([]string{"-name", "test-ca", "-duration", "1h", "-out-crt", "-", "-out-key", keyF.Name()}, ob, eb, nopw))
-	assert.Empty(t, eb.String())
-	c, _, err := cert.UnmarshalCertificateFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.True(t, c.IsCA())
-	assert.Equal(t, "test-ca", c.Name())
-
-	// out-key on stdout, out-crt on disk
-	os.Remove(keyF.Name())
-	ob.Reset()
-	eb.Reset()
-	require.NoError(t, ca([]string{"-name", "test-ca", "-duration", "1h", "-out-crt", crtF.Name(), "-out-key", "-"}, ob, eb, nopw))
-	assert.Empty(t, eb.String())
-	_, _, curve, err := cert.UnmarshalSigningPrivateKeyFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.Equal(t, cert.Curve_CURVE25519, curve)
-
-	// dual stdout is rejected up front
-	os.Remove(crtF.Name())
-	ob.Reset()
-	eb.Reset()
-	require.EqualError(t,
-		ca([]string{"-name", "test-ca", "-duration", "1h", "-out-crt", "-", "-out-key", "-"}, ob, eb, nopw),
-		`-out-key and -out-crt both set to "-", only one output may write to stdout`)
-	assert.Empty(t, ob.String())
-
-	// an output conflict combined with -encrypt must error BEFORE prompting
-	// for a passphrase; pr would record any read attempt
-	tracker := &trackingPasswordReader{}
-	ob.Reset()
-	eb.Reset()
-	require.EqualError(t,
-		ca([]string{"-name", "test-ca", "-duration", "1h", "-encrypt", "-out-crt", "-", "-out-key", "-"}, ob, eb, tracker),
-		`-out-key and -out-crt both set to "-", only one output may write to stdout`)
-	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
-	assert.Zero(t, tracker.calls, "passphrase prompt should not have been called")
-}
-
-type trackingPasswordReader struct {
-	calls int
-}
-
-func (pr *trackingPasswordReader) ReadPassword() ([]byte, error) {
-	pr.calls++
-	return []byte(""), nil
-}

cmd/nebula-cert/keygen.go

@@ -42,8 +42,6 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 		if err = mustFlagString("out-key", cf.outKeyPath); err != nil {
 			return err
 		}
-	} else if *cf.outKeyPath != "" {
-		return newHelpErrorf("cannot set -out-key with -pkcs11")
 	}
 	if err = mustFlagString("out-pub", cf.outPubPath); err != nil {
 		return err
@@ -71,14 +69,6 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 		}
 	}
 
-	var claims ioClaims
-	if err := reserveOutputs(&claims,
-		"out-key", *cf.outKeyPath,
-		"out-pub", *cf.outPubPath,
-	); err != nil {
-		return err
-	}
-
 	if isP11 {
 		p11Client, err := pkclient.FromUrl(*cf.p11url)
 		if err != nil {
@@ -92,12 +82,12 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while getting public key: %w", err)
 		}
 	} else {
-		err = writeOutput(*cf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600, out)
+		err = os.WriteFile(*cf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
 	}
-	err = writeOutput(*cf.outPubPath, cert.MarshalPublicKeyToPEM(curve, pub), 0600, out)
+	err = os.WriteFile(*cf.outPubPath, cert.MarshalPublicKeyToPEM(curve, pub), 0600)
 	if err != nil {
 		return fmt.Errorf("error while writing out-pub: %s", err)
 	}
@@ -112,7 +102,6 @@ func keygenSummary() string {
 func keygenHelp(out io.Writer) {
 	cf := newKeygenFlags()
 	_, _ = out.Write([]byte("Usage of " + os.Args[0] + " " + keygenSummary() + "\n"))
-	_, _ = out.Write([]byte(stdioHelpText))
 	cf.set.SetOutput(out)
 	cf.set.PrintDefaults()
 }

cmd/nebula-cert/keygen_test.go

@@ -20,7 +20,6 @@ func Test_keygenHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" keygen <flags>: create a public/private key pair. the public key can be passed to `nebula-cert sign`\n"+
-			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -curve string\n"+
 			"    \tECDH Curve (25519, P256) (default \"25519\")\n"+
 			"  -out-key string\n"+
@@ -94,43 +93,3 @@ func Test_keygen(t *testing.T) {
 	require.NoError(t, err)
 	assert.Len(t, lPub, 32)
 }
-
-func Test_keygen_stdio(t *testing.T) {
-	keyF, err := os.CreateTemp("", "test.key")
-	require.NoError(t, err)
-	os.Remove(keyF.Name())
-	defer os.Remove(keyF.Name())
-
-	pubF, err := os.CreateTemp("", "test.pub")
-	require.NoError(t, err)
-	os.Remove(pubF.Name())
-	defer os.Remove(pubF.Name())
-
-	// out-pub on stdout, out-key on disk
-	ob := &bytes.Buffer{}
-	eb := &bytes.Buffer{}
-	require.NoError(t, keygen([]string{"-out-pub", "-", "-out-key", keyF.Name()}, ob, eb))
-	assert.Empty(t, eb.String())
-	lPub, _, curve, err := cert.UnmarshalPublicKeyFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.Equal(t, cert.Curve_CURVE25519, curve)
-	assert.Len(t, lPub, 32)
-
-	// out-key on stdout, out-pub on disk
-	os.Remove(keyF.Name())
-	ob.Reset()
-	eb.Reset()
-	require.NoError(t, keygen([]string{"-out-pub", pubF.Name(), "-out-key", "-"}, ob, eb))
-	assert.Empty(t, eb.String())
-	lKey, _, curve, err := cert.UnmarshalPrivateKeyFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.Equal(t, cert.Curve_CURVE25519, curve)
-	assert.Len(t, lKey, 32)
-
-	// both on stdout is a conflict caught up front
-	ob.Reset()
-	eb.Reset()
-	require.EqualError(t, keygen([]string{"-out-pub", "-", "-out-key", "-"}, ob, eb),
-		`-out-key and -out-pub both set to "-", only one output may write to stdout`)
-	assert.Empty(t, ob.String())
-}

cmd/nebula-cert/passwords.go

@@ -22,9 +22,7 @@ func (pr StdinPasswordReader) ReadPassword() ([]byte, error) {
 	}
 
 	password, err := term.ReadPassword(int(os.Stdin.Fd()))
-	// Terminal echo is off while reading, so the user's Enter key does not
-	// produce a visible newline. Emit one on stderr to match the prompt.
-	fmt.Fprintln(os.Stderr)
+	fmt.Println()
 
 	return password, err
 }

cmd/nebula-cert/print.go

@@ -40,23 +40,11 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}
 
-	var claims ioClaims
-	if err := reserveInputs(&claims, "path", *pf.path); err != nil {
-		return err
-	}
-	if err := reserveOutputs(&claims, "out-qr", *pf.outQRPath); err != nil {
-		return err
-	}
-
-	rawCert, err := readInput("path", *pf.path, &claims)
+	rawCert, err := os.ReadFile(*pf.path)
 	if err != nil {
 		return fmt.Errorf("unable to read cert; %s", err)
 	}
 
-	// When the QR is going to stdout, suppress the human-readable text/json
-	// output so the binary stream is not contaminated.
-	qrToStdout := isStdio(*pf.outQRPath)
-
 	var c cert.Certificate
 	var qrBytes []byte
 	part := 0
@@ -69,13 +57,11 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while unmarshaling cert: %s", err)
 		}
 
-		if !qrToStdout {
-			if *pf.json {
-				jsonCerts = append(jsonCerts, c)
-			} else {
-				_, _ = out.Write([]byte(c.String()))
-				_, _ = out.Write([]byte("\n"))
-			}
+		if *pf.json {
+			jsonCerts = append(jsonCerts, c)
+		} else {
+			_, _ = out.Write([]byte(c.String()))
+			_, _ = out.Write([]byte("\n"))
 		}
 
 		if *pf.outQRPath != "" {
@@ -93,7 +79,7 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 		part++
 	}
 
-	if *pf.json && !qrToStdout {
+	if *pf.json {
 		b, _ := json.Marshal(jsonCerts)
 		_, _ = out.Write(b)
 		_, _ = out.Write([]byte("\n"))
@@ -105,7 +91,7 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}
 
-		err = writeOutput(*pf.outQRPath, b, 0600, out)
+		err = os.WriteFile(*pf.outQRPath, b, 0600)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -121,7 +107,6 @@ func printSummary() string {
 func printHelp(out io.Writer) {
 	pf := newPrintFlags()
 	out.Write([]byte("Usage of " + os.Args[0] + " " + printSummary() + "\n"))
-	out.Write([]byte(stdioHelpText))
 	pf.set.SetOutput(out)
 	pf.set.PrintDefaults()
 }

cmd/nebula-cert/print_test.go

@@ -25,7 +25,6 @@ func Test_printHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" print <flags>: prints details about a certificate\n"+
-			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -json\n"+
 			"    \tOptional: outputs certificates in json format\n"+
 			"  -out-qr string\n"+
@@ -179,44 +178,6 @@ func Test_printCert(t *testing.T) {
 		ob.String(),
 	)
 	assert.Empty(t, eb.String())
-
-	// read cert from stdin
-	ob.Reset()
-	eb.Reset()
-	withStdin(t, bytes.NewReader(p))
-	err = printCert([]string{"-json", "-path", "-"}, ob, eb)
-	require.NoError(t, err)
-	assert.Equal(
-		t,
-		`[{"details":{"curve":"CURVE25519","groups":["hi"],"isCa":false,"issuer":"`+c.Issuer()+`","name":"test","networks":["10.0.0.123/8"],"notAfter":"0001-01-01T00:00:00Z","notBefore":"0001-01-01T00:00:00Z","publicKey":"`+pk+`","unsafeNetworks":[]},"fingerprint":"`+fp+`","signature":"`+sig+`","version":1}]
-`,
-		ob.String(),
-	)
-	assert.Empty(t, eb.String())
-
-	// -out-qr - sends only the PNG to stdout, suppressing the cert dump
-	ob.Reset()
-	eb.Reset()
-	withStdin(t, bytes.NewReader(p))
-	err = printCert([]string{"-path", "-", "-out-qr", "-"}, ob, eb)
-	require.NoError(t, err)
-	assert.Empty(t, eb.String())
-	stdout := ob.Bytes()
-	require.NotEmpty(t, stdout)
-	// PNG magic, no PEM/JSON noise prepended
-	assert.Equal(t, []byte{0x89, 'P', 'N', 'G', 0x0d, 0x0a, 0x1a, 0x0a}, stdout[:8])
-	assert.NotContains(t, string(stdout), "NebulaCertificate")
-	assert.NotContains(t, string(stdout), `"details"`)
-
-	// json + out-qr - still suppresses json
-	ob.Reset()
-	eb.Reset()
-	withStdin(t, bytes.NewReader(p))
-	err = printCert([]string{"-json", "-path", "-", "-out-qr", "-"}, ob, eb)
-	require.NoError(t, err)
-	assert.Empty(t, eb.String())
-	assert.Equal(t, []byte{0x89, 'P', 'N', 'G'}, ob.Bytes()[:4])
-	assert.NotContains(t, ob.String(), `"details"`)
 }
 
 // NewTestCaCert will generate a CA cert

cmd/nebula-cert/sign.go

@@ -85,9 +85,6 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 	if !isP11 && *sf.inPubPath != "" && *sf.outKeyPath != "" {
 		return newHelpErrorf("cannot set both -in-pub and -out-key")
 	}
-	if isP11 && *sf.outKeyPath != "" {
-		return newHelpErrorf("cannot set -out-key with -pkcs11")
-	}
 
 	var v4Networks []netip.Prefix
 	var v6Networks []netip.Prefix
@@ -105,35 +102,13 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		return newHelpErrorf("-version must be either %v or %v", cert.Version1, cert.Version2)
 	}
 
-	if *sf.outKeyPath == "" {
-		*sf.outKeyPath = *sf.name + ".key"
-	}
-	if *sf.outCertPath == "" {
-		*sf.outCertPath = *sf.name + ".crt"
-	}
-
-	var claims ioClaims
-	if err := reserveInputs(&claims,
-		"ca-key", *sf.caKeyPath,
-		"ca-crt", *sf.caCertPath,
-		"in-pub", *sf.inPubPath,
-	); err != nil {
-		return err
-	}
-	if err := reserveOutputs(&claims,
-		"out-key", *sf.outKeyPath,
-		"out-crt", *sf.outCertPath,
-		"out-qr", *sf.outQRPath,
-	); err != nil {
-		return err
-	}
-
 	var curve cert.Curve
 	var caKey []byte
 
 	if !isP11 {
 		var rawCAKey []byte
-		rawCAKey, err = readInput("ca-key", *sf.caKeyPath, &claims)
+		rawCAKey, err := os.ReadFile(*sf.caKeyPath)
+
 		if err != nil {
 			return fmt.Errorf("error while reading ca-key: %s", err)
 		}
@@ -146,7 +121,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 			if len(passphrase) == 0 {
 				// ask for a passphrase until we get one
 				for i := 0; i < 5; i++ {
-					errOut.Write([]byte("Enter passphrase: "))
+					out.Write([]byte("Enter passphrase: "))
 					passphrase, err = pr.ReadPassword()
 
 					if errors.Is(err, ErrNoTerminal) {
@@ -172,7 +147,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		}
 	}
 
-	rawCACert, err := readInput("ca-crt", *sf.caCertPath, &claims)
+	rawCACert, err := os.ReadFile(*sf.caCertPath)
 	if err != nil {
 		return fmt.Errorf("error while reading ca-crt: %s", err)
 	}
@@ -270,7 +245,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 
 	if *sf.inPubPath != "" {
 		var pubCurve cert.Curve
-		rawPub, err := readInput("in-pub", *sf.inPubPath, &claims)
+		rawPub, err := os.ReadFile(*sf.inPubPath)
 		if err != nil {
 			return fmt.Errorf("error while reading in-pub: %s", err)
 		}
@@ -291,10 +266,16 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		pub, rawPriv = newKeypair(curve)
 	}
 
-	if !isStdio(*sf.outCertPath) {
-		if _, err := os.Stat(*sf.outCertPath); err == nil {
-			return fmt.Errorf("refusing to overwrite existing cert: %s", *sf.outCertPath)
-		}
+	if *sf.outKeyPath == "" {
+		*sf.outKeyPath = *sf.name + ".key"
+	}
+
+	if *sf.outCertPath == "" {
+		*sf.outCertPath = *sf.name + ".crt"
+	}
+
+	if _, err := os.Stat(*sf.outCertPath); err == nil {
+		return fmt.Errorf("refusing to overwrite existing cert: %s", *sf.outCertPath)
 	}
 
 	var crts []cert.Certificate
@@ -379,13 +360,11 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 	}
 
 	if !isP11 && *sf.inPubPath == "" {
-		if !isStdio(*sf.outKeyPath) {
-			if _, err := os.Stat(*sf.outKeyPath); err == nil {
-				return fmt.Errorf("refusing to overwrite existing key: %s", *sf.outKeyPath)
-			}
+		if _, err := os.Stat(*sf.outKeyPath); err == nil {
+			return fmt.Errorf("refusing to overwrite existing key: %s", *sf.outKeyPath)
 		}
 
-		err = writeOutput(*sf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600, out)
+		err = os.WriteFile(*sf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
@@ -400,7 +379,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		b = append(b, sb...)
 	}
 
-	err = writeOutput(*sf.outCertPath, b, 0600, out)
+	err = os.WriteFile(*sf.outCertPath, b, 0600)
 	if err != nil {
 		return fmt.Errorf("error while writing out-crt: %s", err)
 	}
@@ -411,7 +390,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}
 
-		err = writeOutput(*sf.outQRPath, b, 0600, out)
+		err = os.WriteFile(*sf.outQRPath, b, 0600)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -461,7 +440,6 @@ func signSummary() string {
 func signHelp(out io.Writer) {
 	sf := newSignFlags()
 	out.Write([]byte("Usage of " + os.Args[0] + " " + signSummary() + "\n"))
-	out.Write([]byte(stdioHelpText))
 	sf.set.SetOutput(out)
 	sf.set.PrintDefaults()
 }

cmd/nebula-cert/sign_test.go

@@ -27,7 +27,6 @@ func Test_signHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" sign <flags>: create and sign a certificate\n"+
-			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -ca-crt string\n"+
 			"    \tOptional: path to the signing CA cert (default \"ca.crt\")\n"+
 			"  -ca-key string\n"+
@@ -377,18 +376,15 @@ func Test_signCert(t *testing.T) {
 	// test with the proper password
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.NoError(t, signCert(args, ob, eb, testpw))
-	assert.Empty(t, ob.String())
-	assert.Equal(t, "Enter passphrase: ", eb.String())
+	assert.Equal(t, "Enter passphrase: ", ob.String())
+	assert.Empty(t, eb.String())
 
 	// test with the proper password in the environment
 	os.Remove(crtF.Name())
 	os.Remove(keyF.Name())
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	os.Setenv("NEBULA_CA_PASSPHRASE", string(passphrase))
-	ob.Reset()
-	eb.Reset()
 	require.NoError(t, signCert(args, ob, eb, testpw))
-	assert.Empty(t, ob.String())
 	assert.Empty(t, eb.String())
 	os.Setenv("NEBULA_CA_PASSPHRASE", "")
 
@@ -399,8 +395,8 @@ func Test_signCert(t *testing.T) {
 	testpw.password = []byte("invalid password")
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.Error(t, signCert(args, ob, eb, testpw))
-	assert.Empty(t, ob.String())
-	assert.Equal(t, "Enter passphrase: ", eb.String())
+	assert.Equal(t, "Enter passphrase: ", ob.String())
+	assert.Empty(t, eb.String())
 
 	// test with the wrong password in environment
 	ob.Reset()
@@ -420,8 +416,8 @@ func Test_signCert(t *testing.T) {
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.Error(t, signCert(args, ob, eb, nopw))
 	// normally the user hitting enter on the prompt would add newlines between these
-	assert.Empty(t, ob.String())
-	assert.Equal(t, "Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: ", eb.String())
+	assert.Equal(t, "Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: ", ob.String())
+	assert.Empty(t, eb.String())
 
 	// test an error condition
 	ob.Reset()
@@ -429,106 +425,6 @@ func Test_signCert(t *testing.T) {
 
 	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", keyF.Name(), "-duration", "100m", "-subnets", "10.1.1.1/32, ,   10.2.2.2/32   ,   ,  ,, 10.5.5.5/32", "-groups", "1,,   2    ,        ,,,3,4,5"}
 	require.Error(t, signCert(args, ob, eb, errpw))
-	assert.Empty(t, ob.String())
-	assert.Equal(t, "Enter passphrase: ", eb.String())
-}
-
-func Test_signCert_stdio(t *testing.T) {
-	nopw := &StubPasswordReader{
-		password: []byte(""),
-		err:      nil,
-	}
-
-	caPub, caPriv, _ := ed25519.GenerateKey(rand.Reader)
-	rawCAKey := cert.MarshalSigningPrivateKeyToPEM(cert.Curve_CURVE25519, caPriv)
-
-	ca, _ := NewTestCaCert("ca", caPub, caPriv, time.Now(), time.Now().Add(time.Minute*200), nil, nil, nil)
-	rawCACrt, _ := ca.MarshalPEM()
-
-	caCrtF, err := os.CreateTemp("", "sign-cert.crt")
-	require.NoError(t, err)
-	defer os.Remove(caCrtF.Name())
-	caCrtF.Write(rawCACrt)
-
-	caKeyF, err := os.CreateTemp("", "sign-cert.key")
-	require.NoError(t, err)
-	defer os.Remove(caKeyF.Name())
-	caKeyF.Write(rawCAKey)
-
-	keyF, err := os.CreateTemp("", "sign.key")
-	require.NoError(t, err)
-	os.Remove(keyF.Name())
-	defer os.Remove(keyF.Name())
-
-	// ca-key on stdin, cert to stdout
-	withStdin(t, bytes.NewReader(rawCAKey))
-	ob := &bytes.Buffer{}
-	eb := &bytes.Buffer{}
-	args := []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", "-", "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", "-", "-out-key", keyF.Name(), "-duration", "100m"}
-	require.NoError(t, signCert(args, ob, eb, nopw))
-	assert.Empty(t, eb.String())
-
-	lCrt, _, err := cert.UnmarshalCertificateFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.Equal(t, "stdin-test", lCrt.Name())
-	assert.True(t, lCrt.CheckSignature(caPub))
-
-	// two flags reading from stdin should error before any read attempt;
-	// otherwise an interactive shell would hang on io.ReadAll
-	stdinIn := bytes.NewReader(rawCAKey)
-	withStdin(t, stdinIn)
-	ob.Reset()
-	eb.Reset()
-	args = []string{"-version", "1", "-ca-crt", "-", "-ca-key", "-", "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", "nope", "-out-key", "nope", "-duration", "100m"}
-	require.EqualError(t, signCert(args, ob, eb, nopw),
-		`-ca-key and -ca-crt both set to "-", only one input may read from stdin`)
-	assert.Equal(t, len(rawCAKey), stdinIn.Len(), "stdin should be untouched when conflict is caught up front")
-
-	// two flags writing to stdout should error before any output is written
-	// AND before stdin is consumed
-	stdinR := bytes.NewReader(rawCAKey)
-	withStdin(t, stdinR)
-	ob.Reset()
-	eb.Reset()
-	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", "-", "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", "-", "-out-key", "-", "-duration", "100m"}
-	require.EqualError(t, signCert(args, ob, eb, nopw),
-		`-out-key and -out-crt both set to "-", only one output may write to stdout`)
-	assert.Empty(t, ob.String())
-	// stdin should be untouched because the conflict was caught up front
-	assert.Equal(t, len(rawCAKey), stdinR.Len())
-
-	// out-key on stdout, cert on disk
-	keyF2, err := os.CreateTemp("", "sign.key")
-	require.NoError(t, err)
-	os.Remove(keyF2.Name())
-	defer os.Remove(keyF2.Name())
-	crtF, err := os.CreateTemp("", "sign.crt")
-	require.NoError(t, err)
-	os.Remove(crtF.Name())
-	defer os.Remove(crtF.Name())
-
-	ob.Reset()
-	eb.Reset()
-	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "stdin-test", "-ip", "1.1.1.1/24", "-out-crt", crtF.Name(), "-out-key", "-", "-duration", "100m"}
-	require.NoError(t, signCert(args, ob, eb, nopw))
-	assert.Empty(t, eb.String())
-	_, _, curve, err := cert.UnmarshalPrivateKeyFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.Equal(t, cert.Curve_CURVE25519, curve)
-
-	// in-pub on stdin (caller already has a keypair, only the cert is generated)
-	inPub, _ := x25519Keypair()
-	rawInPub := cert.MarshalPublicKeyToPEM(cert.Curve_CURVE25519, inPub)
-
-	withStdin(t, bytes.NewReader(rawInPub))
-	os.Remove(crtF.Name())
-	ob.Reset()
-	eb.Reset()
-	args = []string{"-version", "1", "-ca-crt", caCrtF.Name(), "-ca-key", caKeyF.Name(), "-name", "in-pub-test", "-ip", "1.1.1.1/24", "-in-pub", "-", "-out-crt", "-", "-duration", "100m"}
-	require.NoError(t, signCert(args, ob, eb, nopw))
-	assert.Empty(t, eb.String())
-	stdinCrt, _, err := cert.UnmarshalCertificateFromPEM(ob.Bytes())
-	require.NoError(t, err)
-	assert.Equal(t, "in-pub-test", stdinCrt.Name())
-	assert.Equal(t, inPub, stdinCrt.PublicKey())
+	assert.Equal(t, "Enter passphrase: ", ob.String())
+	assert.Empty(t, eb.String())
 }

cmd/nebula-cert/stdio.go

@@ -1,117 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"io"
-	"os"
-)
-
-// stdioPath is the special path value that selects stdin (for inputs) or
-// stdout (for outputs) instead of a file on disk.
-const stdioPath = "-"
-
-// stdioHelpText is rendered just under the Usage line of each subcommand
-// help so the - convention is documented once instead of on every flag.
-const stdioHelpText = "  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"
-
-// stdinReader is the source used when an input flag is set to "-".
-// It is a package level var so tests can swap in a deterministic reader.
-// Tests that mutate stdinReader cannot run with t.Parallel().
-var stdinReader io.Reader = os.Stdin
-
-// ioClaims tracks which flags have claimed stdin and stdout during a single
-// command invocation so we can refuse a second flag asking for the same
-// stream.
-type ioClaims struct {
-	in  string
-	out string
-}
-
-func (c *ioClaims) claimIn(flagName string) error {
-	if c.in != "" && c.in != flagName {
-		return fmt.Errorf("-%s and -%s both set to %q, only one input may read from stdin", c.in, flagName, stdioPath)
-	}
-	c.in = flagName
-	return nil
-}
-
-func (c *ioClaims) claimOut(flagName string) error {
-	if c.out != "" && c.out != flagName {
-		return fmt.Errorf("-%s and -%s both set to %q, only one output may write to stdout", c.out, flagName, stdioPath)
-	}
-	c.out = flagName
-	return nil
-}
-
-// reserveInputs walks alternating (flagName, path) pairs and claims stdin
-// for any path equal to stdioPath. It must be called before any input is
-// read so a conflict can be reported immediately instead of blocking on
-// io.ReadAll while waiting for input that will never arrive.
-func reserveInputs(claims *ioClaims, pairs ...string) error {
-	return reserveStdio(claims, "reserveInputs", (*ioClaims).claimIn, pairs)
-}
-
-// reserveOutputs walks alternating (flagName, path) pairs and claims stdout
-// for any path equal to stdioPath. It must be called before any output is
-// written so a conflict cannot leave one stream half written before the
-// second flag fails.
-func reserveOutputs(claims *ioClaims, pairs ...string) error {
-	return reserveStdio(claims, "reserveOutputs", (*ioClaims).claimOut, pairs)
-}
-
-func reserveStdio(claims *ioClaims, who string, claim func(*ioClaims, string) error, pairs []string) error {
-	if len(pairs)%2 != 0 {
-		panic(who + " requires alternating name, path pairs")
-	}
-	for i := 0; i < len(pairs); i += 2 {
-		name, path := pairs[i], pairs[i+1]
-		if path != stdioPath {
-			continue
-		}
-		if err := claim(claims, name); err != nil {
-			return err
-		}
-	}
-	return nil
-}
-
-// readInput returns the bytes referenced by path, reading from stdin when
-// path is stdioPath.
-func readInput(flagName, path string, claims *ioClaims) ([]byte, error) {
-	if path == stdioPath {
-		if err := claims.claimIn(flagName); err != nil {
-			return nil, err
-		}
-		return io.ReadAll(stdinReader)
-	}
-	return os.ReadFile(path)
-}
-
-// openInput returns a reader for path. When path is stdioPath the returned
-// reader wraps stdin and Close is a no-op.
-func openInput(flagName, path string, claims *ioClaims) (io.ReadCloser, error) {
-	if path == stdioPath {
-		if err := claims.claimIn(flagName); err != nil {
-			return nil, err
-		}
-		return io.NopCloser(stdinReader), nil
-	}
-	return os.Open(path)
-}
-
-// writeOutput writes data to path, or to stdout when path is stdioPath. perm
-// is only used for file output. The caller must have already claimed stdout
-// via reserveOutputs before invoking with stdioPath.
-func writeOutput(path string, data []byte, perm os.FileMode, stdout io.Writer) error {
-	if path == stdioPath {
-		_, err := stdout.Write(data)
-		return err
-	}
-	return os.WriteFile(path, data, perm)
-}
-
-// isStdio reports whether path is the stdio sentinel and so should skip
-// existence checks like "refuse to overwrite".
-func isStdio(path string) bool {
-	return path == stdioPath
-}

cmd/nebula-cert/stdio_test.go

@@ -1,167 +0,0 @@
-package main
-
-import (
-	"bytes"
-	"io"
-	"os"
-	"path/filepath"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// withStdin temporarily replaces stdinReader for the duration of t.
-func withStdin(t *testing.T, r io.Reader) {
-	t.Helper()
-	prev := stdinReader
-	stdinReader = r
-	t.Cleanup(func() { stdinReader = prev })
-}
-
-func Test_readInput_stdin(t *testing.T) {
-	withStdin(t, bytes.NewBufferString("hello"))
-	var claims ioClaims
-
-	got, err := readInput("path", "-", &claims)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("hello"), got)
-	assert.Equal(t, "path", claims.in)
-}
-
-func Test_readInput_file(t *testing.T) {
-	dir := t.TempDir()
-	p := filepath.Join(dir, "f")
-	require.NoError(t, os.WriteFile(p, []byte("file"), 0600))
-	var claims ioClaims
-
-	got, err := readInput("path", p, &claims)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("file"), got)
-	assert.Empty(t, claims.in)
-}
-
-func Test_readInput_doubleStdinErrors(t *testing.T) {
-	withStdin(t, bytes.NewBufferString("hello"))
-	var claims ioClaims
-
-	_, err := readInput("ca-key", "-", &claims)
-	require.NoError(t, err)
-
-	_, err = readInput("ca-crt", "-", &claims)
-	require.EqualError(t, err, `-ca-key and -ca-crt both set to "-", only one input may read from stdin`)
-}
-
-func Test_openInput_stdin(t *testing.T) {
-	withStdin(t, bytes.NewBufferString("hi"))
-	var claims ioClaims
-
-	r, err := openInput("ca", "-", &claims)
-	require.NoError(t, err)
-	defer r.Close()
-	b, err := io.ReadAll(r)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("hi"), b)
-}
-
-func Test_openInput_doubleStdinErrors(t *testing.T) {
-	withStdin(t, bytes.NewBufferString("hi"))
-	var claims ioClaims
-
-	r, err := openInput("ca", "-", &claims)
-	require.NoError(t, err)
-	r.Close()
-
-	_, err = openInput("crt", "-", &claims)
-	require.EqualError(t, err, `-ca and -crt both set to "-", only one input may read from stdin`)
-}
-
-func Test_writeOutput_stdout(t *testing.T) {
-	out := &bytes.Buffer{}
-
-	err := writeOutput("-", []byte("payload"), 0600, out)
-	require.NoError(t, err)
-	assert.Equal(t, "payload", out.String())
-}
-
-func Test_writeOutput_file(t *testing.T) {
-	dir := t.TempDir()
-	p := filepath.Join(dir, "f")
-	out := &bytes.Buffer{}
-
-	err := writeOutput(p, []byte("payload"), 0600, out)
-	require.NoError(t, err)
-	assert.Empty(t, out.String())
-	got, err := os.ReadFile(p)
-	require.NoError(t, err)
-	assert.Equal(t, []byte("payload"), got)
-}
-
-func Test_reserveOutputs_noConflict(t *testing.T) {
-	var claims ioClaims
-	require.NoError(t, reserveOutputs(&claims,
-		"out-key", "/tmp/key",
-		"out-crt", "-",
-		"out-qr", "",
-	))
-	assert.Equal(t, "out-crt", claims.out)
-}
-
-func Test_reserveOutputs_conflict(t *testing.T) {
-	var claims ioClaims
-	err := reserveOutputs(&claims,
-		"out-key", "-",
-		"out-crt", "-",
-	)
-	require.EqualError(t, err, `-out-key and -out-crt both set to "-", only one output may write to stdout`)
-}
-
-func Test_reserveOutputs_panicsOnOddPairs(t *testing.T) {
-	defer func() {
-		r := recover()
-		require.NotNil(t, r)
-	}()
-	var claims ioClaims
-	_ = reserveOutputs(&claims, "out-key")
-}
-
-func Test_reserveInputs_noConflict(t *testing.T) {
-	var claims ioClaims
-	require.NoError(t, reserveInputs(&claims,
-		"ca-key", "/tmp/ca.key",
-		"ca-crt", "-",
-		"in-pub", "",
-	))
-	assert.Equal(t, "ca-crt", claims.in)
-}
-
-func Test_reserveInputs_conflict(t *testing.T) {
-	var claims ioClaims
-	err := reserveInputs(&claims,
-		"ca-key", "-",
-		"ca-crt", "-",
-	)
-	require.EqualError(t, err, `-ca-key and -ca-crt both set to "-", only one input may read from stdin`)
-}
-
-func Test_claimIn_idempotent(t *testing.T) {
-	// pre-claim then a lazy re-claim of the same flag should be a no-op
-	var claims ioClaims
-	require.NoError(t, claims.claimIn("ca-key"))
-	require.NoError(t, claims.claimIn("ca-key"))
-	assert.Equal(t, "ca-key", claims.in)
-}
-
-func Test_claimOut_idempotent(t *testing.T) {
-	var claims ioClaims
-	require.NoError(t, claims.claimOut("out-crt"))
-	require.NoError(t, claims.claimOut("out-crt"))
-	assert.Equal(t, "out-crt", claims.out)
-}
-
-func Test_isStdio(t *testing.T) {
-	assert.True(t, isStdio("-"))
-	assert.False(t, isStdio(""))
-	assert.False(t, isStdio("./-"))
-	assert.False(t, isStdio("foo"))
-}

cmd/nebula-cert/verify.go

@@ -39,26 +39,18 @@ func verify(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}
 
-	var claims ioClaims
-	if err := reserveInputs(&claims,
-		"ca", *vf.caPath,
-		"crt", *vf.certPath,
-	); err != nil {
-		return err
-	}
-
-	caReader, err := openInput("ca", *vf.caPath, &claims)
+	caFile, err := os.Open(*vf.caPath)
 	if err != nil {
 		return fmt.Errorf("error while reading ca: %w", err)
 	}
-	defer caReader.Close()
+	defer caFile.Close()
 
-	caPool, err := cert.NewCAPoolFromPEMReader(caReader)
+	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 := readInput("crt", *vf.certPath, &claims)
+	rawCert, err := os.ReadFile(*vf.certPath)
 	if err != nil {
 		return fmt.Errorf("unable to read crt: %w", err)
 	}
@@ -93,7 +85,6 @@ func verifySummary() string {
 func verifyHelp(out io.Writer) {
 	vf := newVerifyFlags()
 	_, _ = out.Write([]byte("Usage of " + os.Args[0] + " " + verifySummary() + "\n"))
-	_, _ = out.Write([]byte(stdioHelpText))
 	vf.set.SetOutput(out)
 	vf.set.PrintDefaults()
 }

cmd/nebula-cert/verify_test.go

@@ -23,7 +23,6 @@ func Test_verifyHelp(t *testing.T) {
 	assert.Equal(
 		t,
 		"Usage of "+os.Args[0]+" verify <flags>: verifies a certificate isn't expired and was signed by a trusted authority.\n"+
-			"  Pass \"-\" to any path flag to read from stdin or write to stdout.\n"+
 			"  -ca string\n"+
 			"    \tRequired: path to a file containing one or more ca certificates\n"+
 			"  -crt string\n"+
@@ -123,46 +122,3 @@ func Test_verify(t *testing.T) {
 	assert.Empty(t, eb.String())
 	require.NoError(t, err)
 }
-
-func Test_verify_stdio(t *testing.T) {
-	ob := &bytes.Buffer{}
-	eb := &bytes.Buffer{}
-
-	caPub, caPriv, _ := ed25519.GenerateKey(rand.Reader)
-	ca, _ := NewTestCaCert("test-ca", caPub, caPriv, time.Now().Add(time.Hour*-1), time.Now().Add(time.Hour*2), nil, nil, nil)
-	caPEM, _ := ca.MarshalPEM()
-
-	crt, _ := NewTestCert(ca, caPriv, "test-cert", time.Now().Add(time.Hour*-1), time.Now().Add(time.Hour), nil, nil, nil)
-	crtPEM, _ := crt.MarshalPEM()
-
-	caFile, err := os.CreateTemp("", "verify-ca")
-	require.NoError(t, err)
-	defer os.Remove(caFile.Name())
-	caFile.Write(caPEM)
-
-	// crt on stdin, ca on disk
-	withStdin(t, bytes.NewReader(crtPEM))
-	require.NoError(t, verify([]string{"-ca", caFile.Name(), "-crt", "-"}, ob, eb))
-	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
-
-	// ca on stdin, crt on disk
-	certFile, err := os.CreateTemp("", "verify-cert")
-	require.NoError(t, err)
-	defer os.Remove(certFile.Name())
-	certFile.Write(crtPEM)
-
-	withStdin(t, bytes.NewReader(caPEM))
-	ob.Reset()
-	eb.Reset()
-	require.NoError(t, verify([]string{"-ca", "-", "-crt", certFile.Name()}, ob, eb))
-	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
-
-	// both flags on stdin should error
-	withStdin(t, bytes.NewReader(caPEM))
-	ob.Reset()
-	eb.Reset()
-	require.EqualError(t, verify([]string{"-ca", "-", "-crt", "-"}, ob, eb),
-		`-ca and -crt both set to "-", only one input may read from stdin`)
-}

cmd/nebula-service/main.go

@@ -61,12 +61,9 @@ func main() {
 	}
 
 	if *configPath == "" {
-		p, err := config.DefaultPath()
-		if err != nil {
-			fmt.Println(err)
-			os.Exit(1)
-		}
-		*configPath = p
+		fmt.Println("-config flag must be set")
+		flag.Usage()
+		os.Exit(1)
 	}
 
 	c := config.NewC(l)

cmd/nebula-service/service.go

@@ -3,6 +3,8 @@ package main
 import (
 	"fmt"
 	"log"
+	"os"
+	"path/filepath"
 
 	"github.com/kardianos/service"
 	"github.com/slackhq/nebula"
@@ -55,13 +57,24 @@ func (p *program) Stop(s service.Service) error {
 	return nil
 }
 
+func fileExists(filename string) bool {
+	_, err := os.Stat(filename)
+	if os.IsNotExist(err) {
+		return false
+	}
+	return true
+}
+
 func doService(configPath *string, configTest *bool, build string, serviceFlag *string) error {
 	if *configPath == "" {
-		p, err := config.DefaultPath()
+		ex, err := os.Executable()
 		if err != nil {
 			return err
 		}
-		*configPath = p
+		*configPath = filepath.Dir(ex) + "/config.yaml"
+		if !fileExists(*configPath) {
+			*configPath = filepath.Dir(ex) + "/config.yml"
+		}
 	}
 
 	svcConfig := &service.Config{

cmd/nebula/main.go

@@ -50,12 +50,9 @@ func main() {
 	}
 
 	if *configPath == "" {
-		p, err := config.DefaultPath()
-		if err != nil {
-			fmt.Println(err)
-			os.Exit(1)
-		}
-		*configPath = p
+		fmt.Println("-config flag must be set")
+		flag.Usage()
+		os.Exit(1)
 	}
 
 	l := logging.NewLogger(os.Stdout)

config/default.go

@@ -1,29 +0,0 @@
-package config
-
-import (
-	"fmt"
-	"os"
-	"path/filepath"
-)
-
-// DefaultPath returns a path to a config file alongside the running executable, preferring config.yaml over config.yml.
-// If neither file exists an error is returned that names both paths checked.
-func DefaultPath() (string, error) {
-	ex, err := os.Executable()
-	if err != nil {
-		return "", err
-	}
-	return defaultPathInDir(filepath.Dir(ex))
-}
-
-func defaultPathInDir(dir string) (string, error) {
-	yamlPath := filepath.Join(dir, "config.yaml")
-	if _, err := os.Stat(yamlPath); err == nil {
-		return yamlPath, nil
-	}
-	ymlPath := filepath.Join(dir, "config.yml")
-	if _, err := os.Stat(ymlPath); err == nil {
-		return ymlPath, nil
-	}
-	return "", fmt.Errorf("no default config found at %s or %s", yamlPath, ymlPath)
-}

config/default_test.go

@@ -1,67 +0,0 @@
-package config
-
-import (
-	"os"
-	"path/filepath"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestDefaultPathInDir(t *testing.T) {
-	t.Run("prefers config.yaml when both exist", func(t *testing.T) {
-		dir := t.TempDir()
-		want := filepath.Join(dir, "config.yaml")
-		other := filepath.Join(dir, "config.yml")
-		require.NoError(t, os.WriteFile(want, []byte("a: 1"), 0644))
-		require.NoError(t, os.WriteFile(other, []byte("a: 2"), 0644))
-
-		got, err := defaultPathInDir(dir)
-		require.NoError(t, err)
-		assert.Equal(t, want, got)
-	})
-
-	t.Run("returns config.yaml when only it exists", func(t *testing.T) {
-		dir := t.TempDir()
-		want := filepath.Join(dir, "config.yaml")
-		require.NoError(t, os.WriteFile(want, []byte("a: 1"), 0644))
-
-		got, err := defaultPathInDir(dir)
-		require.NoError(t, err)
-		assert.Equal(t, want, got)
-	})
-
-	t.Run("falls back to config.yml when only it exists", func(t *testing.T) {
-		dir := t.TempDir()
-		want := filepath.Join(dir, "config.yml")
-		require.NoError(t, os.WriteFile(want, []byte("a: 1"), 0644))
-
-		got, err := defaultPathInDir(dir)
-		require.NoError(t, err)
-		assert.Equal(t, want, got)
-	})
-
-	t.Run("errors when neither exists and names both paths", func(t *testing.T) {
-		dir := t.TempDir()
-		got, err := defaultPathInDir(dir)
-		assert.Empty(t, got)
-		require.Error(t, err)
-		assert.Contains(t, err.Error(), filepath.Join(dir, "config.yaml"))
-		assert.Contains(t, err.Error(), filepath.Join(dir, "config.yml"))
-	})
-}
-
-func TestDefaultPath(t *testing.T) {
-	got, err := DefaultPath()
-	if err != nil {
-		ex, exErr := os.Executable()
-		require.NoError(t, exErr)
-		assert.Contains(t, err.Error(), filepath.Dir(ex))
-		return
-	}
-	ex, err := os.Executable()
-	require.NoError(t, err)
-	assert.Equal(t, filepath.Dir(ex), filepath.Dir(got))
-	assert.Contains(t, []string{"config.yaml", "config.yml"}, filepath.Base(got))
-}

connection_manager.go

@@ -136,6 +136,14 @@ func (cm *connectionManager) getAndResetTrafficCheck(h *HostInfo, now time.Time)
 	return in, out
 }
 
+// AddTrafficWatch must be called for every new HostInfo.
+// We will continue to monitor the HostInfo until the tunnel is dropped.
+func (cm *connectionManager) AddTrafficWatch(h *HostInfo) {
+	if h.out.Swap(true) == false {
+		cm.trafficTimer.Add(h.localIndexId, cm.checkInterval)
+	}
+}
+
 func (cm *connectionManager) Start(ctx context.Context) {
 	clockSource := time.NewTicker(cm.trafficTimer.t.tickDuration)
 	defer clockSource.Stop()
@@ -298,8 +306,8 @@ func (cm *connectionManager) migrateRelayUsed(oldhostinfo, newhostinfo *HostInfo
 		} else {
 			cm.intf.SendMessageToHostInfo(header.Control, 0, newhostinfo, msg, make([]byte, 12), make([]byte, mtu))
 			cm.l.Info("send CreateRelayRequest",
-				"relayFrom", relayFrom,
-				"relayTo", relayTo,
+				"relayFrom", req.RelayFromAddr,
+				"relayTo", req.RelayToAddr,
 				"initiatorRelayIndex", req.InitiatorRelayIndex,
 				"responderRelayIndex", req.ResponderRelayIndex,
 				"vpnAddrs", newhostinfo.vpnAddrs,

connection_state.go

@@ -10,7 +10,7 @@ import (
 	"github.com/slackhq/nebula/noiseutil"
 )
 
-const ReplayWindow = 1024
+const ReplayWindow = 8192
 
 type ConnectionState struct {
 	eKey           noiseutil.CipherState

dist/wireshark/nebula.lua

@@ -62,7 +62,7 @@ function nebula.dissector(tvbuf, pktinfo, root)
     tree:add(pf_version, tvbuf:range(0,1))
     local type = tree:add(pf_type,    tvbuf:range(0,1))
 
-    local nebula_type = bit.band(tvbuf:range(0,1):uint(), 0x0F)
+    local nebula_type = bit32.band(tvbuf:range(0,1):uint(), 0x0F)
     if nebula_type == 0 then
         local stage = tvbuf(8,8):uint64()
         tree:add(pf_subtype_handshake, tvbuf:range(1,1))

dns_server.go

@@ -11,21 +11,19 @@ import (
 	"sync"
 	"sync/atomic"
 
+	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
 	"github.com/slackhq/nebula/config"
 )
 
 type dnsServer struct {
 	sync.RWMutex
-	l       *slog.Logger
-	ctx     context.Context
-	dnsMap4 map[string]netip.Addr
-	dnsMap6 map[string]netip.Addr
-	hostMap *HostMap
-	pki     *PKI
-
-	// selfHost is the cached FQDN we last seeded for ourselves
-	selfHost string
+	l               *slog.Logger
+	ctx             context.Context
+	dnsMap4         map[string]netip.Addr
+	dnsMap6         map[string]netip.Addr
+	hostMap         *HostMap
+	myVpnAddrsTable *bart.Lite
 
 	mux *dns.ServeMux
 
@@ -57,14 +55,14 @@ type dnsServer struct {
 // they no-op when DNS isn't enabled. Each Start invocation owns a ctx-cancel
 // watcher that tears the listener down on nebula shutdown. The returned
 // pointer is always non-nil, even on error.
-func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, pki *PKI, hostMap *HostMap, c *config.C) (*dnsServer, error) {
+func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, cs *CertState, hostMap *HostMap, c *config.C) (*dnsServer, error) {
 	ds := &dnsServer{
-		l:       l,
-		ctx:     ctx,
-		dnsMap4: make(map[string]netip.Addr),
-		dnsMap6: make(map[string]netip.Addr),
-		hostMap: hostMap,
-		pki:     pki,
+		l:               l,
+		ctx:             ctx,
+		dnsMap4:         make(map[string]netip.Addr),
+		dnsMap6:         make(map[string]netip.Addr),
+		hostMap:         hostMap,
+		myVpnAddrsTable: cs.myVpnAddrsTable,
 	}
 	ds.mux = dns.NewServeMux()
 	ds.mux.HandleFunc(".", ds.handleDnsRequest)
@@ -78,7 +76,6 @@ func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, pki *PKI, hostM
 	if err := ds.reload(c, true); err != nil {
 		return ds, err
 	}
-	ds.seedSelf()
 	return ds, nil
 }
 
@@ -116,7 +113,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 			d.Stop()
 		}
 		// Drop any records that accumulated while enabled; a later re-enable
-		// will repopulate from fresh handshakes and a fresh seedSelf.
+		// will repopulate from fresh handshakes.
 		d.clearRecords()
 		return nil
 	}
@@ -124,14 +121,17 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	if running == nil {
 		// Was disabled (or never started); bring it up now.
 		go d.Start()
-	} else if !sameAddr {
-		d.shutdownServer(running, runningStarted, "reload")
-		// Old Start goroutine has now exited; bring up a fresh listener on the new address.
-		go d.Start()
+		return nil
 	}
 
-	// Refresh the self entry every enabled reload so cert renewals that change our name or VPN addresses are picked up.
-	d.seedSelf()
+	if sameAddr {
+		return nil
+	}
+
+	d.shutdownServer(running, runningStarted, "reload")
+	// Old Start goroutine has now exited; bring up a fresh listener on the
+	// new address.
+	go d.Start()
 	return nil
 }
 
@@ -249,20 +249,6 @@ func (d *dnsServer) QueryCert(data string) string {
 		return ""
 	}
 
-	// The hostmap only ever contains peers we have handshaked with, so it never carries an entry for ourselves.
-	// Answer self lookups straight from the local cert state.
-	if cs := d.certState(); cs != nil && cs.myVpnAddrsTable != nil && cs.myVpnAddrsTable.Contains(ip) {
-		c := cs.GetDefaultCertificate()
-		if c == nil {
-			return ""
-		}
-		b, err := c.MarshalJSON()
-		if err != nil {
-			return ""
-		}
-		return string(b)
-	}
-
 	hostinfo := d.hostMap.QueryVpnAddr(ip)
 	if hostinfo == nil {
 		return ""
@@ -280,60 +266,12 @@ func (d *dnsServer) QueryCert(data string) string {
 	return string(b)
 }
 
-// clearRecords drops all DNS records, including the self entry.
+// clearRecords drops all DNS records.
 func (d *dnsServer) clearRecords() {
 	d.Lock()
 	defer d.Unlock()
 	clear(d.dnsMap4)
 	clear(d.dnsMap6)
-	d.selfHost = ""
-}
-
-// seedSelf inserts (or refreshes) a record for our own cert name pointing at our VPN addresses,
-// so a single-lighthouse network can resolve the lighthouse's own hostname without the two-process workaround.
-func (d *dnsServer) seedSelf() {
-	if !d.enabled.Load() {
-		return
-	}
-	cs := d.certState()
-	if cs == nil {
-		return
-	}
-	c := cs.GetDefaultCertificate()
-	if c == nil {
-		return
-	}
-	newHost := strings.ToLower(c.Name()) + "."
-
-	d.Lock()
-	defer d.Unlock()
-	if d.selfHost != "" && d.selfHost != newHost {
-		delete(d.dnsMap4, d.selfHost)
-		delete(d.dnsMap6, d.selfHost)
-	}
-	d.selfHost = newHost
-	delete(d.dnsMap4, newHost)
-	delete(d.dnsMap6, newHost)
-	haveV4, haveV6 := false, false
-	for _, addr := range cs.myVpnAddrs {
-		if addr.Is4() && !haveV4 {
-			d.dnsMap4[newHost] = addr
-			haveV4 = true
-		} else if addr.Is6() && !haveV6 {
-			d.dnsMap6[newHost] = addr
-			haveV6 = true
-		}
-		if haveV4 && haveV6 {
-			break
-		}
-	}
-}
-
-func (d *dnsServer) certState() *CertState {
-	if d.pki == nil {
-		return nil
-	}
-	return d.pki.getCertState()
 }
 
 // Add adds the first IPv4 and IPv6 address that appears in `addresses` as the record for `host`
@@ -371,12 +309,8 @@ func (d *dnsServer) isSelfNebulaOrLocalhost(addr string) bool {
 		return true
 	}
 
-	cs := d.certState()
-	if cs == nil || cs.myVpnAddrsTable == nil {
-		return false
-	}
 	//if we found it in this table, it's good
-	return cs.myVpnAddrsTable.Contains(b)
+	return d.myVpnAddrsTable.Contains(b)
 }
 
 func (d *dnsServer) parseQuery(m *dns.Msg, w dns.ResponseWriter) {

dns_server_test.go

@@ -9,10 +9,7 @@ import (
 	"testing"
 	"time"
 
-	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/cert_test"
 	"github.com/slackhq/nebula/config"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -279,92 +276,6 @@ func TestDnsServer_Stop_beforeBind_doesNotHang(t *testing.T) {
 	}
 }
 
-// newTestPKI builds a minimal *PKI with a single v1 cert whose name and
-// VPN addresses are caller-provided, suitable for exercising seedSelf and
-// QueryCert self handling.
-func newTestPKI(t *testing.T, name string, addrs []netip.Addr) *PKI {
-	t.Helper()
-	networks := make([]netip.Prefix, 0, len(addrs))
-	for _, a := range addrs {
-		bits := 32
-		if a.Is6() {
-			bits = 128
-		}
-		networks = append(networks, netip.PrefixFrom(a, bits))
-	}
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Time{}, time.Time{}, nil, nil, nil)
-	c, _, _, _ := cert_test.NewTestCert(cert.Version2, cert.Curve_CURVE25519, ca, caKey, name, time.Time{}, time.Time{}, networks, nil, nil)
-
-	addrsTable := new(bart.Lite)
-	for _, a := range addrs {
-		addrsTable.Insert(netip.PrefixFrom(a, a.BitLen()))
-	}
-
-	cs := &CertState{
-		v2Cert:            c,
-		initiatingVersion: cert.Version2,
-		myVpnAddrs:        addrs,
-		myVpnAddrsTable:   addrsTable,
-	}
-	pki := &PKI{}
-	pki.cs.Store(cs)
-	return pki
-}
-
-func TestDnsServer_seedSelf_addsOwnRecord(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	myV4 := netip.MustParseAddr("10.0.0.1")
-	myV6 := netip.MustParseAddr("fd00::1")
-	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{myV4, myV6})
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
-	require.NoError(t, ds.reload(c, true))
-
-	ds.seedSelf()
-	got4, exists := ds.Query(dns.TypeA, "lighthouse.")
-	assert.True(t, exists)
-	assert.Equal(t, myV4, got4)
-	got6, exists := ds.Query(dns.TypeAAAA, "lighthouse.")
-	assert.True(t, exists)
-	assert.Equal(t, myV6, got6)
-}
-
-func TestDnsServer_seedSelf_disabled_noOp(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{netip.MustParseAddr("10.0.0.1")})
-	setDnsConfig(c, "127.0.0.1", "0", true, false)
-	require.NoError(t, ds.reload(c, true))
-
-	ds.seedSelf()
-	_, exists := ds.Query(dns.TypeA, "lighthouse.")
-	assert.False(t, exists)
-}
-
-func TestDnsServer_clearRecords_dropsSelfHost(t *testing.T) {
-	ds, c := newTestDnsServer(t)
-	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{netip.MustParseAddr("10.0.0.1")})
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
-	require.NoError(t, ds.reload(c, true))
-	ds.seedSelf()
-	require.NotEmpty(t, ds.selfHost)
-
-	ds.clearRecords()
-	assert.Empty(t, ds.selfHost)
-	_, exists := ds.Query(dns.TypeA, "lighthouse.")
-	assert.False(t, exists)
-}
-
-func TestDnsServer_QueryCert_returnsOwnCert(t *testing.T) {
-	ds, _ := newTestDnsServer(t)
-	myV4 := netip.MustParseAddr("10.0.0.1")
-	ds.pki = newTestPKI(t, "lighthouse", []netip.Addr{myV4})
-
-	got := ds.QueryCert(myV4.String() + ".")
-	assert.NotEmpty(t, got, "TXT lookup of our own VPN address should return our cert")
-
-	other := netip.MustParseAddr("10.0.0.99")
-	assert.Empty(t, ds.QueryCert(other.String()+"."), "unknown peer IP should return nothing")
-}
-
 func TestDnsServer_reload_disable_stopsRunningServer(t *testing.T) {
 	port := freeUDPPort(t)
 	ds, c := newTestDnsServer(t)

docker/Dockerfile

@@ -1,16 +1,6 @@
 FROM gcr.io/distroless/static:latest
 
 ARG TARGETOS TARGETARCH
-
-ARG VERSION=dev
-ARG REVISION=unknown
-LABEL org.opencontainers.image.title="nebula" \
-      org.opencontainers.image.description="A scalable overlay networking tool with a focus on performance, simplicity and security" \
-      org.opencontainers.image.vendor="Nebula OSS" \
-      org.opencontainers.image.source="https://github.com/slackhq/nebula" \
-      org.opencontainers.image.version="${VERSION}" \
-      org.opencontainers.image.revision="${REVISION}"
-
 COPY build/$TARGETOS-$TARGETARCH/nebula /nebula
 COPY build/$TARGETOS-$TARGETARCH/nebula-cert /nebula-cert
 

e2e/helpers_test.go

@@ -4,15 +4,13 @@
 package e2e
 
 import (
-	"io"
+	"log/slog"
 	"net/netip"
 	"os"
 	"strings"
 	"testing"
 	"time"
 
-	"log/slog"
-
 	"dario.cat/mergo"
 	"github.com/google/gopacket"
 	"github.com/google/gopacket/layers"
@@ -382,7 +380,7 @@ func getAddrs(ns []netip.Prefix) []netip.Addr {
 func NewTestLogger() *slog.Logger {
 	v := os.Getenv("TEST_LOGS")
 	if v == "" {
-		return slog.New(slog.NewTextHandler(io.Discard, nil))
+		return slog.New(slog.DiscardHandler)
 	}
 
 	level := slog.LevelInfo

e2e/tunnels_test.go

@@ -15,7 +15,6 @@ import (
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/udp"
 	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
 	"gopkg.in/yaml.v3"
 )
 
@@ -374,100 +373,6 @@ func TestCrossStackRelaysWork(t *testing.T) {
 	//relayControl.Stop()
 }
 
-// TestRelayReplayProtection asserts that a relay (forwarding-type) node rejects
-// replayed relay frames. A captured relay frame, re-injected with the same
-// message counter, must be dropped by the replay window rather than re-forwarded
-// to the relay target. Before the fix, handleOutsideRelayPacket authenticated the
-// frame but never advanced the replay window, so every replay was re-forwarded.
-func TestRelayReplayProtection(t *testing.T) {
-	t.Parallel()
-	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
-	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version2, ca, caKey, "me     ", "10.128.0.1/24,fc00::1/64", m{"relay": m{"use_relays": true}})
-	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "relay  ", "10.128.0.128/24,fc00::128/64", m{"relay": m{"am_relay": true}})
-	theirUdp := netip.MustParseAddrPort("10.0.0.2:4242")
-	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServerWithUdp(cert.Version2, ca, caKey, "them   ", "fc00::2/64", theirUdp, m{"relay": m{"use_relays": true}})
-
-	myVpnV6 := myVpnIpNet[1]
-	relayVpnV4 := relayVpnIpNet[0]
-	relayVpnV6 := relayVpnIpNet[1]
-	theirVpnV6 := theirVpnIpNet[0]
-
-	// Teach me how to reach the relay and that them is reachable via the relay
-	myControl.InjectLightHouseAddr(relayVpnV4.Addr(), relayUdpAddr)
-	myControl.InjectLightHouseAddr(relayVpnV6.Addr(), relayUdpAddr)
-	myControl.InjectRelays(theirVpnV6.Addr(), []netip.Addr{relayVpnV6.Addr()})
-	relayControl.InjectLightHouseAddr(theirVpnV6.Addr(), theirUdpAddr)
-
-	r := router.NewR(t, myControl, relayControl, theirControl)
-	defer r.RenderFlow()
-
-	myControl.Start()
-	relayControl.Start()
-	theirControl.Start()
-
-	// Establish the relayed tunnel in both directions so all handshakes complete.
-	t.Log("Establish the relayed tunnel")
-	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnV6.Addr(), 80, myVpnV6.Addr(), 80, []byte("Hi from me")))
-	p := r.RouteForAllUntilTxTun(theirControl)
-	assertUdpPacket(t, []byte("Hi from me"), p, myVpnV6.Addr(), theirVpnV6.Addr(), 80, 80)
-	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnV6.Addr(), 80, theirVpnV6.Addr(), 80, []byte("Hi from them")))
-	p = r.RouteForAllUntilTxTun(myControl)
-	assertUdpPacket(t, []byte("Hi from them"), p, theirVpnV6.Addr(), myVpnV6.Addr(), 80, 80)
-
-	// Drain anything still queued on me's UDP tx so the next packet we pull is the
-	// relay frame we are about to generate.
-	for myControl.GetFromUDP(false) != nil {
-	}
-
-	// Capture a single legitimate relay frame that me transmits toward the relay.
-	t.Log("Capture a relay frame from me -> relay")
-	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnV6.Addr(), 80, myVpnV6.Addr(), 80, []byte("replay me")))
-	relayFrame := myControl.GetFromUDP(true)
-	require.Equal(t, relayUdpAddr, relayFrame.To, "captured frame should be addressed to the relay")
-	var fh header.H
-	require.NoError(t, fh.Parse(relayFrame.Data))
-	require.Equal(t, header.Message, fh.Type)
-	require.Equal(t, header.MessageRelay, fh.Subtype)
-
-	// drainForwards counts relay frames the relay forwards toward them within the
-	// settle window. We match on destination + (Message, MessageRelay) so the
-	// relay's own direct traffic to them can't be miscounted.
-	drainForwards := func(settle time.Duration) int {
-		ch := relayControl.GetUDPTxChan()
-		count := 0
-		for {
-			select {
-			case pkt := <-ch:
-				var ph header.H
-				if pkt.To == theirUdpAddr && ph.Parse(pkt.Data) == nil &&
-					ph.Type == header.Message && ph.Subtype == header.MessageRelay {
-					count++
-				}
-				pkt.Release()
-			case <-time.After(settle):
-				return count
-			}
-		}
-	}
-
-	// First delivery of the captured frame: the relay should forward it once.
-	t.Log("Deliver the captured frame once; relay forwards it to them")
-	relayControl.InjectUDPPacket(relayFrame)
-	require.Equal(t, 1, drainForwards(200*time.Millisecond), "relay should forward the first, legitimate copy")
-
-	// Replay the exact same frame several times. A correct replay window rejects
-	// these duplicates so the relay forwards none of them.
-	t.Log("Replay the captured frame; relay must drop the duplicates")
-	const replays = 3
-	for i := 0; i < replays; i++ {
-		relayControl.InjectUDPPacket(relayFrame)
-	}
-	forwarded := drainForwards(200 * time.Millisecond)
-	assert.Equal(t, 0, forwarded, "relay re-forwarded %d/%d replayed relay frames; replay protection is ineffective on relay tunnels", forwarded, replays)
-
-	r.RenderHostmaps("Final hostmaps", myControl, relayControl, theirControl)
-}
-
 func TestCloseTunnelAuthenticated(t *testing.T) {
 	t.Parallel()
 	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})

ecn_inner_test.go

@@ -0,0 +1,124 @@
+package nebula
+
+import (
+	"log/slog"
+	"testing"
+)
+
+func TestInnerECN(t *testing.T) {
+	cases := []struct {
+		name string
+		pkt  []byte
+		want byte
+	}{
+		{"empty", nil, 0},
+		{"v4_NotECT", v4WithToS(0x00), 0x00},
+		{"v4_ECT0", v4WithToS(0x02), 0x02},
+		{"v4_ECT1", v4WithToS(0x01), 0x01},
+		{"v4_CE", v4WithToS(0x03), 0x03},
+		{"v4_DSCP_then_NotECT", v4WithToS(0x88 | 0x00), 0x00},
+		{"v4_DSCP_then_CE", v4WithToS(0x88 | 0x03), 0x03},
+		{"v6_NotECT", v6WithTC(0x00), 0x00},
+		{"v6_ECT0", v6WithTC(0x02), 0x02},
+		{"v6_CE", v6WithTC(0x03), 0x03},
+		{"v6_DSCP_then_CE", v6WithTC(0x88 | 0x03), 0x03},
+		{"unknown_version", []byte{0xa5, 0xff}, 0},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			got := innerECN(c.pkt)
+			if got != c.want {
+				t.Errorf("innerECN=0x%02x want 0x%02x", got, c.want)
+			}
+		})
+	}
+}
+
+// v4WithToS returns a 2-byte slice tall enough for innerECN: byte 0 carries
+// version=4 in the high nibble, byte 1 is the full ToS so we exercise both
+// the DSCP and ECN portions through the byte 1 mask.
+func v4WithToS(tos byte) []byte {
+	return []byte{0x45, tos}
+}
+
+// v6WithTC builds a 2-byte slice that places a known traffic class value
+// across bytes 0 (high nibble of TC) and 1 (low nibble of TC). innerECN
+// extracts ECN as (b[1]>>4)&0x03, which corresponds to TC[1:0].
+func v6WithTC(tc byte) []byte {
+	return []byte{0x60 | (tc>>4)&0x0f, (tc & 0x0f) << 4}
+}
+
+func TestApplyOuterECN(t *testing.T) {
+	silent := slog.New(slog.DiscardHandler)
+	hi := &HostInfo{}
+
+	// Build a v4 packet helper with a given inner ECN field.
+	v4 := func(innerECN byte) []byte {
+		// 20-byte minimal IPv4 header with ToS = innerECN (DSCP zeroed).
+		return []byte{
+			0x45, innerECN, 0, 28,
+			0, 0, 0x40, 0,
+			64, 6, 0, 0,
+			10, 0, 0, 1,
+			10, 0, 0, 2,
+		}
+	}
+	// Build a v6 packet helper with a given inner ECN field. ECN occupies
+	// TC[1:0] which sit at byte 1 mask 0x30.
+	v6 := func(innerECN byte) []byte {
+		// 40-byte minimal IPv6 header with TC[1:0] = innerECN.
+		pkt := make([]byte, 40)
+		pkt[0] = 0x60                   // version=6, TC[7:4]=0
+		pkt[1] = (innerECN & 0x03) << 4 // TC[3:0]: low 2 bits = ECN, top 2 = DSCP-low (0)
+		return pkt
+	}
+
+	type cell struct {
+		outer    byte
+		inner    byte
+		wantECN  byte
+		wantSame bool // expect inner unchanged (true => verify the byte didn't move)
+	}
+
+	// RFC 6040 normal-mode combine table. Only outer==CE causes mutation.
+	table := []cell{
+		{ecnNotECT, ecnNotECT, ecnNotECT, true},
+		{ecnNotECT, ecnECT0, ecnECT0, true},
+		{ecnNotECT, ecnECT1, ecnECT1, true},
+		{ecnNotECT, ecnCE, ecnCE, true},
+
+		{ecnECT0, ecnNotECT, ecnNotECT, true},
+		{ecnECT0, ecnECT0, ecnECT0, true},
+		{ecnECT0, ecnECT1, ecnECT1, true},
+		{ecnECT0, ecnCE, ecnCE, true},
+
+		{ecnECT1, ecnNotECT, ecnNotECT, true},
+		{ecnECT1, ecnECT0, ecnECT0, true},
+		{ecnECT1, ecnECT1, ecnECT1, true},
+		{ecnECT1, ecnCE, ecnCE, true},
+
+		{ecnCE, ecnNotECT, ecnNotECT, true}, // legacy: log, leave alone
+		{ecnCE, ecnECT0, ecnCE, false},      // CE folded in
+		{ecnCE, ecnECT1, ecnCE, false},
+		{ecnCE, ecnCE, ecnCE, true},
+	}
+
+	for _, c := range table {
+		t.Run("v4", func(t *testing.T) {
+			pkt := v4(c.inner)
+			applyOuterECN(pkt, c.outer, hi, silent)
+			got := pkt[1] & 0x03
+			if got != c.wantECN {
+				t.Errorf("v4 outer=0x%02x inner=0x%02x: got 0x%02x want 0x%02x", c.outer, c.inner, got, c.wantECN)
+			}
+		})
+		t.Run("v6", func(t *testing.T) {
+			pkt := v6(c.inner)
+			applyOuterECN(pkt, c.outer, hi, silent)
+			got := (pkt[1] >> 4) & 0x03
+			if got != c.wantECN {
+				t.Errorf("v6 outer=0x%02x inner=0x%02x: got 0x%02x want 0x%02x", c.outer, c.inner, got, c.wantECN)
+			}
+		})
+	}
+}

examples/config.yml

@@ -397,7 +397,7 @@ firewall:
   #   `drop` (default): silently drop the packet.
   #   `reject`: send a reject reply.
   #     - For TCP, this will be a RST "Connection Reset" packet.
-  #     - For other protocols, this will be an ICMP "Destination unreachable: Communication administratively prohibited" packet.
+  #     - For other protocols, this will be an ICMP port unreachable packet.
   outbound_action: drop
   inbound_action: drop
 

firewall.go

@@ -58,9 +58,8 @@ type Firewall struct {
 	routableNetworks *bart.Lite
 
 	// assignedNetworks is a list of vpn networks assigned to us in the certificate.
-	assignedNetworks []netip.Prefix
-	// unsafeNetworks is the list of unsafe networks issued to us in the certificate
-	unsafeNetworks []netip.Prefix
+	assignedNetworks  []netip.Prefix
+	hasUnsafeNetworks bool
 
 	rules        string
 	rulesVersion uint16
@@ -159,9 +158,10 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 		assignedNetworks = append(assignedNetworks, network)
 	}
 
-	unsafeNetworks := c.UnsafeNetworks()
-	for _, n := range unsafeNetworks {
+	hasUnsafeNetworks := false
+	for _, n := range c.UnsafeNetworks() {
 		routableNetworks.Insert(n)
+		hasUnsafeNetworks = true
 	}
 
 	return &Firewall{
@@ -169,15 +169,15 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 			Conns:      make(map[firewall.Packet]*conn),
 			TimerWheel: NewTimerWheel[firewall.Packet](tmin, tmax),
 		},
-		InRules:          newFirewallTable(),
-		OutRules:         newFirewallTable(),
-		TCPTimeout:       tcpTimeout,
-		UDPTimeout:       UDPTimeout,
-		DefaultTimeout:   defaultTimeout,
-		routableNetworks: routableNetworks,
-		assignedNetworks: assignedNetworks,
-		unsafeNetworks:   unsafeNetworks,
-		l:                l,
+		InRules:           newFirewallTable(),
+		OutRules:          newFirewallTable(),
+		TCPTimeout:        tcpTimeout,
+		UDPTimeout:        UDPTimeout,
+		DefaultTimeout:    defaultTimeout,
+		routableNetworks:  routableNetworks,
+		assignedNetworks:  assignedNetworks,
+		hasUnsafeNetworks: hasUnsafeNetworks,
+		l:                 l,
 
 		incomingMetrics: firewallMetrics{
 			droppedLocalAddr:  metrics.GetOrRegisterCounter("firewall.incoming.dropped.local_addr", nil),
@@ -897,7 +897,7 @@ func (flc *firewallLocalCIDR) addRule(f *Firewall, localCidr string) error {
 	}
 
 	if localCidr == "" {
-		if len(f.unsafeNetworks) == 0 || f.defaultLocalCIDRAny {
+		if !f.hasUnsafeNetworks || f.defaultLocalCIDRAny {
 			flc.Any = true
 			return nil
 		}
@@ -1055,6 +1055,7 @@ func (r *rule) sanity() error {
 }
 
 func parsePort(s string) (int32, int32, error) {
+	var err error
 	const notAPort int32 = -2
 	if s == "any" {
 		return firewall.PortAny, firewall.PortAny, nil
@@ -1063,11 +1064,11 @@ func parsePort(s string) (int32, int32, error) {
 		return firewall.PortFragment, firewall.PortFragment, nil
 	}
 	if !strings.Contains(s, `-`) {
-		rPort, err := parsePortValue("", s)
+		rPort, err := strconv.Atoi(s)
 		if err != nil {
-			return notAPort, notAPort, err
+			return notAPort, notAPort, fmt.Errorf("was not a number; `%s`", s)
 		}
-		return rPort, rPort, nil
+		return int32(rPort), int32(rPort), nil
 	}
 
 	sPorts := strings.SplitN(s, `-`, 2)
@@ -1078,40 +1079,22 @@ func parsePort(s string) (int32, int32, error) {
 		return notAPort, notAPort, fmt.Errorf("appears to be a range but could not be parsed; `%s`", s)
 	}
 
-	startPort, err := parsePortValue("beginning range ", sPorts[0])
+	rStartPort, err := strconv.Atoi(sPorts[0])
 	if err != nil {
-		return notAPort, notAPort, err
+		return notAPort, notAPort, fmt.Errorf("beginning range was not a number; `%s`", sPorts[0])
 	}
 
-	endPort, err := parsePortValue("ending range ", sPorts[1])
+	rEndPort, err := strconv.Atoi(sPorts[1])
 	if err != nil {
-		return notAPort, notAPort, err
+		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
 }
-
-// parsePortValue accepts a base-10 decimal in [0, 65535] and returns it
-// widened to int32. Using strconv.ParseUint with bitSize 16 rejects
-// negative input, out-of-range input (>65535), and any non-decimal byte
-// by construction, so the int32 widening that follows is provably safe
-// and cannot collide with firewall.PortAny (0) or firewall.PortFragment
-// (-1) via integer truncation.
-//
-// prefix is prepended to both error messages so callers can disambiguate
-// the single-port path (prefix="") from the range bounds (prefix="beginning
-// range " / "ending range "), preserving the historical error strings.
-func parsePortValue(prefix, s string) (int32, error) {
-	n, err := strconv.ParseUint(s, 10, 16)
-	if err == nil {
-		return int32(n), nil
-	}
-	if errors.Is(err, strconv.ErrRange) {
-		return 0, fmt.Errorf("%sout of range [0,65535]; `%s`", prefix, s)
-	}
-	return 0, fmt.Errorf("%swas not a number; `%s`", prefix, s)
-}

firewall/cache.go

@@ -5,6 +5,8 @@ import (
 	"log/slog"
 	"sync/atomic"
 	"time"
+
+	"github.com/slackhq/nebula/logging"
 )
 
 // ConntrackCache is used as a local routine cache to know if a given flow
@@ -56,8 +58,8 @@ func (c *ConntrackCacheTicker) Get() ConntrackCache {
 	if tick := c.cacheTick.Load(); tick != c.cacheV {
 		c.cacheV = tick
 		if ll := len(c.cache); ll > 0 {
-			if c.l.Enabled(context.Background(), slog.LevelDebug) {
-				c.l.Debug("resetting conntrack cache", "len", ll)
+			if c.l.Enabled(context.Background(), logging.LevelTrace) {
+				c.l.Log(context.Background(), logging.LevelTrace, "resetting conntrack cache", "len", ll)
 			}
 			c.cache = make(ConntrackCache, ll)
 		}

firewall/cache_test.go

@@ -6,6 +6,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/slackhq/nebula/logging"
 	"github.com/slackhq/nebula/test"
 	"github.com/stretchr/testify/assert"
 )
@@ -30,27 +31,27 @@ func newFixedTicker(t *testing.T, l *slog.Logger, cacheLen int) *ConntrackCacheT
 
 func TestConntrackCacheTicker_Get_TextFormat(t *testing.T) {
 	buf := &bytes.Buffer{}
-	l := test.NewLoggerWithOutputAndLevel(buf, slog.LevelDebug)
+	l := test.NewLoggerWithOutputAndLevel(buf, logging.LevelTrace)
 
 	c := newFixedTicker(t, l, 3)
 	c.Get()
 
-	assert.Equal(t, "level=DEBUG msg=\"resetting conntrack cache\" len=3\n", buf.String())
+	assert.Equal(t, "level=DEBUG-4 msg=\"resetting conntrack cache\" len=3\n", buf.String())
 }
 
 func TestConntrackCacheTicker_Get_JSONFormat(t *testing.T) {
 	buf := &bytes.Buffer{}
-	l := test.NewJSONLoggerWithOutput(buf, slog.LevelDebug)
+	l := test.NewJSONLoggerWithOutput(buf, logging.LevelTrace)
 
 	c := newFixedTicker(t, l, 2)
 	c.Get()
 
-	assert.JSONEq(t, `{"level":"DEBUG","msg":"resetting conntrack cache","len":2}`, strings.TrimSpace(buf.String()))
+	assert.JSONEq(t, `{"level":"DEBUG-4","msg":"resetting conntrack cache","len":2}`, strings.TrimSpace(buf.String()))
 }
 
-func TestConntrackCacheTicker_Get_QuietBelowDebug(t *testing.T) {
+func TestConntrackCacheTicker_Get_QuietBelowTrace(t *testing.T) {
 	buf := &bytes.Buffer{}
-	l := test.NewLoggerWithOutputAndLevel(buf, slog.LevelInfo)
+	l := test.NewLoggerWithOutputAndLevel(buf, slog.LevelDebug)
 
 	c := newFixedTicker(t, l, 5)
 	c.Get()
@@ -60,7 +61,7 @@ func TestConntrackCacheTicker_Get_QuietBelowDebug(t *testing.T) {
 
 func TestConntrackCacheTicker_Get_QuietWhenCacheEmpty(t *testing.T) {
 	buf := &bytes.Buffer{}
-	l := test.NewLoggerWithOutputAndLevel(buf, slog.LevelDebug)
+	l := test.NewLoggerWithOutputAndLevel(buf, logging.LevelTrace)
 
 	c := newFixedTicker(t, l, 0)
 	c.Get()

firewall_test.go

@@ -1029,75 +1029,6 @@ func Test_parsePort(t *testing.T) {
 	require.NoError(t, err)
 }
 
-// Test_parsePort_invalid covers inputs that must error. The named bug is
-// that int32(strconv.Atoi("4294967296")) truncates to 0 == firewall.PortAny,
-// silently turning a typo into a match-all-ports rule; the rest are
-// representative syntax/range probes.
-func Test_parsePort_invalid(t *testing.T) {
-	tests := []struct {
-		name            string
-		input           string
-		wantErrContains string
-	}{
-		// Numeric overflow (the named bug + boundary).
-		{"named bug: 2^32 truncates to PortAny", "4294967296", "out of range"},
-		{"just above max real port", "65536", "out of range"},
-
-		// Negatives route through the range branch and hit the empty-half
-		// guard; included as defense in depth so a future refactor cannot
-		// accidentally reach the int32 cast.
-		{"negative", "-1", "could not be parsed"},
-
-		// Syntax probes.
-		{"NUL between digits", "4\x002", "was not a number"},
-		{"hex notation", "0x10", "was not a number"},
-		{"scientific notation", "1e3", "was not a number"},
-		{"leading whitespace", " 42", "was not a number"},
-		{"fullwidth digits", "４２", "was not a number"},
-
-		// Range branch.
-		{"range upper out of range", "1-65536", "ending range out of range"},
-		{"range lower out of range", "65536-65537", "beginning range out of range"},
-		{"range with negative upper", "1--1", "ending range was not a number"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			_, _, err := parsePort(tc.input)
-			require.Error(t, err, "input %q must error", tc.input)
-			require.ErrorContains(t, err, tc.wantErrContains)
-		})
-	}
-}
-
-// Test_parsePort_valid_boundaries locks in success cases at 0, 1, and 65535
-// so a future refactor cannot regress the boundaries.
-func Test_parsePort_valid_boundaries(t *testing.T) {
-	tests := []struct {
-		name      string
-		input     string
-		wantStart int32
-		wantEnd   int32
-	}{
-		{"zero is PortAny", "0", 0, 0},
-		{"min real port", "1", 1, 1},
-		{"max real port", "65535", 65535, 65535},
-		{"range zero to max forces end to zero", "0-65535", 0, 0},
-		{"range max to max", "65535-65535", 65535, 65535},
-		{"range one to max", "1-65535", 1, 65535},
-		{"range with whitespace inside", " 1 - 2    ", 1, 2},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			s, e, err := parsePort(tc.input)
-			require.NoError(t, err)
-			assert.Equal(t, tc.wantStart, s, "start port")
-			assert.Equal(t, tc.wantEnd, e, "end port")
-		})
-	}
-}
-
 func TestNewFirewallFromConfig(t *testing.T) {
 	l := test.NewLogger()
 	// Test a bad rule definition

go.mod

@@ -9,7 +9,7 @@ require (
 	github.com/armon/go-radix v1.0.0
 	github.com/cyberdelia/go-metrics-graphite v0.0.0-20161219230853-39f87cc3b432
 	github.com/flynn/noise v1.1.0
-	github.com/gaissmai/bart v0.28.0
+	github.com/gaissmai/bart v0.26.1
 	github.com/gogo/protobuf v1.3.2
 	github.com/google/gopacket v1.1.19
 	github.com/kardianos/service v1.2.4
@@ -24,12 +24,12 @@ require (
 	github.com/vishvananda/netlink v1.3.1
 	go.uber.org/goleak v1.3.0
 	go.yaml.in/yaml/v3 v3.0.4
-	golang.org/x/crypto v0.53.0
+	golang.org/x/crypto v0.50.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.56.0
-	golang.org/x/sync v0.21.0
-	golang.org/x/sys v0.46.0
-	golang.org/x/term v0.44.0
+	golang.org/x/net v0.53.0
+	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 v0.6.1

go.sum

@@ -26,8 +26,8 @@ github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/flynn/noise v1.1.0 h1:KjPQoQCEFdZDiP03phOvGi11+SVVhBG2wOWAorLsstg=
 github.com/flynn/noise v1.1.0/go.mod h1:xbMo+0i6+IGbYdJhF31t2eR1BIU0CYc12+BNAKwUTag=
-github.com/gaissmai/bart v0.28.0 h1:89yZLo8NmyqD0RYgJ3QO9HhqqGGw+oWhf90cZm69Lko=
-github.com/gaissmai/bart v0.28.0/go.mod h1:GREWQfTLRWz/c5FTOsIw+KkscuFkIV5t8Rp7Nd1Td5c=
+github.com/gaissmai/bart v0.26.1 h1:+w4rnLGNlA2GDVn382Tfe3jOsK5vOr5n4KmigJ9lbTo=
+github.com/gaissmai/bart v0.26.1/go.mod h1:GREWQfTLRWz/c5FTOsIw+KkscuFkIV5t8Rp7Nd1Td5c=
 github.com/go-kit/kit v0.8.0/go.mod h1:xBxKIO96dXMWWy0MnWVtmwkA9/13aqxPnvrjFYMA2as=
 github.com/go-kit/kit v0.9.0/go.mod h1:xBxKIO96dXMWWy0MnWVtmwkA9/13aqxPnvrjFYMA2as=
 github.com/go-kit/log v0.1.0/go.mod h1:zbhenjAZHb184qTLMA9ZjW7ThYL0H2mk7Q6pNt4vbaY=
@@ -162,8 +162,8 @@ 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 v0.53.0 h1:QZ4Muo8THX6CizN2vPPd5fBGHyogrdK9fG4wLPFUsto=
-golang.org/x/crypto v0.53.0/go.mod h1:DNLU434OwVakk9PzuwV8w62mAJpRJL3vsgcfp4Qnsio=
+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=
@@ -182,8 +182,8 @@ 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 v0.56.0 h1:Rw8j/hFzGvJUZwNBXnAtf5sVDVt+65SK2C7IxCxZt5o=
-golang.org/x/net v0.56.0/go.mod h1:D3Ku6r+V6JROoZK144D2XfMHFcMq/0zSfLelVTCFKec=
+golang.org/x/net v0.53.0 h1:d+qAbo5L0orcWAr0a9JweQpjXF19LMXJE8Ey7hwOdUA=
+golang.org/x/net v0.53.0/go.mod h1:JvMuJH7rrdiCfbeHoo3fCQU24Lf5JJwT9W3sJFulfgs=
 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=
@@ -191,8 +191,8 @@ 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 v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+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=
@@ -208,11 +208,11 @@ 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 v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+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 v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=
-golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=
+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=

handshake/errors.go

@@ -13,7 +13,6 @@ var (
 	ErrUnknownSubtype          = errors.New("unknown handshake subtype")
 	ErrMissingContent          = errors.New("expected handshake content but message was empty")
 	ErrUnexpectedContent       = errors.New("received unexpected handshake content")
-	ErrInvalidRemoteIndex      = errors.New("peer sent an invalid index in handshake payload")
 	ErrIndexAllocation         = errors.New("failed to allocate local index")
 	ErrNoCredential            = errors.New("no handshake credential available for cert version")
 	ErrAsymmetricCipherKeys    = errors.New("noise produced only one cipher key")

handshake/machine.go

@@ -312,19 +312,11 @@ func (m *Machine) processPayload(msg []byte, flags msgFlags) error {
 
 	// Process payload
 	if flags.expectsPayload {
-		var remoteIndex uint32
 		if m.result.Initiator {
-			remoteIndex = payload.ResponderIndex
+			m.result.RemoteIndex = payload.ResponderIndex
 		} else {
-			remoteIndex = payload.InitiatorIndex
+			m.result.RemoteIndex = payload.InitiatorIndex
 		}
-		// The payload presence check above can be satisfied by Time alone, so a payload
-		// could still carry a zero index here. We need to reject it.
-		if remoteIndex == 0 {
-			m.failed = true
-			return ErrInvalidRemoteIndex
-		}
-		m.result.RemoteIndex = remoteIndex
 		m.result.HandshakeTime = payload.Time
 		m.payloadSet = true
 	}

handshake/machine_test.go

@@ -229,24 +229,6 @@ func TestMachineProcessPayload(t *testing.T) {
 		require.ErrorIs(t, err, ErrUnexpectedContent)
 		assert.True(t, m.Failed())
 	})
-
-	t.Run("zero initiator index on responder is fatal", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, false, 100)
-		bytes := MarshalPayload(nil, Payload{InitiatorIndex: 0, Time: 1})
-		err := m.processPayload(bytes, msgFlags{expectsPayload: true})
-		require.ErrorIs(t, err, ErrInvalidRemoteIndex)
-		assert.True(t, m.Failed())
-		assert.Zero(t, m.result.RemoteIndex)
-	})
-
-	t.Run("zero responder index on initiator is fatal", func(t *testing.T) {
-		m := newTestMachine(t, cs, v, true, 100)
-		bytes := MarshalPayload(nil, Payload{InitiatorIndex: 100, ResponderIndex: 0, Time: 1})
-		err := m.processPayload(bytes, msgFlags{expectsPayload: true})
-		require.ErrorIs(t, err, ErrInvalidRemoteIndex)
-		assert.True(t, m.Failed())
-		assert.Zero(t, m.result.RemoteIndex)
-	})
 }
 
 // TestMachineRequireComplete checks the fail-on-incomplete-handshake path

handshake_manager.go

@@ -83,7 +83,6 @@ type HandshakeHostInfo struct {
 	initiatingVersionOverride cert.Version     // Should we use a non-default cert version for this handshake?
 	counter                   int64            // How many attempts have we made so far
 	lastRemotes               []netip.AddrPort // Remotes that we sent to during the previous attempt
-	lastRelays                []netip.Addr     // Relays we attempted to use during the previous attempt
 	packetStore               []*cachedPacket  // A set of packets to be transmitted once the handshake completes
 
 	hostinfo *HostInfo
@@ -218,6 +217,7 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		fields := []any{
 			"udpAddrs", hh.hostinfo.remotes.CopyAddrs(hm.mainHostMap.GetPreferredRanges()),
 			"initiatorIndex", hh.hostinfo.localIndexId,
+			"remoteIndex", hh.hostinfo.remoteIndexId,
 			"durationNs", time.Since(hh.startTime).Nanoseconds(),
 		}
 		// hh.machine can be nil here if buildStage0Packet never succeeded
@@ -323,7 +323,7 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		)
 	}
 
-	hm.f.relayManager.StartRelays(hm.f, vpnIp, hh, stage0)
+	hm.f.relayManager.StartRelays(hm.f, vpnIp, hostinfo, stage0)
 
 	// If a lighthouse triggered this attempt then we are still in the timer wheel and do not need to re-add
 	if !lighthouseTriggered {
@@ -465,6 +465,7 @@ func (hm *HandshakeManager) CheckAndComplete(hostinfo *HostInfo, handshakePacket
 		// We have a collision, but this can happen since we can't control
 		// the remote ID. Just log about the situation as a note.
 		hostinfo.logger(hm.l).Info("New host shadows existing host remoteIndex",
+			"remoteIndex", hostinfo.remoteIndexId,
 			"collision", existingRemoteIndex.vpnAddrs,
 		)
 	}
@@ -487,6 +488,7 @@ func (hm *HandshakeManager) Complete(hostinfo *HostInfo, f *Interface) {
 		// We have a collision, but this can happen since we can't control
 		// the remote ID. Just log about the situation as a note.
 		hostinfo.logger(hm.l).Info("New host shadows existing host remoteIndex",
+			"remoteIndex", hostinfo.remoteIndexId,
 			"collision", existingRemoteIndex.vpnAddrs,
 		)
 	}
@@ -796,6 +798,7 @@ func (hm *HandshakeManager) beginHandshake(via ViaSender, packet []byte, h *head
 	}
 
 	hm.sendHandshakeResponse(via, response, hostinfo, false)
+	f.connectionManager.AddTrafficWatch(hostinfo)
 	hostinfo.remotes.RefreshFromHandshake(vpnAddrs)
 
 	// Don't wait for UpdateWorker
@@ -962,6 +965,7 @@ func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostIn
 	hostinfo.buildNetworks(f.myVpnNetworksTable, remoteCert.Certificate)
 
 	hm.Complete(hostinfo, f)
+	f.connectionManager.AddTrafficWatch(hostinfo)
 
 	if len(hh.packetStore) > 0 {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -970,6 +974,7 @@ func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostIn
 		nb := make([]byte, 12, 12)
 		out := make([]byte, mtu)
 		for _, cp := range hh.packetStore {
+			//todo use a sendbatcher
 			cp.callback(cp.messageType, cp.messageSubType, hostinfo, cp.packet, nb, out)
 		}
 		f.cachedPacketMetrics.sent.Inc(int64(len(hh.packetStore)))

hostmap.go

@@ -138,9 +138,9 @@ func (rs *RelayState) InsertRelayTo(ip netip.Addr) {
 }
 
 func (rs *RelayState) CopyRelayIps() []netip.Addr {
+	ret := make([]netip.Addr, len(rs.relays))
 	rs.RLock()
 	defer rs.RUnlock()
-	ret := make([]netip.Addr, len(rs.relays))
 	copy(ret, rs.relays)
 	return ret
 }
@@ -623,11 +623,6 @@ func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
 	hm.Indexes[hostinfo.localIndexId] = hostinfo
 	hm.RemoteIndexes[hostinfo.remoteIndexId] = hostinfo
 
-	hostinfo.out.Store(true)
-	if f.connectionManager != nil { // f.connectionManager is only nil in some unit tests
-		f.connectionManager.trafficTimer.Add(hostinfo.localIndexId, f.connectionManager.checkInterval)
-	}
-
 	if hm.l.Enabled(context.Background(), slog.LevelDebug) {
 		hm.l.Debug("Hostmap vpnIp added",
 			"hostMap", m{"vpnAddrs": hostinfo.vpnAddrs, "mapTotalSize": len(hm.Hosts),

inside.go

@@ -2,6 +2,7 @@ package nebula
 
 import (
 	"context"
+	"io"
 	"log/slog"
 	"net/netip"
 
@@ -9,10 +10,17 @@ import (
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/iputil"
 	"github.com/slackhq/nebula/noiseutil"
+	"github.com/slackhq/nebula/overlay/batch"
 	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/wire"
 )
 
-func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) consumeInsidePacket(pkt wire.TunPacket, fwPacket *firewall.Packet, nb []byte, sendBatch *batch.SendBatch, rejectBuf []byte, q int, localCache firewall.ConntrackCache) {
+	// pkt.Bytes is either one IP datagram (GSO zero) or a TSO/USO
+	// superpacket. In both cases the L3+L4 headers at the start describe
+	// the same 5-tuple every segment will share, so a single newPacket /
+	// firewall check covers the whole superpacket.
+	packet := pkt.Bytes
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -37,7 +45,10 @@ 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 := pkt.PerSegment(func(seg []byte) error {
+				_, werr := f.readers[q].Write(seg)
+				return werr
+			})
 			if err != nil {
 				f.l.Error("Failed to forward to tun", "error", err)
 			}
@@ -53,11 +64,23 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	}
 
 	hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
-		hh.cachePacket(f.l, header.Message, 0, packet, f.sendMessageNow, f.cachedPacketMetrics)
+		// borrowed: PerSegment builds each segment in the kernel-supplied pkt
+		// bytes underneath. cachePacket explicitly copies its argument (handshake_manager.go cachePacket),
+		// so retaining segments past the loop is safe.
+		err := pkt.PerSegment(func(seg []byte) error {
+			hh.cachePacket(f.l, header.Message, 0, seg, f.sendMessageNow, f.cachedPacketMetrics)
+			return nil
+		})
+		if err != nil && f.l.Enabled(context.Background(), slog.LevelDebug) {
+			f.l.Debug("Failed to segment superpacket for handshake cache",
+				"error", err,
+				"vpnAddr", fwPacket.RemoteAddr,
+			)
+		}
 	})
 
 	if hostinfo == nil {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			f.l.Debug("dropping outbound packet, vpnAddr not in our vpn networks or in unsafe networks",
 				"vpnAddr", fwPacket.RemoteAddr,
@@ -73,10 +96,9 @@ 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, pkt, nb, sendBatch)
 	} else {
-		f.rejectInside(packet, out, q)
+		f.rejectInside(packet, rejectBuf, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("dropping outbound packet",
 				"fwPacket", fwPacket,
@@ -86,6 +108,149 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 	}
 }
 
+func (f *Interface) sendInsideEncrypt(hostinfo *HostInfo, ci *ConnectionState, seg, scratch, nb []byte) []byte {
+	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)
+
+	out, encErr := ci.eKey.EncryptDanger(out, out, seg, c, nb)
+	if noiseutil.EncryptLockNeeded {
+		ci.writeLock.Unlock()
+	}
+	if encErr != nil {
+		hostinfo.logger(f.l).Error("Failed to encrypt outgoing packet",
+			"error", encErr,
+			"udpAddr", hostinfo.remote,
+			"counter", c,
+		)
+		// Skip this segment; the rest of the superpacket can still
+		// go out — TCP will retransmit anything we drop here.
+		return nil
+	}
+
+	return out
+}
+
+// sendInsideMessage encrypts a firewall-approved inside packet (or every
+// segment of a TSO/USO superpacket) into the caller's batch slot for
+// later sendmmsg flush. Segmentation is fused with encryption here so the
+// kernel-supplied superpacket bytes never get written into a separate
+// scratch arena: PerSegment builds each segment's plaintext in
+// segScratch[:segLen] in turn, and we encrypt directly into a fresh
+// SendBatch slot.
+func (f *Interface) sendInsideMessage(hostinfo *HostInfo, pkt wire.TunPacket, nb []byte, sendBatch *batch.SendBatch) {
+	ci := hostinfo.ConnectionState
+	if ci.eKey == nil {
+		return
+	}
+
+	ecnEnabled := f.ecnEnabled.Load()
+	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.Enabled(context.Background(), slog.LevelDebug) {
+			hostinfo.logger(f.l).Debug("Lighthouse update triggered for punch due to rebind counter",
+				"vpnAddrs", hostinfo.vpnAddrs,
+			)
+		}
+	}
+
+	if !hostinfo.remote.IsValid() { //the relay path
+		//first, find our relay hostinfo:
+		var relayHostInfo *HostInfo
+		var relay *Relay
+		var err error
+		for _, relayIP := range hostinfo.relayState.CopyRelayIps() {
+			relayHostInfo, relay, err = f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relayIP)
+			if err != nil {
+				hostinfo.relayState.DeleteRelay(relayIP)
+				hostinfo.logger(f.l).Info("sendNoMetrics failed to find HostInfo",
+					"relay", relayIP,
+					"error", err,
+				)
+				continue
+			}
+			break
+		}
+		if relayHostInfo == nil || relay == nil {
+			//failure already logged
+			return
+		}
+
+		err = pkt.PerSegment(func(seg []byte) error {
+			//relay header + header + plaintext + AEAD tag (16 bytes for both AES-GCM and ChaCha20-Poly1305) + relay tag
+			scratch := sendBatch.Reserve(header.Len + header.Len + len(seg) + 16 + 16)
+
+			innerPacket := f.sendInsideEncrypt(hostinfo, ci, seg, scratch[header.Len:], nb)
+			if innerPacket == nil {
+				return nil
+			}
+
+			//now we need to do a relay-encrypt:
+			toSend, err := f.prepareSendVia(relayHostInfo, relay, innerPacket, nb, scratch, true)
+			if err != nil {
+				//already logged
+				return nil
+			}
+
+			var ecn byte
+			if ecnEnabled {
+				ecn = innerECN(seg)
+			}
+			sendBatch.Commit(toSend, relayHostInfo.remote, ecn)
+			return nil
+		})
+		if err != nil {
+			hostinfo.logger(f.l).Error("Failed to segment superpacket for relay send", "error", err)
+		}
+		return
+	}
+
+	err := pkt.PerSegment(func(seg []byte) error {
+		// header + plaintext + AEAD tag (16 bytes for both AES-GCM and ChaCha20-Poly1305)
+		scratch := sendBatch.Reserve(header.Len + len(seg) + 16)
+
+		out := f.sendInsideEncrypt(hostinfo, ci, seg, scratch, nb)
+		if out == nil {
+			return nil
+		}
+
+		var ecn byte
+		if ecnEnabled {
+			ecn = innerECN(seg)
+		}
+		sendBatch.Commit(out, hostinfo.remote, ecn)
+		return nil
+	})
+	if err != nil {
+		hostinfo.logger(f.l).Error("Failed to segment superpacket for send",
+			"error", err,
+		)
+	}
+}
+
+// innerECN returns the 2-bit IP-level ECN codepoint of an inner IPv4 or IPv6
+// packet, or 0 if pkt is too short or its IP version is unrecognized. Used at
+// encap to copy the inner codepoint onto the outer carrier per RFC 6040.
+func innerECN(pkt []byte) byte {
+	if len(pkt) < 2 {
+		return 0
+	}
+	switch pkt[0] >> 4 {
+	case 4:
+		return pkt[1] & 0x03
+	case 6:
+		return (pkt[1] >> 4) & 0x03
+	}
+	return 0
+}
+
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 	if !f.firewall.InSendReject {
 		return
@@ -275,21 +440,13 @@ func (f *Interface) sendTo(t header.MessageType, st header.MessageSubType, ci *C
 	f.sendNoMetrics(t, st, ci, hostinfo, remote, p, nb, out, 0)
 }
 
-// SendVia sends a payload through a Relay tunnel. No authentication or encryption is done
-// to the payload for the ultimate target host, making this a useful method for sending
-// handshake messages to peers through relay tunnels.
-// via is the HostInfo through which the message is relayed.
-// ad is the plaintext data to authenticate, but not encrypt
-// nb is a buffer used to store the nonce value, re-used for performance reasons.
-// out is a buffer used to store the result of the Encrypt operation
-// q indicates which writer to use to send the packet.
-func (f *Interface) SendVia(via *HostInfo,
+func (f *Interface) prepareSendVia(via *HostInfo,
 	relay *Relay,
 	ad,
 	nb,
 	out []byte,
 	nocopy bool,
-) {
+) ([]byte, error) {
 	if noiseutil.EncryptLockNeeded {
 		// NOTE: for goboring AESGCMTLS we need to lock because of the nonce check
 		via.ConnectionState.writeLock.Lock()
@@ -311,7 +468,7 @@ func (f *Interface) SendVia(via *HostInfo,
 			"headerLen", len(out),
 			"cipherOverhead", via.ConnectionState.eKey.Overhead(),
 		)
-		return
+		return nil, io.ErrShortBuffer
 	}
 
 	// The header bytes are written to the 'out' slice; Grow the slice to hold the header and associated data payload.
@@ -331,13 +488,37 @@ func (f *Interface) SendVia(via *HostInfo,
 	}
 	if err != nil {
 		via.logger(f.l).Info("Failed to EncryptDanger in sendVia", "error", err)
+		return nil, err
+	}
+	f.connectionManager.RelayUsed(relay.LocalIndex)
+	return out, nil
+}
+
+// SendVia sends a payload through a Relay tunnel. No authentication or encryption is done
+// to the payload for the ultimate target host, making this a useful method for sending
+// handshake messages to peers through relay tunnels.
+// via is the HostInfo through which the message is relayed.
+// ad is the plaintext data to authenticate, but not encrypt
+// nb is a buffer used to store the nonce value, re-used for performance reasons.
+// out is a buffer used to store the result of the Encrypt operation.
+// The write goes through writers[0] — SendVia is called from contexts
+// without a per-queue index (handshake, async control paths).
+func (f *Interface) SendVia(via *HostInfo,
+	relay *Relay,
+	ad,
+	nb,
+	out []byte,
+	nocopy bool,
+) {
+	toSend, err := f.prepareSendVia(via, relay, ad, nb, out, nocopy)
+	if err != nil {
+		via.logger(f.l).Info("Failed to prepareSendVia", "error", err)
 		return
 	}
-	err = f.writers[0].WriteTo(out, via.remote)
+	err = f.writers[0].WriteTo(toSend, via.remote)
 	if err != nil {
 		via.logger(f.l).Info("Failed to WriteTo in sendVia", "error", err)
 	}
-	f.connectionManager.RelayUsed(relay.LocalIndex)
 }
 
 func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType, ci *ConnectionState, hostinfo *HostInfo, remote netip.AddrPort, p, nb, out []byte, q int) {
@@ -391,6 +572,7 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 			"error", err,
 			"udpAddr", remote,
 			"counter", c,
+			"attemptedCounter", c,
 		)
 		return
 	}

interface.go

@@ -4,22 +4,24 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
-	"slices"
+	"runtime"
 	"sync"
 	"sync/atomic"
 	"time"
 
 	"github.com/gaissmai/bart"
 	"github.com/rcrowley/go-metrics"
+	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
 
-	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/overlay"
+	"github.com/slackhq/nebula/overlay/batch"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/udp"
 )
 
@@ -39,6 +41,7 @@ type InterfaceConfig struct {
 	DropLocalBroadcast bool
 	DropMulticast      bool
 	routines           int
+	batchSize          int
 	MessageMetrics     *MessageMetrics
 	version            string
 	relayManager       *relayManager
@@ -49,7 +52,14 @@ type InterfaceConfig struct {
 	reQueryWait     time.Duration
 
 	ConntrackCacheTimeout time.Duration
-	l                     *slog.Logger
+
+	// CpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
+	// should pin to. Queue i pins to CpuAffinity[i % len(CpuAffinity)] —
+	// shorter lists than `routines` cycle. Empty list keeps the default
+	// pin-to-(i % NumCPU) behavior.
+	CpuAffinity []int
+
+	l *slog.Logger
 }
 
 type Interface struct {
@@ -71,9 +81,19 @@ type Interface struct {
 	dropLocalBroadcast    bool
 	dropMulticast         bool
 	routines              int
+	batchSize             int
 	disconnectInvalid     atomic.Bool
 	closed                atomic.Bool
-	relayManager          *relayManager
+	// cpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
+	// should pin to. Queue i pins to cpuAffinity[i % len(cpuAffinity)].
+	// Empty falls back to the default pin-to-(i % NumCPU) behavior.
+	cpuAffinity []int
+	// ecnEnabled gates RFC 6040 underlay ECN propagation. When true,
+	// inside.go copies the inner ECN onto the outer carrier on encap and
+	// decryptToTun folds outer CE into the inner header on decap. Toggle
+	// via tunnels.ecn (default true).
+	ecnEnabled   atomic.Bool
+	relayManager *relayManager
 
 	tryPromoteEvery atomic.Uint32
 	reQueryEvery    atomic.Uint32
@@ -90,8 +110,12 @@ type Interface struct {
 
 	ctx     context.Context
 	writers []udp.Conn
-	readers []io.ReadWriteCloser
-	wg      sync.WaitGroup
+	readers []tio.Queue
+	// batchers is one per tun queue, wrapping readers[i].
+	// decryptToTun sends plaintext into the batch.RxBatcher;
+	// listenOut calls its Flush at the end of each UDP recvmmsg batch.
+	batchers []batch.RxBatcher
+	wg       sync.WaitGroup
 
 	// fatalErr holds the first unexpected reader error that caused shutdown.
 	// nil means "no fatal error" (yet)
@@ -187,9 +211,11 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		dropLocalBroadcast:    c.DropLocalBroadcast,
 		dropMulticast:         c.DropMulticast,
 		routines:              c.routines,
+		batchSize:             c.batchSize,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		readers:               make([]io.ReadWriteCloser, c.routines),
+		readers:               make([]tio.Queue, c.routines),
+		batchers:              make([]batch.RxBatcher, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -198,6 +224,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		relayManager:          c.relayManager,
 		connectionManager:     c.connectionManager,
 		conntrackCacheTimeout: c.ConntrackCacheTimeout,
+		cpuAffinity:           c.CpuAffinity,
 
 		metricHandshakes: metrics.GetOrRegisterHistogram("handshakes", nil, metrics.NewExpDecaySample(1028, 0.015)),
 		messageMetrics:   c.MessageMetrics,
@@ -247,15 +274,27 @@ func (f *Interface) activate() error {
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
 	// Prepare n tun queues
-	var reader io.ReadWriteCloser = f.inside
 	for i := 0; i < f.routines; i++ {
 		if i > 0 {
-			reader, err = f.inside.NewMultiQueueReader()
-			if err != nil {
+			if err = f.inside.NewMultiQueueReader(); err != nil {
 				return err
 			}
 		}
-		f.readers[i] = reader
+	}
+	f.readers = f.inside.Readers()
+	for i := range f.readers {
+		caps := f.readers[i].Capabilities()
+		if caps.TSO || caps.USO {
+			// Multi-lane: TCP gets coalesced when TSO is on, UDP when USO
+			// is on, everything else (and either lane disabled) falls
+			// through to passthrough so non-IP / non-TCP-UDP traffic still
+			// reaches the TUN.
+			arena := util.NewArena(max(f.batchSize, 1) * 65535)
+			f.batchers[i] = batch.NewMultiCoalescer(f.readers[i], f.l, arena, caps.TSO, caps.USO)
+		} else {
+			arena := util.NewArena(max(f.batchSize, 1) * udp.MTU)
+			f.batchers[i] = batch.NewPassthrough(f.readers[i], f.batchSize, arena)
+		}
 	}
 
 	f.wg.Add(1) // for us to wait on Close() to return
@@ -313,14 +352,21 @@ func (f *Interface) listenOut(i int) {
 
 	ctCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 	lhh := f.lightHouse.NewRequestHandler()
-	plaintext := make([]byte, udp.MTU)
 	h := &header.H{}
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
-	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
-		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get())
-	})
+	listener := func(fromUdpAddr netip.AddrPort, payload []byte, meta udp.RxMeta) {
+		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, payload, h, fwPacket, lhh, nb, i, ctCache.Get(), meta)
+	}
+
+	flusher := func() {
+		if err := f.batchers[i].Flush(); err != nil {
+			f.l.Error("Failed to flush tun coalescer", "error", err)
+		}
+	}
+
+	err := li.ListenOut(listener, flusher)
 
 	if err != nil && !f.closed.Load() {
 		f.l.Error("Error while reading inbound packet, closing", "error", err)
@@ -330,28 +376,57 @@ func (f *Interface) listenOut(i int) {
 	f.l.Debug("underlay reader is done", "reader", i)
 }
 
-func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
-	packet := make([]byte, mtu)
-	out := make([]byte, mtu)
+func (f *Interface) listenIn(reader tio.Queue, q int) {
+	// Pinning this thread (and goroutine) to a single CPU keeps every sendmmsg from this goroutine going through the
+	// same TX ring on the nic, so the wire sees per-flow order.
+	cpu := q % runtime.NumCPU()
+	if n := len(f.cpuAffinity); n > 0 {
+		cpu = f.cpuAffinity[q%n]
+	}
+	if err := util.PinThreadToCPU(cpu); err != nil {
+		f.l.Warn("failed to pin tun reader to CPU", "queue", q, "cpu", cpu, "err", err)
+	}
+
+	const bonusInfo = 16
+	bufferScale := udp.MTU + bonusInfo
+	numTunPackets := 1
+	caps := reader.Capabilities()
+	if caps.TSO || caps.USO {
+		bufferScale = 65535 + bonusInfo
+		numTunPackets = f.batchSize
+	}
+
+	rejectBuf := make([]byte, mtu)
+	tunPackets := make([]wire.TunPacket, numTunPackets)
+	packetMem := make([]byte, bufferScale*numTunPackets)
+
+	arenaSize := batch.SendBatchCap * (udp.MTU + 32)
+	sb := batch.NewSendBatch(f.writers[q], batch.SendBatchCap, util.NewArena(arenaSize))
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
 	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.Read(packet)
+		n, err := reader.Read(tunPackets, packetMem)
 		if err != nil {
 			if !f.closed.Load() {
-				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", i)
+				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", q)
 				f.onFatal(err)
 			}
 			break
 		}
 
-		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get())
+		ctCache := conntrackCache.Get()
+		for i := range n {
+			f.consumeInsidePacket(tunPackets[i], fwPacket, nb, sb, rejectBuf, q, ctCache)
+		}
+		if err := sb.Flush(); err != nil {
+			f.l.Error("Failed to write outgoing batch", "error", err, "writer", q)
+		}
 	}
 
-	f.l.Debug("overlay reader is done", "reader", i)
+	f.l.Debug("overlay reader is done", "reader", q)
 }
 
 func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
@@ -360,6 +435,7 @@ func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
 	c.RegisterReloadCallback(f.reloadAcceptRecvError)
 	c.RegisterReloadCallback(f.reloadDisconnectInvalid)
 	c.RegisterReloadCallback(f.reloadMisc)
+	c.RegisterReloadCallback(f.reloadEcn)
 
 	for _, udpConn := range f.writers {
 		c.RegisterReloadCallback(udpConn.ReloadConfig)
@@ -377,22 +453,13 @@ func (f *Interface) reloadDisconnectInvalid(c *config.C) {
 }
 
 func (f *Interface) reloadFirewall(c *config.C) {
-	cs := f.pki.getCertState()
-	curCert := cs.getCertificate(cert.Version2)
-	if curCert == nil {
-		curCert = cs.getCertificate(cert.Version1)
-	}
-
-	// The firewall builds its routableNetworks set from the certificate's UnsafeNetworks at construction.
-	// Check to see if that set has changed, and if so, rebuild the firewall.
-	certUnsafeChanged := curCert != nil && !slices.Equal(curCert.UnsafeNetworks(), f.firewall.unsafeNetworks)
-
-	if !c.HasChanged("firewall") && !certUnsafeChanged {
+	//TODO: need to trigger/detect if the certificate changed too
+	if c.HasChanged("firewall") == false {
 		f.l.Debug("No firewall config change detected")
 		return
 	}
 
-	fw, err := NewFirewallFromConfig(f.l, cs, c)
+	fw, err := NewFirewallFromConfig(f.l, f.pki.getCertState(), c)
 	if err != nil {
 		f.l.Error("Error while creating firewall during reload", "error", err)
 		return
@@ -492,6 +559,20 @@ func (f *Interface) reloadMisc(c *config.C) {
 	}
 }
 
+// reloadEcn syncs Interface.ecnEnabled with the tunnels.ecn config knob.
+// Default is enabled (RFC 6040 normal mode); set false on the rare path
+// where an underlay middlebox rewrites or drops ECN bits unpredictably.
+func (f *Interface) reloadEcn(c *config.C) {
+	initial := c.InitialLoad()
+	if initial || c.HasChanged("tunnels.ecn") {
+		v := c.GetBool("tunnels.ecn", true)
+		f.ecnEnabled.Store(v)
+		if !initial {
+			f.l.Info("tunnels.ecn changed", "enabled", v)
+		}
+	}
+}
+
 func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 	ticker := time.NewTicker(i)
 	defer ticker.Stop()
@@ -502,34 +583,26 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 	certInitiatingVersion := metrics.GetOrRegisterGauge("certificate.initiating_version", nil)
 	certMaxVersion := metrics.GetOrRegisterGauge("certificate.max_version", nil)
 
-	emit := func() {
-		f.firewall.EmitStats()
-		f.handshakeManager.EmitStats()
-		udpStats()
-
-		certState := f.pki.getCertState()
-		defaultCrt := certState.GetDefaultCertificate()
-		certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
-		certInitiatingVersion.Update(int64(defaultCrt.Version()))
-
-		// Report the max certificate version we are capable of using
-		if certState.v2Cert != nil {
-			certMaxVersion.Update(int64(certState.v2Cert.Version()))
-		} else {
-			certMaxVersion.Update(int64(certState.v1Cert.Version()))
-		}
-	}
-
-	// Prime gauges so a Prometheus scrape that lands before the first tick
-	// sees real values instead of the zero defaults (issue #907).
-	emit()
-
 	for {
 		select {
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			emit()
+			f.firewall.EmitStats()
+			f.handshakeManager.EmitStats()
+			udpStats()
+
+			certState := f.pki.getCertState()
+			defaultCrt := certState.GetDefaultCertificate()
+			certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
+			certInitiatingVersion.Update(int64(defaultCrt.Version()))
+
+			// Report the max certificate version we are capable of using
+			if certState.v2Cert != nil {
+				certMaxVersion.Update(int64(certState.v2Cert.Version()))
+			} else {
+				certMaxVersion.Update(int64(certState.v1Cert.Version()))
+			}
 		}
 	}
 }

interface_emit_test.go

@@ -1,73 +0,0 @@
-//go:build linux || darwin
-
-package nebula
-
-import (
-	"context"
-	"net/netip"
-	"testing"
-	"time"
-
-	"github.com/rcrowley/go-metrics"
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/firewall"
-	"github.com/slackhq/nebula/overlay/overlaytest"
-	"github.com/slackhq/nebula/test"
-	"github.com/slackhq/nebula/udp"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// Test_emitStats_primesGauges covers issue #907: a Prometheus scrape that
-// landed before the first ticker fire used to read 0 for the cert gauges.
-// emitStats now primes the gauges before entering the ticker loop. We assert
-// the gauge is zero before the first call and non-zero after.
-func Test_emitStats_primesGauges(t *testing.T) {
-	defer metrics.DefaultRegistry.UnregisterAll()
-
-	l := test.NewLogger()
-	hostMap := newHostMap(l)
-	preferredRanges := []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")}
-	hostMap.preferredRanges.Store(&preferredRanges)
-
-	notAfter := time.Now().Add(time.Hour)
-	cs := &CertState{
-		initiatingVersion: cert.Version1,
-		privateKey:        []byte{},
-		v1Cert:            &dummyCert{version: cert.Version1, notAfter: notAfter},
-		v1Credential:      nil,
-	}
-
-	lh := newTestLighthouse()
-	ifce := &Interface{
-		hostMap:          hostMap,
-		inside:           &overlaytest.NoopTun{},
-		outside:          &udp.NoopConn{},
-		firewall:         &Firewall{Conntrack: &FirewallConntrack{Conns: map[firewall.Packet]*conn{}}},
-		lightHouse:       lh,
-		pki:              &PKI{},
-		handshakeManager: NewHandshakeManager(l, hostMap, lh, &udp.NoopConn{}, defaultHandshakeConfig),
-		l:                l,
-		// On linux, udp.NewUDPStatsEmitter indexes writers[0] and asserts to
-		// *udp.StdConn. A zero value works: getMemInfo sees a nil rawConn,
-		// returns an error, and the emitter falls through to a no-op.
-		writers: []udp.Conn{&udp.StdConn{}},
-	}
-	ifce.pki.cs.Store(cs)
-
-	ttlGauge := metrics.GetOrRegisterGauge("certificate.ttl_seconds", nil)
-	require.Zero(t, ttlGauge.Value(), "gauge should be zero before emitStats runs")
-
-	// Pre-cancel the context so emitStats returns after priming the gauges
-	// without ever reading from ticker.C. The one hour interval is just a
-	// belt-and-suspenders, the test does not expect the ticker to fire.
-	ctx, cancel := context.WithCancel(context.Background())
-	cancel()
-	ifce.emitStats(ctx, time.Hour)
-
-	ttl := ttlGauge.Value()
-	assert.Positive(t, ttl, "ttl gauge should be primed by emitStats before its first tick")
-	assert.LessOrEqual(t, ttl, int64(3600))
-	assert.Equal(t, int64(cert.Version1), metrics.GetOrRegisterGauge("certificate.initiating_version", nil).Value())
-	assert.Equal(t, int64(cert.Version1), metrics.GetOrRegisterGauge("certificate.max_version", nil).Value())
-}

interface_test.go

@@ -1,120 +0,0 @@
-package nebula
-
-import (
-	"net/netip"
-	"testing"
-
-	"github.com/slackhq/nebula/cert"
-	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/test"
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// TestReloadFirewall_CertUnsafeNetworksChanged verifies that reloadFirewall
-// rebuilds the firewall when only the certificate's UnsafeNetworks have changed,
-// even if the firewall section of the YAML has not.
-func TestReloadFirewall_CertUnsafeNetworksChanged(t *testing.T) {
-	l := test.NewLogger()
-
-	vpnNet := netip.MustParsePrefix("10.0.0.1/24")
-	initialUnsafe := []netip.Prefix{netip.MustParsePrefix("198.51.100.0/24")}
-
-	// dummyCert avoids dragging the real signing pipeline into a unit test.
-	c1 := &dummyCert{
-		version:        cert.Version2,
-		networks:       []netip.Prefix{vpnNet},
-		unsafeNetworks: initialUnsafe,
-	}
-	pki := &PKI{}
-	pki.cs.Store(&CertState{v2Cert: c1, initiatingVersion: cert.Version2})
-
-	rawYAML := `firewall:
-  outbound:
-    - port: any
-      proto: any
-      host: any
-  inbound:
-    - port: any
-      proto: any
-      host: any
-`
-	cfg := config.NewC(l)
-	require.NoError(t, cfg.LoadString(rawYAML))
-
-	fw, err := NewFirewallFromConfig(l, pki.getCertState(), cfg)
-	require.NoError(t, err)
-	require.Equal(t, initialUnsafe, fw.unsafeNetworks)
-
-	f := &Interface{
-		pki:      pki,
-		firewall: fw,
-		l:        l,
-	}
-
-	// Swap the cert with a different UnsafeNetworks set.
-	newUnsafe := []netip.Prefix{
-		netip.MustParsePrefix("198.51.100.0/24"),
-		netip.MustParsePrefix("203.0.113.0/24"),
-	}
-	c2 := &dummyCert{
-		version:        cert.Version2,
-		networks:       []netip.Prefix{vpnNet},
-		unsafeNetworks: newUnsafe,
-	}
-	pki.cs.Store(&CertState{v2Cert: c2, initiatingVersion: cert.Version2})
-
-	// Reload with the same YAML so HasChanged("firewall") reports false.
-	require.NoError(t, cfg.ReloadConfigString(rawYAML))
-	require.False(t, cfg.HasChanged("firewall"))
-
-	f.reloadFirewall(cfg)
-
-	assert.NotSame(t, fw, f.firewall, "firewall pointer should have been replaced")
-	assert.Equal(t, newUnsafe, f.firewall.unsafeNetworks)
-	assert.True(t, f.firewall.routableNetworks.Contains(netip.MustParseAddr("203.0.113.5")))
-}
-
-// TestReloadFirewall_NoChange verifies that reloadFirewall is a no-op when
-// neither the firewall config nor the cert's UnsafeNetworks have changed.
-func TestReloadFirewall_NoChange(t *testing.T) {
-	l := test.NewLogger()
-
-	vpnNet := netip.MustParsePrefix("10.0.0.1/24")
-	unsafe := []netip.Prefix{netip.MustParsePrefix("198.51.100.0/24")}
-
-	c1 := &dummyCert{
-		version:        cert.Version2,
-		networks:       []netip.Prefix{vpnNet},
-		unsafeNetworks: unsafe,
-	}
-	pki := &PKI{}
-	pki.cs.Store(&CertState{v2Cert: c1, initiatingVersion: cert.Version2})
-
-	rawYAML := `firewall:
-  outbound:
-    - port: any
-      proto: any
-      host: any
-  inbound:
-    - port: any
-      proto: any
-      host: any
-`
-	cfg := config.NewC(l)
-	require.NoError(t, cfg.LoadString(rawYAML))
-
-	fw, err := NewFirewallFromConfig(l, pki.getCertState(), cfg)
-	require.NoError(t, err)
-
-	f := &Interface{
-		pki:      pki,
-		firewall: fw,
-		l:        l,
-	}
-
-	require.NoError(t, cfg.ReloadConfigString(rawYAML))
-	f.reloadFirewall(cfg)
-
-	assert.Same(t, fw, f.firewall, "firewall should not have been replaced")
-}

iputil/packet.go

@@ -4,54 +4,26 @@ import (
 	"encoding/binary"
 
 	"golang.org/x/net/ipv4"
-	"golang.org/x/net/ipv6"
 )
 
 const (
-	// MaxIPv4RejectPacketSize is the largest IPv4 reject packet:
+	// Need 96 bytes for the largest reject packet:
 	// - 20 byte ipv4 header
 	// - 8 byte icmpv4 header
 	// - 68 byte body (60 byte max orig ipv4 header + 8 byte orig icmpv4 header)
-	maxIPv4RejectPacketSize = ipv4.HeaderLen + 8 + 60 + 8
-
-	// MaxRejectPacketSize is sized for the largest possible reject packet (IPv6):
-	// - 40 byte ipv6 header
-	// - 8 byte icmpv6 header
-	// - up to 1000 byte body (original packet, possibly truncated. We want to stay
-	//   under the MTU with Nebula overhead included)
-	maxIPv6RejectPacketSize = ipv6.HeaderLen + 8 + 1000
-
-	MaxRejectPacketSize = maxIPv6RejectPacketSize
+	MaxRejectPacketSize = ipv4.HeaderLen + 8 + 60 + 8
 )
 
 func CreateRejectPacket(packet []byte, out []byte) []byte {
-	if len(packet) < 1 {
+	if len(packet) < ipv4.HeaderLen || int(packet[0]>>4) != ipv4.Version {
 		return nil
 	}
 
-	version := int(packet[0] >> 4)
-	switch version {
-	case ipv4.Version:
-		if len(packet) < ipv4.HeaderLen {
-			return nil
-		}
-		// Do not send reject packets for non-first fragments
-		if packet[6]&0x1f != 0 || packet[7] != 0 {
-			return nil
-		}
-		switch packet[9] {
-		case 6: // tcp
-			return ipv4CreateRejectTCPPacket(packet, out)
-		default:
-			return ipv4CreateRejectICMPPacket(packet, out)
-		}
-	case ipv6.Version:
-		if len(packet) < ipv6.HeaderLen {
-			return nil
-		}
-		return ipv6CreateRejectPacket(packet, out)
+	switch packet[9] {
+	case 6: // tcp
+		return ipv4CreateRejectTCPPacket(packet, out)
 	default:
-		return nil
+		return ipv4CreateRejectICMPPacket(packet, out)
 	}
 }
 
@@ -63,16 +35,11 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
 		return nil
 	}
 
-	// Do not generate ICMP errors in response to ICMP error packets
-	if packet[9] == 1 && len(packet) > ihl {
-		icmpType := packet[ihl]
-		if icmpType == 3 || icmpType == 4 || icmpType == 5 || icmpType == 11 || icmpType == 12 {
-			return nil
-		}
-	}
-
 	// ICMP reply includes original header and first 8 bytes of the packet
-	packetLen := min(len(packet), ihl+8)
+	packetLen := len(packet)
+	if packetLen > ihl+8 {
+		packetLen = ihl + 8
+	}
 
 	outLen := ipv4.HeaderLen + 8 + packetLen
 	if outLen > cap(out) {
@@ -104,14 +71,14 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
 
 	// ICMP Destination Unreachable
 	icmpOut := out[ipv4.HeaderLen:]
-	icmpOut[0] = 3  // type (Destination unreachable)
-	icmpOut[1] = 13 // code (Communication administratively prohibited)
-	icmpOut[2] = 0  // checksum
-	icmpOut[3] = 0  //  .
-	icmpOut[4] = 0  // unused
-	icmpOut[5] = 0  //  .
-	icmpOut[6] = 0  //  .
-	icmpOut[7] = 0  //  .
+	icmpOut[0] = 3 // type (Destination unreachable)
+	icmpOut[1] = 3 // code (Port unreachable error)
+	icmpOut[2] = 0 // checksum
+	icmpOut[3] = 0 //  .
+	icmpOut[4] = 0 // unused
+	icmpOut[5] = 0 //  .
+	icmpOut[6] = 0 //  .
+	icmpOut[7] = 0 //  .
 
 	// Copy original IP header and first 8 bytes as body
 	copy(icmpOut[8:], packet[:packetLen])
@@ -198,193 +165,7 @@ func ipv4CreateRejectTCPPacket(packet []byte, out []byte) []byte {
 	return out
 }
 
-func ipv6CreateRejectPacket(packet []byte, out []byte) []byte {
-	proto, offset, isFragment := ipv6FindUpperProtocol(packet)
-	if isFragment {
-		return nil
-	}
-	switch proto {
-	case 6: // tcp
-		return ipv6CreateRejectTCPPacket(packet, out, offset)
-	default:
-		return ipv6CreateRejectICMPPacket(packet, out, proto, offset)
-	}
-}
-
-func ipv6CreateRejectICMPPacket(packet []byte, out []byte, proto uint8, offset int) []byte {
-	// Do not generate ICMPv6 errors in response to ICMPv6 error packets
-	if proto == 58 && len(packet) > offset {
-		icmpType := packet[offset]
-		if icmpType >= 1 && icmpType <= 4 {
-			return nil
-		}
-	}
-
-	// Include as much of the original packet as possible, up to 1000 bytes,
-	// so the response fits comfortably within any tunnel MTU.
-	packetLen := min(len(packet), 1000)
-
-	outLen := ipv6.HeaderLen + 8 + packetLen
-	if outLen > cap(out) {
-		return nil
-	}
-
-	out = out[:outLen]
-
-	// IPv6 header
-	ipHdr := out[0:ipv6.HeaderLen]
-	ipHdr[0] = ipv6.Version << 4 // version, traffic class (high bits)
-	ipHdr[1] = 0                 // traffic class (low bits), flow label (high bits)
-	ipHdr[2] = 0                 // flow label
-	ipHdr[3] = 0                 // flow label
-
-	payloadLen := uint16(outLen - ipv6.HeaderLen)
-	binary.BigEndian.PutUint16(ipHdr[4:], payloadLen) // payload length
-	ipHdr[6] = 58                                     // next header (ICMPv6)
-	ipHdr[7] = 64                                     // hop limit
-
-	// Swap dest / src IPs (each 16 bytes, src at 8, dst at 24)
-	copy(ipHdr[8:24], packet[24:40])
-	copy(ipHdr[24:40], packet[8:24])
-
-	// ICMPv6 Destination Unreachable
-	icmpOut := out[ipv6.HeaderLen:]
-	icmpOut[0] = 1 // type (Destination Unreachable)
-	icmpOut[1] = 1 // code (Communication with destination administratively prohibited)
-	icmpOut[2] = 0 // checksum
-	icmpOut[3] = 0 //  .
-	icmpOut[4] = 0 // unused
-	icmpOut[5] = 0 //  .
-	icmpOut[6] = 0 //  .
-	icmpOut[7] = 0 //  .
-
-	copy(icmpOut[8:], packet[:packetLen])
-
-	// ICMPv6 checksum uses a pseudo-header
-	csum := ipv6PseudoheaderChecksum(ipHdr[8:24], ipHdr[24:40], 58, uint32(payloadLen))
-	binary.BigEndian.PutUint16(icmpOut[2:], tcpipChecksum(icmpOut, csum))
-
-	return out
-}
-
-func ipv6CreateRejectTCPPacket(packet []byte, out []byte, offset int) []byte {
-	const tcpLen = 20
-
-	if len(packet) < offset+tcpLen {
-		return nil
-	}
-
-	outLen := ipv6.HeaderLen + tcpLen
-	if outLen > cap(out) {
-		return nil
-	}
-
-	out = out[:outLen]
-
-	// IPv6 header
-	ipHdr := out[0:ipv6.HeaderLen]
-	ipHdr[0] = ipv6.Version << 4 // version, traffic class (high bits)
-	ipHdr[1] = 0                 // traffic class (low bits), flow label (high bits)
-	ipHdr[2] = 0                 // flow label
-	ipHdr[3] = 0                 // flow label
-
-	binary.BigEndian.PutUint16(ipHdr[4:], tcpLen) // payload length
-	ipHdr[6] = 6                                  // next header (TCP)
-	ipHdr[7] = 64                                 // hop limit
-
-	// Swap dest / src IPs
-	copy(ipHdr[8:24], packet[24:40])
-	copy(ipHdr[24:40], packet[8:24])
-
-	// TCP RST
-	tcpIn := packet[offset:]
-	var ackSeq, seq uint32
-	outFlags := byte(0b00000100) // RST
-
-	inAck := tcpIn[13]&0b00010000 != 0
-	if inAck {
-		seq = binary.BigEndian.Uint32(tcpIn[8:])
-	} else {
-		inSyn := uint32((tcpIn[13] & 0b00000010) >> 1)
-		inFin := uint32(tcpIn[13] & 0b00000001)
-		ackSeq = binary.BigEndian.Uint32(tcpIn[4:]) + inSyn + inFin + uint32(len(tcpIn)) - uint32(tcpIn[12]>>4)<<2
-		outFlags |= 0b00010000 // ACK
-	}
-
-	tcpOut := out[ipv6.HeaderLen:]
-	// Swap dest / src ports
-	copy(tcpOut[0:2], tcpIn[2:4])
-	copy(tcpOut[2:4], tcpIn[0:2])
-	binary.BigEndian.PutUint32(tcpOut[4:], seq)
-	binary.BigEndian.PutUint32(tcpOut[8:], ackSeq)
-	tcpOut[12] = (tcpLen >> 2) << 4 // data offset, reserved, NS
-	tcpOut[13] = outFlags           // CWR, ECE, URG, ACK, PSH, RST, SYN, FIN
-	tcpOut[14] = 0                  // window size
-	tcpOut[15] = 0                  //  .
-	tcpOut[16] = 0                  // checksum
-	tcpOut[17] = 0                  //  .
-	tcpOut[18] = 0                  // URG Pointer
-	tcpOut[19] = 0                  //  .
-
-	// Calculate checksum with IPv6 pseudo-header
-	csum := ipv6PseudoheaderChecksum(ipHdr[8:24], ipHdr[24:40], 6, tcpLen)
-	binary.BigEndian.PutUint16(tcpOut[16:], tcpipChecksum(tcpOut, csum))
-
-	return out
-}
-
-func ipv6FindUpperProtocol(packet []byte) (nextHeader uint8, offset int, isFragment bool) {
-	nextHeader = packet[6]
-	offset = ipv6.HeaderLen
-
-	for {
-		switch nextHeader {
-		case 0, 43, 60: // Hop-by-Hop, Routing, Destination
-			if len(packet) < offset+2 {
-				return nextHeader, offset, isFragment
-			}
-			nextHeader = packet[offset]
-			offset += int(packet[offset+1]+1) << 3
-
-		case 44: // Fragment
-			if len(packet) < offset+8 {
-				return nextHeader, offset, isFragment
-			}
-			if packet[offset+2] != 0 || packet[offset+3]&0xf8 != 0 {
-				isFragment = true
-			}
-			nextHeader = packet[offset]
-			offset += 8
-
-		case 51: // AH
-			if len(packet) < offset+2 {
-				return nextHeader, offset, isFragment
-			}
-			nextHeader = packet[offset]
-			offset += int(packet[offset+1]+2) << 2
-
-		default:
-			return nextHeader, offset, isFragment
-		}
-	}
-}
-
 func CreateICMPEchoResponse(packet, out []byte) []byte {
-	if len(packet) < 1 {
-		return nil
-	}
-
-	switch packet[0] >> 4 {
-	case 4:
-		return createICMPv4EchoResponse(packet, out)
-	case 6:
-		return createICMPv6EchoResponse(packet, out)
-	default:
-		return nil
-	}
-}
-
-func createICMPv4EchoResponse(packet, out []byte) []byte {
 	// Return early if this is not a simple ICMP Echo Request
 	//TODO: make constants out of these
 	if !(len(packet) >= 28 && len(packet) <= 9001 && packet[0] == 0x45 && packet[9] == 0x01 && packet[20] == 0x08) {
@@ -418,43 +199,6 @@ func createICMPv4EchoResponse(packet, out []byte) []byte {
 	return out
 }
 
-func createICMPv6EchoResponse(packet, out []byte) []byte {
-	// IPv6 header (40 bytes) + ICMPv6 header (8 bytes minimum)
-	if len(packet) < ipv6.HeaderLen+8 || len(packet) > 9001 {
-		return nil
-	}
-
-	// Next Header must be ICMPv6 (58)
-	if packet[6] != 58 {
-		return nil
-	}
-
-	// ICMPv6 type must be Echo Request (128)
-	if packet[ipv6.HeaderLen] != 128 {
-		return nil
-	}
-
-	out = out[:len(packet)]
-	copy(out, packet)
-
-	// Swap src/dst addresses (bytes 8-23 and 24-39)
-	copy(out[8:24], packet[24:40])
-	copy(out[24:40], packet[8:24])
-
-	// Change ICMPv6 type to Echo Reply (129)
-	icmp := out[ipv6.HeaderLen:]
-	icmp[0] = 129
-	icmp[2] = 0
-	icmp[3] = 0
-
-	// ICMPv6 checksum uses a pseudo-header with src, dst, length, and next header
-	payloadLen := uint32(len(icmp))
-	csum := ipv6PseudoheaderChecksum(out[8:24], out[24:40], 58, payloadLen)
-	binary.BigEndian.PutUint16(icmp[2:], tcpipChecksum(icmp, csum))
-
-	return out
-}
-
 // calculates the TCP/IP checksum defined in rfc1071. The passed-in
 // csum is any initial checksum data that's already been computed.
 //
@@ -492,18 +236,3 @@ func ipv4PseudoheaderChecksum(src, dst []byte, proto, length uint32) (csum uint3
 	csum += length >> 16
 	return csum
 }
-
-// based on:
-// - https://github.com/google/gopacket/blob/v1.1.19/layers/tcpip.go#L37-L48
-func ipv6PseudoheaderChecksum(src, dst []byte, proto, length uint32) (csum uint32) {
-	for i := 0; i < 16; i += 2 {
-		csum += uint32(src[i]) << 8
-		csum += uint32(src[i+1])
-		csum += uint32(dst[i]) << 8
-		csum += uint32(dst[i+1])
-	}
-	csum += proto
-	csum += length & 0xffff
-	csum += length >> 16
-	return csum
-}

iputil/packet_test.go

@@ -1,13 +1,11 @@
 package iputil
 
 import (
-	"encoding/binary"
 	"net"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
 	"golang.org/x/net/ipv4"
-	"golang.org/x/net/ipv6"
 )
 
 func Test_CreateRejectPacket(t *testing.T) {
@@ -45,7 +43,7 @@ func Test_CreateRejectPacket(t *testing.T) {
 	}
 	b = append(b, []byte{0, 3, 0, 4, 0, 0, 0, 0}...)
 
-	expectedLen = maxIPv4RejectPacketSize
+	expectedLen = MaxRejectPacketSize
 	out = make([]byte, MaxRejectPacketSize)
 	rejectPacket = CreateRejectPacket(b, out)
 	assert.NotNil(t, rejectPacket)
@@ -73,404 +71,3 @@ func Test_CreateRejectPacket(t *testing.T) {
 	assert.NotNil(t, rejectPacket)
 	assert.Len(t, rejectPacket, expectedLen)
 }
-
-func Test_CreateRejectPacket_NoFragment(t *testing.T) {
-	out := make([]byte, MaxRejectPacketSize)
-
-	// IPv4: non-zero fragment offset should not generate reject packet
-	h := ipv4.Header{
-		Len:      20,
-		Src:      net.IPv4(10, 0, 0, 1),
-		Dst:      net.IPv4(10, 0, 0, 2),
-		Protocol: 17, // UDP
-	}
-	b, err := h.Marshal()
-	if err != nil {
-		t.Fatalf("h.Marshal: %v", err)
-	}
-	b = append(b, make([]byte, 8)...)
-	// Set fragment offset to non-zero (byte 6-7, offset in 8-byte units)
-	b[6] = 0x00
-	b[7] = 0x01
-	assert.Nil(t, CreateRejectPacket(b, out))
-
-	// MF flag with zero offset (first fragment) should still generate reject
-	b[6] = 0x20 // MF flag set
-	b[7] = 0x00
-	assert.NotNil(t, CreateRejectPacket(b, out))
-
-	// Non-fragment should still generate reject packet
-	b[6] = 0x00
-	b[7] = 0x00
-	assert.NotNil(t, CreateRejectPacket(b, out))
-
-	// DF flag only (not a fragment) should still generate reject packet
-	b[6] = 0x40
-	b[7] = 0x00
-	assert.NotNil(t, CreateRejectPacket(b, out))
-}
-
-func Test_CreateRejectPacketIPv6_NoFragment(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-	out := make([]byte, MaxRejectPacketSize)
-
-	// IPv6 with Fragment header and non-zero offset should not generate reject
-	fragHeader := []byte{
-		17,   // next header: UDP
-		0,    // reserved
-		0, 9, // fragment offset=1 (shifted left 3), M=1
-		0, 0, 0, 1, // identification
-	}
-	udpPayload := make([]byte, 8)
-	payload := append(fragHeader, udpPayload...)
-	packet := makeIPv6Packet(src, dst, 44, payload) // next header 44 = Fragment
-	assert.Nil(t, CreateRejectPacket(packet, out))
-
-	// Fragment header with zero offset (first fragment) should still generate reject
-	fragHeader[2] = 0
-	fragHeader[3] = 1 // offset=0, M=1
-	payload = append(fragHeader, udpPayload...)
-	packet = makeIPv6Packet(src, dst, 44, payload)
-	assert.NotNil(t, CreateRejectPacket(packet, out))
-}
-
-func Test_CreateRejectPacket_NoICMPError(t *testing.T) {
-	out := make([]byte, MaxRejectPacketSize)
-
-	// ICMP error types should not generate reject packets
-	icmpErrorTypes := []byte{3, 4, 5, 11, 12}
-	for _, icmpType := range icmpErrorTypes {
-		h := ipv4.Header{
-			Len:      20,
-			Src:      net.IPv4(10, 0, 0, 1),
-			Dst:      net.IPv4(10, 0, 0, 2),
-			Protocol: 1, // ICMP
-		}
-
-		b, err := h.Marshal()
-		if err != nil {
-			t.Fatalf("h.Marshal: %v", err)
-		}
-		b = append(b, icmpType, 0, 0, 0, 0, 0, 0, 0)
-
-		rejectPacket := CreateRejectPacket(b, out)
-		assert.Nil(t, rejectPacket, "ICMP type %d should not generate a reject packet", icmpType)
-	}
-
-	// ICMP non-error types should still generate reject packets
-	icmpNonErrorTypes := []byte{0, 8, 13, 14}
-	for _, icmpType := range icmpNonErrorTypes {
-		h := ipv4.Header{
-			Len:      20,
-			Src:      net.IPv4(10, 0, 0, 1),
-			Dst:      net.IPv4(10, 0, 0, 2),
-			Protocol: 1, // ICMP
-		}
-
-		b, err := h.Marshal()
-		if err != nil {
-			t.Fatalf("h.Marshal: %v", err)
-		}
-		b = append(b, icmpType, 0, 0, 0, 0, 0, 0, 0)
-
-		rejectPacket := CreateRejectPacket(b, out)
-		assert.NotNil(t, rejectPacket, "ICMP type %d should generate a reject packet", icmpType)
-	}
-}
-
-func makeIPv6Packet(src, dst net.IP, nextHeader uint8, payload []byte) []byte {
-	b := make([]byte, ipv6.HeaderLen+len(payload))
-	b[0] = ipv6.Version << 4
-	binary.BigEndian.PutUint16(b[4:], uint16(len(payload)))
-	b[6] = nextHeader
-	b[7] = 64
-	copy(b[8:24], src.To16())
-	copy(b[24:40], dst.To16())
-	copy(b[ipv6.HeaderLen:], payload)
-	return b
-}
-
-func Test_CreateRejectPacketIPv6_ICMP(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-
-	// Small UDP packet: entire original included in body
-	udpPayload := make([]byte, 20)
-	udpPayload[0] = 0x00 // src port high
-	udpPayload[1] = 0x50 // src port low (80)
-	udpPayload[2] = 0x01 // dst port high
-	udpPayload[3] = 0xBB // dst port low (443)
-	packet := makeIPv6Packet(src, dst, 17, udpPayload)
-
-	out := make([]byte, MaxRejectPacketSize)
-	rejectPacket := CreateRejectPacket(packet, out)
-	assert.NotNil(t, rejectPacket)
-
-	// Small packet fits entirely: 40 (ipv6 hdr) + 8 (icmpv6 hdr) + 60 (original)
-	expectedLen := ipv6.HeaderLen + 8 + len(packet)
-	assert.Len(t, rejectPacket, expectedLen)
-
-	// Verify version
-	assert.Equal(t, byte(ipv6.Version<<4), rejectPacket[0]&0xf0)
-	// Verify next header is ICMPv6 (58)
-	assert.Equal(t, byte(58), rejectPacket[6])
-	// Verify src/dst are swapped
-	assert.Equal(t, dst.To16(), net.IP(rejectPacket[8:24]))
-	assert.Equal(t, src.To16(), net.IP(rejectPacket[24:40]))
-	// Verify ICMPv6 type=1 (Dest Unreachable), code=1 (Administratively prohibited)
-	assert.Equal(t, byte(1), rejectPacket[ipv6.HeaderLen])
-	assert.Equal(t, byte(1), rejectPacket[ipv6.HeaderLen+1])
-	// Verify entire original packet is included in body
-	assert.Equal(t, packet, rejectPacket[ipv6.HeaderLen+8:])
-
-	// Large packet: body is truncated to 1000 bytes
-	largePkt := makeIPv6Packet(src, dst, 17, make([]byte, 1200))
-	rejectPacket = CreateRejectPacket(largePkt, out)
-	assert.NotNil(t, rejectPacket)
-	assert.Len(t, rejectPacket, ipv6.HeaderLen+8+1000)
-	assert.Equal(t, largePkt[:1000], rejectPacket[ipv6.HeaderLen+8:])
-}
-
-func Test_CreateRejectPacketIPv6_TCP(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-
-	// TCP SYN packet (next header 6)
-	tcpPayload := make([]byte, 20)
-	tcpPayload[0] = 0x00                             // src port high
-	tcpPayload[1] = 0x50                             // src port low (80)
-	tcpPayload[2] = 0x01                             // dst port high
-	tcpPayload[3] = 0xBB                             // dst port low (443)
-	binary.BigEndian.PutUint32(tcpPayload[4:], 1000) // seq
-	binary.BigEndian.PutUint32(tcpPayload[8:], 0)    // ack seq
-	tcpPayload[12] = (20 >> 2) << 4                  // data offset
-	tcpPayload[13] = 0b00000010                      // SYN flag
-
-	packet := makeIPv6Packet(src, dst, 6, tcpPayload)
-
-	out := make([]byte, MaxRejectPacketSize)
-	rejectPacket := CreateRejectPacket(packet, out)
-	assert.NotNil(t, rejectPacket)
-
-	// Expected: 40 (ipv6 hdr) + 20 (tcp RST)
-	expectedLen := ipv6.HeaderLen + 20
-	assert.Len(t, rejectPacket, expectedLen)
-
-	// Verify version
-	assert.Equal(t, byte(ipv6.Version<<4), rejectPacket[0]&0xf0)
-	// Verify next header is TCP (6)
-	assert.Equal(t, byte(6), rejectPacket[6])
-	// Verify src/dst are swapped
-	assert.Equal(t, dst.To16(), net.IP(rejectPacket[8:24]))
-	assert.Equal(t, src.To16(), net.IP(rejectPacket[24:40]))
-	// Verify ports are swapped
-	tcpOut := rejectPacket[ipv6.HeaderLen:]
-	assert.Equal(t, uint16(443), binary.BigEndian.Uint16(tcpOut[0:2]))
-	assert.Equal(t, uint16(80), binary.BigEndian.Uint16(tcpOut[2:4]))
-	// RST+ACK flags (since input was SYN without ACK)
-	assert.Equal(t, byte(0b00010100), tcpOut[13])
-	// ack_seq = original seq (1000) + SYN (1) + FIN (0) + segment data (0)
-	assert.Equal(t, uint32(1001), binary.BigEndian.Uint32(tcpOut[8:]))
-}
-
-func Test_CreateRejectPacketIPv6_TCPWithACK(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-
-	// TCP packet with ACK set
-	tcpPayload := make([]byte, 20)
-	tcpPayload[0] = 0x00
-	tcpPayload[1] = 0x50
-	tcpPayload[2] = 0x01
-	tcpPayload[3] = 0xBB
-	binary.BigEndian.PutUint32(tcpPayload[4:], 1000) // seq
-	binary.BigEndian.PutUint32(tcpPayload[8:], 2000) // ack seq
-	tcpPayload[12] = (20 >> 2) << 4                  // data offset
-	tcpPayload[13] = 0b00010000                      // ACK flag
-
-	packet := makeIPv6Packet(src, dst, 6, tcpPayload)
-
-	out := make([]byte, MaxRejectPacketSize)
-	rejectPacket := CreateRejectPacket(packet, out)
-	assert.NotNil(t, rejectPacket)
-
-	tcpOut := rejectPacket[ipv6.HeaderLen:]
-	// RST only (no ACK) since input had ACK
-	assert.Equal(t, byte(0b00000100), tcpOut[13])
-	// seq = original ack_seq
-	assert.Equal(t, uint32(2000), binary.BigEndian.Uint32(tcpOut[4:]))
-}
-
-func Test_CreateRejectPacketIPv6_NoICMPError(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-	out := make([]byte, MaxRejectPacketSize)
-
-	// ICMPv6 error types (1-4) should not generate reject packets
-	for icmpType := byte(1); icmpType <= 4; icmpType++ {
-		payload := make([]byte, 8)
-		payload[0] = icmpType
-		packet := makeIPv6Packet(src, dst, 58, payload)
-
-		rejectPacket := CreateRejectPacket(packet, out)
-		assert.Nil(t, rejectPacket, "ICMPv6 type %d should not generate a reject packet", icmpType)
-	}
-
-	// ICMPv6 non-error types should still generate reject packets
-	nonErrorTypes := []byte{128, 129, 133, 134}
-	for _, icmpType := range nonErrorTypes {
-		payload := make([]byte, 8)
-		payload[0] = icmpType
-		packet := makeIPv6Packet(src, dst, 58, payload)
-
-		rejectPacket := CreateRejectPacket(packet, out)
-		assert.NotNil(t, rejectPacket, "ICMPv6 type %d should generate a reject packet", icmpType)
-	}
-}
-
-func Test_CreateRejectPacketIPv6_TooShort(t *testing.T) {
-	// Packet too short to be valid IPv6
-	out := make([]byte, MaxRejectPacketSize)
-	assert.Nil(t, CreateRejectPacket([]byte{0x60}, out))
-	assert.Nil(t, CreateRejectPacket(make([]byte, 39), out))
-}
-
-func Test_CreateRejectPacketIPv6_ExtensionHeaders(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-
-	// IPv6 + Hop-by-Hop extension header + TCP
-	hopByHop := []byte{
-		6,    // next header: TCP
-		0,    // length (8 bytes total)
-		0, 0, // padding
-		0, 0, 0, 0,
-	}
-	tcpPayload := make([]byte, 20)
-	tcpPayload[0] = 0x00
-	tcpPayload[1] = 0x50
-	tcpPayload[2] = 0x01
-	tcpPayload[3] = 0xBB
-	binary.BigEndian.PutUint32(tcpPayload[4:], 1000)
-	binary.BigEndian.PutUint32(tcpPayload[8:], 2000)
-	tcpPayload[12] = (20 >> 2) << 4
-	tcpPayload[13] = 0b00010000 // ACK
-
-	payload := append(hopByHop, tcpPayload...)
-	packet := makeIPv6Packet(src, dst, 0, payload) // next header 0 = Hop-by-Hop
-
-	out := make([]byte, MaxRejectPacketSize)
-	rejectPacket := CreateRejectPacket(packet, out)
-	assert.NotNil(t, rejectPacket)
-
-	// Should produce TCP RST
-	expectedLen := ipv6.HeaderLen + 20
-	assert.Len(t, rejectPacket, expectedLen)
-	assert.Equal(t, byte(6), rejectPacket[6]) // next header is TCP
-	tcpOut := rejectPacket[ipv6.HeaderLen:]
-	assert.Equal(t, byte(0b00000100), tcpOut[13]) // RST only
-}
-
-func TestCreateICMPEchoResponse_IPv4(t *testing.T) {
-	// Build a simple IPv4 ICMP Echo Request
-	packet := make([]byte, 28)
-	packet[0] = 0x45                                   // version 4, IHL 5
-	binary.BigEndian.PutUint16(packet[2:], uint16(28)) // total length
-	packet[8] = 64                                     // TTL
-	packet[9] = 1                                      // protocol ICMP
-	copy(packet[12:16], net.IPv4(10, 0, 0, 1).To4())   // src
-	copy(packet[16:20], net.IPv4(10, 0, 0, 2).To4())   // dst
-	packet[20] = 8                                     // ICMP Echo Request
-
-	out := make([]byte, len(packet))
-	result := CreateICMPEchoResponse(packet, out)
-	assert.NotNil(t, result)
-	assert.Equal(t, byte(0x45), result[0])
-	// src/dst swapped
-	assert.Equal(t, net.IPv4(10, 0, 0, 2).To4(), net.IP(result[12:16]))
-	assert.Equal(t, net.IPv4(10, 0, 0, 1).To4(), net.IP(result[16:20]))
-	// ICMP Echo Reply
-	assert.Equal(t, byte(0), result[20])
-}
-
-func TestCreateICMPEchoResponse_IPv6(t *testing.T) {
-	src := net.ParseIP("fd00::1").To16()
-	dst := net.ParseIP("fd00::2").To16()
-
-	// Build an IPv6 ICMPv6 Echo Request packet
-	// IPv6 header (40 bytes) + ICMPv6 (8 bytes)
-	packet := make([]byte, 48)
-	packet[0] = 0x60        // version 6
-	payloadLen := uint16(8) // ICMPv6 header only
-	binary.BigEndian.PutUint16(packet[4:], payloadLen)
-	packet[6] = 58           // Next Header: ICMPv6
-	packet[7] = 64           // Hop Limit
-	copy(packet[8:24], src)  // src address
-	copy(packet[24:40], dst) // dst address
-
-	// ICMPv6 Echo Request
-	icmp := packet[40:]
-	icmp[0] = 128                           // type: Echo Request
-	icmp[1] = 0                             // code
-	binary.BigEndian.PutUint16(icmp[4:], 1) // identifier
-	binary.BigEndian.PutUint16(icmp[6:], 1) // sequence number
-
-	// Compute correct checksum for the request
-	csum := ipv6PseudoheaderChecksum(src, dst, 58, uint32(payloadLen))
-	binary.BigEndian.PutUint16(icmp[2:], tcpipChecksum(icmp, csum))
-
-	out := make([]byte, len(packet))
-	result := CreateICMPEchoResponse(packet, out)
-	assert.NotNil(t, result)
-
-	// Version should still be 6
-	assert.Equal(t, byte(6), result[0]>>4)
-	// src/dst swapped
-	assert.Equal(t, dst, net.IP(result[8:24]))
-	assert.Equal(t, src, net.IP(result[24:40]))
-	// ICMPv6 Echo Reply type
-	assert.Equal(t, byte(129), result[40])
-
-	// Verify checksum is valid (tcpipChecksum returns 0 when data+checksum is correct)
-	respIcmp := result[40:]
-	verifyCsum := ipv6PseudoheaderChecksum(result[8:24], result[24:40], 58, uint32(payloadLen))
-	assert.Equal(t, uint16(0), tcpipChecksum(respIcmp, verifyCsum))
-}
-
-func TestCreateICMPEchoResponse_IPv6_NotEchoRequest(t *testing.T) {
-	src := net.ParseIP("fd00::1").To16()
-	dst := net.ParseIP("fd00::2").To16()
-
-	packet := make([]byte, 48)
-	packet[0] = 0x60
-	binary.BigEndian.PutUint16(packet[4:], 8)
-	packet[6] = 58
-	packet[7] = 64
-	copy(packet[8:24], src)
-	copy(packet[24:40], dst)
-
-	// ICMPv6 type 1 (Destination Unreachable) - not Echo Request
-	packet[40] = 1
-
-	out := make([]byte, len(packet))
-	result := CreateICMPEchoResponse(packet, out)
-	assert.Nil(t, result)
-}
-
-func TestCreateICMPEchoResponse_IPv6_NotICMPv6(t *testing.T) {
-	src := net.ParseIP("fd00::1").To16()
-	dst := net.ParseIP("fd00::2").To16()
-
-	packet := make([]byte, 48)
-	packet[0] = 0x60
-	binary.BigEndian.PutUint16(packet[4:], 8)
-	packet[6] = 6 // TCP, not ICMPv6
-	packet[7] = 64
-	copy(packet[8:24], src)
-	copy(packet[24:40], dst)
-
-	out := make([]byte, len(packet))
-	result := CreateICMPEchoResponse(packet, out)
-	assert.Nil(t, result)
-}

lighthouse.go

@@ -272,18 +272,16 @@ func (lh *LightHouse) reload(c *config.C, initial bool) error {
 	//NOTE: many things will get much simpler when we combine static_host_map and lighthouse.hosts in config
 	if initial || c.HasChanged("static_host_map") || c.HasChanged("static_map.cadence") || c.HasChanged("static_map.network") || c.HasChanged("static_map.lookup_timeout") {
 		// Clean up. Entries still in the static_host_map will be re-built.
-		ourselves := lh.myVpnNetworks[0].Addr()
-		oldStaticList := lh.staticList.Load()
-		if oldStaticList != nil {
+		// Entries no longer present must have their (possible) background DNS goroutines stopped.
+		if existingStaticList := lh.staticList.Load(); existingStaticList != nil {
 			lh.RLock()
-			for staticVpnAddr := range *oldStaticList {
+			for staticVpnAddr := range *existingStaticList {
 				if am, ok := lh.addrMap[staticVpnAddr]; ok && am != nil {
-					am.ResetForOwner(ourselves)
+					am.hr.Cancel()
 				}
 			}
 			lh.RUnlock()
 		}
-
 		// Build a new list based on current config.
 		staticList := make(map[netip.Addr]struct{})
 		err := lh.loadStaticMap(c, staticList)
@@ -291,21 +289,6 @@ func (lh *LightHouse) reload(c *config.C, initial bool) error {
 			return err
 		}
 
-		// For entries removed from static_host_map, stop the DNS goroutine and drop the cached addrs.
-		// All addrs must come from the lighthouses now that it's no longer a static host.
-		if oldStaticList != nil {
-			lh.RLock()
-			for staticVpnAddr := range *oldStaticList {
-				if _, stillStatic := staticList[staticVpnAddr]; stillStatic {
-					continue
-				}
-				if am, ok := lh.addrMap[staticVpnAddr]; ok && am != nil {
-					am.ClearHostnameResults()
-				}
-			}
-			lh.RUnlock()
-		}
-
 		lh.staticList.Store(&staticList)
 		if !initial {
 			if c.HasChanged("static_host_map") {

lighthouse_test.go

@@ -303,132 +303,6 @@ func TestLighthouse_reload(t *testing.T) {
 	require.NoError(t, err)
 }
 
-// TestLighthouse_reloadStaticHostMap verifies that reloading static_host_map applies the new
-// config rather than appending to it. See issue #718.
-func TestLighthouse_reloadStaticHostMap(t *testing.T) {
-	l := test.NewLogger()
-	c := config.NewC(l)
-	c.Settings["lighthouse"] = map[string]any{"am_lighthouse": true}
-	c.Settings["listen"] = map[string]any{"port": 4242}
-	c.Settings["static_host_map"] = map[string]any{
-		"10.128.0.2": []any{"1.1.1.1:4242"},
-	}
-
-	myVpnNet := netip.MustParsePrefix("10.128.0.1/24")
-	nt := new(bart.Lite)
-	nt.Insert(myVpnNet)
-	cs := &CertState{
-		myVpnNetworks:      []netip.Prefix{myVpnNet},
-		myVpnNetworksTable: nt,
-	}
-
-	lh, err := NewLightHouseFromConfig(t.Context(), l, c, cs, nil, nil)
-	require.NoError(t, err)
-
-	staticHost := netip.MustParseAddr("10.128.0.2")
-	otherHost := netip.MustParseAddr("10.128.0.3")
-
-	// Capture the RemoteList pointer up front; an in-flight handshake would hold the same one
-	// on hostinfo.remotes, so it must reflect every reload below.
-	pinned := lh.Query(staticHost)
-	require.NotNil(t, pinned)
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("1.1.1.1:4242")}, pinned.CopyAddrs([]netip.Prefix{}))
-
-	// Replace the remote address. The new address should be the only entry.
-	nc := map[string]any{
-		"static_host_map": map[string]any{
-			"10.128.0.2": []any{"2.2.2.2:4242"},
-		},
-	}
-	rc, err := yaml.Marshal(nc)
-	require.NoError(t, err)
-	require.NoError(t, c.ReloadConfigString(string(rc)))
-
-	rl := lh.Query(staticHost)
-	require.NotNil(t, rl)
-	assert.Same(t, pinned, rl, "RemoteList pointer must stay stable so in-flight handshakes pick up the change")
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("2.2.2.2:4242")}, rl.CopyAddrs([]netip.Prefix{}))
-
-	// Reload back to the original IP. Mirrors the round-trip in issue #718 step 6-8 where
-	// the buggy reload produced [1.1.1.1, 2.2.2.2, 1.1.1.1] instead of [1.1.1.1].
-	nc = map[string]any{
-		"static_host_map": map[string]any{
-			"10.128.0.2": []any{"1.1.1.1:4242"},
-		},
-	}
-	rc, err = yaml.Marshal(nc)
-	require.NoError(t, err)
-	require.NoError(t, c.ReloadConfigString(string(rc)))
-
-	rl = lh.Query(staticHost)
-	require.NotNil(t, rl)
-	assert.Same(t, pinned, rl)
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("1.1.1.1:4242")}, rl.CopyAddrs([]netip.Prefix{}))
-
-	// Reload with the same config. An unchanged entry must not duplicate.
-	require.NoError(t, c.ReloadConfigString(string(rc)))
-
-	rl = lh.Query(staticHost)
-	require.NotNil(t, rl)
-	assert.Same(t, pinned, rl)
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("1.1.1.1:4242")}, rl.CopyAddrs([]netip.Prefix{}))
-
-	// Switch back to 2.2.2.2 so the rest of the test continues against a known address.
-	nc = map[string]any{
-		"static_host_map": map[string]any{
-			"10.128.0.2": []any{"2.2.2.2:4242"},
-		},
-	}
-	rc, err = yaml.Marshal(nc)
-	require.NoError(t, err)
-	require.NoError(t, c.ReloadConfigString(string(rc)))
-
-	// Add a second host alongside the first. Both should be present, neither duplicated.
-	nc = map[string]any{
-		"static_host_map": map[string]any{
-			"10.128.0.2": []any{"2.2.2.2:4242"},
-			"10.128.0.3": []any{"3.3.3.3:4242"},
-		},
-	}
-	rc, err = yaml.Marshal(nc)
-	require.NoError(t, err)
-	require.NoError(t, c.ReloadConfigString(string(rc)))
-
-	rl = lh.Query(staticHost)
-	require.NotNil(t, rl)
-	assert.Same(t, pinned, rl, "adding a sibling entry must not displace the existing RemoteList")
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("2.2.2.2:4242")}, rl.CopyAddrs([]netip.Prefix{}))
-
-	rl = lh.Query(otherHost)
-	require.NotNil(t, rl)
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("3.3.3.3:4242")}, rl.CopyAddrs([]netip.Prefix{}))
-
-	// Drop the first host entirely. The vpnAddr is no longer marked static, our owner
-	// contribution is cleared, but the addrMap entry stays in place so non-static cache
-	// data (from lighthouse queries) on the same RemoteList isn't lost. In-flight handshakes
-	// that already had the pointer see an empty address list rather than retrying stale ones.
-	nc = map[string]any{
-		"static_host_map": map[string]any{
-			"10.128.0.3": []any{"3.3.3.3:4242"},
-		},
-	}
-	rc, err = yaml.Marshal(nc)
-	require.NoError(t, err)
-	require.NoError(t, c.ReloadConfigString(string(rc)))
-
-	_, isStatic := lh.GetStaticHostList()[staticHost]
-	assert.False(t, isStatic)
-
-	rl = lh.Query(staticHost)
-	require.NotNil(t, rl)
-	assert.Same(t, pinned, rl)
-	assert.Empty(t, rl.CopyAddrs([]netip.Prefix{}))
-
-	rl = lh.Query(otherHost)
-	require.NotNil(t, rl)
-	assert.Equal(t, []netip.AddrPort{netip.MustParseAddrPort("3.3.3.3:4242")}, rl.CopyAddrs([]netip.Prefix{}))
-}
-
 func newLHHostRequest(fromAddr netip.AddrPort, myVpnIp, queryVpnIp netip.Addr, lhh *LightHouseHandler) testLhReply {
 	req := &NebulaMeta{
 		Type:    NebulaMeta_HostQuery,

main.go

@@ -5,7 +5,10 @@ import (
 	"fmt"
 	"log/slog"
 	"net"
+	"net/http"
+	_ "net/http/pprof"
 	"net/netip"
+	"runtime"
 	"runtime/debug"
 	"strings"
 	"time"
@@ -33,6 +36,9 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		buildVersion = moduleVersion()
 	}
 
+	//todo no merge
+	go http.ListenAndServe(":6060", nil)
+
 	// Print the config if in test, the exit comes later
 	if configTest {
 		b, err := yaml.Marshal(c.Settings)
@@ -194,7 +200,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	handshakeManager := NewHandshakeManager(l, hostMap, lightHouse, udpConns[0], handshakeConfig)
 	lightHouse.handshakeTrigger = handshakeManager.trigger
 
-	ds, err := newDnsServerFromConfig(ctx, l, pki, hostMap, c)
+	ds, err := newDnsServerFromConfig(ctx, l, pki.getCertState(), hostMap, c)
 	if err != nil {
 		l.Warn("Failed to start DNS responder", "error", err)
 	}
@@ -215,11 +221,13 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		DropLocalBroadcast:    c.GetBool("tun.drop_local_broadcast", false),
 		DropMulticast:         c.GetBool("tun.drop_multicast", false),
 		routines:              routines,
+		batchSize:             c.GetInt("listen.batch", 64),
 		MessageMetrics:        messageMetrics,
 		version:               buildVersion,
 		relayManager:          NewRelayManager(ctx, l, hostMap, c),
 		punchy:                punchy,
 		ConntrackCacheTimeout: conntrackCacheTimeout,
+		CpuAffinity:           parseCpuAffinity(c, l, routines),
 		l:                     l,
 	}
 
@@ -237,6 +245,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		ifce.reloadDisconnectInvalid(c)
 		ifce.reloadSendRecvError(c)
 		ifce.reloadAcceptRecvError(c)
+		ifce.reloadEcn(c)
 
 		handshakeManager.f = ifce
 		go handshakeManager.Run(ctx)
@@ -271,6 +280,53 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	}, nil
 }
 
+// parseCpuAffinity reads `tun.cpu_affinity` from the config — a list of
+// integer CPU IDs, one per TUN reader goroutine. Empty / unset returns nil
+// (listenIn falls back to its default `i % NumCPU` pinning). Length
+// mismatch with `routines` is a warning, not an error: shorter lists are
+// modulo-cycled across queues, longer lists' tail is ignored. Invalid
+// entries (non-integer, out of range) are also a warning and disable the
+// override entirely so we don't silently pin to the wrong CPU.
+func parseCpuAffinity(c *config.C, l *slog.Logger, routines int) []int {
+	raw := c.Get("tun.cpu_affinity")
+	if raw == nil {
+		return nil
+	}
+	rv, ok := raw.([]any)
+	if !ok {
+		l.Warn("tun.cpu_affinity must be a list of integers; ignoring", "value", raw)
+		return nil
+	}
+	nCPU := runtime.NumCPU()
+	cpus := make([]int, 0, len(rv))
+	for i, e := range rv {
+		var cpu int
+		switch v := e.(type) {
+		case int:
+			cpu = v
+		case int64:
+			cpu = int(v)
+		case float64:
+			cpu = int(v)
+		default:
+			l.Warn("tun.cpu_affinity entry not an integer; ignoring affinity",
+				"index", i, "value", e)
+			return nil
+		}
+		if cpu < 0 || cpu >= nCPU {
+			l.Warn("tun.cpu_affinity entry out of range; ignoring affinity",
+				"index", i, "cpu", cpu, "num_cpu", nCPU)
+			return nil
+		}
+		cpus = append(cpus, cpu)
+	}
+	if len(cpus) != routines {
+		l.Warn("tun.cpu_affinity length doesn't match routines; queues will modulo-cycle through the list",
+			"affinity_len", len(cpus), "routines", routines)
+	}
+	return cpus
+}
+
 func moduleVersion() string {
 	info, ok := debug.ReadBuildInfo()
 	if !ok {

noiseutil/cipher_state_test.go

@@ -147,6 +147,62 @@ func buildCipherStatesB(b *testing.B, c noise.CipherFunc) (*noise.CipherState, *
 	return eI, dR
 }
 
+// TestDecryptDangerRelayShapeNoAlloc covers the AD-only relay path used in
+// outside.go's handleOutsideRelayPacket: the body is AD, the trailing 16 bytes
+// are the AEAD tag, the plaintext is empty, and the caller passes nil as the
+// destination because it only needs the auth side-effect. The call must
+// succeed, return an empty plaintext, and not allocate on the hot path.
+func TestDecryptDangerRelayShapeNoAlloc(t *testing.T) {
+	cases := []struct {
+		name string
+		c    noise.CipherFunc
+		wrap func(*noise.CipherState) CipherState
+	}{
+		{"AESGCM", CipherAESGCM, func(cs *noise.CipherState) CipherState { return NewCipherStateAESGCM(cs) }},
+		{"ChaChaPoly", noise.CipherChaChaPoly, func(cs *noise.CipherState) CipherState { return NewCipherStateChaChaPoly(cs) }},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			encCS, decCS := buildCipherStates(t, tc.c)
+			enc, dec := tc.wrap(encCS), tc.wrap(decCS)
+
+			ad := make([]byte, 1200) // typical relay packet body size
+			for i := range ad {
+				ad[i] = byte(i)
+			}
+			nb := make([]byte, 12)
+
+			// Build the "signature value" the way handleOutsideRelayPacket sees it:
+			// empty plaintext encrypted with the body as AD yields just the 16-byte tag.
+			tag, err := enc.EncryptDanger(nil, ad, nil, 1, nb)
+			require.NoError(t, err)
+			require.Len(t, tag, dec.Overhead())
+
+			// Sanity: the relay-shaped call returns empty plaintext, no error.
+			out, err := dec.DecryptDanger(nil, ad, tag, 1, nb)
+			require.NoError(t, err)
+			assert.Empty(t, out)
+
+			// Tampering with the AD must fail authentication.
+			ad[0] ^= 0xff
+			_, err = dec.DecryptDanger(nil, ad, tag, 1, nb)
+			require.Error(t, err)
+			ad[0] ^= 0xff
+
+			// The hot path must not allocate. AllocsPerRun does a warm-up run, so any
+			// one-time setup is excluded. Counter has to advance so the AEAD nonce is
+			// unique per call, but we don't care whether the auth succeeds — we only
+			// care about whether the call path allocates.
+			var counter uint64 = 2
+			allocs := testing.AllocsPerRun(100, func() {
+				_, _ = dec.DecryptDanger(nil, ad, tag, counter, nb)
+				counter++
+			})
+			assert.Equal(t, 0.0, allocs, "DecryptDanger(nil, ...) must not allocate")
+		})
+	}
+}
+
 func TestCipherStateNilSafety(t *testing.T) {
 	var aes *CipherStateAESGCM
 	_, err := aes.EncryptDanger(nil, nil, nil, 0, make([]byte, 12))

outside.go

@@ -13,6 +13,7 @@ import (
 
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
 	"golang.org/x/net/ipv4"
 )
 
@@ -22,7 +23,7 @@ const (
 
 var ErrOutOfWindow = errors.New("out of window packet")
 
-func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) readOutsidePackets(via ViaSender, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache, meta udp.RxMeta) {
 	err := h.Parse(packet)
 	if err != nil {
 		// Hole punch packets are 0 or 1 byte big, so lets ignore printing those errors
@@ -110,11 +111,11 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 
 	// Relay packets are special
 	if isMessageRelay {
-		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache)
-
+		f.handleOutsideRelayPacket(hostinfo, via, packet, h, fwPacket, lhf, nb, q, localCache, meta)
 		return
 	}
 
+	out := f.batchers[q].Reserve(len(packet))[:0]
 	out, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -135,7 +136,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 	case header.Message:
 		switch h.Subtype {
 		case header.MessageNone:
-			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache)
+			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache, meta)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected message subtype seen", "from", via, "header", h)
 			return
@@ -168,7 +169,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 	}
 }
 
-func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache, meta udp.RxMeta) {
 	// The entire body is sent as AD, not encrypted.
 	// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
 	// The packet is guaranteed to be at least 16 bytes at this point, b/c it got past the h.Parse() call above. If it's
@@ -176,16 +177,10 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 	// which will gracefully fail in the DecryptDanger call.
 	signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
 	signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
-	var err error
-	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, signedPayload, signatureValue, h.MessageCounter, nb)
-	if err != nil {
-		return
-	}
-	// Advance the replay window now that the frame is authenticated
-	if !hostinfo.ConnectionState.window.Update(f.l, h.MessageCounter) {
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("dropping out of window relay packet", "header", h)
-		}
+	// The decrypted output is empty (relay packets carry their payload as AD) and unused.
+	// The recursive readOutsidePackets call below operates on signedPayload. Passing
+	// nil avoids reserving an arena slot.
+	if _, err := hostinfo.ConnectionState.dKey.DecryptDanger(nil, signedPayload, signatureValue, h.MessageCounter, nb); err != nil {
 		return
 	}
 	// Successfully validated the thing. Get rid of the Relay header.
@@ -201,7 +196,8 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		// The only way this happens is if hostmap has an index to the correct HostInfo, but the HostInfo is missing
 		// its internal mapping. This should never happen.
 		hostinfo.logger(f.l).Error("HostInfo missing remote relay index",
-			"relayRemoteIndex", h.RemoteIndex,
+			"vpnAddrs", hostinfo.vpnAddrs,
+			"remoteIndex", h.RemoteIndex,
 		)
 		return
 	}
@@ -217,15 +213,15 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 			relay:     relay,
 			IsRelayed: true,
 		}
-		f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache)
+		f.readOutsidePackets(via, signedPayload, h, fwPacket, lhf, nb, q, localCache, meta)
 	case ForwardingType:
 		// Find the target HostInfo relay object
 		targetHI, targetRelay, err := f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relay.PeerAddr)
 		if err != nil {
 			hostinfo.logger(f.l).Info("Failed to find target host info by ip",
 				"relayTo", relay.PeerAddr,
-				"relayFrom", hostinfo.vpnAddrs[0],
 				"error", err,
+				"hostinfo.vpnAddrs", hostinfo.vpnAddrs,
 			)
 			return
 		}
@@ -235,7 +231,8 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 			switch targetRelay.Type {
 			case ForwardingType:
 				// Forward this packet through the relay tunnel
-				// Find the target HostInfo
+				// Find the target HostInfo //todo it would potentially be nice to batch these
+				out := f.batchers[q].Reserve(len(packet) + header.Len + hostinfo.ConnectionState.dKey.Overhead())[:0]
 				f.SendVia(targetHI, targetRelay, signedPayload, nb, out, false)
 			case TerminalType:
 				hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
@@ -518,7 +515,61 @@ func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, out []byte, packet []
 	return out, nil
 }
 
-func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache) {
+// 2-bit IP-level ECN codepoints (lower bits of IPv4 ToS / IPv6 TC).
+const (
+	ecnNotECT = 0x00
+	ecnECT1   = 0x01
+	ecnECT0   = 0x02
+	ecnCE     = 0x03
+)
+
+// applyOuterECN folds an outer CE mark from the underlay into the inner
+// IP header per RFC 6040 normal mode. It mutates pkt[1] in place. Other
+// codepoints are advisory only and leave the inner unchanged.
+//
+// Merge cases (outer × inner → action):
+//
+//	outer != CE                : no-op (inner is authoritative)
+//	outer == CE, inner Not-ECT : log; cannot propagate to a non-ECN host
+//	outer == CE, inner ECT/CE  : rewrite inner ECN to CE
+func applyOuterECN(pkt []byte, outerECN byte, hostinfo *HostInfo, l *slog.Logger) {
+	if outerECN&ecnCE != ecnCE || len(pkt) < 2 {
+		return
+	}
+	switch pkt[0] >> 4 {
+	case 4:
+		switch pkt[1] & 0x03 {
+		case ecnNotECT:
+			if l.Enabled(context.Background(), slog.LevelDebug) {
+				hostinfo.logger(l).Debug("RFC 6040: outer CE on inner Not-ECT, leaving inner unchanged")
+			}
+		case ecnCE:
+			// Already CE.
+		default:
+			pkt[1] = (pkt[1] &^ 0x03) | ecnCE
+		}
+	case 6:
+		switch (pkt[1] >> 4) & 0x03 {
+		case ecnNotECT:
+			if l.Enabled(context.Background(), slog.LevelDebug) {
+				hostinfo.logger(l).Debug("RFC 6040: outer CE on inner Not-ECT, leaving inner unchanged")
+			}
+		case ecnCE:
+			// Already CE.
+		default:
+			pkt[1] = (pkt[1] &^ 0x30) | (ecnCE << 4)
+		}
+	}
+}
+
+func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache, meta udp.RxMeta) {
+	// RFC 6040 normal-mode combine: fold any outer CE mark stamped by the
+	// underlay into the inner header before firewall + TUN write. Other
+	// outer codepoints are advisory only — we keep the inner unchanged.
+	if f.ecnEnabled.Load() {
+		applyOuterECN(out, meta.OuterECN, hostinfo, f.l)
+	}
+
 	err := newPacket(out, true, fwPacket)
 	if err != nil {
 		hostinfo.logger(f.l).Warn("Error while validating inbound packet",
@@ -542,7 +593,7 @@ func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, p
 		return
 	}
 
-	_, err = f.readers[q].Write(out)
+	err = f.batchers[q].Commit(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}

overlay/batch/coalesce_core.go

@@ -0,0 +1,149 @@
+package batch
+
+import (
+	"bytes"
+	"encoding/binary"
+)
+
+// flowKey identifies a transport flow by {src, dst, sport, dport, family}.
+// Comparable, so map lookups and linear scans over the slot list stay tight.
+// Shared by the TCP and UDP coalescers; each coalescer keeps its own
+// openSlots map, so a TCP and UDP flow on the same 5-tuple-without-proto
+// never alias.
+type flowKey struct {
+	src, dst     [16]byte
+	sport, dport uint16
+	isV6         bool
+}
+
+// initialSlots is the starting capacity of the slot pool. One flow per
+// packet is the worst case so this matches a typical carrier-side
+// recvmmsg batch on the encrypted UDP socket.
+const initialSlots = 64
+
+// parsedIP is the IP-level result of parseIPPrologue. The caller layers
+// L4-specific parsing (TCP / UDP) on top.
+type parsedIP struct {
+	fk       flowKey
+	ipHdrLen int
+	// pkt is the original buffer trimmed to the IP-declared total length.
+	// Anything below the IP layer (transport parsers) should slice into
+	// pkt rather than the unbounded original.
+	pkt []byte
+}
+
+// parseIPPrologue extracts the IP-level fields the coalescers care about:
+// IHL/payload length, version, src/dst addresses, and the L4 protocol byte.
+// Returns ok=false for malformed input, IPv4 with options or fragmentation,
+// or IPv6 with extension headers (all rejected by both coalescers in
+// identical ways before this refactor).
+//
+// On success, p.pkt is len-trimmed to the IP-declared length so callers
+// don't have to repeat the trim. wantProto is the IANA protocol number to
+// require (6 for TCP, 17 for UDP); ok=false for any other value.
+func parseIPPrologue(pkt []byte, wantProto byte) (parsedIP, bool) {
+	var p parsedIP
+	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] != wantProto {
+			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])
+		p.pkt = pkt[:totalLen]
+	case 6:
+		if len(pkt) < 40 {
+			return p, false
+		}
+		if pkt[6] != wantProto {
+			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])
+		p.pkt = pkt[:40+payloadLen]
+	default:
+		return p, false
+	}
+	return p, true
+}
+
+// ipHeadersMatch compares the IP portion of two packet header prefixes for
+// byte-for-byte equality on every field that must be identical across
+// coalesced segments. Size/IPID/IPCsum and the 2-bit IP-level ECN field are
+// masked out — the appendPayload step merges CE into the seed.
+//
+// The transport (L4) portion of the header is checked separately by the
+// per-protocol matcher.
+func ipHeadersMatch(a, b []byte, isV6 bool) bool {
+	if isV6 {
+		// IPv6: byte 0 = version/TC[7:4], byte 1 = TC[3:0]/flow[19:16],
+		// bytes [2:4] = flow[15:0], [6:8] = next_hdr/hop, [8:40] = src+dst.
+		// ECN lives in TC[1:0] = byte 1 mask 0x30. Skip [4:6] payload_len.
+		if a[0] != b[0] {
+			return false
+		}
+		if a[1]&^0x30 != b[1]&^0x30 {
+			return false
+		}
+		if !bytes.Equal(a[2:4], b[2:4]) {
+			return false
+		}
+		if !bytes.Equal(a[6:40], b[6:40]) {
+			return false
+		}
+		return true
+	}
+	// IPv4: byte 0 = version/IHL, byte 1 = DSCP(6)|ECN(2),
+	// [6:10] flags/fragoff/TTL/proto, [12:20] src+dst.
+	// Skip [2:4] total len, [4:6] id, [10:12] csum.
+	if a[0] != b[0] {
+		return false
+	}
+	if a[1]&^0x03 != b[1]&^0x03 {
+		return false
+	}
+	if !bytes.Equal(a[6:10], b[6:10]) {
+		return false
+	}
+	if !bytes.Equal(a[12:20], b[12:20]) {
+		return false
+	}
+	return true
+}
+
+// mergeECNIntoSeed ORs the 2-bit IP-level ECN field of pkt's IP header
+// onto the seed's IP header, so a CE mark on any coalesced segment
+// propagates to the final superpacket. (CE is 0b11; ORing yields CE if
+// any segment carried it.) Used by both TCP and UDP coalescers, so the
+// invariant lives in one place.
+func mergeECNIntoSeed(seedHdr, pktHdr []byte, isV6 bool) {
+	if isV6 {
+		seedHdr[1] |= pktHdr[1] & 0x30
+	} else {
+		seedHdr[1] |= pktHdr[1] & 0x03
+	}
+}

overlay/batch/multi_coalesce.go

@@ -0,0 +1,136 @@
+package batch
+
+import (
+	"errors"
+	"log/slog"
+
+	"github.com/slackhq/nebula/overlay/tio"
+	"github.com/slackhq/nebula/util"
+)
+
+// MultiCoalescer fans plaintext packets out to lane-specific batchers based
+// on the IP/L4 protocol of the packet, sharing a single Reserve arena
+// across lanes so the caller's allocation pattern is unchanged.
+//
+// Lanes are processed independently: the TCP coalescer only sees TCP, the
+// UDP coalescer only sees UDP, and the passthrough lane handles everything
+// else. Per-flow arrival order is preserved because a single 5-tuple only
+// ever lands in one lane and each lane preserves its own slot order.
+//
+// Cross-lane order is NOT preserved across the TCP/UDP/passthrough split.
+// This is acceptable because the carrier-side recvmmsg path already
+// stable-sorts by (peer, message counter) before delivering plaintext
+// here, so replay-window invariants are unaffected, and apps observe
+// correct per-flow ordering — which is all the IP layer guarantees anyway.
+// Do not "fix" this by interleaving lane outputs at flush time; that
+// negates the entire point of coalescing (each lane needs to see runs of
+// adjacent same-flow packets to coalesce them).
+type MultiCoalescer struct {
+	tcp *TCPCoalescer
+	udp *UDPCoalescer
+	pt  *Passthrough
+
+	// arena is shared across every lane (constructor hands the same
+	// *Arena to TCP, UDP, and Passthrough), so there's exactly one
+	// backing slab per MultiCoalescer instance. Each lane's Flush calls
+	// Reset; the resets are idempotent because Multi.Flush drains lanes
+	// sequentially and never Reserves in between, so a later lane's
+	// slots stay readable across an earlier lane's Reset (the underlying
+	// bytes are still alive — Reset only re-slices len to 0).
+	arena *util.Arena
+}
+
+// NewMultiCoalescer builds a multi-lane batcher. tcpEnabled lets the caller
+// opt out of TCP coalescing (e.g. when the queue can't do TSO); udpEnabled
+// likewise gates UDP coalescing (only enable when USO was negotiated).
+// Either lane disabled redirects its traffic into the passthrough lane.
+// arena is the single backing slab shared across every lane; the caller
+// pre-sizes it via NewArena so the hot path never allocates.
+func NewMultiCoalescer(w tio.Queue, l *slog.Logger, arena *util.Arena, tcpEnabled, udpEnabled bool) *MultiCoalescer {
+	m := &MultiCoalescer{
+		pt:    NewPassthrough(w, initialSlots, arena),
+		arena: arena,
+	}
+	if tcpEnabled {
+		m.tcp = NewTCPCoalescer(w, l, arena)
+	}
+	if udpEnabled {
+		m.udp = NewUDPCoalescer(w, arena)
+	}
+	return m
+}
+
+func (m *MultiCoalescer) Reserve(sz int) []byte {
+	return m.arena.Reserve(sz)
+}
+
+// Commit dispatches pkt to the appropriate lane based on IP version + L4
+// proto. Borrowed slice contract is identical to the single-lane batchers,
+// pkt must remain valid until the next Flush.
+//
+// On the success path the IP/TCP-or-UDP parse happens here once and the
+// parsed struct is handed to the lane via commitParsed so the lane doesn't
+// re-walk the header.
+func (m *MultiCoalescer) Commit(pkt []byte) error {
+	if len(pkt) < 20 {
+		return m.pt.Commit(pkt)
+	}
+	v := pkt[0] >> 4
+	var proto byte
+	switch v {
+	case 4:
+		proto = pkt[9]
+	case 6:
+		if len(pkt) < 40 {
+			return m.pt.Commit(pkt)
+		}
+		proto = pkt[6]
+	default:
+		return m.pt.Commit(pkt)
+	}
+	switch proto {
+	case ipProtoTCP:
+		if m.tcp != nil {
+			info, ok := parseTCPBase(pkt)
+			if !ok {
+				// Malformed/unsupported TCP shape (IP options, fragments, ...).
+				// Handle this via passthrough support in the TCP coalescer, to attempt to preserve flow order.
+				m.tcp.addPassthrough(pkt)
+				return nil
+			}
+			return m.tcp.commitParsed(pkt, info)
+		}
+	case ipProtoUDP:
+		if m.udp != nil {
+			info, ok := parseUDP(pkt)
+			if !ok {
+				m.udp.addPassthrough(pkt) //we could also m.pt.Commit() here I guess?
+				return nil
+			}
+			return m.udp.commitParsed(pkt, info)
+		}
+	}
+	return m.pt.Commit(pkt)
+}
+
+// Flush drains every lane in a fixed order: TCP, UDP, passthrough. Errors
+// from a lane do not stop subsequent lanes from flushing, we keep
+// draining and return the first observed error so a single bad packet
+// doesn't strand the others.
+func (m *MultiCoalescer) Flush() error {
+	var errs []error
+	if m.tcp != nil {
+		if err := m.tcp.Flush(); err != nil {
+			errs = append(errs, err)
+		}
+	}
+	if m.udp != nil {
+		if err := m.udp.Flush(); err != nil {
+			errs = append(errs, err)
+		}
+	}
+	if err := m.pt.Flush(); err != nil {
+		errs = append(errs, err)
+	}
+	return errors.Join(errs...)
+}

overlay/batch/multi_coalesce_test.go

@@ -0,0 +1,97 @@
+package batch
+
+import (
+	"testing"
+
+	"github.com/slackhq/nebula/test"
+	"github.com/slackhq/nebula/util"
+)
+
+// TestMultiCoalescerRoutesByProto confirms TCP/UDP/other land in the right
+// lane: TCP and UDP get coalesced when their lanes are enabled, anything
+// else (ICMP here) falls through to plain Write.
+func TestMultiCoalescerRoutesByProto(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	m := NewMultiCoalescer(w, test.NewLogger(), util.NewArena(0), true, true)
+
+	tcpPay := make([]byte, 1200)
+	udpPay := make([]byte, 1200)
+	icmp := make([]byte, 28)
+	icmp[0] = 0x45
+	icmp[2] = 0
+	icmp[3] = 28
+	icmp[9] = 1
+
+	if err := m.Commit(buildTCPv4(1000, tcpAck, tcpPay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Commit(buildTCPv4(2200, tcpAck, tcpPay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Commit(buildUDPv4(2000, 53, udpPay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Commit(buildUDPv4(2000, 53, udpPay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Commit(icmp); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// 1 TCP super (2 segments) + 1 UDP super (2 segments) = 2 gso writes.
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes (one TCP + one UDP), got %d", len(w.gsoWrites))
+	}
+	if len(w.writes) != 1 {
+		t.Fatalf("want 1 plain write (ICMP), got %d", len(w.writes))
+	}
+}
+
+// TestMultiCoalescerDisabledUDPFallsThrough verifies that when the UDP lane
+// is disabled (e.g. kernel doesn't support USO), UDP packets still reach
+// the kernel via the passthrough lane rather than being lost.
+func TestMultiCoalescerDisabledUDPFallsThrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	m := NewMultiCoalescer(w, test.NewLogger(), util.NewArena(0), true, false) // TSO on, USO off
+
+	if err := m.Commit(buildUDPv4(1000, 53, make([]byte, 800))); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Commit(buildUDPv4(1000, 53, make([]byte, 800))); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 0 {
+		t.Errorf("UDP must NOT be coalesced when USO disabled, got %d gso writes", len(w.gsoWrites))
+	}
+	if len(w.writes) != 2 {
+		t.Errorf("UDP must pass through as 2 plain writes, got %d", len(w.writes))
+	}
+}
+
+// TestMultiCoalescerDisabledTCPFallsThrough mirrors the TSO=off case.
+func TestMultiCoalescerDisabledTCPFallsThrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	m := NewMultiCoalescer(w, test.NewLogger(), util.NewArena(0), false, true) // TSO off, USO on
+
+	pay := make([]byte, 1200)
+	if err := m.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Commit(buildTCPv4(2200, tcpAck, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 0 {
+		t.Errorf("TCP must NOT be coalesced when TSO disabled, got %d gso writes", len(w.gsoWrites))
+	}
+	if len(w.writes) != 2 {
+		t.Errorf("TCP must pass through as 2 plain writes, got %d", len(w.writes))
+	}
+}

overlay/batch/passthrough.go

@@ -0,0 +1,47 @@
+package batch
+
+import (
+	"io"
+
+	"github.com/slackhq/nebula/util"
+)
+
+// Passthrough is a RxBatcher that doesn't batch anything, it just accumulates and then sends packets.
+type Passthrough struct {
+	out   io.Writer
+	slots [][]byte
+	// arena is injected; see TCPCoalescer.arena for the contract.
+	arena  *util.Arena
+	cursor int
+}
+
+func NewPassthrough(w io.Writer, slots int, arena *util.Arena) *Passthrough {
+	return &Passthrough{
+		out:   w,
+		slots: make([][]byte, 0, slots),
+		arena: arena,
+	}
+}
+
+func (p *Passthrough) Reserve(sz int) []byte {
+	return p.arena.Reserve(sz)
+}
+
+func (p *Passthrough) Commit(pkt []byte) error {
+	p.slots = append(p.slots, pkt)
+	return nil
+}
+
+func (p *Passthrough) Flush() error {
+	var firstErr error
+	for _, s := range p.slots {
+		_, err := p.out.Write(s)
+		if err != nil && firstErr == nil {
+			firstErr = err
+		}
+	}
+	clear(p.slots)
+	p.slots = p.slots[:0]
+	p.arena.Reset()
+	return firstErr
+}

overlay/batch/rx_batch.go

@@ -0,0 +1,12 @@
+package batch
+
+type RxBatcher interface {
+	// Reserve creates a pkt to borrow
+	Reserve(sz int) []byte
+	// Commit borrows pkt. The caller must keep pkt valid until the next Flush
+	Commit(pkt []byte) error
+	// Flush emits every queued packet in arrival order. 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.
+	Flush() error
+}

overlay/batch/tcp_coalesce.go

@@ -0,0 +1,743 @@
+package batch
+
+import (
+	"bytes"
+	"context"
+	"encoding/binary"
+	"io"
+	"log/slog"
+	"net/netip"
+	"slices"
+
+	"github.com/slackhq/nebula/overlay/tio"
+	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
+)
+
+// 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
+
+// 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 tio.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   tio.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
+	// lastSlot caches the most recently touched open slot. Steady-state
+	// bulk traffic is dominated by a single flow, so comparing the
+	// incoming key against the cached slot's own fk lets the hot path
+	// skip the map lookup (and the aeshash of a 38-byte key) entirely.
+	// Kept in lockstep with openSlots: nil whenever the slot it pointed
+	// at is removed/sealed.
+	lastSlot *coalesceSlot
+	pool     []*coalesceSlot // free list for reuse
+
+	// arena is injected; the coalescer borrows slices from it via Reserve
+	// and tells it to release them via Reset on Flush. When wrapped in
+	// MultiCoalescer the same *Arena is shared with the other lanes so
+	// there's exactly one backing slab per Multi instance.
+	arena *util.Arena
+	l     *slog.Logger
+}
+
+func NewTCPCoalescer(w tio.Queue, l *slog.Logger, arena *util.Arena) *TCPCoalescer {
+	c := &TCPCoalescer{
+		plainW:    w,
+		slots:     make([]*coalesceSlot, 0, initialSlots),
+		openSlots: make(map[flowKey]*coalesceSlot, initialSlots),
+		pool:      make([]*coalesceSlot, 0, initialSlots),
+		arena:     arena,
+		l:         l,
+	}
+	if gw, ok := tio.SupportsGSO(w, wire.GSOProtoTCP); ok {
+		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
+	ip, ok := parseIPPrologue(pkt, ipProtoTCP)
+	if !ok {
+		return p, false
+	}
+	pkt = ip.pkt
+	p.fk = ip.fk
+	p.ipHdrLen = ip.ipHdrLen
+
+	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
+}
+
+// TCP flag bits (byte 13 of the TCP header). Only the bits actually consulted
+// by the coalescer are named; FIN/SYN/RST/URG/CWR are rejected via the
+// negative mask in coalesceable, not by name.
+const (
+	tcpFlagPsh = 0x08
+	tcpFlagAck = 0x10
+	tcpFlagEce = 0x40
+)
+
+// coalesceable reports whether a parsed TCP segment is eligible for
+// coalescing. Accepts ACK, ACK|PSH, ACK|ECE, ACK|PSH|ECE with a
+// non-empty payload. CWR is excluded because it marks a one-shot
+// congestion-window-reduced transition the receiver must observe at a
+// segment boundary.
+func (p parsedTCP) coalesceable() bool {
+	if p.flags&tcpFlagAck == 0 {
+		return false
+	}
+	if p.flags&^(tcpFlagAck|tcpFlagPsh|tcpFlagEce) != 0 {
+		return false
+	}
+	return p.payLen > 0
+}
+
+func (c *TCPCoalescer) Reserve(sz int) []byte {
+	return c.arena.Reserve(sz)
+}
+
+// Commit 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) Commit(pkt []byte) error {
+	if c.gsoW == nil {
+		c.addPassthrough(pkt)
+		return nil
+	}
+	info, ok := parseTCPBase(pkt)
+	if !ok {
+		c.addPassthrough(pkt)
+		return nil
+	}
+	return c.commitParsed(pkt, info)
+}
+
+// commitParsed is the post-parse half of Commit. The caller must have
+// already verified parseTCPBase succeeded (info is a valid TCP parse).
+// Used by MultiCoalescer.Commit to avoid re-walking the IP/TCP header
+// after the dispatcher has already done so.
+func (c *TCPCoalescer) commitParsed(pkt []byte, info parsedTCP) error {
+	if c.gsoW == nil {
+		c.addPassthrough(pkt)
+		return nil
+	}
+	if !info.coalesceable() {
+		// TCP but not admissible (SYN/FIN/RST/URG/CWR or zero-payload).
+		// Seal this flow's open slot so later in-flow packets don't extend
+		// it and accidentally reorder past this passthrough.
+		if last := c.lastSlot; last != nil && last.fk == info.fk {
+			c.lastSlot = nil
+		}
+		delete(c.openSlots, info.fk)
+		c.addPassthrough(pkt)
+		return nil
+	}
+
+	// Single-flow fast path: with only one open flow the cache hits every
+	// packet, and len(openSlots)==1 lets us skip the 38-byte fk compare
+	// when there are multiple flows in flight (where the hit rate would
+	// be ~0 and the compare is pure overhead).
+	var open *coalesceSlot
+	if last := c.lastSlot; last != nil && len(c.openSlots) == 1 && last.fk == info.fk {
+		open = last
+	} else {
+		open = c.openSlots[info.fk]
+	}
+	if open != nil {
+		if c.canAppend(open, pkt, info) {
+			c.appendPayload(open, pkt, info)
+			if open.psh {
+				delete(c.openSlots, info.fk)
+				c.lastSlot = nil
+			} else {
+				c.lastSlot = open
+			}
+			return nil
+		}
+		// Can't extend — seal it and fall through to seed a fresh slot.
+		delete(c.openSlots, info.fk)
+		if c.lastSlot == open {
+			c.lastSlot = nil
+		}
+	}
+	c.seed(pkt, info)
+	return nil
+}
+
+// Flush emits every queued event in (per-flow) seq order. Coalesced slots
+// go out via WriteGSO; passthrough slots go out via plainW.Write.
+// reorderForFlush first sorts each flow's slots into TCP-seq order within
+// passthrough-bounded segments and merges contiguous adjacent slots, so
+// any wire-side reorder that crossed an rxOrder batch boundary doesn't
+// get amplified into kernel-visible reorder by the slot machinery.
+// 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 {
+	c.reorderForFlush()
+	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)
+	}
+	clear(c.slots)
+	c.slots = c.slots[:0]
+	clear(c.openSlots)
+	c.lastSlot = nil
+
+	c.arena.Reset()
+	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&tcpFlagPsh != 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
+		c.lastSlot = s
+	} else if last := c.lastSlot; last != nil && last.fk == info.fk {
+		// PSH-on-seed seals the slot immediately. Any prior cached open
+		// slot for this flow has just been sealed-and-replaced by this
+		// passthrough-shaped seed, so drop the cache too.
+		c.lastSlot = nil
+	}
+}
+
+// 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
+	}
+	// ECE state must be stable across a burst — receivers expect the
+	// flag set on every segment of a CE-echoing window or none.
+	seedFlags := s.hdrBuf[s.ipHdrLen+13]
+	if (seedFlags^info.flags)&tcpFlagEce != 0 {
+		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.flags&tcpFlagPsh != 0 {
+		// Propagate PSH into the seed header so kernel TSO sets it on the
+		// last segment. Without this the sender's push signal is dropped.
+		s.hdrBuf[s.ipHdrLen+13] |= tcpFlagPsh
+	}
+	// Merge IP-level CE marks into the seed: headersMatch ignores ECN, so
+	// this is the one place the signal is preserved.
+	mergeECNIntoSeed(s.hdrBuf[:s.ipHdrLen], pkt[:s.ipHdrLen], s.isV6)
+	if info.payLen < s.gsoSize || info.flags&tcpFlagPsh != 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
+	clear(s.payIovs)
+	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.ipHdrLen], hdr[s.ipHdrLen:], s.payIovs, wire.GSOProtoTCP)
+}
+
+// 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, as is the
+// 2-bit IP-level ECN field — appendPayload merges CE into the seed.
+func headersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	if !ipHeadersMatch(a, b, isV6) {
+		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
+}
+
+// reorderForFlush neutralizes wire-side reorder that the rxOrder buffer
+// couldn't catch (anything crossing a recvmmsg batch boundary). Without
+// this pass a small wire reorder — counter 250 arriving in batch K when
+// 200..249 are coming in batch K+1 — would seed an out-of-seq slot first
+// and emit it ahead of the lower-seq slot, manifesting at the inner TCP
+// receiver as a much larger reorder than the wire actually had.
+//
+// Two phases:
+//  1. Sort each passthrough-bounded segment of c.slots by (flow, seq).
+//     Cross-flow ordering inside a segment isn't preserved (it never was
+//     and doesn't matter for any single flow's TCP correctness).
+//  2. Sweep once and merge adjacent same-flow slots whose ranges are now
+//     contiguous AND whose tail is gsoSize-aligned. The tail constraint
+//     matters because the kernel TSO splitter chops at gsoSize from the
+//     start of the merged payload — a short segment in the middle would
+//     desynchronize every later segment.
+//
+// Passthrough slots act as barriers: the merge check skips them on either
+// side, so a SYN/FIN/RST/CWR is never reordered relative to its flow's
+// data.
+func (c *TCPCoalescer) reorderForFlush() {
+	if len(c.slots) <= 1 {
+		return
+	}
+	runStart := 0
+	for i := 0; i <= len(c.slots); i++ {
+		if i < len(c.slots) && !c.slots[i].passthrough {
+			continue
+		}
+		c.sortRun(c.slots[runStart:i])
+		runStart = i + 1
+	}
+	out := c.slots[:0]
+	logged := false
+	for _, s := range c.slots {
+		if n := len(out); n > 0 {
+			prev := out[n-1]
+			if !prev.passthrough && !s.passthrough && prev.fk == s.fk {
+				// Same-flow neighbors after sort. If they aren't seq-
+				// contiguous it's a real gap — packets the wire reordered
+				// across batches, or actual loss before nebula. Log it so
+				// the operator can quantify how often it happens; the data
+				// itself still emits in seq order, kernel TCP handles the
+				// gap via its OOO queue.
+				if c.l.Enabled(context.Background(), slog.LevelDebug) {
+					if prev.nextSeq != slotSeedSeq(s) {
+						logged = true
+						gap := int64(slotSeedSeq(s)) - int64(prev.nextSeq)
+						c.l.Debug("tcp coalesce: cross-slot seq gap",
+							"src", flowKeyAddr(s.fk, false),
+							"dst", flowKeyAddr(s.fk, true),
+							"sport", s.fk.sport,
+							"dport", s.fk.dport,
+							"prev_seed_seq", slotSeedSeq(prev),
+							"prev_next_seq", prev.nextSeq,
+							"this_seed_seq", slotSeedSeq(s),
+							"gap_bytes", gap,
+							"prev_seg_count", prev.numSeg,
+							"prev_total_pay", prev.totalPay,
+						)
+					}
+				}
+
+				if canMergeSlots(prev, s) {
+					mergeSlots(prev, s)
+					c.release(s)
+					continue
+				}
+			}
+		}
+		out = append(out, s)
+	}
+	if logged {
+		c.l.Warn("==== end of batch ====")
+	}
+	c.slots = out
+}
+
+// flowKeyAddr returns the src or dst address from fk as a netip.Addr for
+// logging. Only used on the cold gap-log path so the netip allocation
+// doesn't matter.
+func flowKeyAddr(fk flowKey, dst bool) netip.Addr {
+	src := fk.src
+	if dst {
+		src = fk.dst
+	}
+	if fk.isV6 {
+		return netip.AddrFrom16(src)
+	}
+	var v4 [4]byte
+	copy(v4[:], src[:4])
+	return netip.AddrFrom4(v4)
+}
+
+// sortRun stable-sorts run by (flowKey, seedSeq) so each flow's slots
+// cluster together in seq order, ready for the merge sweep. Stable so
+// equal-key slots keep their original relative position (defensive — a
+// duplicate seedSeq would already mean something's wrong upstream).
+func (c *TCPCoalescer) sortRun(run []*coalesceSlot) {
+	if len(run) <= 1 {
+		return
+	}
+	// slices.SortStableFunc with a free, non-capturing comparator avoids the
+	// reflection + closure-escape allocations that sort.SliceStable forces.
+	slices.SortStableFunc(run, compareCoalesceSlots)
+}
+
+func compareCoalesceSlots(a, b *coalesceSlot) int {
+	if cmp := flowKeyCompare(a.fk, b.fk); cmp != 0 {
+		return cmp
+	}
+	aSeq, bSeq := slotSeedSeq(a), slotSeedSeq(b)
+	if aSeq == bSeq {
+		return 0
+	}
+	if tcpSeqLess(aSeq, bSeq) {
+		return -1
+	}
+	return 1
+}
+
+// slotSeedSeq returns the TCP seq of the slot's seed (first segment).
+// nextSeq tracks the seq just past the last appended byte; subtracting
+// totalPay walks back to the seed. uint32 wraparound is the right TCP
+// arithmetic so no special-casing is needed.
+func slotSeedSeq(s *coalesceSlot) uint32 {
+	return s.nextSeq - uint32(s.totalPay)
+}
+
+// tcpSeqLess reports whether a precedes b in TCP serial-number arithmetic
+// (RFC 1323 §2.3). The signed int32 cast turns the modular subtraction
+// into the right comparison even across the 2^32 wrap.
+func tcpSeqLess(a, b uint32) bool {
+	return int32(a-b) < 0
+}
+
+// flowKeyCompare orders flowKeys deterministically. The exact ordering
+// is irrelevant — only that same-flow slots cluster together so the
+// post-sort sweep can merge contiguous pairs.
+func flowKeyCompare(a, b flowKey) int {
+	// Cheap scalar fields first so most non-matching keys short-circuit
+	// without ever calling bytes.Compare. sport is the ephemeral port on
+	// egress flows and discriminates fastest. For matching keys (same
+	// flow), array equality on src/dst inlines to word-sized compares,
+	// so we only pay bytes.Compare when the arrays actually differ.
+	if a.sport != b.sport {
+		if a.sport < b.sport {
+			return -1
+		}
+		return 1
+	}
+	if a.dport != b.dport {
+		if a.dport < b.dport {
+			return -1
+		}
+		return 1
+	}
+	if a.dst != b.dst {
+		return bytes.Compare(a.dst[:], b.dst[:])
+	}
+	if a.src != b.src {
+		return bytes.Compare(a.src[:], b.src[:])
+	}
+	if a.isV6 != b.isV6 {
+		if !a.isV6 {
+			return -1
+		}
+		return 1
+	}
+	return 0
+}
+
+// canMergeSlots reports whether s can fold into prev as one merged TSO
+// superpacket. Same flow, contiguous TCP byte range, equal gsoSize, and
+// fits within the kernel TSO limits. The tail-of-prev check rejects any
+// merge whose first slot ended on a sub-gsoSize segment — kernel TSO
+// would split the merged skb at gsoSize boundaries from the start, so a
+// short segment in the middle would corrupt every later segment. PSH and
+// ECE state must agree across both slots: PSH is a semantic delimiter
+// (preserving the sender's push boundary) and ECE state must be uniform
+// across a window (the same rule canAppend enforces for in-flow appends).
+//
+// Note: a slot sealed by reorder (canAppend returned false on seq
+// mismatch) keeps psh=false, so this restriction does not block the
+// reorder-fix merge — only legitimate PSH-set seals.
+func canMergeSlots(prev, s *coalesceSlot) bool {
+	if prev.psh {
+		return false
+	}
+	if prev.fk != s.fk {
+		return false
+	}
+	if prev.gsoSize != s.gsoSize {
+		return false
+	}
+	if prev.nextSeq != slotSeedSeq(s) {
+		return false
+	}
+	if prev.numSeg+s.numSeg > tcpCoalesceMaxSegs {
+		return false
+	}
+	if prev.hdrLen+prev.totalPay+s.totalPay > tcpCoalesceBufSize {
+		return false
+	}
+	if len(prev.payIovs[len(prev.payIovs)-1]) != prev.gsoSize {
+		return false
+	}
+	prevFlags := prev.hdrBuf[prev.ipHdrLen+13]
+	sFlags := s.hdrBuf[s.ipHdrLen+13]
+	if (prevFlags^sFlags)&tcpFlagEce != 0 {
+		return false
+	}
+	if !headersMatch(prev.hdrBuf[:prev.hdrLen], s.hdrBuf[:s.hdrLen], prev.isV6, prev.ipHdrLen) {
+		return false
+	}
+	return true
+}
+
+// mergeSlots folds src into dst in place: payIovs concatenated, counters
+// and totals updated, PSH and IP-level CE bits OR'd into the seed header
+// so neither the push signal nor a CE mark is lost. The seed header's
+// seq, gsoSize, and fk are unchanged. Caller is responsible for releasing
+// src (it's no longer in c.slots after this call).
+func mergeSlots(dst, src *coalesceSlot) {
+	dst.payIovs = append(dst.payIovs, src.payIovs...)
+	dst.numSeg += src.numSeg
+	dst.totalPay += src.totalPay
+	dst.nextSeq = src.nextSeq
+	if src.psh {
+		dst.psh = true
+		dst.hdrBuf[dst.ipHdrLen+13] |= tcpFlagPsh
+	}
+	mergeECNIntoSeed(dst.hdrBuf[:dst.ipHdrLen], src.hdrBuf[:src.ipHdrLen], dst.isV6)
+}
+
+// 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 L4 pseudo-header partial sum
+// expected by the virtio NEEDS_CSUM kernel path: the 32-bit accumulator
+// before folding. proto selects the L4 (TCP or UDP); the UDP coalescer
+// reuses these helpers.
+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)
+}

overlay/batch/tcp_coalesce_bench_test.go

@@ -0,0 +1,244 @@
+package batch
+
+import (
+	"encoding/binary"
+	"runtime"
+	"testing"
+
+	"github.com/slackhq/nebula/overlay/tio"
+	"github.com/slackhq/nebula/test"
+	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
+)
+
+// nopTunWriter is a zero-alloc tio.GSOWriter for benchmarks. Discards
+// everything but satisfies the interface the coalescer detects.
+type nopTunWriter struct{}
+
+func (nopTunWriter) Write(p []byte) (int, error)                    { return len(p), nil }
+func (nopTunWriter) Read(_ []wire.TunPacket, _ []byte) (int, error) { return 0, nil }
+func (nopTunWriter) Close() error                                   { return nil }
+func (nopTunWriter) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte, _ wire.GSOProto) error {
+	return nil
+}
+func (nopTunWriter) Capabilities() tio.Capabilities {
+	return tio.Capabilities{TSO: true, USO: true}
+}
+
+// buildTCPv4BulkFlow returns a slice of N adjacent ACK-only TCP segments
+// on a single 5-tuple, each carrying payloadLen bytes. Seq numbers are
+// contiguous so every packet is coalesceable onto the previous one.
+func buildTCPv4BulkFlow(n, payloadLen int) [][]byte {
+	pkts := make([][]byte, n)
+	pay := make([]byte, payloadLen)
+	seq := uint32(1000)
+	for i := range n {
+		pkts[i] = buildTCPv4(seq, tcpAck, pay)
+		seq += uint32(payloadLen)
+	}
+	return pkts
+}
+
+// buildTCPv4Interleaved returns nFlows * perFlow packets with per-flow
+// seq continuity but round-robin across flows — worst case for any
+// "last-slot" cache.
+func buildTCPv4Interleaved(nFlows, perFlow, payloadLen int) [][]byte {
+	pay := make([]byte, payloadLen)
+	seqs := make([]uint32, nFlows)
+	for i := range seqs {
+		seqs[i] = uint32(1000 + i*1000000)
+	}
+	pkts := make([][]byte, 0, nFlows*perFlow)
+	for range perFlow {
+		for f := range nFlows {
+			sport := uint16(10000 + f)
+			pkts = append(pkts, buildTCPv4Ports(sport, 2000, seqs[f], tcpAck, pay))
+			seqs[f] += uint32(payloadLen)
+		}
+	}
+	return pkts
+}
+
+// buildICMPv4 returns a minimal non-TCP packet that takes the passthrough
+// branch in Commit.
+func buildICMPv4() []byte {
+	pkt := make([]byte, 28)
+	pkt[0] = 0x45
+	binary.BigEndian.PutUint16(pkt[2:4], 28)
+	pkt[9] = 1 // ICMP
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+	return pkt
+}
+
+// runCommitBench drives Commit over pkts batchSize at a time, flushing
+// between batches, and reports per-packet cost.
+func runCommitBench(b *testing.B, pkts [][]byte, batchSize int) {
+	b.Helper()
+	c := NewTCPCoalescer(nopTunWriter{}, test.NewLogger(), util.NewArena(0))
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkts[0])))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		pkt := pkts[i%len(pkts)]
+		if err := c.Commit(pkt); err != nil {
+			b.Fatal(err)
+		}
+		if (i+1)%batchSize == 0 {
+			if err := c.Flush(); err != nil {
+				b.Fatal(err)
+			}
+		}
+	}
+	// Drain any trailing partial batch so slot state doesn't leak across runs.
+	_ = c.Flush()
+}
+
+// BenchmarkCommitSingleFlow is the bulk-TCP steady state: one flow,
+// contiguous seq, 1200-byte payloads. Every packet past the seed should
+// append onto the open slot. This is the case we most care about.
+func BenchmarkCommitSingleFlow(b *testing.B) {
+	pkts := buildTCPv4BulkFlow(tcpCoalesceMaxSegs, 1200)
+	runCommitBench(b, pkts, tcpCoalesceMaxSegs)
+}
+
+// BenchmarkCommitInterleaved4 has 4 concurrent bulk flows round-robined.
+// A single-entry fast-path cache will miss on every packet; an N-way
+// cache or map lookup carries the weight.
+func BenchmarkCommitInterleaved4(b *testing.B) {
+	pkts := buildTCPv4Interleaved(4, tcpCoalesceMaxSegs, 1200)
+	runCommitBench(b, pkts, len(pkts))
+}
+
+// BenchmarkCommitInterleaved16 stresses the map at higher flow counts.
+func BenchmarkCommitInterleaved16(b *testing.B) {
+	pkts := buildTCPv4Interleaved(16, tcpCoalesceMaxSegs, 1200)
+	runCommitBench(b, pkts, len(pkts))
+}
+
+// BenchmarkCommitPassthrough exercises the non-TCP branch: parseTCPBase
+// bails early and addPassthrough is the only work.
+func BenchmarkCommitPassthrough(b *testing.B) {
+	pkt := buildICMPv4()
+	pkts := make([][]byte, 64)
+	for i := range pkts {
+		pkts[i] = pkt
+	}
+	runCommitBench(b, pkts, 64)
+}
+
+// BenchmarkCommitNonCoalesceableTCP sends SYN|ACK packets on one flow.
+// Each packet takes the "TCP but not admissible" branch which does a
+// map delete + passthrough. Measures the seal-without-slot cost.
+func BenchmarkCommitNonCoalesceableTCP(b *testing.B) {
+	pay := make([]byte, 0)
+	pkts := make([][]byte, 64)
+	for i := range pkts {
+		pkts[i] = buildTCPv4(uint32(1000+i), tcpSyn|tcpAck, pay)
+	}
+	runCommitBench(b, pkts, 64)
+}
+
+// runMultiCommitBench drives MultiCoalescer.Commit. The dispatcher does
+// the IP/L4 parse once and passes the parsed struct to the lane, so this
+// is the bench that shows the savings of skipping the lane's re-parse.
+func runMultiCommitBench(b *testing.B, pkts [][]byte, batchSize int) {
+	b.Helper()
+	m := NewMultiCoalescer(nopTunWriter{}, test.NewLogger(), util.NewArena(0), true, true)
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkts[0])))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		pkt := pkts[i%len(pkts)]
+		if err := m.Commit(pkt); err != nil {
+			b.Fatal(err)
+		}
+		if (i+1)%batchSize == 0 {
+			if err := m.Flush(); err != nil {
+				b.Fatal(err)
+			}
+		}
+	}
+	_ = m.Flush()
+}
+
+// BenchmarkMultiCommitSingleFlow is the multi-lane analogue of
+// BenchmarkCommitSingleFlow — same workload but routed through the
+// dispatcher. The delta vs the single-lane bench measures dispatcher
+// overhead.
+func BenchmarkMultiCommitSingleFlow(b *testing.B) {
+	pkts := buildTCPv4BulkFlow(tcpCoalesceMaxSegs, 1200)
+	runMultiCommitBench(b, pkts, tcpCoalesceMaxSegs)
+}
+
+// BenchmarkMultiCommitInterleaved4 mirrors BenchmarkCommitInterleaved4
+// through the dispatcher.
+func BenchmarkMultiCommitInterleaved4(b *testing.B) {
+	pkts := buildTCPv4Interleaved(4, tcpCoalesceMaxSegs, 1200)
+	runMultiCommitBench(b, pkts, len(pkts))
+}
+
+// flowKeyPair is one comparison input for the flowKeyCompare bench.
+type flowKeyPair struct{ a, b flowKey }
+
+// makeFlowKey builds an IPv4 flowKey from compact inputs.
+func makeFlowKey(srcLow, dstLow uint32, sport, dport uint16) flowKey {
+	var fk flowKey
+	binary.BigEndian.PutUint32(fk.src[12:16], srcLow)
+	binary.BigEndian.PutUint32(fk.dst[12:16], dstLow)
+	fk.sport = sport
+	fk.dport = dport
+	return fk
+}
+
+// flowKeyCases are the workload mixes flowKeyCompare sees in practice.
+//   - sameFlow: equal keys; tests the equal-path cost (sort runs hit this
+//     repeatedly when many segments share a flow).
+//   - sportDiffers: same src/dst/dport, different sport — the typical
+//     "sibling flows from one host to one server" pattern.
+//   - dstDiffers: same src/sport/dport, different dst — outbound to many
+//     servers from a fixed local port.
+//   - allDiffer: every field differs; worst case for short-circuiting.
+func flowKeyCases() map[string][]flowKeyPair {
+	const n = 64
+	cases := map[string][]flowKeyPair{
+		"sameFlow":     make([]flowKeyPair, n),
+		"sportDiffers": make([]flowKeyPair, n),
+		"dstDiffers":   make([]flowKeyPair, n),
+		"allDiffer":    make([]flowKeyPair, n),
+	}
+	for i := range n {
+		base := makeFlowKey(0x0a000001, 0x0a000002, 40000, 443)
+		cases["sameFlow"][i] = flowKeyPair{a: base, b: base}
+		cases["sportDiffers"][i] = flowKeyPair{
+			a: base,
+			b: makeFlowKey(0x0a000001, 0x0a000002, uint16(40001+i), 443),
+		}
+		cases["dstDiffers"][i] = flowKeyPair{
+			a: base,
+			b: makeFlowKey(0x0a000001, uint32(0x0a000002+i+1), 40000, 443),
+		}
+		cases["allDiffer"][i] = flowKeyPair{
+			a: makeFlowKey(uint32(0x0a000001+i), uint32(0x0a000002+i), uint16(40000+i), uint16(80+i)),
+			b: makeFlowKey(uint32(0x0b000001+i), uint32(0x0b000002+i), uint16(50000+i), uint16(443+i)),
+		}
+	}
+	return cases
+}
+
+// BenchmarkFlowKeyCompare measures flowKeyCompare across the workloads
+// the sort step actually sees. Use this to compare reorderings.
+func BenchmarkFlowKeyCompare(b *testing.B) {
+	for name, pairs := range flowKeyCases() {
+		b.Run(name, func(b *testing.B) {
+			b.ReportAllocs()
+			b.ResetTimer()
+			var sink int
+			for i := 0; i < b.N; i++ {
+				p := pairs[i&(len(pairs)-1)]
+				sink += flowKeyCompare(p.a, p.b)
+			}
+			runtime.KeepAlive(sink)
+		})
+	}
+}

overlay/batch/tx_batch.go

@@ -0,0 +1,60 @@
+package batch
+
+import (
+	"net/netip"
+
+	"github.com/slackhq/nebula/util"
+)
+
+const SendBatchCap = 128
+
+// batchWriter is the minimal subset of udp.Conn needed by SendBatch to flush.
+type batchWriter interface {
+	WriteBatch(bufs [][]byte, addrs []netip.AddrPort, outerECNs []byte) error
+}
+
+// SendBatch accumulates encrypted UDP packets and flushes them via WriteBatch.
+// One SendBatch is owned by each listenIn goroutine; no locking is needed.
+// Slot bytes are borrowed from the injected Arena and remain valid until
+// Flush, which Resets the arena.
+type SendBatch struct {
+	out   batchWriter
+	bufs  [][]byte
+	dsts  []netip.AddrPort
+	ecns  []byte
+	arena *util.Arena
+}
+
+// NewSendBatch makes a SendBatch with batchCap slots backed by arena.
+func NewSendBatch(out batchWriter, batchCap int, arena *util.Arena) *SendBatch {
+	return &SendBatch{
+		out:   out,
+		bufs:  make([][]byte, 0, batchCap),
+		dsts:  make([]netip.AddrPort, 0, batchCap),
+		ecns:  make([]byte, 0, batchCap),
+		arena: arena,
+	}
+}
+
+func (b *SendBatch) Reserve(sz int) []byte {
+	return b.arena.Reserve(sz)
+}
+
+func (b *SendBatch) Commit(pkt []byte, dst netip.AddrPort, outerECN byte) {
+	b.bufs = append(b.bufs, pkt)
+	b.dsts = append(b.dsts, dst)
+	b.ecns = append(b.ecns, outerECN)
+}
+
+func (b *SendBatch) Flush() error {
+	var err error
+	if len(b.bufs) > 0 {
+		err = b.out.WriteBatch(b.bufs, b.dsts, b.ecns)
+	}
+	clear(b.bufs)
+	b.bufs = b.bufs[:0]
+	b.dsts = b.dsts[:0]
+	b.ecns = b.ecns[:0]
+	b.arena.Reset()
+	return err
+}

overlay/batch/tx_batch_test.go

@@ -0,0 +1,126 @@
+package batch
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/slackhq/nebula/util"
+)
+
+type fakeBatchWriter struct {
+	bufs  [][]byte
+	addrs []netip.AddrPort
+	ecns  []byte
+}
+
+func (w *fakeBatchWriter) WriteBatch(bufs [][]byte, addrs []netip.AddrPort, ecns []byte) error {
+	// Snapshot — SendBatch.Flush nils its slot pointers right after WriteBatch
+	// returns, so tests must capture data before that happens.
+	w.bufs = make([][]byte, len(bufs))
+	for i, b := range bufs {
+		cp := make([]byte, len(b))
+		copy(cp, b)
+		w.bufs[i] = cp
+	}
+	w.addrs = append(w.addrs[:0], addrs...)
+	w.ecns = append(w.ecns[:0], ecns...)
+	return nil
+}
+
+func TestSendBatchReserveCommitFlush(t *testing.T) {
+	fw := &fakeBatchWriter{}
+	b := NewSendBatch(fw, 4, util.NewArena(32))
+
+	ap := netip.MustParseAddrPort("10.0.0.1:4242")
+	for i := 0; i < 4; i++ {
+		slot := b.Reserve(32)
+		if cap(slot) != 32 {
+			t.Fatalf("slot %d: cap=%d want 32", i, cap(slot))
+		}
+		pkt := append(slot[:0], byte(i), byte(i+1), byte(i+2))
+		b.Commit(pkt, ap, 0)
+	}
+	if err := b.Flush(); err != nil {
+		t.Fatalf("Flush: %v", err)
+	}
+	if len(fw.bufs) != 4 {
+		t.Fatalf("WriteBatch got %d bufs want 4", len(fw.bufs))
+	}
+	for i, buf := range fw.bufs {
+		if len(buf) != 3 || buf[0] != byte(i) {
+			t.Errorf("buf %d: %x", i, buf)
+		}
+		if fw.addrs[i] != ap {
+			t.Errorf("addr %d: got %v want %v", i, fw.addrs[i], ap)
+		}
+	}
+
+	// Flush again with nothing committed — should be a no-op.
+	fw.bufs = nil
+	if err := b.Flush(); err != nil {
+		t.Fatalf("empty Flush: %v", err)
+	}
+	if fw.bufs != nil {
+		t.Fatalf("empty Flush triggered WriteBatch")
+	}
+
+	// Reuse after Flush.
+	slot := b.Reserve(32)
+	if cap(slot) != 32 {
+		t.Fatalf("after Flush Reserve wrong cap: %d", cap(slot))
+	}
+}
+
+func TestSendBatchSlotsDoNotOverlap(t *testing.T) {
+	fw := &fakeBatchWriter{}
+	b := NewSendBatch(fw, 3, util.NewArena(8))
+	ap := netip.MustParseAddrPort("10.0.0.1:80")
+
+	for i := 0; i < 3; i++ {
+		s := b.Reserve(8)
+		pkt := append(s[:0], byte(0xA0+i), byte(0xB0+i))
+		b.Commit(pkt, ap, 0)
+	}
+	if err := b.Flush(); err != nil {
+		t.Fatalf("Flush: %v", err)
+	}
+
+	for i, buf := range fw.bufs {
+		if buf[0] != byte(0xA0+i) || buf[1] != byte(0xB0+i) {
+			t.Errorf("slot %d corrupted: %x", i, buf)
+		}
+	}
+}
+
+func TestSendBatchGrowPreservesCommitted(t *testing.T) {
+	fw := &fakeBatchWriter{}
+	// Tiny initial backing forces a grow on the second Reserve.
+	b := NewSendBatch(fw, 1, util.NewArena(4))
+	ap := netip.MustParseAddrPort("10.0.0.1:80")
+
+	s1 := b.Reserve(4)
+	pkt1 := append(s1[:0], 0x11, 0x22, 0x33, 0x44)
+	b.Commit(pkt1, ap, 0)
+
+	s2 := b.Reserve(8) // exceeds remaining cap, triggers grow
+	pkt2 := append(s2[:0], 0xA, 0xB, 0xC, 0xD, 0xE)
+	b.Commit(pkt2, ap, 0)
+
+	// pkt1 must still be intact even though backing reallocated.
+	if pkt1[0] != 0x11 || pkt1[3] != 0x44 {
+		t.Fatalf("first packet corrupted by grow: %x", pkt1)
+	}
+
+	if err := b.Flush(); err != nil {
+		t.Fatalf("Flush: %v", err)
+	}
+	if len(fw.bufs) != 2 {
+		t.Fatalf("got %d bufs want 2", len(fw.bufs))
+	}
+	if fw.bufs[0][0] != 0x11 || fw.bufs[0][3] != 0x44 {
+		t.Errorf("first packet on the wire: %x", fw.bufs[0])
+	}
+	if fw.bufs[1][0] != 0xA || fw.bufs[1][4] != 0xE {
+		t.Errorf("second packet on the wire: %x", fw.bufs[1])
+	}
+}

overlay/batch/udp_coalesce.go

@@ -0,0 +1,339 @@
+package batch
+
+import (
+	"encoding/binary"
+	"io"
+
+	"github.com/slackhq/nebula/overlay/tio"
+	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
+)
+
+// ipProtoUDP is the IANA protocol number for UDP.
+const ipProtoUDP = 17
+
+// udpCoalesceBufSize caps total bytes per UDP superpacket. Mirrors the
+// kernel's gso_max_size; payloads beyond this are emitted as-is.
+const udpCoalesceBufSize = 65535
+
+// udpCoalesceMaxSegs caps how many segments we'll coalesce. Kernel UDP-GSO
+// accepts up to 64 segments per skb (UDP_MAX_SEGMENTS); stay under that.
+const udpCoalesceMaxSegs = 64
+
+// udpCoalesceHdrCap is the scratch space we copy a seed's IP+UDP header
+// into. IPv6 (40) + UDP (8) = 48; round up for safety.
+const udpCoalesceHdrCap = 64
+
+// udpSlot is one entry in the UDPCoalescer's ordered event queue. Same
+// passthrough-vs-coalesced shape as the TCP coalescer's slot, but no
+// seq/PSH/CWR bookkeeping — UDP segments only need 5-tuple + length
+// matching to coalesce.
+type udpSlot struct {
+	passthrough bool
+	rawPkt      []byte // borrowed when passthrough
+
+	fk       flowKey
+	hdrBuf   [udpCoalesceHdrCap]byte
+	hdrLen   int
+	ipHdrLen int
+	isV6     bool
+	gsoSize  int // per-segment UDP payload length
+	numSeg   int
+	totalPay int
+	// sealed closes the chain: set when a sub-gsoSize segment is appended
+	// (kernel UDP-GSO requires every segment but the last to be exactly
+	// gsoSize) or when limits are hit. No further appends after.
+	sealed  bool
+	payIovs [][]byte
+}
+
+// UDPCoalescer accumulates adjacent in-flow UDP datagrams across multiple
+// concurrent flows and emits each flow's run as a single GSO_UDP_L4
+// superpacket via tio.GSOWriter. Falls back to per-packet writes when the
+// underlying writer doesn't support USO.
+//
+// All output — coalesced or not — is deferred until Flush so per-flow
+// arrival order is preserved on the wire. Cross-flow order is NOT preserved
+// across the TCP/UDP/passthrough split when this coalescer runs alongside
+// others — see multi_coalesce.go. Per-flow order is preserved because a
+// single 5-tuple only ever lands in one lane and each lane preserves its
+// own slot order.
+//
+// Owns no locks; one coalescer per TUN write queue.
+type UDPCoalescer struct {
+	plainW io.Writer
+	gsoW   tio.GSOWriter // nil when the queue can't accept GSO_UDP_L4
+
+	slots     []*udpSlot
+	openSlots map[flowKey]*udpSlot
+	pool      []*udpSlot
+
+	// arena is injected; see TCPCoalescer.arena for the contract.
+	arena *util.Arena
+}
+
+// NewUDPCoalescer wraps w. The caller is responsible for only constructing
+// this when the underlying Queue's Capabilities advertise USO; otherwise
+// the kernel may reject GSO_UDP_L4 writes. If w does not implement
+// tio.GSOWriter at all (single-packet Queue), the coalescer degrades to
+// plain Writes — same defensive shape as the TCP coalescer.
+func NewUDPCoalescer(w tio.Queue, arena *util.Arena) *UDPCoalescer {
+	c := &UDPCoalescer{
+		plainW:    w,
+		slots:     make([]*udpSlot, 0, initialSlots),
+		openSlots: make(map[flowKey]*udpSlot, initialSlots),
+		pool:      make([]*udpSlot, 0, initialSlots),
+		arena:     arena,
+	}
+	if gw, ok := tio.SupportsGSO(w, wire.GSOProtoUDP); ok {
+		c.gsoW = gw
+	}
+	return c
+}
+
+// parsedUDP holds the fields extracted from a single parse so later steps
+// (admission, slot lookup, canAppend) don't re-walk the header.
+type parsedUDP struct {
+	fk       flowKey
+	ipHdrLen int
+	hdrLen   int // ipHdrLen + 8
+	payLen   int
+}
+
+// parseUDP extracts the flow key and IP/UDP offsets for a UDP packet.
+// Returns ok=false for non-UDP, malformed, or unsupported header shapes
+// (IPv4 with options/fragmentation, IPv6 with extension headers).
+func parseUDP(pkt []byte) (parsedUDP, bool) {
+	var p parsedUDP
+	ip, ok := parseIPPrologue(pkt, ipProtoUDP)
+	if !ok {
+		return p, false
+	}
+	pkt = ip.pkt
+	p.fk = ip.fk
+	p.ipHdrLen = ip.ipHdrLen
+
+	if len(pkt) < p.ipHdrLen+8 {
+		return p, false
+	}
+	p.hdrLen = p.ipHdrLen + 8
+	// UDP `length` field: must equal IP-derived length-of-UDP-header-plus-payload.
+	udpLen := int(binary.BigEndian.Uint16(pkt[p.ipHdrLen+4 : p.ipHdrLen+6]))
+	if udpLen < 8 || udpLen > len(pkt)-p.ipHdrLen {
+		return p, false
+	}
+	p.payLen = udpLen - 8
+	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
+}
+
+func (c *UDPCoalescer) Reserve(sz int) []byte {
+	return c.arena.Reserve(sz)
+}
+
+// Commit borrows pkt. The caller must keep pkt valid until the next Flush.
+func (c *UDPCoalescer) Commit(pkt []byte) error {
+	if c.gsoW == nil {
+		c.addPassthrough(pkt)
+		return nil
+	}
+	info, ok := parseUDP(pkt)
+	if !ok {
+		c.addPassthrough(pkt)
+		return nil
+	}
+	return c.commitParsed(pkt, info)
+}
+
+// commitParsed is the post-parse half of Commit. The caller must have
+// already verified parseUDP succeeded. Used by MultiCoalescer.Commit to
+// avoid re-walking the IP/UDP header.
+func (c *UDPCoalescer) commitParsed(pkt []byte, info parsedUDP) error {
+	if c.gsoW == nil {
+		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.sealed {
+				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
+}
+
+func (c *UDPCoalescer) 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)
+	}
+	clear(c.slots)
+	c.slots = c.slots[:0]
+	clear(c.openSlots)
+	c.arena.Reset()
+	return first
+}
+
+func (c *UDPCoalescer) addPassthrough(pkt []byte) {
+	s := c.take()
+	s.passthrough = true
+	s.rawPkt = pkt
+	c.slots = append(c.slots, s)
+}
+
+func (c *UDPCoalescer) seed(pkt []byte, info parsedUDP) {
+	if info.hdrLen > udpCoalesceHdrCap || info.hdrLen+info.payLen > udpCoalesceBufSize {
+		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.sealed = false
+	s.payIovs = append(s.payIovs[:0], pkt[info.hdrLen:info.hdrLen+info.payLen])
+	c.slots = append(c.slots, s)
+	c.openSlots[info.fk] = s
+}
+
+// canAppend reports whether info's packet extends the slot's seed.
+// Kernel UDP-GSO requires every segment except possibly the last to be
+// exactly gsoSize, and the last may be shorter (≤ gsoSize).
+func (c *UDPCoalescer) canAppend(s *udpSlot, pkt []byte, info parsedUDP) bool {
+	if s.sealed {
+		return false
+	}
+	if info.hdrLen != s.hdrLen {
+		return false
+	}
+	if s.numSeg >= udpCoalesceMaxSegs {
+		return false
+	}
+	if info.payLen > s.gsoSize {
+		return false
+	}
+	if s.hdrLen+s.totalPay+info.payLen > udpCoalesceBufSize {
+		return false
+	}
+	if !udpHeadersMatch(s.hdrBuf[:s.hdrLen], pkt[:info.hdrLen], s.isV6, s.ipHdrLen) {
+		return false
+	}
+	return true
+}
+
+func (c *UDPCoalescer) appendPayload(s *udpSlot, pkt []byte, info parsedUDP) {
+	s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
+	s.numSeg++
+	s.totalPay += info.payLen
+	// Merge IP-level CE marks into the seed (same trick TCP coalescer uses).
+	mergeECNIntoSeed(s.hdrBuf[:s.ipHdrLen], pkt[:s.ipHdrLen], s.isV6)
+	if info.payLen < s.gsoSize {
+		// Last-segment-can-be-shorter: this seals the chain.
+		s.sealed = true
+	}
+}
+
+func (c *UDPCoalescer) take() *udpSlot {
+	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 &udpSlot{}
+}
+
+func (c *UDPCoalescer) release(s *udpSlot) {
+	s.passthrough = false
+	s.rawPkt = nil
+	clear(s.payIovs)
+	s.payIovs = s.payIovs[:0]
+	s.numSeg = 0
+	s.totalPay = 0
+	s.sealed = false
+	c.pool = append(c.pool, s)
+}
+
+// flushSlot patches the IP header total length / IPv6 payload length and
+// the UDP length to the *total* across all coalesced segments, then seeds
+// the UDP checksum field with the pseudo-header partial (single-fold, not
+// inverted) per virtio NEEDS_CSUM. The kernel's ip_rcv_core (v4) and
+// ip6_rcv_core (v6) trim the skb to those length fields, so per-segment
+// values would silently drop everything but the first segment. The kernel
+// then walks each segment in __udp_gso_segment, recomputing per-segment
+// uh->len / iph->tot_len / IPv6 plen and adjusting the checksum via
+// `check = csum16_add(csum16_sub(uh->check, uh->len), newlen)` — meaning
+// our seed's uh->check must be consistent with the seed's uh->len, which
+// is what passing the total to both pseudoSum and the UDP length field
+// guarantees.
+func (c *UDPCoalescer) flushSlot(s *udpSlot) error {
+	hdr := s.hdrBuf[:s.hdrLen]
+	total := s.hdrLen + s.totalPay // full IP+UDP+all_payloads bytes
+	l4Len := total - s.ipHdrLen    // total UDP (8 + sum of payloads)
+
+	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]))
+	}
+
+	// UDP length field (offset 4 inside the UDP header) = total UDP size.
+	binary.BigEndian.PutUint16(hdr[s.ipHdrLen+4:s.ipHdrLen+6], uint16(l4Len))
+
+	var psum uint32
+	if s.isV6 {
+		psum = pseudoSumIPv6(hdr[8:24], hdr[24:40], ipProtoUDP, l4Len)
+	} else {
+		psum = pseudoSumIPv4(hdr[12:16], hdr[16:20], ipProtoUDP, l4Len)
+	}
+	udpCsumOff := s.ipHdrLen + 6
+	binary.BigEndian.PutUint16(hdr[udpCsumOff:udpCsumOff+2], foldOnceNoInvert(psum))
+
+	return c.gsoW.WriteGSO(hdr[:s.ipHdrLen], hdr[s.ipHdrLen:], s.payIovs, wire.GSOProtoUDP)
+}
+
+// udpHeadersMatch compares two IP+UDP header prefixes for byte-equality on
+// every field that must be identical across coalesced segments. Length
+// fields and the ECN bits in IP TOS/TC are masked out — appendPayload
+// merges CE into the seed; flushSlot rewrites lengths.
+func udpHeadersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	if !ipHeadersMatch(a, b, isV6) {
+		return false
+	}
+	// UDP: compare sport+dport ([0:4]). Skip length [4:6] and checksum [6:8] —
+	// length varies (we rewrite at flush) and the checksum will be redone.
+	udp := ipHdrLen
+	if a[udp] != b[udp] || a[udp+1] != b[udp+1] || a[udp+2] != b[udp+2] || a[udp+3] != b[udp+3] {
+		return false
+	}
+	return true
+}

overlay/batch/udp_coalesce_test.go

@@ -0,0 +1,385 @@
+package batch
+
+import (
+	"encoding/binary"
+	"testing"
+
+	"github.com/slackhq/nebula/util"
+)
+
+// buildUDPv4 builds a minimal IPv4+UDP packet with the given payload and ports.
+func buildUDPv4(sport, dport uint16, payload []byte) []byte {
+	const ipHdrLen = 20
+	const udpHdrLen = 8
+	total := ipHdrLen + udpHdrLen + 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] = ipProtoUDP
+	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.PutUint16(pkt[24:26], uint16(udpHdrLen+len(payload)))
+	binary.BigEndian.PutUint16(pkt[26:28], 0)
+
+	copy(pkt[28:], payload)
+	return pkt
+}
+
+// buildUDPv6 builds a minimal IPv6+UDP packet.
+func buildUDPv6(sport, dport uint16, payload []byte) []byte {
+	const ipHdrLen = 40
+	const udpHdrLen = 8
+	total := ipHdrLen + udpHdrLen + len(payload)
+	pkt := make([]byte, total)
+
+	pkt[0] = 0x60
+	binary.BigEndian.PutUint16(pkt[4:6], uint16(udpHdrLen+len(payload)))
+	pkt[6] = ipProtoUDP
+	pkt[7] = 64
+	pkt[8] = 0xfe
+	pkt[9] = 0x80
+	pkt[23] = 1
+	pkt[24] = 0xfe
+	pkt[25] = 0x80
+	pkt[39] = 2
+
+	binary.BigEndian.PutUint16(pkt[40:42], sport)
+	binary.BigEndian.PutUint16(pkt[42:44], dport)
+	binary.BigEndian.PutUint16(pkt[44:46], uint16(udpHdrLen+len(payload)))
+	binary.BigEndian.PutUint16(pkt[46:48], 0)
+
+	copy(pkt[48:], payload)
+	return pkt
+}
+
+func TestUDPCoalescerPassthroughWhenGSOUnavailable(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: false}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pkt := buildUDPv4(1000, 53, make([]byte, 100))
+	if err := c.Commit(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
+		t.Fatalf("no Add-time writes: 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 TestUDPCoalescerNonUDPPassthrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	// ICMP packet
+	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.Commit(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 must pass through unchanged: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+func TestUDPCoalescerSeedThenFlushAlone(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pkt := buildUDPv4(1000, 53, make([]byte, 800))
+	if err := c.Commit(pkt); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Single-segment flush goes through WriteGSO; the writer infers GSO_NONE
+	// from len(pays)==1 and the kernel fills in the UDP csum (NEEDS_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))
+	}
+}
+
+func TestUDPCoalescerCoalescesEqualSized(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pay := make([]byte, 1200)
+	for i := 0; i < 3; i++ {
+		if err := c.Commit(buildUDPv4(1000, 53, 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.pays) != 3 {
+		t.Errorf("pay count=%d want 3", len(g.pays))
+	}
+	if g.csumStart != 20 {
+		t.Errorf("csumStart=%d want 20", g.csumStart)
+	}
+	// IP totalLen and UDP length must be the TOTAL across all segments —
+	// the kernel's ip_rcv_core trims skbs to iph->tot_len, so a per-segment
+	// value would silently drop everything but the first segment. Total =
+	// IP(20) + UDP(8) + 3*1200 = 3628.
+	gotTotalLen := binary.BigEndian.Uint16(g.hdr[2:4])
+	if gotTotalLen != 3628 {
+		t.Errorf("ipv4 total_len=%d want 3628 (must be total across segments)", gotTotalLen)
+	}
+	gotUDPLen := binary.BigEndian.Uint16(g.hdr[20+4 : 20+6])
+	if gotUDPLen != 8+3*1200 {
+		t.Errorf("udp len=%d want %d", gotUDPLen, 8+3*1200)
+	}
+}
+
+// Last segment may be shorter, sealing the chain.
+func TestUDPCoalescerShortLastSegmentSeals(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	full := make([]byte, 1200)
+	tail := make([]byte, 600)
+	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildUDPv4(1000, 53, tail)); err != nil {
+		t.Fatal(err)
+	}
+	// A 4th packet, even same-sized, must NOT join — chain is sealed.
+	if err := c.Commit(buildUDPv4(1000, 53, full)); 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 (sealed + new seed), got %d", len(w.gsoWrites))
+	}
+	if len(w.gsoWrites[0].pays) != 3 {
+		t.Errorf("first super: want 3 pays, got %d", len(w.gsoWrites[0].pays))
+	}
+	if len(w.gsoWrites[1].pays) != 1 {
+		t.Errorf("second super: want 1 pay (re-seed), got %d", len(w.gsoWrites[1].pays))
+	}
+}
+
+// A larger-than-gsoSize packet cannot extend the slot — it reseeds.
+func TestUDPCoalescerLargerThanSeedReseeds(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	if err := c.Commit(buildUDPv4(1000, 53, make([]byte, 800))); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildUDPv4(1000, 53, make([]byte, 1200))); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 separate seeds, got %d", len(w.gsoWrites))
+	}
+}
+
+// Different 5-tuples must not coalesce.
+func TestUDPCoalescerDifferentFlowsKeepSeparate(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pay := make([]byte, 800)
+	if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildUDPv4(2000, 53, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(buildUDPv4(2000, 53, pay)); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// Two flows × 2 datagrams each = 2 superpackets of 2 segments.
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes (one per flow), got %d", len(w.gsoWrites))
+	}
+	for i, g := range w.gsoWrites {
+		if len(g.pays) != 2 {
+			t.Errorf("super %d: want 2 pays, got %d", i, len(g.pays))
+		}
+	}
+}
+
+// Caps at udpCoalesceMaxSegs.
+func TestUDPCoalescerCapsAtMaxSegs(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pay := make([]byte, 100)
+	for i := 0; i < udpCoalesceMaxSegs+5; i++ {
+		if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	// First superpacket holds udpCoalesceMaxSegs segments; the spillover
+	// reseeds a new one.
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 gso writes (cap then reseed), got %d", len(w.gsoWrites))
+	}
+	if len(w.gsoWrites[0].pays) != udpCoalesceMaxSegs {
+		t.Errorf("first super: pays=%d want %d", len(w.gsoWrites[0].pays), udpCoalesceMaxSegs)
+	}
+	if len(w.gsoWrites[1].pays) != 5 {
+		t.Errorf("second super: pays=%d want 5", len(w.gsoWrites[1].pays))
+	}
+}
+
+// CE marks on appended segments must be merged into the seed's IP TOS.
+func TestUDPCoalescerMergesCEMark(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pay := make([]byte, 800)
+	pkt0 := buildUDPv4(1000, 53, pay) // ECN=00
+	pkt1 := buildUDPv4(1000, 53, pay)
+	pkt1[1] = 0x03 // CE
+	pkt2 := buildUDPv4(1000, 53, pay)
+	if err := c.Commit(pkt0); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(pkt1); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(pkt2); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 1 {
+		t.Fatalf("want 1 merged gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
+	}
+	if w.gsoWrites[0].hdr[1]&0x03 != 0x03 {
+		t.Errorf("CE not merged into seed (tos=%#x)", w.gsoWrites[0].hdr[1])
+	}
+}
+
+// IPv6 path: same flow, equal-sized → coalesced.
+func TestUDPCoalescerIPv6Coalesces(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pay := make([]byte, 1200)
+	for i := 0; i < 3; i++ {
+		if err := c.Commit(buildUDPv6(1000, 53, 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", len(w.gsoWrites))
+	}
+	g := w.gsoWrites[0]
+	if !g.isV6 {
+		t.Errorf("expected v6 write")
+	}
+	if g.csumStart != 40 {
+		t.Errorf("csumStart=%d want 40", g.csumStart)
+	}
+	// IPv6 payload_len and UDP length must be TOTAL — kernel's
+	// ip6_rcv_core trims to payload_len + ipv6 hdr size. Total UDP = 8 +
+	// 3*1200 = 3608.
+	gotPlen := binary.BigEndian.Uint16(g.hdr[4:6])
+	if gotPlen != 8+3*1200 {
+		t.Errorf("ipv6 payload_len=%d want %d (must be total)", gotPlen, 8+3*1200)
+	}
+	gotUDPLen := binary.BigEndian.Uint16(g.hdr[40+4 : 40+6])
+	if gotUDPLen != 8+3*1200 {
+		t.Errorf("udp len=%d want %d", gotUDPLen, 8+3*1200)
+	}
+}
+
+// DSCP differences must reseed (headers don't match outside ECN).
+func TestUDPCoalescerDSCPMismatchReseeds(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pay := make([]byte, 800)
+	pkt0 := buildUDPv4(1000, 53, pay)
+	pkt1 := buildUDPv4(1000, 53, pay)
+	pkt1[1] = 0xb8 // EF DSCP, ECN=0
+	if err := c.Commit(pkt0); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Commit(pkt1); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	if len(w.gsoWrites) != 2 {
+		t.Fatalf("want 2 separate seeds (different DSCP), got %d", len(w.gsoWrites))
+	}
+}
+
+// Fragmented IPv4 must not be coalesced.
+func TestUDPCoalescerFragmentedIPv4PassesThrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pkt := buildUDPv4(1000, 53, make([]byte, 200))
+	binary.BigEndian.PutUint16(pkt[6:8], 0x2000) // MF=1
+	if err := c.Commit(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("frag must pass through plain, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}
+
+// IPv4 with options is not admissible (we require IHL=5).
+func TestUDPCoalescerIPv4WithOptionsPassesThrough(t *testing.T) {
+	w := &fakeTunWriter{gsoEnabled: true}
+	c := NewUDPCoalescer(w, util.NewArena(0))
+	pkt := buildUDPv4(1000, 53, make([]byte, 200))
+	pkt[0] = 0x46 // IHL = 6 (24-byte IPv4 header — has options)
+	if err := c.Commit(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("ipv4-with-options must pass through plain, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
+	}
+}

overlay/checksum/checksum_amd64.go

@@ -0,0 +1,23 @@
+package checksum
+
+import (
+	"golang.org/x/sys/cpu"
+	gvisorchecksum "gvisor.dev/gvisor/pkg/tcpip/checksum"
+)
+
+//go:noescape
+func checksumAVX2(buf []byte, initial uint16) uint16
+
+var hasAVX2 = cpu.X86.HasAVX2
+
+// Checksum computes the RFC 1071 ones-complement sum of buf, seeded with
+// initial. It is a drop-in replacement for gvisor's checksum.Checksum that
+// dispatches to a hand-written AVX2 routine on amd64 CPUs that support it,
+// falling back to gvisor's pure-Go implementation otherwise. The result
+// matches gvisor's bit-for-bit for any buffer length and initial seed.
+func Checksum(buf []byte, initial uint16) uint16 {
+	if hasAVX2 {
+		return checksumAVX2(buf, initial)
+	}
+	return gvisorchecksum.Checksum(buf, initial)
+}

overlay/checksum/checksum_amd64.s

@@ -0,0 +1,157 @@
+#include "textflag.h"
+
+// func checksumAVX2(buf []byte, initial uint16) uint16
+//
+// Computes the RFC 1071 ones-complement sum of buf, seeded with initial.
+//
+// Algorithm: sum the buffer treating it as a stream of uint32s in machine
+// (little-endian) byte order, accumulating into 64-bit lanes (top 32 bits
+// hold cross-add carries — at 1 byte / lane / iter we have 32 bits of
+// headroom which is far more than the 16 KB/64 KB max practical inputs).
+// At the end we fold to 16 bits and byte-swap once to recover the on-wire
+// (big-endian) result. RFC 1071 §1.2.B byte-order independence makes this
+// equivalent to summing as 16-bit big-endian words.
+//
+// The ymm accumulators (Y4..Y7) hold 4 uint64 lanes each = 16 parallel
+// partial sums. The main loop loads 64 bytes per iter as four 16-byte
+// chunks, zero-extending each chunk's four uint32s into a ymm via
+// VPMOVZXDQ-from-memory, then VPADDQ into a separate accumulator per
+// chunk to break the dep chain. After the vector loop the lane sums are
+// horizontally reduced and merged with a scalar accumulator that handles
+// the trailing 0..63 bytes plus the (byte-swapped) initial seed.
+TEXT ·checksumAVX2(SB), NOSPLIT, $0-34
+	MOVQ    buf_base+0(FP), SI
+	MOVQ    buf_len+8(FP), CX
+	MOVWQZX initial+24(FP), AX
+
+	// Pre-byteswap initial into the LE-summing space so it merges directly
+	// with the rest of the accumulator. The final fold's bswap16 will undo
+	// this and convert the whole result back to BE.
+	XCHGB AH, AL
+
+	CMPQ CX, $32
+	JLT  scalar_tail
+
+	VPXOR Y4, Y4, Y4
+	VPXOR Y5, Y5, Y5
+	VPXOR Y6, Y6, Y6
+	VPXOR Y7, Y7, Y7
+
+	CMPQ CX, $64
+	JLT  loop32
+
+loop64:
+	VPMOVZXDQ (SI), Y0
+	VPMOVZXDQ 16(SI), Y1
+	VPMOVZXDQ 32(SI), Y2
+	VPMOVZXDQ 48(SI), Y3
+	VPADDQ    Y0, Y4, Y4
+	VPADDQ    Y1, Y5, Y5
+	VPADDQ    Y2, Y6, Y6
+	VPADDQ    Y3, Y7, Y7
+	ADDQ      $64, SI
+	SUBQ      $64, CX
+	CMPQ      CX, $64
+	JGE       loop64
+
+loop32:
+	CMPQ      CX, $32
+	JLT       reduce_vec
+	VPMOVZXDQ (SI), Y0
+	VPMOVZXDQ 16(SI), Y1
+	VPADDQ    Y0, Y4, Y4
+	VPADDQ    Y1, Y5, Y5
+	ADDQ      $32, SI
+	SUBQ      $32, CX
+	JMP       loop32
+
+reduce_vec:
+	// Combine the four ymm accumulators into Y4.
+	VPADDQ Y5, Y4, Y4
+	VPADDQ Y7, Y6, Y6
+	VPADDQ Y6, Y4, Y4
+
+	// Horizontally reduce Y4's four uint64 lanes to a single scalar.
+	VEXTRACTI128 $1, Y4, X5
+	VPADDQ       X5, X4, X4
+	VPSHUFD      $0x4e, X4, X5
+	VPADDQ       X5, X4, X4
+	VMOVQ        X4, R8
+	VZEROUPPER
+
+	ADDQ R8, AX
+	ADCQ $0, AX
+
+scalar_tail:
+	// Handle remaining 0..63 bytes (or the entire buffer if it was < 32).
+	CMPQ CX, $8
+	JLT  tail4
+
+loop8:
+	ADDQ (SI), AX
+	ADCQ $0, AX
+	ADDQ $8, SI
+	SUBQ $8, CX
+	CMPQ CX, $8
+	JGE  loop8
+
+tail4:
+	CMPQ CX, $4
+	JLT  tail2
+	MOVL (SI), R8
+	ADDQ R8, AX
+	ADCQ $0, AX
+	ADDQ $4, SI
+	SUBQ $4, CX
+
+tail2:
+	CMPQ    CX, $2
+	JLT     tail1
+	MOVWQZX (SI), R8
+	ADDQ    R8, AX
+	ADCQ    $0, AX
+	ADDQ    $2, SI
+	SUBQ    $2, CX
+
+tail1:
+	TESTQ   CX, CX
+	JZ      fold
+	MOVBQZX (SI), R8
+	ADDQ    R8, AX
+	ADCQ    $0, AX
+
+fold:
+	// Fold the 64-bit accumulator to 16 bits via four rounds, mirroring
+	// gvisor's reduce(). Each pair (split, add) halves the live width;
+	// the truncation steps absorb the single bit that may be left over
+	// after each add so the next round's bound holds.
+
+	// 64 → 33 bits.
+	MOVQ AX, R8
+	SHRQ $32, R8
+	MOVL AX, AX
+	ADDQ R8, AX
+
+	// 33 → 32 bits. AX += (AX>>32); truncate to 32. AX is now ≤ 0xFFFF_FFFF.
+	MOVQ AX, R8
+	SHRQ $32, R8
+	ADDQ R8, AX
+	MOVL AX, AX
+
+	// 32 → 17 bits.
+	MOVQ    AX, R8
+	SHRQ    $16, R8
+	MOVWQZX AX, AX
+	ADDQ    R8, AX
+
+	// 17 → 16 bits. AX += (AX>>16); the trailing MOVW truncates bit 16.
+	MOVQ AX, R8
+	SHRQ $16, R8
+	ADDQ R8, AX
+
+	// AX low 16 bits hold the 16-bit sum in machine (LE) byte order; flip
+	// to big-endian to match the gvisor API contract.
+	XCHGB AH, AL
+
+	MOVW AX, ret+32(FP)
+	RET

overlay/checksum/checksum_arm64.go

@@ -0,0 +1,12 @@
+package checksum
+
+//go:noescape
+func checksumNEON(buf []byte, initial uint16) uint16
+
+// Checksum computes the RFC 1071 ones-complement sum of buf, seeded with
+// initial. It is a drop-in replacement for gvisor's checksum.Checksum
+// that dispatches to a hand-written NEON routine. NEON is mandatory in
+// armv8 so no feature check is needed.
+func Checksum(buf []byte, initial uint16) uint16 {
+	return checksumNEON(buf, initial)
+}

overlay/checksum/checksum_arm64.s

@@ -0,0 +1,143 @@
+#include "textflag.h"
+
+// func checksumNEON(buf []byte, initial uint16) uint16
+//
+// Mirrors the algorithm in checksum_amd64.s: sum the buffer treating it as
+// a stream of uint32s in machine (little-endian) byte order, accumulating
+// into 64-bit lanes that have ample carry headroom; fold and byte-swap once
+// at the very end to recover the on-wire (big-endian) result.
+//
+// Each loop iteration loads 64 bytes via VLD1.P into V0..V3 (4 Q regs).
+// VUADDW takes the low two uint32 lanes of a Q reg, zero-extends them to
+// uint64, and adds them into a 2×uint64 accumulator; VUADDW2 does the same
+// for the high two lanes. Four ymm-equivalent accumulators (V8..V11) get
+// updated twice per iter to break the dep chain. Tail bytes go through a
+// scalar ADCS chain seeded with the byte-swapped initial.
+TEXT ·checksumNEON(SB), NOSPLIT, $0-34
+	MOVD  buf_base+0(FP), R0
+	MOVD  buf_len+8(FP), R1
+	MOVHU initial+24(FP), R2
+
+	// Pre-byteswap initial into the LE-summing space so it merges directly
+	// with the rest of the accumulator.
+	REV16W R2, R2
+
+	MOVD ZR, R3 // scalar accumulator
+
+	CMP $32, R1
+	BLT scalar_tail
+
+	VEOR V8.B16, V8.B16, V8.B16
+	VEOR V9.B16, V9.B16, V9.B16
+	VEOR V10.B16, V10.B16, V10.B16
+	VEOR V11.B16, V11.B16, V11.B16
+
+	CMP $64, R1
+	BLT loop16_init
+
+loop64:
+	VLD1.P  64(R0), [V0.B16, V1.B16, V2.B16, V3.B16]
+	VUADDW  V0.S2, V8.D2, V8.D2
+	VUADDW2 V0.S4, V9.D2, V9.D2
+	VUADDW  V1.S2, V10.D2, V10.D2
+	VUADDW2 V1.S4, V11.D2, V11.D2
+	VUADDW  V2.S2, V8.D2, V8.D2
+	VUADDW2 V2.S4, V9.D2, V9.D2
+	VUADDW  V3.S2, V10.D2, V10.D2
+	VUADDW2 V3.S4, V11.D2, V11.D2
+	SUB     $64, R1, R1
+	CMP     $64, R1
+	BGE     loop64
+
+loop16_init:
+	CMP $16, R1
+	BLT reduce_vec
+
+loop16:
+	VLD1.P  16(R0), [V0.B16]
+	VUADDW  V0.S2, V8.D2, V8.D2
+	VUADDW2 V0.S4, V9.D2, V9.D2
+	SUB     $16, R1, R1
+	CMP     $16, R1
+	BGE     loop16
+
+reduce_vec:
+	// Combine the four accumulators into V8.
+	VADD V9.D2, V8.D2, V8.D2
+	VADD V11.D2, V10.D2, V10.D2
+	VADD V10.D2, V8.D2, V8.D2
+
+	// Horizontal-add the two lanes of V8.D2 into a single uint64.
+	VADDP V8.D2, V8.D2, V8.D2
+	VMOV  V8.D[0], R8
+
+	ADDS R8, R3, R3
+	ADC  ZR, R3, R3
+
+scalar_tail:
+	CMP $8, R1
+	BLT tail4
+
+loop8:
+	MOVD.P 8(R0), R8
+	ADDS   R8, R3, R3
+	ADC    ZR, R3, R3
+	SUB    $8, R1, R1
+	CMP    $8, R1
+	BGE    loop8
+
+tail4:
+	CMP    $4, R1
+	BLT    tail2
+	MOVWU.P 4(R0), R8
+	ADDS   R8, R3, R3
+	ADC    ZR, R3, R3
+	SUB    $4, R1, R1
+
+tail2:
+	CMP    $2, R1
+	BLT    tail1
+	MOVHU.P 2(R0), R8
+	ADDS   R8, R3, R3
+	ADC    ZR, R3, R3
+	SUB    $2, R1, R1
+
+tail1:
+	CBZ R1, fold
+	MOVBU (R0), R8
+	ADDS  R8, R3, R3
+	ADC   ZR, R3, R3
+
+fold:
+	// Merge the byte-swapped initial into our LE-form accumulator.
+	ADDS R2, R3, R3
+	ADC  ZR, R3, R3
+
+	// 64 → 33 bits.
+	LSR $32, R3, R8
+	AND $0xffffffff, R3, R3
+	ADD R8, R3, R3
+
+	// 33 → 32 (truncate after adding bit 32 back).
+	LSR $32, R3, R8
+	ADD R8, R3, R3
+	AND $0xffffffff, R3, R3
+
+	// 32 → 17.
+	LSR $16, R3, R8
+	AND $0xffff, R3, R3
+	ADD R8, R3, R3
+
+	// 17 → 16 (truncation absorbs bit 16 below).
+	LSR $16, R3, R8
+	ADD R8, R3, R3
+
+	// AX low 16 bits hold the 16-bit sum in machine (LE) byte order; flip
+	// to big-endian to match the gvisor API contract. REV16W swaps bytes
+	// within each 16-bit halfword of the low 32 bits, so it acts as a
+	// 16-bit byte-swap on the live low 16.
+	REV16W R3, R3
+	AND    $0xffff, R3, R3
+
+	MOVH R3, ret+32(FP)
+	RET

overlay/checksum/checksum_fallback.go

@@ -0,0 +1,10 @@
+//go:build !amd64 && !arm64
+
+package checksum
+
+import gvisorchecksum "gvisor.dev/gvisor/pkg/tcpip/checksum"
+
+// Checksum delegates to gvisor on architectures without a hand-written body.
+func Checksum(buf []byte, initial uint16) uint16 {
+	return gvisorchecksum.Checksum(buf, initial)
+}

overlay/checksum/checksum_test.go

@@ -0,0 +1,190 @@
+package checksum
+
+import (
+	"fmt"
+	"math/rand/v2"
+	"testing"
+
+	gvisorchecksum "gvisor.dev/gvisor/pkg/tcpip/checksum"
+)
+
+// TestChecksumMatchesGvisor walks lengths from 0 to 4096, with several initial
+// seeds and a handful of starting alignments, asserting that our local
+// Checksum matches gvisor's reference bit-for-bit.
+func TestChecksumMatchesGvisor(t *testing.T) {
+	rng := rand.New(rand.NewPCG(1, 2))
+	const padFront = 16
+
+	// Random pool large enough for the longest case + alignment slop.
+	pool := make([]byte, 4096+padFront)
+	for i := range pool {
+		pool[i] = byte(rng.Uint32())
+	}
+
+	seeds := []uint16{0, 0x0001, 0xabcd, 0xffff, 0x1234, 0xfedc}
+	offsets := []int{0, 1, 2, 3, 4, 5, 7, 8, 15, 16}
+
+	for length := 0; length <= 4096; length++ {
+		for _, seed := range seeds {
+			for _, off := range offsets {
+				if off+length > len(pool) {
+					continue
+				}
+				buf := pool[off : off+length]
+				want := gvisorchecksum.Checksum(buf, seed)
+				got := Checksum(buf, seed)
+				if got != want {
+					t.Fatalf("len=%d off=%d seed=%#x: got %#04x want %#04x",
+						length, off, seed, got, want)
+				}
+			}
+		}
+	}
+}
+
+// TestChecksumPatternedBuffers exercises specific byte patterns that have
+// historically tripped up checksum implementations: all-zero, all-0xff,
+// alternating, and ascending sequences.
+func TestChecksumPatternedBuffers(t *testing.T) {
+	for length := 0; length <= 256; length++ {
+		patterns := map[string][]byte{
+			"zeros":       make([]byte, length),
+			"ones":        bytes(length, 0xff),
+			"alternating": pattern(length, []byte{0xa5, 0x5a}),
+			"ascending":   ascending(length),
+		}
+		for name, buf := range patterns {
+			for _, seed := range []uint16{0, 0xffff, 0x8000} {
+				want := gvisorchecksum.Checksum(buf, seed)
+				got := Checksum(buf, seed)
+				if got != want {
+					t.Fatalf("%s len=%d seed=%#x: got %#04x want %#04x",
+						name, length, seed, got, want)
+				}
+			}
+		}
+	}
+}
+
+func bytes(n int, v byte) []byte {
+	b := make([]byte, n)
+	for i := range b {
+		b[i] = v
+	}
+	return b
+}
+
+func pattern(n int, p []byte) []byte {
+	b := make([]byte, n)
+	for i := range b {
+		b[i] = p[i%len(p)]
+	}
+	return b
+}
+
+func ascending(n int) []byte {
+	b := make([]byte, n)
+	for i := range b {
+		b[i] = byte(i)
+	}
+	return b
+}
+
+// TestChecksumTailPaths targets every combination of (SIMD body iterations,
+// trailing tail bytes) the asm handlers walk through. The tail handlers
+// peel off 8 → 4 → 2 → 1 byte chunks in turn; this test exercises each by
+// constructing lengths of the form 64*k + tail for tail ∈ [0, 63] and a
+// representative spread of k values, including k=0 (no main loop, all tail)
+// and k=1 (one main loop iter, then tail). It's explicit coverage for
+// payload sizes that are odd, not divisible by 4, by 8, or by 32.
+func TestChecksumTailPaths(t *testing.T) {
+	rng := rand.New(rand.NewPCG(42, 17))
+	const padFront = 16
+	const maxK = 8
+
+	pool := make([]byte, 64*maxK+padFront+64)
+	for i := range pool {
+		pool[i] = byte(rng.Uint32())
+	}
+
+	seeds := []uint16{0, 0xffff, 0xabcd}
+	offsets := []int{0, 1, 3, 7, 15} // mix of aligned and odd starts
+
+	for k := 0; k <= maxK; k++ {
+		for tail := 0; tail < 64; tail++ {
+			length := 64*k + tail
+			for _, seed := range seeds {
+				for _, off := range offsets {
+					if off+length > len(pool) {
+						continue
+					}
+					buf := pool[off : off+length]
+					want := gvisorchecksum.Checksum(buf, seed)
+					got := Checksum(buf, seed)
+					if got != want {
+						t.Fatalf("k=%d tail=%d (len=%d) off=%d seed=%#x: got %#04x want %#04x",
+							k, tail, length, off, seed, got, want)
+					}
+				}
+			}
+		}
+	}
+}
+
+// BenchmarkChecksumTailSizes covers payload sizes that aren't clean multiples
+// of the SIMD body's 32-byte (amd64) or 16-byte (arm64) chunks, so the tail
+// handler is meaningfully on the hot path. Sizes are picked to either exercise
+// every tail branch (tiny lengths) or sit slightly off realistic packet
+// boundaries (e.g. 1499 = MTU − 1).
+func BenchmarkChecksumTailSizes(b *testing.B) {
+	sizes := []int{
+		1, 3, 7, 15, 31, // sub-SIMD; entire work is scalar tail
+		33, 35, 47, 63, // one loop32 + assorted tails
+		65, 95, 127, // one loop64 + assorted tails
+		1447, 1471, 1499, 1501, // around MTU
+		8191, 8193, // around USO
+		65531, 65533, // near the kernel max
+	}
+	for _, size := range sizes {
+		buf := make([]byte, size)
+		for i := range buf {
+			buf[i] = byte(i)
+		}
+		b.Run(fmt.Sprintf("size=%d/local", size), func(b *testing.B) {
+			b.SetBytes(int64(size))
+			for i := 0; i < b.N; i++ {
+				_ = Checksum(buf, 0)
+			}
+		})
+		b.Run(fmt.Sprintf("size=%d/gvisor", size), func(b *testing.B) {
+			b.SetBytes(int64(size))
+			for i := 0; i < b.N; i++ {
+				_ = gvisorchecksum.Checksum(buf, 0)
+			}
+		})
+	}
+}
+
+// BenchmarkChecksum compares the local Checksum to gvisor's at sizes that
+// match real traffic: a TCP/IP header (60), a typical MSS (1448), a typical
+// USO size (8192), and the kernel's max GSO superpacket (65535).
+func BenchmarkChecksum(b *testing.B) {
+	for _, size := range []int{60, 1448, 8192, 65535} {
+		buf := make([]byte, size)
+		for i := range buf {
+			buf[i] = byte(i)
+		}
+		b.Run(fmt.Sprintf("size=%d/local", size), func(b *testing.B) {
+			b.SetBytes(int64(size))
+			for i := 0; i < b.N; i++ {
+				_ = Checksum(buf, 0)
+			}
+		})
+		b.Run(fmt.Sprintf("size=%d/gvisor", size), func(b *testing.B) {
+			b.SetBytes(int64(size))
+			for i := 0; i < b.N; i++ {
+				_ = gvisorchecksum.Checksum(buf, 0)
+			}
+		})
+	}
+}

overlay/device.go

@@ -4,15 +4,21 @@ import (
 	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"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
+
 type Device interface {
-	io.ReadWriteCloser
+	io.Closer
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
 	SupportsMultiqueue() bool
-	NewMultiQueueReader() (io.ReadWriteCloser, error)
+	NewMultiQueueReader() error
+	Readers() []tio.Queue
 }

overlay/overlaytest/noop.go

@@ -4,10 +4,11 @@ package overlaytest
 
 import (
 	"errors"
-	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/wire"
 )
 
 // NoopTun is an overlay.Device that silently discards every read and write.
@@ -15,6 +16,10 @@ import (
 // exercise the datapath.
 type NoopTun struct{}
 
+func (NoopTun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
+}
+
 func (NoopTun) RoutesFor(addr netip.Addr) routing.Gateways {
 	return routing.Gateways{}
 }
@@ -31,7 +36,7 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read([]byte) (int, error) {
+func (NoopTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
 	return 0, nil
 }
 
@@ -43,8 +48,12 @@ func (NoopTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, errors.New("unsupported")
+func (NoopTun) NewMultiQueueReader() error {
+	return errors.New("unsupported")
+}
+
+func (NoopTun) Readers() []tio.Queue {
+	return []tio.Queue{NoopTun{}}
 }
 
 func (NoopTun) Close() error {

overlay/tio/queueset_gso_linux.go

@@ -0,0 +1,90 @@
+package tio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+type offloadQueueSet struct {
+	pq []*Offload
+	// pqi is exactly the same as pq, but stored as the interface type
+	pqi        []Queue
+	shutdownFd int
+	// usoEnabled is true when newTun successfully negotiated TUN_F_USO4|6
+	// with the kernel. Queues created by Add inherit this and surface it
+	// via Offload.USOSupported so coalescers can gate USO emission.
+	usoEnabled bool
+}
+
+// NewOffloadQueueSet creates a QueueSet that uses virtio_net_hdr to do
+// TSO segmentation in userspace. usoEnabled tells downstream queues whether
+// the kernel agreed to deliver/accept GSO_UDP_L4 superpackets — coalescers
+// should fall back to per-packet writes when this is false.
+func NewOffloadQueueSet(usoEnabled bool) (QueueSet, error) {
+	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 := &offloadQueueSet{
+		pq:         []*Offload{},
+		pqi:        []Queue{},
+		shutdownFd: shutdownFd,
+		usoEnabled: usoEnabled,
+	}
+
+	return out, nil
+}
+
+func (c *offloadQueueSet) Queues() []Queue {
+	return c.pqi
+}
+
+func (c *offloadQueueSet) Add(fd int) error {
+	x, err := newOffload(fd, c.shutdownFd, c.usoEnabled)
+	if err != nil {
+		return err
+	}
+	c.pq = append(c.pq, x)
+	c.pqi = append(c.pqi, x)
+
+	return nil
+}
+
+func (c *offloadQueueSet) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(c.shutdownFd, buf[:])
+	return err
+}
+
+func (c *offloadQueueSet) Close() error {
+	if c.shutdownFd < 0 {
+		return nil
+	}
+
+	errs := []error{}
+
+	// Signal all readers blocked in poll to wake up and exit
+	if err := c.wakeForShutdown(); err != nil {
+		errs = append(errs, err)
+	}
+
+	for _, x := range c.pq {
+		if err := x.Close(); err != nil {
+			errs = append(errs, err)
+		}
+	}
+
+	// All Offloads reference shutdownFd in their pollfd arrays, so close it
+	// only after every Offload.Close has returned.
+	if err := unix.Close(c.shutdownFd); err != nil {
+		errs = append(errs, err)
+	}
+	c.shutdownFd = -1
+
+	return errors.Join(errs...)
+}

overlay/tio/queueset_poll_linux.go

@@ -0,0 +1,80 @@
+package tio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+)
+
+type pollQueueSet struct {
+	pq []*Poll
+	// pqi is exactly the same as pq, but stored as the interface type
+	pqi        []Queue
+	shutdownFd int
+}
+
+func NewPollQueueSet() (QueueSet, error) {
+	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 := &pollQueueSet{
+		pq:         []*Poll{},
+		pqi:        []Queue{},
+		shutdownFd: shutdownFd,
+	}
+
+	return out, nil
+}
+
+func (c *pollQueueSet) Queues() []Queue {
+	return c.pqi
+}
+
+func (c *pollQueueSet) Add(fd int) error {
+	x, err := newPoll(fd, c.shutdownFd)
+	if err != nil {
+		return err
+	}
+	c.pq = append(c.pq, x)
+	c.pqi = append(c.pqi, x)
+
+	return nil
+}
+
+func (c *pollQueueSet) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(c.shutdownFd, buf[:])
+	return err
+}
+
+func (c *pollQueueSet) Close() error {
+	if c.shutdownFd < 0 {
+		return nil
+	}
+
+	errs := []error{}
+
+	if err := c.wakeForShutdown(); err != nil {
+		errs = append(errs, err)
+	}
+
+	for _, x := range c.pq {
+		if err := x.Close(); err != nil {
+			errs = append(errs, err)
+		}
+	}
+
+	// All Polls reference shutdownFd in their pollfd arrays, so close it
+	// only after every Poll.Close has returned.
+	if err := unix.Close(c.shutdownFd); err != nil {
+		errs = append(errs, err)
+	}
+	c.shutdownFd = -1
+
+	return errors.Join(errs...)
+}

overlay/tio/segment_bench_test.go

@@ -0,0 +1,69 @@
+//go:build linux && !android && !e2e_testing
+
+package tio
+
+import (
+	"testing"
+
+	"github.com/slackhq/nebula/wire"
+)
+
+// fakeBatch stands in for batch.TxBatcher inside the bench — same shape
+// of pointer-capturing closure that sendInsideMessage builds.
+type fakeBatch struct{ buf [65536]byte }
+
+func (b *fakeBatch) Reserve(sz int) []byte { return b.buf[:sz] }
+func (b *fakeBatch) Commit([]byte)         {}
+
+type fakeHostInfo struct {
+	remoteIndexId uint32
+	counter       uint64
+}
+type fakeIface struct {
+	rebindCount uint8
+	hi          *fakeHostInfo
+}
+
+// BenchmarkSegmentSuperpacketAllocsTSO measures allocation per
+// PerSegment call when a closure captures pointer-bearing receivers — the
+// realistic shape of sendInsideMessage's closure.
+func BenchmarkSegmentSuperpacketAllocsTSO(b *testing.B) {
+	const mss = 1400
+	const numSeg = 32
+	pkt := buildTSOv6(mss*numSeg, mss)
+	gso := wire.GSOInfo{
+		Size:      mss,
+		HdrLen:    60, // 40 (IPv6) + 20 (TCP)
+		CsumStart: 40,
+		Proto:     wire.GSOProtoTCP,
+	}
+	p := wire.TunPacket{Bytes: pkt, Meta: gso}
+
+	hi := &fakeHostInfo{remoteIndexId: 0xdeadbeef}
+	f := &fakeIface{rebindCount: 7, hi: hi}
+	fb := &fakeBatch{}
+
+	// PerSegment consumes pkt destructively; refresh from a master copy
+	// each iter (matches the production pattern where every TUN read hands
+	// the segmenter a fresh kernel-supplied buffer).
+	master := append([]byte(nil), pkt...)
+	work := make([]byte, len(pkt))
+	p.Bytes = work
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		copy(work, master)
+		err := p.PerSegment(func(seg []byte) error {
+			out := fb.Reserve(16 + len(seg) + 16)
+			out[0] = byte(f.rebindCount)
+			out[1] = byte(hi.counter)
+			hi.counter++
+			fb.Commit(out)
+			return nil
+		})
+		if err != nil {
+			b.Fatalf("PerSegment: %v", err)
+		}
+	}
+}

overlay/tio/tio.go

@@ -0,0 +1,90 @@
+package tio
+
+import (
+	"io"
+
+	"github.com/slackhq/nebula/wire"
+)
+
+// QueueSet holds one or many Queue objects and helps close them in an orderly way.
+type QueueSet interface {
+	io.Closer
+	Queues() []Queue
+
+	// Add takes a tun fd, adds it to the set, and prepares it for use as a Queue.
+	Add(fd int) error
+}
+
+// Capabilities advertises which kernel offload features a Queue successfully negotiated.
+// Callers consult this to decide which coalescers to wire onto the write path.
+type Capabilities struct {
+	// TSO means the FD was opened with IFF_VNET_HDR and the kernel agreed
+	// to TUN_F_TSO4|TSO6 — i.e. WriteGSO with GSOProtoTCP is safe.
+	TSO bool
+	// USO means the kernel additionally agreed to TUN_F_USO4|USO6, so
+	// WriteGSO with GSOProtoUDP is safe. Linux ≥ 6.2.
+	USO bool
+}
+
+// Queue is a readable/writable Poll queue. One Queue is driven by a single
+// read goroutine plus a single writer (see Write below).
+type Queue interface {
+	io.Closer
+
+	// Read will read at least 1 packet from the tun (up to len(p)).
+	// mem will be used to provide the backing for each of p[n].Bytes.
+	// Callers should size mem and p to avoid exhausting mem before p.
+	// Returns the number of packets actually read, or error.
+	Read(p []wire.TunPacket, mem []byte) (int, error)
+
+	// Write emits a single packet on the plaintext (outside→inside)
+	// delivery path.
+	Write(p []byte) (int, error)
+
+	// Capabilities returns the Queue's negotiated offload capabilities,
+	// or the zero value when q does not advertise any.
+	Capabilities() Capabilities
+}
+
+// GSOWriter is implemented by Queues that can emit a TCP or UDP 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 do not implement this interface and coalescing is skipped.
+//
+// hdr contains the IPv4/IPv6 header prefix (mutable - callers will have
+// filled in total length and IP csum). transportHdr is the TCP or UDP
+// header (mutable - the L4 checksum field must hold the pseudo-header
+// partial, single-fold not inverted, per virtio NEEDS_CSUM semantics).
+// 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. Every segment
+// in pays except possibly the last is exactly the same size. proto picks
+// the L4 protocol so the writer knows which GSOType / CsumOffset to set.
+//
+// Callers should also consult Queue.Capabilities (via SupportsGSO) for
+// the per-protocol negotiated capability; an implementation of GSOWriter
+// is necessary but not sufficient since USO may not have been negotiated
+// even when TSO was.
+type GSOWriter interface {
+	WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte, proto wire.GSOProto) error
+}
+
+// SupportsGSO reports whether w implements GSOWriter and the underlying
+// queue advertises the negotiated capability for `want` via Capabilities.
+func SupportsGSO(w Queue, want wire.GSOProto) (GSOWriter, bool) {
+	gw, ok := w.(GSOWriter)
+	if !ok {
+		return nil, false
+	}
+	caps := w.Capabilities()
+	switch want {
+	case wire.GSOProtoTCP:
+		return gw, caps.TSO
+	case wire.GSOProtoUDP:
+		return gw, caps.USO
+	default:
+		return gw, false
+	}
+}

overlay/tio/tio_gso_linux.go

@@ -0,0 +1,447 @@
+package tio
+
+import (
+	"fmt"
+	"io"
+	"log/slog"
+	"os"
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+
+	"github.com/slackhq/nebula/wire"
+	"golang.org/x/sys/unix"
+
+	"github.com/slackhq/nebula/overlay/tio/virtio"
+)
+
+// tunRxBufSize is the per-Read worst-case footprint for one kernel-supplied
+// packet body, which is at most ~64 KiB (tunReadBufSize). Segmentation
+// happens at encrypt time via wire.TunPacket.PerSegment on a per-routine
+// MTU-sized scratch, so the caller-supplied read buffer only holds raw
+// kernel-supplied bytes. Used by Read's drain loop to gate further reads
+// on whether the remaining buffer can still hold one worst-case packet.
+const tunRxBufSize = 64 * 1024
+
+// gsoMaxIovs caps the iovec budget WriteGSO assembles per call: 3 fixed
+// entries (virtio_net_hdr, IP hdr, transport hdr) plus up to gsoMaxIovs-3
+// payload fragments. Sized comfortably above the typical kernel GSO
+// segment cap (Linux UDP_GRO is 64) so realistic coalesced bursts never
+// touch the limit. iovecs are tiny (16 bytes), so the entire scratch is
+// 4 KiB — fine to keep resident on every queue. WriteGSO returns an error
+// rather than reallocating when a caller exceeds this budget.
+const gsoMaxIovs = 256
+
+// 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 paths already carry
+// a valid L4 checksum (either supplied by a remote peer whose ciphertext we
+// AEAD-authenticated, produced by virtio.SegmentTCP/SegmentUDP during
+// superpacket segmentation, or built locally by CreateRejectPacket), so
+// trusting them is safe.
+var validVnetHdr = [virtio.Size]byte{unix.VIRTIO_NET_HDR_F_DATA_VALID}
+
+// Offload 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 Offload struct {
+	fd         int
+	shutdownFd int
+	readPoll   [2]unix.PollFd
+	writePoll  [2]unix.PollFd
+	// writeLock serializes blockOnWrite's read+clear of writePoll[*].Revents.
+	// Any goroutine that calls Write may end up parked in poll(2); without
+	// the lock concurrent waiters could race the Revents reset and lose
+	// events.
+	writeLock sync.Mutex
+	closed    atomic.Bool
+
+	// readVnetScratch holds the 10-byte virtio_net_hdr split off the front of
+	// every TUN read via readv(2). Decoupling the header from the packet body
+	// lets us read the body directly into the caller-supplied mem at the
+	// current rxOff with no userspace copy on the GSO_NONE fast path.
+	readVnetScratch [virtio.Size]byte
+	// readIovs is the readv(2) iovec scratch wired once at construction —
+	// iovec[0] points at readVnetScratch; iovec[1].Base/Len is updated per
+	// read to address the caller-supplied mem slot.
+	readIovs [2]unix.Iovec
+
+	// usoEnabled records whether the kernel agreed to TUN_F_USO* on this FD,
+	// so writers can decide whether emitting GSO_UDP_L4 superpackets is safe.
+	usoEnabled bool
+
+	// gsoHdrBuf is a per-queue 10-byte scratch for the virtio_net_hdr emitted
+	// by WriteGSO. Kept separate from the read-only package-level validVnetHdr
+	// so non-GSO Writes can ship that constant directly while WriteGSO
+	// rewrites this scratch on every call.
+	gsoHdrBuf [virtio.Size]byte
+	// gsoIovs is the writev iovec scratch for WriteGSO. Pre-sized to
+	// gsoMaxIovs at construction; never grown. WriteGSO returns an error
+	// (and drops the call) if a caller hands it more fragments than fit.
+	gsoIovs []unix.Iovec
+}
+
+func newOffload(fd int, shutdownFd int, usoEnabled bool) (*Offload, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+
+	out := &Offload{
+		fd:         fd,
+		shutdownFd: shutdownFd,
+		usoEnabled: usoEnabled,
+		closed:     atomic.Bool{},
+		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},
+		},
+		writeLock: sync.Mutex{},
+		gsoIovs:   make([]unix.Iovec, 2, gsoMaxIovs),
+	}
+
+	out.gsoIovs[0].Base = &out.gsoHdrBuf[0]
+	out.gsoIovs[0].SetLen(virtio.Size)
+
+	// readIovs[0] is wired once to the virtio_net_hdr scratch; per-read we
+	// only repoint readIovs[1] at the next caller-supplied mem slot
+	// (see readPacket).
+	out.readIovs[0].Base = &out.readVnetScratch[0]
+	out.readIovs[0].SetLen(virtio.Size)
+
+	return out, nil
+}
+
+func (r *Offload) 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 *Offload) 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!
+	r.writeLock.Lock()
+	tunEvents := r.writePoll[0].Revents
+	shutdownEvents := r.writePoll[1].Revents
+	r.writePoll[0].Revents = 0
+	r.writePoll[1].Revents = 0
+	r.writeLock.Unlock()
+	//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
+}
+
+// readPacket issues a single readv(2) splitting the virtio_net_hdr off
+// into readVnetScratch and reading the packet body directly into mem.
+// Returns the body length (zero virtio header bytes, just the IP
+// packet/superpacket). block controls whether EAGAIN is retried via poll:
+// the initial read of a drain blocks; subsequent drain reads do not.
+//
+// The body iovec capacity is always tunReadBufSize; the Read drain loop
+// gates entry on len(mem)-rxOff >= tunRxBufSize, sized to hold one
+// worst-case kernel-supplied packet body. Without that gate the body
+// iovec could be smaller than the next inbound packet and the kernel
+// would truncate.
+func (r *Offload) readPacket(mem []byte, block bool) (int, error) {
+	for {
+		r.readIovs[1].Base = &mem[0]
+		r.readIovs[1].SetLen(tunReadBufSize)
+		n, _, errno := syscall.Syscall(unix.SYS_READV, uintptr(r.fd), uintptr(unsafe.Pointer(&r.readIovs[0])), uintptr(len(r.readIovs)))
+		if errno == 0 {
+			if int(n) < virtio.Size {
+				return 0, io.ErrShortWrite
+			}
+			return int(n) - virtio.Size, nil
+		}
+		if errno == unix.EAGAIN {
+			if !block {
+				return 0, errno
+			}
+			if err := r.blockOnRead(); err != nil {
+				return 0, err
+			}
+			continue
+		}
+		if errno == unix.EINTR {
+			continue
+		}
+		if errno == unix.EBADF {
+			return 0, os.ErrClosed
+		}
+		return 0, errno
+	}
+}
+
+// Read returns one or more packets from the tun. Each wire.TunPacket
+// either carries a single ready-to-use IP datagram (GSO zero) or a TSO/USO
+// superpacket plus the wire.GSOInfo a caller needs to segment it (see
+// wire.TunPacket.PerSegment). 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), p is full, or mem no longer has room
+// for another worst-case packet (tunRxBufSize). This amortizes the poll
+// wake over bursts of small packets (e.g. TCP ACKs). The Bytes slices on
+// returned packets point into the caller-supplied mem and are only valid
+// until the next Read or Close on this Queue.
+func (r *Offload) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	maxP := len(p)
+	maxM := len(mem)
+	p = p[:0]
+	rxOff := 0
+
+	// Initial (blocking) read. Retry on decode errors so a single bad
+	// packet does not stall the reader.
+	for {
+		n, err := r.readPacket(mem, true)
+		if err != nil {
+			return 0, err
+		}
+		if p, err = r.decodeRead(p, mem, n); err != nil {
+			// Drop and read again — a bad packet should not kill the reader.
+			continue
+		}
+
+		rxOff += n
+		break
+	}
+
+	// Drain: non-blocking reads until the kernel queue is empty, p is full,
+	// or mem no longer has room for another worst-case kernel-supplied
+	// packet (tunRxBufSize).
+	for len(p) < maxP && maxM-rxOff >= tunRxBufSize {
+		n, err := r.readPacket(mem[rxOff:], false)
+		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 p, err = r.decodeRead(p, mem[rxOff:], n); err != nil {
+			// Drop this packet and stop the drain; we'd rather hand off
+			// what we have than keep spinning here.
+			break
+		}
+		rxOff += n
+	}
+
+	return len(p), nil
+}
+
+// decodeRead processes the packet sitting at mem[:pktLen]. The bytes stay
+// in mem — for GSO_NONE we slice them as a regular IP datagram (running
+// finishChecksum if NEEDS_CSUM is set); for TSO/USO superpackets we attach
+// the corrected GSO metadata so the caller can segment lazily at encrypt
+// time. The caller advances its own rxOff past the kernel-supplied body
+// and nothing else, since segmentation no longer writes back into mem.
+func (r *Offload) decodeRead(p []wire.TunPacket, mem []byte, pktLen int) ([]wire.TunPacket, error) {
+	if pktLen <= 0 {
+		return p, fmt.Errorf("short tun read: %d", pktLen)
+	}
+	var hdr virtio.Hdr
+	hdr.Decode(r.readVnetScratch[:])
+
+	body := mem[:pktLen]
+
+	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_NONE {
+		if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
+			if err := virtio.FinishChecksum(body, hdr); err != nil {
+				return p, err
+			}
+		}
+		p = append(p, wire.TunPacket{Bytes: body})
+		return p, nil
+	}
+
+	// GSO superpacket: validate, fix the kernel-supplied HdrLen on the
+	// FORWARD path (CorrectHdrLen), pick the L4 protocol, and attach
+	// the metadata. The bytes stay in mem untouched; segmentation
+	// happens in wire.TunPacket.PerSegment at encrypt time.
+	if err := virtio.CheckValid(body, hdr); err != nil {
+		return p, err
+	}
+	if err := virtio.CorrectHdrLen(body, &hdr); err != nil {
+		return p, err
+	}
+	proto, err := protoFromGSOType(hdr.GSOType)
+	if err != nil {
+		return p, err
+	}
+	p = append(p, wire.TunPacket{
+		Bytes: body,
+		Meta: wire.GSOInfo{
+			Size:      hdr.GSOSize,
+			HdrLen:    hdr.HdrLen,
+			CsumStart: hdr.CsumStart,
+			Proto:     proto,
+		},
+	})
+	return p, nil
+}
+
+func (r *Offload) Write(buf []byte) (int, error) {
+	iovs := [2]unix.Iovec{
+		{Base: &validVnetHdr[0]},
+		{Base: &buf[0]},
+	}
+	iovs[0].SetLen(virtio.Size)
+	iovs[1].SetLen(len(buf))
+	return r.writeWithScratch(buf, &iovs)
+}
+
+func (r *Offload) writeWithScratch(buf []byte, iovs *[2]unix.Iovec) (int, error) {
+	if len(buf) == 0 {
+		return 0, nil
+	}
+	iovs[1].Base = &buf[0]
+	iovs[1].SetLen(len(buf))
+	return r.rawWrite(unsafe.Slice(&iovs[0], len(iovs)))
+}
+
+func (r *Offload) rawWrite(iovs []unix.Iovec) (int, error) {
+	for {
+		n, _, errno := syscall.Syscall(unix.SYS_WRITEV, uintptr(r.fd), uintptr(unsafe.Pointer(&iovs[0])), uintptr(len(iovs)))
+		if errno == 0 {
+			if int(n) < virtio.Size {
+				return 0, io.ErrShortWrite
+			}
+			return int(n) - virtio.Size, nil
+		}
+		if errno == unix.EAGAIN {
+			if err := r.blockOnWrite(); err != nil {
+				return 0, err
+			}
+			continue
+		}
+		if errno == unix.EINTR {
+			continue
+		}
+		if errno == unix.EBADF {
+			return 0, os.ErrClosed
+		}
+		return 0, errno
+	}
+}
+
+// Capabilities reports the offload features negotiated for this Queue. TSO
+// is always true for Offload (we only construct it on IFF_VNET_HDR FDs);
+// USO is true only when the kernel agreed to TUN_F_USO4|6 at open time
+// (Linux ≥ 6.2).
+func (r *Offload) Capabilities() Capabilities {
+	return Capabilities{TSO: true, USO: r.usoEnabled}
+}
+
+func (r *Offload) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte, proto wire.GSOProto) error {
+	if len(hdr) == 0 || len(pays) == 0 || len(transportHdr) == 0 {
+		return nil
+	}
+	// L4 checksum offset inside transportHdr: TCP=16 (the `check` field after
+	// seq/ack/dataoff/flags/window), UDP=6 (after sport/dport/length).
+	var csumOff uint16
+	switch proto {
+	case wire.GSOProtoUDP:
+		csumOff = 6
+	default:
+		csumOff = 16
+	}
+	vhdr := virtio.Hdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		HdrLen:     uint16(len(hdr) + len(transportHdr)),
+		GSOSize:    uint16(len(pays[0])),
+		CsumStart:  uint16(len(hdr)),
+		CsumOffset: csumOff,
+	}
+	if len(pays) > 1 {
+		ipVer := hdr[0] >> 4
+		switch {
+		case proto == wire.GSOProtoUDP && (ipVer == 4 || ipVer == 6):
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_UDP_L4
+		case ipVer == 6:
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV6
+		case ipVer == 4:
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV4
+		default:
+			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+			vhdr.GSOSize = 0
+		}
+	} else {
+		vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
+		vhdr.GSOSize = 0
+	}
+	vhdr.Encode(r.gsoHdrBuf[:])
+
+	// Build the iovec array: [virtio_hdr, hdr, transportHdr, pays...]. r.gsoIovs[0] is
+	// wired to gsoHdrBuf at construction and never changes.
+	need := 3 + len(pays)
+	if need > cap(r.gsoIovs) {
+		slog.Default().Warn("tio: WriteGSO iovec budget exceeded; dropping superpacket",
+			"need", need, "cap", cap(r.gsoIovs), "segments", len(pays))
+		return fmt.Errorf("tio: WriteGSO needs %d iovecs but cap is %d", need, cap(r.gsoIovs))
+	}
+	r.gsoIovs = r.gsoIovs[:need]
+	r.gsoIovs[1].Base = &hdr[0]
+	r.gsoIovs[1].SetLen(len(hdr))
+	r.gsoIovs[2].Base = &transportHdr[0]
+	r.gsoIovs[2].SetLen(len(transportHdr))
+	for i, p := range pays {
+		r.gsoIovs[3+i].Base = &p[0]
+		r.gsoIovs[3+i].SetLen(len(p))
+	}
+
+	_, err := r.rawWrite(r.gsoIovs)
+	return err
+}
+
+func (r *Offload) Close() error {
+	if r.closed.Swap(true) {
+		return nil
+	}
+
+	//shutdownFd is owned by the container, so we should not close it
+	var err error
+	if r.fd >= 0 {
+		err = unix.Close(r.fd)
+		r.fd = -1
+	}
+
+	return err
+}

overlay/tio/tio_poll_linux.go

@@ -0,0 +1,169 @@
+package tio
+
+import (
+	"fmt"
+	"os"
+	"sync/atomic"
+
+	"github.com/slackhq/nebula/wire"
+	"golang.org/x/sys/unix"
+)
+
+// 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
+
+type Poll struct {
+	fd int
+
+	readPoll  [2]unix.PollFd
+	writePoll [2]unix.PollFd
+	closed    atomic.Bool
+}
+
+func newPoll(fd int, shutdownFd int) (*Poll, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		_ = unix.Close(fd)
+		return nil, fmt.Errorf("failed to set Poll device as nonblocking: %w", err)
+	}
+
+	out := &Poll{
+		fd: fd,
+		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},
+		},
+	}
+	return out, nil
+}
+
+// blockOnRead waits until the Poll fd is readable or shutdown has been signaled.
+// Returns os.ErrClosed if Close was called.
+func (t *Poll) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(t.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := t.readPoll[0].Revents
+	shutdownEvents := t.readPoll[1].Revents
+	t.readPoll[0].Revents = 0
+	t.readPoll[1].Revents = 0
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (t *Poll) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(t.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := t.writePoll[0].Revents
+	shutdownEvents := t.writePoll[1].Revents
+	t.writePoll[0].Revents = 0
+	t.writePoll[1].Revents = 0
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (t *Poll) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) == 0 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = wire.GSOInfo{}
+	n, err := t.readOne(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[:n]
+	return 1, nil
+}
+
+func (t *Poll) readOne(to []byte) (int, error) {
+	for {
+		n, errno := unix.Read(t.fd, to)
+		if errno == nil {
+			return n, nil
+		}
+		switch errno {
+		case unix.EAGAIN:
+			if err := t.blockOnRead(); err != nil {
+				return 0, err
+			}
+		case unix.EINTR:
+			// retry
+		case unix.EBADF:
+			return 0, os.ErrClosed
+		default:
+			return 0, errno
+		}
+	}
+}
+
+// Write is only valid for single threaded use
+func (t *Poll) Write(from []byte) (int, error) {
+	for {
+		n, errno := unix.Write(t.fd, from)
+		if errno == nil {
+			return n, nil
+		}
+		switch errno {
+		case unix.EAGAIN:
+			if err := t.blockOnWrite(); err != nil {
+				return 0, err
+			}
+		case unix.EINTR:
+			// retry
+		case unix.EBADF:
+			return 0, os.ErrClosed
+		default:
+			return 0, errno
+		}
+	}
+}
+
+func (t *Poll) Close() error {
+	if t.closed.Swap(true) {
+		return nil
+	}
+	//shutdownFd is owned by the container, so we should not close it
+	var err error
+	if t.fd >= 0 {
+		err = unix.Close(t.fd)
+		t.fd = -1
+	}
+
+	return err
+}
+
+func (t *Poll) Capabilities() Capabilities {
+	return Capabilities{}
+}

overlay/tio/tun_file_linux_test.go

@@ -0,0 +1,106 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"errors"
+	"os"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula/wire"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/unix"
+)
+
+// newReadPipe returns a read fd. The matching write fd is registered for cleanup.
+// The caller takes ownership of the read fd (pass it into a QueueSet).
+func newReadPipe(t *testing.T) int {
+	t.Helper()
+	var fds [2]int
+	if err := unix.Pipe2(fds[:], unix.O_CLOEXEC); err != nil {
+		t.Fatalf("pipe2: %v", err)
+	}
+	t.Cleanup(func() { _ = unix.Close(fds[1]) })
+	return fds[0]
+}
+
+func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
+	parent, err := NewPollQueueSet()
+	require.NoError(t, err)
+	require.NoError(t, parent.Add(newReadPipe(t)))
+	require.NoError(t, parent.Add(newReadPipe(t)))
+	// QueueSet.Close owns the read fds we Added — don't register a separate
+	// Cleanup to close them or we'll double-close whatever fd the kernel
+	// has since reused.
+
+	readers := parent.Queues()
+	errs := make([]error, len(readers))
+	var wg sync.WaitGroup
+	for i, r := range readers {
+		wg.Add(1)
+		go func(i int, r Queue) {
+			defer wg.Done()
+			pkts := make([]wire.TunPacket, 1)
+			_, errs[i] = r.Read(pkts, make([]byte, 64))
+		}(i, r)
+	}
+
+	time.Sleep(50 * time.Millisecond)
+
+	if err := parent.Close(); err != nil {
+		t.Fatalf("Close: %v", err)
+	}
+
+	done := make(chan struct{})
+	go func() { wg.Wait(); close(done) }()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("readers did not wake")
+	}
+
+	for i, err := range errs {
+		if !errors.Is(err, os.ErrClosed) {
+			t.Errorf("reader %d: expected os.ErrClosed, got %v", i, err)
+		}
+	}
+}
+
+func TestPoll_Close_Idempotent(t *testing.T) {
+	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = unix.Close(shutdownFd) })
+
+	tf, err := newPoll(newReadPipe(t), shutdownFd)
+	require.NoError(t, err)
+	if err := tf.Close(); err != nil {
+		t.Fatalf("first Close: %v", err)
+	}
+	if err := tf.Close(); err != nil {
+		t.Fatalf("second Close should be a no-op, got %v", err)
+	}
+}
+
+func TestPollQueueSet_Close_ClosesEventfd(t *testing.T) {
+	qs, err := NewPollQueueSet()
+	require.NoError(t, err)
+	require.NoError(t, qs.Add(newReadPipe(t)))
+
+	fd := qs.(*pollQueueSet).shutdownFd
+	require.NoError(t, qs.Close())
+
+	// Closing the eventfd again should fail with EBADF, proving Close
+	// actually released it.
+	if err := unix.Close(fd); err == nil {
+		t.Fatalf("eventfd %d still open after QueueSet.Close", fd)
+	}
+
+	// Second Close must be a no-op (and must not double-close the eventfd
+	// in case the kernel handed it out to another caller in the meantime).
+	if err := qs.Close(); err != nil {
+		t.Fatalf("second Close: %v", err)
+	}
+}

overlay/tio/tun_linux_offload.go

@@ -0,0 +1,25 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"fmt"
+
+	"github.com/slackhq/nebula/wire"
+	"golang.org/x/sys/unix"
+)
+
+// protoFromGSOType maps a virtio_net_hdr GSOType to the GSOProto value the
+// segment-time helpers use. Returns an error for GSO_NONE or any unknown
+// value — the caller should only invoke this on a confirmed superpacket.
+func protoFromGSOType(t uint8) (wire.GSOProto, error) {
+	switch t {
+	case unix.VIRTIO_NET_HDR_GSO_TCPV4, unix.VIRTIO_NET_HDR_GSO_TCPV6:
+		return wire.GSOProtoTCP, nil
+	case unix.VIRTIO_NET_HDR_GSO_UDP_L4:
+		return wire.GSOProtoUDP, nil
+	default:
+		return 0, fmt.Errorf("unsupported virtio gso type: %d", t)
+	}
+}

overlay/tio/tun_linux_offload_test.go

@@ -0,0 +1,782 @@
+//go:build linux && !android && !e2e_testing
+// +build linux,!android,!e2e_testing
+
+package tio
+
+import (
+	"encoding/binary"
+	"os"
+	"testing"
+
+	"golang.org/x/sys/unix"
+	"gvisor.dev/gvisor/pkg/tcpip/checksum"
+
+	"github.com/slackhq/nebula/overlay/tio/virtio"
+	"github.com/slackhq/nebula/wire"
+)
+
+// testSegScratchSize is a generous segmentation scratch sized to fit any
+// of the synthetic TSO/USO superpackets these tests generate (one
+// worst-case 64 KiB superpacket plus replicated per-segment headers).
+const testSegScratchSize = 192 * 1024
+
+// verifyChecksum confirms that the one's-complement sum across `b`, seeded
+// with a folded pseudo-header sum, equals all-ones (valid).
+func verifyChecksum(b []byte, pseudo uint16) bool {
+	return checksum.Checksum(b, pseudo) == 0xffff
+}
+
+// segmentForTest is the test-only counterpart to the production
+// wire.TunPacket.PerSegment path. It handles GSO_NONE (with optional
+// finishChecksum) inline and dispatches GSO superpackets through
+// PerSegment, draining each yielded segment into a freshly-copied [][]byte
+// slot so callers can iterate after the call returns. Tests pre-set
+// hdr.HdrLen correctly, so correctHdrLen is not invoked here.
+func segmentForTest(pkt []byte, hdr virtio.Hdr, out *[][]byte, scratch []byte) error {
+	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_NONE {
+		cp := append([]byte(nil), pkt...)
+		if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
+			if err := virtio.FinishChecksum(cp, hdr); err != nil {
+				return err
+			}
+		}
+		*out = append(*out, cp)
+		return nil
+	}
+	proto, err := protoFromGSOType(hdr.GSOType)
+	if err != nil {
+		return err
+	}
+	p := wire.TunPacket{
+		Bytes: pkt,
+		Meta: wire.GSOInfo{
+			Size:      hdr.GSOSize,
+			HdrLen:    hdr.HdrLen,
+			CsumStart: hdr.CsumStart,
+			Proto:     proto,
+		},
+	}
+	return p.PerSegment(func(seg []byte) error {
+		*out = append(*out, append([]byte(nil), seg...))
+		return nil
+	})
+}
+
+// pseudoHeaderIPv4 returns the folded pseudo-header sum used to verify a
+// TCP/UDP segment's checksum in tests. src/dst are 4 bytes each.
+func pseudoHeaderIPv4(src, dst []byte, proto byte, l4Len int) uint16 {
+	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
+	s += uint32(proto) + uint32(l4Len)
+	s = (s & 0xffff) + (s >> 16)
+	s = (s & 0xffff) + (s >> 16)
+	return uint16(s)
+}
+
+// pseudoHeaderIPv6 returns the folded pseudo-header sum used to verify a
+// TCP/UDP segment's checksum in tests. src/dst are 16 bytes each.
+func pseudoHeaderIPv6(src, dst []byte, proto byte, l4Len int) uint16 {
+	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
+	s += uint32(l4Len>>16) + uint32(l4Len&0xffff) + uint32(proto)
+	s = (s & 0xffff) + (s >> 16)
+	s = (s & 0xffff) + (s >> 16)
+	return uint16(s)
+}
+
+// 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, virtio.Hdr) {
+	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, virtio.Hdr{
+		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, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %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, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %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 := virtio.Hdr{
+		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, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %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, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %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))
+	}
+}
+
+// TestSegmentRejectsLegacyUDPGSO ensures the legacy GSO_UDP (UFO) marker is
+// still rejected; only modern GSO_UDP_L4 (USO) is supported.
+func TestSegmentRejectsLegacyUDPGSO(t *testing.T) {
+	hdr := virtio.Hdr{GSOType: unix.VIRTIO_NET_HDR_GSO_UDP}
+	var out [][]byte
+	if err := segmentForTest(nil, hdr, &out, nil); err == nil {
+		t.Fatalf("expected rejection for legacy UDP GSO")
+	}
+}
+
+// buildUSOv4 builds a synthetic IPv4/UDP USO superpacket with payload of
+// payLen bytes, segmented at gsoSize.
+func buildUSOv4(t *testing.T, payLen, gsoSize int) ([]byte, virtio.Hdr) {
+	t.Helper()
+	const ipLen = 20
+	const udpLen = 8
+	pkt := make([]byte, ipLen+udpLen+payLen)
+
+	// IPv4 header
+	pkt[0] = 0x45 // version 4, IHL 5
+	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+udpLen+payLen))
+	binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
+	pkt[8] = 64
+	pkt[9] = unix.IPPROTO_UDP
+	copy(pkt[12:16], []byte{10, 0, 0, 1})
+	copy(pkt[16:20], []byte{10, 0, 0, 2})
+
+	// UDP header (length + checksum filled in per segment by segmentUDPYield)
+	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
+	binary.BigEndian.PutUint16(pkt[22:24], 53)    // dport
+
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+udpLen+i] = byte(i & 0xff)
+	}
+
+	return pkt, virtio.Hdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
+		HdrLen:     uint16(ipLen + udpLen),
+		GSOSize:    uint16(gsoSize),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 6,
+	}
+}
+
+func TestSegmentUDPv4(t *testing.T) {
+	const gso = 100
+	const numSeg = 3
+	pkt, hdr := buildUSOv4(t, gso*numSeg, gso)
+
+	scratch := make([]byte, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
+	}
+
+	for i, seg := range out {
+		if len(seg) != 28+gso {
+			t.Errorf("seg %d: len %d want %d", i, len(seg), 28+gso)
+		}
+		totalLen := binary.BigEndian.Uint16(seg[2:4])
+		if totalLen != uint16(28+gso) {
+			t.Errorf("seg %d: total_len=%d want %d", i, totalLen, 28+gso)
+		}
+		// kernel UDP-GSO does NOT bump the IPv4 ID across segments; every
+		// segment carries the same ID as the seed.
+		id := binary.BigEndian.Uint16(seg[4:6])
+		if id != 0x4242 {
+			t.Errorf("seg %d: ip id=%#x want %#x", i, id, 0x4242)
+		}
+		udpLen := binary.BigEndian.Uint16(seg[24:26])
+		if udpLen != uint16(8+gso) {
+			t.Errorf("seg %d: udp len=%d want %d", i, udpLen, 8+gso)
+		}
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_UDP, 8+gso)
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad UDP checksum", i)
+		}
+	}
+}
+
+func TestSegmentUDPv4OddTail(t *testing.T) {
+	// 250 bytes payload, gsoSize=100 → segments of 100, 100, 50.
+	pkt, hdr := buildUSOv4(t, 250, 100)
+	scratch := make([]byte, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %v", err)
+	}
+	if len(out) != 3 {
+		t.Fatalf("want 3 segments, got %d", len(out))
+	}
+	wantPay := []int{100, 100, 50}
+	for i, seg := range out {
+		if len(seg)-28 != wantPay[i] {
+			t.Errorf("seg %d: pay len %d want %d", i, len(seg)-28, wantPay[i])
+		}
+		udpLen := binary.BigEndian.Uint16(seg[24:26])
+		if udpLen != uint16(8+wantPay[i]) {
+			t.Errorf("seg %d: udp len=%d want %d", i, udpLen, 8+wantPay[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_UDP, 8+wantPay[i])
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad UDP checksum", i)
+		}
+	}
+}
+
+func TestSegmentUDPv6(t *testing.T) {
+	const ipLen = 40
+	const udpLen = 8
+	const gso = 120
+	const numSeg = 2
+	payLen := gso * numSeg
+	pkt := make([]byte, ipLen+udpLen+payLen)
+
+	// IPv6 header
+	pkt[0] = 0x60
+	binary.BigEndian.PutUint16(pkt[4:6], uint16(udpLen+payLen))
+	pkt[6] = unix.IPPROTO_UDP
+	pkt[7] = 64
+	pkt[8] = 0xfe
+	pkt[9] = 0x80
+	pkt[23] = 1
+	pkt[24] = 0xfe
+	pkt[25] = 0x80
+	pkt[39] = 2
+
+	binary.BigEndian.PutUint16(pkt[40:42], 12345)
+	binary.BigEndian.PutUint16(pkt[42:44], 53)
+
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+udpLen+i] = byte(i)
+	}
+
+	hdr := virtio.Hdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
+		HdrLen:     uint16(ipLen + udpLen),
+		GSOSize:    uint16(gso),
+		CsumStart:  uint16(ipLen),
+		CsumOffset: 6,
+	}
+
+	scratch := make([]byte, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %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+udpLen+gso {
+			t.Errorf("seg %d: len %d want %d", i, len(seg), ipLen+udpLen+gso)
+		}
+		pl := binary.BigEndian.Uint16(seg[4:6])
+		if pl != uint16(udpLen+gso) {
+			t.Errorf("seg %d: payload_length=%d want %d", i, pl, udpLen+gso)
+		}
+		ul := binary.BigEndian.Uint16(seg[ipLen+4 : ipLen+6])
+		if ul != uint16(udpLen+gso) {
+			t.Errorf("seg %d: udp len=%d want %d", i, ul, udpLen+gso)
+		}
+		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_UDP, udpLen+gso)
+		if !verifyChecksum(seg[ipLen:], psum) {
+			t.Errorf("seg %d: bad UDP checksum", i)
+		}
+	}
+}
+
+// TestSegmentUDPCEPropagates confirms IP-level CE marks on the seed appear on
+// every segment. UDP has no transport-level CWR/ECE: the IP TOS/TC byte is
+// copied verbatim into every segment by the segment-prefix copy.
+func TestSegmentUDPCEPropagates(t *testing.T) {
+	pkt, hdr := buildUSOv4(t, 200, 100)
+	pkt[1] = 0x03 // CE codepoint in IP-ECN
+
+	scratch := make([]byte, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %v", err)
+	}
+	if len(out) != 2 {
+		t.Fatalf("want 2 segments, got %d", len(out))
+	}
+	for i, seg := range out {
+		if seg[1]&0x03 != 0x03 {
+			t.Errorf("seg %d: CE missing (tos=%#x)", i, seg[1])
+		}
+		if !verifyChecksum(seg[:20], 0) {
+			t.Errorf("seg %d: bad IPv4 header checksum", i)
+		}
+	}
+}
+
+// TestSegmentTCPCwrFirstSegmentOnly confirms RFC 3168 §6.1.2: when a TSO
+// burst's seed has CWR set, only the first emitted segment carries CWR.
+// ECE is preserved on every segment (different signal, persistent state).
+func TestSegmentTCPCwrFirstSegmentOnly(t *testing.T) {
+	const mss = 100
+	const numSeg = 3
+	pkt, hdr := buildTSOv4(t, mss*numSeg, mss)
+	// Seed flags: CWR | ECE | ACK | PSH.
+	pkt[33] = 0x80 | 0x40 | 0x10 | 0x08
+
+	scratch := make([]byte, testSegScratchSize)
+	var out [][]byte
+	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
+		t.Fatalf("segmentForTest: %v", err)
+	}
+	if len(out) != numSeg {
+		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
+	}
+	for i, seg := range out {
+		flags := seg[33]
+		hasCwr := flags&0x80 != 0
+		hasEce := flags&0x40 != 0
+		hasPsh := flags&0x08 != 0
+		wantCwr := i == 0
+		wantPsh := i == numSeg-1
+		if hasCwr != wantCwr {
+			t.Errorf("seg %d: CWR=%v want %v (flags=%#x)", i, hasCwr, wantCwr, flags)
+		}
+		if !hasEce {
+			t.Errorf("seg %d: ECE missing (flags=%#x)", i, flags)
+		}
+		if hasPsh != wantPsh {
+			t.Errorf("seg %d: PSH=%v want %v (flags=%#x)", i, hasPsh, wantPsh, flags)
+		}
+		// IP and TCP checksums must still verify after the flag rewrite.
+		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+mss)
+		if !verifyChecksum(seg[20:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", i)
+		}
+	}
+}
+
+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 := virtio.Hdr{
+				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, testSegScratchSize)
+			out := make([][]byte, 0, 64)
+
+			// PerSegment consumes its input destructively; restore pkt from
+			// a master copy each iteration. The restore mirrors the
+			// kernel→userspace copy that hands a fresh GSO blob to the
+			// segmenter in production, so it's representative cost rather
+			// than bench overhead.
+			master := append([]byte(nil), pkt...)
+			work := make([]byte, len(pkt))
+
+			b.SetBytes(int64(len(pkt)))
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				copy(work, master)
+				out = out[:0]
+				if err := segmentForTest(work, 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 := &Offload{fd: fd}
+
+	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)
+	}
+}
+
+// buildTSOv6 builds a synthetic IPv6/TCP TSO superpacket with payLen bytes
+// of payload, segmented at gso. Returns the packet bytes only; the
+// virtio_net_hdr is the caller's responsibility.
+func buildTSOv6(payLen, gso int) []byte {
+	const ipLen = 40
+	const tcpLen = 20
+	pkt := make([]byte, ipLen+tcpLen+payLen)
+
+	pkt[0] = 0x60 // version 6
+	binary.BigEndian.PutUint16(pkt[4:6], uint16(tcpLen+payLen))
+	pkt[6] = unix.IPPROTO_TCP
+	pkt[7] = 64
+	pkt[8] = 0xfe
+	pkt[9] = 0x80
+	pkt[23] = 1
+	pkt[24] = 0xfe
+	pkt[25] = 0x80
+	pkt[39] = 2
+
+	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] = 0x10 // ACK only
+	binary.BigEndian.PutUint16(pkt[54:56], 65535)
+
+	for i := 0; i < payLen; i++ {
+		pkt[ipLen+tcpLen+i] = byte(i)
+	}
+	return pkt
+}
+
+// TestDecodeReadFitsMaxTSO proves decodeRead can absorb a worst-case
+// 64KiB TSO superpacket without dropping it. With segmentation deferred to
+// encrypt time, decodeRead writes nothing — it just slices the
+// caller-supplied mem and attaches GSO metadata — so the size requirement
+// is just "fit one worst-case input."
+//
+// Regression history: in a prior layout the rx buffer doubled as the
+// segmentation output, a near-threshold drain read returned "scratch too
+// small", the whole 45-segment TSO burst was dropped, and the remote's TCP
+// fast-retransmit collapsed cwnd. Keeping this test guards against
+// re-introducing per-call sizing assumptions inside decodeRead.
+func TestDecodeReadFitsMaxTSO(t *testing.T) {
+	const ipv6HdrLen = 40
+	const tcpHdrLen = 20
+	const headerLen = ipv6HdrLen + tcpHdrLen
+	pktLen := tunReadBufSize
+	payLen := pktLen - headerLen
+	const targetSegs = 64
+	gsoSize := (payLen + targetSegs - 1) / targetSegs
+
+	pkt := buildTSOv6(payLen, gsoSize)
+	if len(pkt) != pktLen {
+		t.Fatalf("buildTSOv6 produced %d bytes, want %d", len(pkt), pktLen)
+	}
+
+	o := &Offload{}
+	// mem is sized exactly to one worst-case packet — the caller-side
+	// invariant the drain loop in Read enforces. decodeRead must process
+	// the burst within that window.
+	mem := make([]byte, pktLen)
+	copy(mem, pkt)
+
+	// Encode the matching virtio_net_hdr.
+	hdr := virtio.Hdr{
+		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
+		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV6,
+		HdrLen:     uint16(headerLen),
+		GSOSize:    uint16(gsoSize),
+		CsumStart:  uint16(ipv6HdrLen),
+		CsumOffset: 16,
+	}
+	hdr.Encode(o.readVnetScratch[:])
+
+	var pkts []wire.TunPacket
+	pkts, err := o.decodeRead(pkts, mem, pktLen)
+	if err != nil {
+		t.Fatalf("decodeRead returned %v — sizing regression: "+
+			"tunRxBufSize=%d must hold one worst-case input (%d)",
+			err, tunRxBufSize, pktLen)
+	}
+
+	if len(pkts) != 1 {
+		t.Fatalf("got %d packets, want 1 superpacket entry", len(pkts))
+	}
+	got := pkts[0]
+	if !got.Meta.IsSuperpacket() {
+		t.Fatalf("expected superpacket GSO metadata, got %+v", got.Meta)
+	}
+	if got.Meta.Proto != wire.GSOProtoTCP {
+		t.Errorf("Meta.Proto=%d want TCP", got.Meta.Proto)
+	}
+	if got.Meta.Size != uint16(gsoSize) {
+		t.Errorf("Meta.Size=%d want %d", got.Meta.Size, gsoSize)
+	}
+	if got.Meta.HdrLen != uint16(headerLen) {
+		t.Errorf("Meta.HdrLen=%d want %d", got.Meta.HdrLen, headerLen)
+	}
+	if got.Meta.CsumStart != uint16(ipv6HdrLen) {
+		t.Errorf("Meta.CsumStart=%d want %d", got.Meta.CsumStart, ipv6HdrLen)
+	}
+	if len(got.Bytes) != pktLen {
+		t.Errorf("len(Bytes)=%d want %d", len(got.Bytes), pktLen)
+	}
+
+	// Validate that segmenting the returned superpacket reproduces the
+	// expected per-segment IPv6 payload length and TCP checksum.
+	wantSegs := (payLen + gsoSize - 1) / gsoSize
+	gotSegs := 0
+	if err := got.PerSegment(func(seg []byte) error {
+		defer func() { gotSegs++ }()
+		if len(seg) < headerLen+1 {
+			t.Errorf("seg %d too short: %d", gotSegs, len(seg))
+			return nil
+		}
+		if seg[0]>>4 != 6 {
+			t.Errorf("seg %d: bad IP version %#x", gotSegs, seg[0])
+		}
+		segPay := len(seg) - headerLen
+		gotPL := binary.BigEndian.Uint16(seg[4:6])
+		if gotPL != uint16(tcpHdrLen+segPay) {
+			t.Errorf("seg %d: payload_len=%d want %d", gotSegs, gotPL, tcpHdrLen+segPay)
+		}
+		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_TCP, tcpHdrLen+segPay)
+		if !verifyChecksum(seg[ipv6HdrLen:], psum) {
+			t.Errorf("seg %d: bad TCP checksum", gotSegs)
+		}
+		return nil
+	}); err != nil {
+		t.Fatalf("PerSegment: %v", err)
+	}
+	if gotSegs != wantSegs {
+		t.Fatalf("got %d segments, want %d", gotSegs, wantSegs)
+	}
+}

overlay/tio/virtio/header_linux.go

@@ -0,0 +1,43 @@
+//go:build linux && !android
+// +build linux,!android
+
+package virtio
+
+import "encoding/binary"
+
+// Size is the on-wire length of struct virtio_net_hdr the kernel
+// prepends/expects on a TUN opened with IFF_VNET_HDR (TUNSETVNETHDRSZ
+// not set).
+const Size = 10
+
+// Hdr is the Go view of the legacy virtio_net_hdr.
+type Hdr 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 *Hdr) 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 Size bytes). Used to emit a TSO superpacket on egress.
+func (h *Hdr) 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)
+}

overlay/tio/virtio/segment_linux.go

@@ -0,0 +1,402 @@
+//go:build linux && !android
+// +build linux,!android
+
+// Package virtio implements the pure validation, header-correction, and
+// per-segment slicing logic for kernel-supplied TSO/USO superpackets on
+// IFF_VNET_HDR TUN devices. It is FD-free and depends only on the byte
+// layout of the virtio_net_hdr and the IP/TCP/UDP headers it describes,
+// so it can be unit-tested in isolation from the tio Queue runtime.
+package virtio
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+
+	"golang.org/x/sys/unix"
+
+	"github.com/slackhq/nebula/overlay/checksum"
+)
+
+// Protocol header size bounds used to validate / cap kernel-supplied offsets.
+const (
+	ipv4HeaderMinLen = 20 // IHL=5, no options
+	ipv4HeaderMaxLen = 60 // IHL=15, max options
+	ipv6FixedLen     = 40 // IPv6 base header; extensions would extend this
+	tcpHeaderMinLen  = 20 // data-offset=5, no options
+	tcpHeaderMaxLen  = 60 // data-offset=15, max options
+)
+
+// Byte offsets inside an IPv4 header.
+const (
+	ipv4TotalLenOff = 2
+	ipv4IDOff       = 4
+	ipv4ChecksumOff = 10
+	ipv4SrcOff      = 12
+	ipv4AddrsEnd    = 20 // end of dst address (ipv4SrcOff + 2*4)
+)
+
+// Byte offsets inside an IPv6 header.
+const (
+	ipv6PayloadLenOff = 4
+	ipv6SrcOff        = 8
+	ipv6AddrsEnd      = 40 // end of dst address (ipv6SrcOff + 2*16)
+)
+
+// Byte offsets inside a TCP header (relative to its start, i.e. csumStart).
+const (
+	tcpSeqOff      = 4
+	tcpDataOffOff  = 12 // upper nibble is header len in 32-bit words
+	tcpFlagsOff    = 13
+	tcpChecksumOff = 16
+)
+
+// UDP header is fixed at 8 bytes: {sport, dport, length, checksum}.
+const (
+	udpHeaderLen   = 8
+	udpLengthOff   = 4
+	udpChecksumOff = 6
+)
+
+// tcpFinPshMask is cleared on every segment except the last of a TSO burst.
+const tcpFinPshMask = 0x09 // FIN(0x01) | PSH(0x08)
+
+// tcpCwrFlag is cleared on every segment except the first. Per RFC 3168
+// §6.1.2 the CWR bit signals a one-shot transition (the sender just halved
+// its window) and must appear on the first segment of a TSO burst only.
+const tcpCwrFlag = 0x80
+
+// CheckValid rejects packets whose virtio_net_hdr/IP combination would
+// cause a downstream miscompute. The TUN should never emit RSC_INFO and
+// the GSO type must agree with the IP version nibble.
+func CheckValid(pkt []byte, hdr Hdr) 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")
+	}
+	if len(pkt) < ipv4HeaderMinLen {
+		return fmt.Errorf("packet too short")
+	}
+	ipVersion := pkt[0] >> 4
+	switch hdr.GSOType {
+	case unix.VIRTIO_NET_HDR_GSO_TCPV4:
+		if ipVersion != 4 {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	case unix.VIRTIO_NET_HDR_GSO_TCPV6:
+		if ipVersion != 6 {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	case unix.VIRTIO_NET_HDR_GSO_UDP_L4:
+		// USO carries either v4 or v6; the leading nibble disambiguates.
+		if !(ipVersion == 4 || ipVersion == 6) {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	default:
+		if !(ipVersion == 6 || ipVersion == 4) {
+			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
+		}
+	}
+
+	return nil
+}
+
+// CorrectHdrLen rewrites hdr.HdrLen based on the actual transport header
+// length read out of pkt. The kernel's hdr.HdrLen on the FORWARD path can
+// be the length of the entire first packet, so we don't trust it.
+func CorrectHdrLen(pkt []byte, hdr *Hdr) error {
+	// Thank you wireguard-go for documenting these edge-cases
+	// Don't trust hdr.hdrLen from the kernel as it can be equal to the length
+	// of the entire first packet when the kernel is handling it as part of a
+	// FORWARD path. Instead, parse the transport header length and add it onto
+	// csumStart, which is synonymous for IP header length.
+
+	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_UDP_L4 {
+		hdr.HdrLen = hdr.CsumStart + 8
+	} else {
+		if len(pkt) <= int(hdr.CsumStart+tcpDataOffOff) {
+			return errors.New("packet is too short")
+		}
+
+		tcpHLen := uint16(pkt[hdr.CsumStart+tcpDataOffOff] >> 4 * 4)
+		if tcpHLen < 20 || tcpHLen > 60 {
+			// A TCP header must be between 20 and 60 bytes in length.
+			return fmt.Errorf("tcp header len is invalid: %d", tcpHLen)
+		}
+		hdr.HdrLen = hdr.CsumStart + tcpHLen
+	}
+
+	if len(pkt) < int(hdr.HdrLen) {
+		return fmt.Errorf("length of packet (%d) < virtioNetHdr.HdrLen (%d)", len(pkt), hdr.HdrLen)
+	}
+
+	if hdr.HdrLen < hdr.CsumStart {
+		return fmt.Errorf("virtioNetHdr.HdrLen (%d) < virtioNetHdr.CsumStart (%d)", hdr.HdrLen, hdr.CsumStart)
+	}
+	cSumAt := int(hdr.CsumStart + hdr.CsumStart)
+	if cSumAt+1 >= len(pkt) {
+		return fmt.Errorf("end of checksum offset (%d) exceeds packet length (%d)", cSumAt+1, len(pkt))
+	}
+	return nil
+}
+
+// SegmentTCP walks a TSO superpacket pkt, yielding each segment as a
+// slice into pkt itself. Per-segment plaintext is laid out by sliding a
+// freshly-patched copy of the L3+L4 header into pkt at offset i*gsoSize,
+// where it sits immediately before that segment's payload chunk in the
+// original buffer. The slide is destructive: iter i's header write overwrites
+// the last hdrLen bytes of seg_{i-1}'s payload, which is dead by the time
+// the next iteration begins. pkt is consumed by this call and must not be
+// inspected by the caller after the final yield.
+func SegmentTCP(pkt []byte, hdrLenU, csumStartU, gsoSizeU uint16, yield func(seg []byte) error) error {
+	if gsoSizeU == 0 {
+		return fmt.Errorf("gso_size is zero")
+	}
+	if csumStartU == 0 {
+		return fmt.Errorf("csum_start is zero")
+	}
+
+	headerLen := int(hdrLenU)
+	csumStart := int(csumStartU)
+	isV4 := pkt[0]>>4 == 4
+
+	tcpHdrLen := int(pkt[csumStart+tcpDataOffOff]>>4) * 4
+	payLen := len(pkt) - headerLen
+	gsoSize := int(gsoSizeU)
+	numSeg := (payLen + gsoSize - 1) / gsoSize
+	if numSeg == 0 {
+		numSeg = 1
+	}
+
+	origSeq := binary.BigEndian.Uint32(pkt[csumStart+tcpSeqOff : csumStart+tcpSeqOff+4])
+	origFlags := pkt[csumStart+tcpFlagsOff]
+
+	var tmp [tcpHeaderMaxLen]byte
+	copy(tmp[:tcpHdrLen], pkt[csumStart:headerLen])
+	tmp[tcpSeqOff], tmp[tcpSeqOff+1], tmp[tcpSeqOff+2], tmp[tcpSeqOff+3] = 0, 0, 0, 0
+	tmp[tcpFlagsOff] = 0
+	tmp[tcpChecksumOff], tmp[tcpChecksumOff+1] = 0, 0
+	baseTcpHdrSum := uint32(checksum.Checksum(tmp[:tcpHdrLen], 0))
+
+	var baseProtoSum uint32
+	if isV4 {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv4SrcOff:ipv4AddrsEnd], 0))
+	} else {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv6SrcOff:ipv6AddrsEnd], 0))
+	}
+	baseProtoSum += uint32(unix.IPPROTO_TCP)
+
+	var origIPID uint16
+	var baseIPHdrSum uint32
+	if isV4 {
+		origIPID = binary.BigEndian.Uint16(pkt[ipv4IDOff : ipv4IDOff+2])
+		ihl := int(pkt[0]&0x0f) * 4
+		if ihl < ipv4HeaderMinLen || ihl > csumStart {
+			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
+		}
+		var ipTmp [ipv4HeaderMaxLen]byte
+		copy(ipTmp[:ihl], pkt[:ihl])
+		ipTmp[ipv4TotalLenOff], ipTmp[ipv4TotalLenOff+1] = 0, 0
+		ipTmp[ipv4IDOff], ipTmp[ipv4IDOff+1] = 0, 0
+		ipTmp[ipv4ChecksumOff], ipTmp[ipv4ChecksumOff+1] = 0, 0
+		baseIPHdrSum = uint32(checksum.Checksum(ipTmp[:ihl], 0))
+	}
+
+	for i := 0; i < numSeg; i++ {
+		segStart := i * gsoSize
+		segEnd := segStart + gsoSize
+		if segEnd > payLen {
+			segEnd = payLen
+		}
+		segPayLen := segEnd - segStart
+		segLen := headerLen + segPayLen
+		headerOff := i * gsoSize
+
+		// Slide the header into place immediately before this segment's
+		// payload. Iter 0's header is already at pkt[:headerLen]; for
+		// i ≥ 1 we copy from there. The constant-byte fields of pkt[:headerLen]
+		// survive iter 0's in-place patches (only seq/flags/cksum/totalLen/id
+		// are touched), and iter 0's stale variable-field values are
+		// overwritten by the per-segment patches below.
+		if i > 0 {
+			copy(pkt[headerOff:headerOff+headerLen], pkt[:headerLen])
+		}
+		seg := pkt[headerOff : headerOff+segLen]
+
+		segSeq := origSeq + uint32(segStart)
+		segFlags := origFlags
+		if i != 0 {
+			segFlags &^= tcpCwrFlag
+		}
+		if i != numSeg-1 {
+			segFlags &^= tcpFinPshMask
+		}
+		totalLen := segLen
+
+		if isV4 {
+			segID := origIPID + uint16(i)
+			binary.BigEndian.PutUint16(seg[ipv4TotalLenOff:ipv4TotalLenOff+2], uint16(totalLen))
+			binary.BigEndian.PutUint16(seg[ipv4IDOff:ipv4IDOff+2], segID)
+			ipSum := baseIPHdrSum + uint32(totalLen) + uint32(segID)
+			binary.BigEndian.PutUint16(seg[ipv4ChecksumOff:ipv4ChecksumOff+2], foldComplement(ipSum))
+		} else {
+			binary.BigEndian.PutUint16(seg[ipv6PayloadLenOff:ipv6PayloadLenOff+2], uint16(headerLen-ipv6FixedLen+segPayLen))
+		}
+
+		binary.BigEndian.PutUint32(seg[csumStart+tcpSeqOff:csumStart+tcpSeqOff+4], segSeq)
+		seg[csumStart+tcpFlagsOff] = segFlags
+
+		tcpLen := tcpHdrLen + segPayLen
+		// Payload bytes still live at their original offset in pkt. The
+		// header slide above only writes into pkt[i*G : i*G+H], which is
+		// the tail of seg_{i-1}'s payload (already consumed) and never
+		// overlaps seg_i's own payload at pkt[H+i*G : H+(i+1)*G].
+		paySum := uint32(checksum.Checksum(pkt[headerLen+segStart:headerLen+segEnd], 0))
+		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+tcpChecksumOff:csumStart+tcpChecksumOff+2], foldComplement(uint32(wide)))
+
+		if err := yield(seg); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// SegmentUDP walks a USO superpacket, sliding a per-segment-patched
+// L3+L4 header into pkt at offset i*gsoSize and yielding pkt[i*G:i*G+segLen]
+// to the caller. Per-segment patches are total_len + IPv4 csum (or IPv6
+// payload_len) plus the UDP length and checksum. pkt is consumed
+// destructively; see SegmentTCP for the layout reasoning.
+//
+// UDP-GSO leaves the IPv4 ID identical across segments (the kernel does not
+// bump it), which is why the IP-level per-segment work is limited to
+// total_len + IPv4 header checksum (v4) or payload_len (v6).
+func SegmentUDP(pkt []byte, hdrLenU, csumStartU, gsoSizeU uint16, yield func(seg []byte) error) error {
+	if gsoSizeU == 0 {
+		return fmt.Errorf("gso_size is zero")
+	}
+	if csumStartU == 0 {
+		return fmt.Errorf("csum_start is zero")
+	}
+
+	isV4 := pkt[0]>>4 == 4
+	headerLen := int(hdrLenU)
+	csumStart := int(csumStartU)
+	if headerLen-csumStart != udpHeaderLen {
+		return fmt.Errorf("udp header len mismatch: %d", headerLen-csumStart)
+	}
+
+	payLen := len(pkt) - headerLen
+	gsoSize := int(gsoSizeU)
+	numSeg := (payLen + gsoSize - 1) / gsoSize
+	if numSeg == 0 {
+		numSeg = 1
+	}
+
+	var udpTmp [udpHeaderLen]byte
+	copy(udpTmp[:], pkt[csumStart:headerLen])
+	udpTmp[udpLengthOff], udpTmp[udpLengthOff+1] = 0, 0
+	udpTmp[udpChecksumOff], udpTmp[udpChecksumOff+1] = 0, 0
+	baseUDPHdrSum := uint32(checksum.Checksum(udpTmp[:], 0))
+
+	var baseProtoSum uint32
+	if isV4 {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv4SrcOff:ipv4AddrsEnd], 0))
+	} else {
+		baseProtoSum = uint32(checksum.Checksum(pkt[ipv6SrcOff:ipv6AddrsEnd], 0))
+	}
+	baseProtoSum += uint32(unix.IPPROTO_UDP)
+
+	var baseIPHdrSum uint32
+	if isV4 {
+		ihl := int(pkt[0]&0x0f) * 4
+		if ihl < ipv4HeaderMinLen || ihl > csumStart {
+			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
+		}
+		var ipTmp [ipv4HeaderMaxLen]byte
+		copy(ipTmp[:ihl], pkt[:ihl])
+		ipTmp[ipv4TotalLenOff], ipTmp[ipv4TotalLenOff+1] = 0, 0
+		ipTmp[ipv4ChecksumOff], ipTmp[ipv4ChecksumOff+1] = 0, 0
+		baseIPHdrSum = uint32(checksum.Checksum(ipTmp[:ihl], 0))
+	}
+
+	for i := 0; i < numSeg; i++ {
+		segStart := i * gsoSize
+		segEnd := segStart + gsoSize
+		if segEnd > payLen {
+			segEnd = payLen
+		}
+		segPayLen := segEnd - segStart
+		segLen := headerLen + segPayLen
+		headerOff := i * gsoSize
+
+		if i > 0 {
+			copy(pkt[headerOff:headerOff+headerLen], pkt[:headerLen])
+		}
+		seg := pkt[headerOff : headerOff+segLen]
+
+		totalLen := segLen
+		udpLen := udpHeaderLen + segPayLen
+
+		if isV4 {
+			binary.BigEndian.PutUint16(seg[ipv4TotalLenOff:ipv4TotalLenOff+2], uint16(totalLen))
+			ipSum := baseIPHdrSum + uint32(totalLen)
+			binary.BigEndian.PutUint16(seg[ipv4ChecksumOff:ipv4ChecksumOff+2], foldComplement(ipSum))
+		} else {
+			binary.BigEndian.PutUint16(seg[ipv6PayloadLenOff:ipv6PayloadLenOff+2], uint16(headerLen-ipv6FixedLen+segPayLen))
+		}
+
+		binary.BigEndian.PutUint16(seg[csumStart+udpLengthOff:csumStart+udpLengthOff+2], uint16(udpLen))
+
+		paySum := uint32(checksum.Checksum(pkt[headerLen+segStart:headerLen+segEnd], 0))
+		wide := uint64(baseUDPHdrSum) + uint64(paySum) + uint64(baseProtoSum)
+		wide += uint64(udpLen) + uint64(udpLen)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		wide = (wide & 0xffffffff) + (wide >> 32)
+		csum := foldComplement(uint32(wide))
+		if csum == 0 {
+			csum = 0xffff
+		}
+		binary.BigEndian.PutUint16(seg[csumStart+udpChecksumOff:csumStart+udpChecksumOff+2], csum)
+
+		if err := yield(seg); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// 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 Hdr) 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 := binary.BigEndian.Uint16(seg[cs+co : cs+co+2])
+	seg[cs+co] = 0
+	seg[cs+co+1] = 0
+	binary.BigEndian.PutUint16(seg[cs+co:cs+co+2], ^checksum.Checksum(seg[cs:], partial))
+	return nil
+}
+
+// foldComplement folds a 32-bit one's-complement partial sum to 16 bits and
+// complements it, yielding the on-wire Internet checksum value.
+func foldComplement(sum uint32) uint16 {
+	sum = (sum & 0xffff) + (sum >> 16)
+	sum = (sum & 0xffff) + (sum >> 16)
+	return ^uint16(sum)
+}

overlay/tun_android.go

@@ -13,12 +13,14 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	fd          int
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
@@ -26,16 +28,37 @@ type tun struct {
 	l           *slog.Logger
 }
 
+func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) == 0 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = struct{}{}
+	n, err := t.rwc.Read(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[:n]
+	return 1, nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
+}
+
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
 	// XXX Android returns an fd in non-blocking mode which is necessary for shutdown to work properly.
 	// Be sure not to call file.Fd() as it will set the fd to blocking mode.
 	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)
@@ -62,7 +85,7 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 	return r
 }
 
-func (t tun) Activate() error {
+func (t *tun) Activate() error {
 	return nil
 }
 
@@ -99,6 +122,14 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
+func (t *tun) NewMultiQueueReader() error {
+	return fmt.Errorf("TODO: multiqueue not implemented for android")
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
+func (t *tun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
 }

overlay/tun_darwin.go

@@ -16,14 +16,16 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
 	netroute "golang.org/x/net/route"
 	"golang.org/x/sys/unix"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	Device      string
 	vpnNetworks []netip.Prefix
 	DefaultMTU  int
@@ -124,11 +126,11 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 	}
 
 	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 +160,8 @@ func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (*tun, e
 }
 
 func (t *tun) Close() error {
-	if t.ReadWriteCloser != nil {
-		return t.ReadWriteCloser.Close()
+	if t.rwc != nil {
+		return t.rwc.Close()
 	}
 	return nil
 }
@@ -502,13 +504,17 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
-	buf := make([]byte, len(to)+4)
-
-	n, err := t.ReadWriteCloser.Read(buf)
-
-	copy(to, buf[4:])
-	return n - 4, err
+func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) <= 4 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = struct{}{}
+	n, err := t.rwc.Read(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[4:n]
+	return 1, nil
 }
 
 // Write is only valid for single threaded use
@@ -536,7 +542,7 @@ 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
 }
 
@@ -552,6 +558,14 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
+func (t *tun) NewMultiQueueReader() error {
+	return fmt.Errorf("TODO: multiqueue not implemented for darwin")
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
+func (t *tun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
 }

overlay/tun_disabled.go

@@ -10,7 +10,9 @@ import (
 
 	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/iputil"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/wire"
 )
 
 type disabledTun struct {
@@ -18,9 +20,10 @@ type disabledTun struct {
 	vpnNetworks []netip.Prefix
 
 	// Track these metrics since we don't have the tun device to do it for us
-	tx metrics.Counter
-	rx metrics.Counter
-	l  *slog.Logger
+	tx         metrics.Counter
+	rx         metrics.Counter
+	numReaders int
+	l          *slog.Logger
 }
 
 func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled bool, l *slog.Logger) *disabledTun {
@@ -28,6 +31,7 @@ func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled boo
 		vpnNetworks: vpnNetworks,
 		read:        make(chan []byte, queueLen),
 		l:           l,
+		numReaders:  1,
 	}
 
 	if metricsEnabled {
@@ -57,7 +61,7 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
-func (t *disabledTun) Read(b []byte) (int, error) {
+func (t *disabledTun) readOne(b []byte) (int, error) {
 	r, ok := <-t.read
 	if !ok {
 		return 0, io.EOF
@@ -75,6 +79,19 @@ func (t *disabledTun) Read(b []byte) (int, error) {
 	return copy(b, r), nil
 }
 
+func (t *disabledTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) == 0 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = wire.GSOInfo{}
+	n, err := t.readOne(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[:n]
+	return 1, nil
+}
+
 func (t *disabledTun) handleICMPEchoRequest(b []byte) bool {
 	out := make([]byte, len(b))
 	out = iputil.CreateICMPEchoResponse(b, out)
@@ -110,8 +127,21 @@ func (t *disabledTun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return t, nil
+func (t *disabledTun) NewMultiQueueReader() error {
+	t.numReaders++
+	return nil
+}
+
+func (t *disabledTun) Readers() []tio.Queue {
+	out := make([]tio.Queue, t.numReaders)
+	for i := range t.numReaders {
+		out[i] = t
+	}
+	return out
+}
+
+func (t *disabledTun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
 }
 
 func (t *disabledTun) Close() error {

overlay/tun_file_linux_test.go

@@ -1,120 +0,0 @@
-//go:build linux && !android && !e2e_testing
-// +build linux,!android,!e2e_testing
-
-package overlay
-
-import (
-	"errors"
-	"os"
-	"sync"
-	"testing"
-	"time"
-
-	"golang.org/x/sys/unix"
-)
-
-// newReadPipe returns a read fd. The matching write fd is registered for cleanup.
-// The caller takes ownership of the read fd (pass it to newTunFd / newFriend).
-func newReadPipe(t *testing.T) int {
-	t.Helper()
-	var fds [2]int
-	if err := unix.Pipe2(fds[:], unix.O_CLOEXEC); err != nil {
-		t.Fatalf("pipe2: %v", err)
-	}
-	t.Cleanup(func() { _ = unix.Close(fds[1]) })
-	return fds[0]
-}
-
-func TestTunFile_WakeForShutdown_UnblocksRead(t *testing.T) {
-	tf, err := newTunFd(newReadPipe(t))
-	if err != nil {
-		t.Fatalf("newTunFd: %v", err)
-	}
-	t.Cleanup(func() { _ = tf.Close() })
-
-	done := make(chan error, 1)
-	go func() {
-		_, err := tf.Read(make([]byte, 64))
-		done <- err
-	}()
-
-	// Verify Read is actually blocked in poll.
-	select {
-	case err := <-done:
-		t.Fatalf("Read returned before shutdown signal: %v", err)
-	case <-time.After(50 * time.Millisecond):
-	}
-
-	if err := tf.wakeForShutdown(); err != nil {
-		t.Fatalf("wakeForShutdown: %v", err)
-	}
-
-	select {
-	case err := <-done:
-		if !errors.Is(err, os.ErrClosed) {
-			t.Fatalf("expected os.ErrClosed, got %v", err)
-		}
-	case <-time.After(2 * time.Second):
-		t.Fatal("Read did not wake on shutdown")
-	}
-}
-
-func TestTunFile_WakeForShutdown_WakesFriends(t *testing.T) {
-	parent, err := newTunFd(newReadPipe(t))
-	if err != nil {
-		t.Fatalf("newTunFd: %v", err)
-	}
-	friend, err := parent.newFriend(newReadPipe(t))
-	if err != nil {
-		_ = parent.Close()
-		t.Fatalf("newFriend: %v", err)
-	}
-	t.Cleanup(func() {
-		_ = friend.Close()
-		_ = parent.Close()
-	})
-
-	readers := []*tunFile{parent, friend}
-	errs := make([]error, len(readers))
-	var wg sync.WaitGroup
-	for i, r := range readers {
-		wg.Add(1)
-		go func(i int, r *tunFile) {
-			defer wg.Done()
-			_, errs[i] = r.Read(make([]byte, 64))
-		}(i, r)
-	}
-
-	time.Sleep(50 * time.Millisecond)
-
-	if err := parent.wakeForShutdown(); err != nil {
-		t.Fatalf("wakeForShutdown: %v", err)
-	}
-
-	done := make(chan struct{})
-	go func() { wg.Wait(); close(done) }()
-	select {
-	case <-done:
-	case <-time.After(2 * time.Second):
-		t.Fatal("readers did not wake")
-	}
-
-	for i, err := range errs {
-		if !errors.Is(err, os.ErrClosed) {
-			t.Errorf("reader %d: expected os.ErrClosed, got %v", i, err)
-		}
-	}
-}
-
-func TestTunFile_Close_Idempotent(t *testing.T) {
-	tf, err := newTunFd(newReadPipe(t))
-	if err != nil {
-		t.Fatalf("newTunFd: %v", err)
-	}
-	if err := tf.Close(); err != nil {
-		t.Fatalf("first Close: %v", err)
-	}
-	if err := tf.Close(); err != nil {
-		t.Fatalf("second Close should be a no-op, got %v", err)
-	}
-}

overlay/tun_freebsd.go

@@ -7,7 +7,6 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
-	"io"
 	"io/fs"
 	"log/slog"
 	"net/netip"
@@ -18,9 +17,10 @@ import (
 	"unsafe"
 
 	"github.com/gaissmai/bart"
+	"github.com/slackhq/nebula/wire"
 
 	"github.com/slackhq/nebula/config"
-
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -157,7 +157,20 @@ func (t *tun) blockOnWrite() error {
 	return nil
 }
 
-func (t *tun) Read(to []byte) (int, error) {
+func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) == 0 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = struct{}{}
+	n, err := t.readOne(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[:n]
+	return 1, nil
+}
+
+func (t *tun) readOne(to []byte) (int, error) {
 	// first 4 bytes is protocol family, in network byte order
 	var head [4]byte
 	iovecs := [2]syscall.Iovec{
@@ -565,8 +578,8 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for freebsd")
+func (t *tun) NewMultiQueueReader() error {
+	return fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {
@@ -593,6 +606,14 @@ func (t *tun) addRoutes(logErrors bool) error {
 	return nil
 }
 
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
+func (t *tun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
+}
+
 func (t *tun) removeRoutes(routes []Route) error {
 	for _, r := range routes {
 		if !r.Install {

overlay/tun_ios.go

@@ -16,18 +16,41 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
 )
 
 type tun struct {
-	io.ReadWriteCloser
+	rwc         io.ReadWriteCloser
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *slog.Logger
 }
 
+func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) <= 4 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = struct{}{}
+	n, err := t.rwc.Read(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[4:n]
+	return 1, nil
+}
+
+func (t *tun) Write(p []byte) (int, error) {
+	return t.rwc.Write(p)
+}
+
+func (t *tun) Close() error {
+	return t.rwc.Close()
+}
+
 func newTun(_ *config.C, _ *slog.Logger, _ []netip.Prefix, _ bool) (*tun, error) {
 	return nil, fmt.Errorf("newTun not supported in iOS")
 }
@@ -35,9 +58,9 @@ func newTun(_ *config.C, _ *slog.Logger, _ []netip.Prefix, _ bool) (*tun, error)
 func newTunFromFd(c *config.C, l *slog.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)
@@ -96,18 +119,9 @@ type tunReadCloser struct {
 	wBuf []byte
 }
 
+// Read returns a packet with the BSD 4-byte header, watch out!
 func (tr *tunReadCloser) Read(to []byte) (int, error) {
-	tr.rMu.Lock()
-	defer tr.rMu.Unlock()
-
-	if cap(tr.rBuf) < len(to)+4 {
-		tr.rBuf = make([]byte, len(to)+4)
-	}
-	tr.rBuf = tr.rBuf[:len(to)+4]
-
-	n, err := tr.f.Read(tr.rBuf)
-	copy(to, tr.rBuf[4:])
-	return n - 4, err
+	return tr.f.Read(to)
 }
 
 func (tr *tunReadCloser) Write(from []byte) (int, error) {
@@ -155,6 +169,14 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
+func (t *tun) NewMultiQueueReader() error {
+	return fmt.Errorf("TODO: multiqueue not implemented for ios")
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
+func (t *tun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
 }

overlay/tun_linux.go

@@ -4,9 +4,7 @@
 package overlay
 
 import (
-	"encoding/binary"
 	"fmt"
-	"io"
 	"log/slog"
 	"net"
 	"net/netip"
@@ -19,180 +17,15 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/vishvananda/netlink"
 	"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
-}
-
-// 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)
-	}
-	return &tunFile{
-		fd:         fd,
-		shutdownFd: r.shutdownFd,
-		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},
-		},
-	}, nil
-}
-
-func newTunFd(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)
-	}
-
-	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,
-		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},
-		},
-	}
-
-	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) Read(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 if err == unix.EINTR {
-			continue
-		} else if err == unix.EBADF {
-			return 0, os.ErrClosed
-		} else {
-			return 0, err
-		}
-	}
-}
-
-func (r *tunFile) Write(buf []byte) (int, error) {
-	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 if err == unix.EBADF {
-			return 0, os.ErrClosed
-		} else {
-			return 0, err
-		}
-	}
-}
-
-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 {
-	*tunFile
-	readers     []*tunFile
+	readers     tio.QueueSet
 	closeLock   sync.Mutex
 	Device      string
 	vpnNetworks []netip.Prefix
@@ -201,6 +34,14 @@ type tun struct {
 	TXQueueLen  int
 	deviceIndex int
 	ioctlFd     uintptr
+	vnetHdr     bool
+	// routeFeatureECN, when true, sets RTAX_FEATURE_ECN on every route we
+	// install for the tun. The kernel then actively negotiates ECN for
+	// connections destined to those prefixes (equivalent to `ip route
+	// change ... features ecn`) regardless of net.ipv4.tcp_ecn, so flows
+	// across the nebula mesh use ECN even when the host default is the
+	// passive setting (=2). Disable via tunnels.ecn=false.
+	routeFeatureECN bool
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -239,7 +80,9 @@ type ifreqQLEN struct {
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	t, err := newTunGeneric(c, l, deviceFd, 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, false, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -249,46 +92,105 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	return t, nil
 }
 
-func newTun(c *config.C, l *slog.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.
+// TSO_ECN tells the kernel we propagate ECN correctly through coalesce and
+// segmentation, so it can deliver superpackets whose seed has CWR/ECE set
+// or whose IP-level codepoint is CE.
+const tsoOffloadFlags = unix.TUN_F_CSUM | unix.TUN_F_TSO4 | unix.TUN_F_TSO6 | unix.TUN_F_TSO_ECN
+
+// usoOffloadFlags adds UDP Segmentation Offload to tsoOffloadFlags. Requires
+// Linux ≥ 6.2; older kernels reject it and we fall back to TCP-only TSO via
+// tsoOffloadFlags.
+const usoOffloadFlags = tsoOffloadFlags | unix.TUN_F_USO4 | unix.TUN_F_USO6
+
+func newTun(c *config.C, l *slog.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 {
+
+	// First try to enable IFF_VNET_HDR via TUNSETIFF and negotiate TUN_F_*
+	// offloads via TUNSETOFFLOAD so we can receive TSO/USO superpackets.
+	// We try TSO+USO first, fall back to TSO-only on kernels without USO
+	// (Linux < 6.2), and finally give up on virtio headers entirely and
+	// reopen as a plain TUN if neither offload mask is accepted.
+	fd, err := openTunDev()
+	if err != nil {
+		return nil, err
+	}
+	vnetHdr := true
+	usoEnabled := false
+	name, err := tunSetIff(fd, nameStr, baseFlags|unix.IFF_VNET_HDR)
+	if err != nil {
 		_ = unix.Close(fd)
-		return nil, &NameError{
-			Name:       nameStr,
-			Underlying: err,
+		vnetHdr = false
+	} else {
+		// Try TSO+USO first. On kernels without USO support (Linux < 6.2)
+		// the ioctl returns EINVAL; fall back to the TCP-only mask before
+		// giving up on VNET_HDR entirely.
+		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(usoOffloadFlags)); err == nil {
+			usoEnabled = true
+		} else if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+			l.Warn("Failed to enable TUN offload (TSO); proceeding without virtio headers", "error", err)
+			_ = unix.Close(fd)
+			vnetHdr = false
 		}
 	}
-	name := strings.Trim(string(req.Name[:]), "\x00")
 
-	t, err := newTunGeneric(c, l, fd, vpnNetworks)
+	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}
+		}
+	}
+
+	if vnetHdr {
+		l.Info("TUN offload enabled", "tso", true, "uso", usoEnabled)
+	}
+
+	t, err := newTunGeneric(c, l, fd, vnetHdr, usoEnabled, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -299,20 +201,34 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 }
 
 // newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
-func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	tfd, err := newTunFd(fd)
+func newTunGeneric(c *config.C, l *slog.Logger, fd int, vnetHdr, usoEnabled bool, vpnNetworks []netip.Prefix) (*tun, error) {
+	var qs tio.QueueSet
+	var err error
+	if vnetHdr {
+		qs, err = tio.NewOffloadQueueSet(usoEnabled)
+	} else {
+		qs, err = tio.NewPollQueueSet()
+	}
+
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, err
 	}
+	err = qs.Add(fd)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, err
+	}
+
 	t := &tun{
-		tunFile:                   tfd,
-		readers:                   []*tunFile{tfd},
+		readers:                   qs,
 		closeLock:                 sync.Mutex{},
+		vnetHdr:                   vnetHdr,
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
 		useSystemRoutesBufferSize: c.GetInt("tun.use_system_route_table_buffer_size", 0),
+		routeFeatureECN:           c.GetBool("tunnels.ecn", true),
 		routesFromSystem:          map[netip.Prefix]routing.Gateways{},
 		l:                         l,
 	}
@@ -410,32 +326,38 @@ func (t *tun) SupportsMultiqueue() bool {
 	return true
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+func (t *tun) NewMultiQueueReader() 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
+		return 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
+	}
+	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	out, err := t.tunFile.newFriend(fd)
+	if t.vnetHdr {
+		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err != nil {
+			_ = unix.Close(fd)
+			return fmt.Errorf("failed to enable offload on multiqueue tun fd: %w", err)
+		}
+	}
+
+	err = t.readers.Add(fd)
 	if err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	t.readers = append(t.readers, out)
-
-	return out, nil
+	return nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -603,6 +525,18 @@ func (t *tun) setDefaultRoute(cidr netip.Prefix) error {
 		Table:     unix.RT_TABLE_MAIN,
 		Type:      unix.RTN_UNICAST,
 	}
+	// Match the metric the kernel uses for its auto-installed connected
+	// route, so RouteReplace overwrites it in place instead of adding a
+	// second route at a worse metric. IPv6 connected routes are installed
+	// at metric 256 (IP6_RT_PRIO_KERN); IPv4 uses 0. Without this, the
+	// kernel route wins lookups and our MTU / AdvMSS / Features never
+	// apply on v6.
+	if cidr.Addr().Is6() {
+		nr.Priority = 256
+	}
+	if t.routeFeatureECN {
+		nr.Features |= unix.RTAX_FEATURE_ECN
+	}
 	err := netlink.RouteReplace(&nr)
 	if err != nil {
 		t.l.Warn("Failed to set default route MTU, retrying", "error", err, "cidr", cidr)
@@ -652,6 +586,9 @@ func (t *tun) addRoutes(logErrors bool) error {
 		if r.Metric > 0 {
 			nr.Priority = r.Metric
 		}
+		if t.routeFeatureECN {
+			nr.Features |= unix.RTAX_FEATURE_ECN
+		}
 
 		err := netlink.RouteReplace(&nr)
 		if err != nil {
@@ -869,6 +806,10 @@ func (t *tun) updateRoutes(r netlink.RouteUpdate) {
 	t.routeTree.Store(newTree)
 }
 
+func (t *tun) Readers() []tio.Queue {
+	return t.readers.Queues()
+}
+
 func (t *tun) Close() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
@@ -878,32 +819,10 @@ func (t *tun) Close() error {
 		t.routeChan = nil
 	}
 
-	// Signal all readers blocked in poll to wake up and exit
-	_ = t.tunFile.wakeForShutdown()
-
 	if t.ioctlFd > 0 {
 		_ = unix.Close(int(t.ioctlFd))
 		t.ioctlFd = 0
 	}
 
-	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.Error("error closing tun reader", "reader", i, "error", err)
-		} else {
-			t.l.Info("closed tun reader", "reader", i)
-		}
-	}
-
-	//this is t.readers[0] too
-	err := t.tunFile.Close()
-	if err != nil {
-		t.l.Error("error closing tun reader", "reader", 0, "error", err)
-	} else {
-		t.l.Info("closed tun reader", "reader", 0)
-	}
-	return err
+	return t.readers.Close()
 }

overlay/tun_linux_test.go

@@ -3,7 +3,9 @@
 
 package overlay
 
-import "testing"
+import (
+	"testing"
+)
 
 var runAdvMSSTests = []struct {
 	name     string

overlay/tun_netbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -17,8 +16,10 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
+	"github.com/slackhq/nebula/wire"
 	netroute "golang.org/x/net/route"
 	"golang.org/x/sys/unix"
 )
@@ -68,6 +69,27 @@ type tun struct {
 	fd          int
 }
 
+func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+	if len(p) == 0 || len(mem) == 0 {
+		return 0, nil //todo should this be an err?
+	}
+	p[0].Meta = struct{}{}
+	n, err := t.readOne(mem)
+	if err != nil {
+		return 0, err
+	}
+	p[0].Bytes = mem[:n]
+	return 1, nil
+}
+
+func (t *tun) Readers() []tio.Queue {
+	return []tio.Queue{t}
+}
+
+func (t *tun) Capabilities() tio.Capabilities {
+	return tio.Capabilities{}
+}
+
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
 
 func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (*tun, error) {
@@ -141,7 +163,7 @@ 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)
@@ -394,8 +416,8 @@ func (t *tun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for netbsd")
+func (t *tun) NewMultiQueueReader() error {
+	return fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {