Group the HostInfo fields the packet paths touch

Co-authored-by: Jack Doan <me@jackdoan.com>

.github/actions/code-sign/action.yml

@@ -25,9 +25,9 @@ inputs:
     required: false
     default: "code-signer"
   key-prefix:
-    description: "S3 key prefix the caller is authorized to write under"
+    description: "S3 key prefix to write under; defaults to code-signing/<owner>/<repo> of the calling repo"
     required: false
-    default: "code-signing/slackhq/nebula"
+    default: ""
 
 runs:
   using: composite
@@ -57,6 +57,9 @@ runs:
         KEY_PREFIX: ${{ inputs.key-prefix }}
       run: |
         set -eu
+        # Default the prefix to this repo so the S3 key attributes the sign correctly.
+        # nebula-nightly runs this same action but writes under its own repo's prefix.
+        KEY_PREFIX="${KEY_PREFIX:-code-signing/$GITHUB_REPOSITORY}"
         RUN="${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
 
         find "$SIGN_PATH" -name '*.exe' -print | while read -r path

.github/workflows/gofmt.yml

@@ -1,34 +0,0 @@
-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

@@ -10,11 +10,11 @@ jobs:
     name: Build Linux/BSD All
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
-          go-version: '1.25'
+          go-version: '1.26'
           check-latest: true
 
       - name: Build
@@ -36,11 +36,11 @@ jobs:
       id-token: write
       contents: read
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
-          go-version: '1.25'
+          go-version: '1.26'
           check-latest: true
 
       - name: Build
@@ -76,11 +76,11 @@ jobs:
       HAS_SIGNING_CREDS: ${{ secrets.AC_USERNAME != '' }}
     runs-on: macos-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
-          go-version: '1.25'
+          go-version: '1.26'
           check-latest: true
 
       - name: Import certificates
@@ -134,7 +134,7 @@ jobs:
       # be overwritten
       - name: Checkout code
         if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Download artifacts
         if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
@@ -163,14 +163,17 @@ jobs:
           mkdir -p build/linux-{amd64,arm64}
           tar -zxvf artifacts/nebula-linux-amd64.tar.gz -C build/linux-amd64/
           tar -zxvf artifacts/nebula-linux-arm64.tar.gz -C build/linux-arm64/
-          docker buildx build . --push -f docker/Dockerfile --platform linux/amd64,linux/arm64 --tag "${DOCKER_IMAGE_REPO}:${DOCKER_IMAGE_TAG}" --tag "${DOCKER_IMAGE_REPO}:${GITHUB_REF#refs/tags/v}"
+          docker buildx build . --push -f docker/Dockerfile --platform linux/amd64,linux/arm64 \
+            --build-arg VERSION="${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}"
 
   release:
     name: Create and Upload Release
     needs: [build-linux, build-darwin, build-windows]
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
       - name: Download artifacts
         uses: actions/download-artifact@v8

.github/workflows/smoke-extra.yml

@@ -30,11 +30,11 @@ jobs:
       VAGRANT_DEFAULT_PROVIDER: libvirt
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: add hashicorp source
@@ -62,11 +62,11 @@ jobs:
       VAGRANT_DEFAULT_PROVIDER: virtualbox
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: add hashicorp source
@@ -88,11 +88,11 @@ jobs:
     runs-on: windows-latest
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     # WSL2 + Ubuntu so the smoke can run a real linux peer with its own

.github/workflows/smoke.yml

@@ -18,11 +18,11 @@ jobs:
     runs-on: ubuntu-latest
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: build
@@ -36,6 +36,14 @@ 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

@@ -5,6 +5,19 @@ 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
@@ -31,24 +44,24 @@ LIGHTHOUSE_IP="203.0.113.2"
         ../genconfig.sh >lighthouse1.yml
 
     HOST="host2" \
-        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
         ../genconfig.sh >host2.yml
 
     HOST="host3" \
-        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
         INBOUND='[{"port": "any", "proto": "icmp", "group": "lighthouse"}]' \
         ../genconfig.sh >host3.yml
 
     HOST="host4" \
-        LIGHTHOUSES="192.168.100.1 $LIGHTHOUSE_IP:4242" \
+        LIGHTHOUSES="$LIGHTHOUSE_NIP $LIGHTHOUSE_IP:4242" \
         OUTBOUND='[{"port": "any", "proto": "icmp", "group": "lighthouse"}]' \
         ../genconfig.sh >host4.yml
 
     ../../../../nebula-cert ca -curve "${CURVE:-25519}" -name "Smoke Test"
-    ../../../../nebula-cert sign -name "lighthouse1" -groups "lighthouse,lighthouse1" -ip "192.168.100.1/24"
+    ../../../../nebula-cert sign -name "lighthouse1" -groups "lighthouse,lighthouse1" -ip "$LIGHTHOUSE_NIP/24"
-    ../../../../nebula-cert sign -name "host2" -groups "host,host2" -ip "192.168.100.2/24"
+    ../../../../nebula-cert sign -name "host2" -groups "host,host2" -ip "$HOST2_NIP/24"
-    ../../../../nebula-cert sign -name "host3" -groups "host,host3" -ip "192.168.100.3/24"
+    ../../../../nebula-cert sign -name "host3" -groups "host,host3" -ip "$HOST3_NIP/24"
-    ../../../../nebula-cert sign -name "host4" -groups "host,host4" -ip "192.168.100.4/24"
+    ../../../../nebula-cert sign -name "host4" -groups "host,host4" -ip "$HOST4_NIP/24"
 )
 
 docker build -t "nebula:${NAME:-smoke}" .

.github/workflows/smoke/smoke.sh

@@ -47,6 +47,19 @@ 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
@@ -80,28 +93,28 @@ 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 0.0.0.0 2000 &
+docker exec host2 ncat -nklv 2000 &
-docker exec host3 ncat -nklv 0.0.0.0 2000 &
+docker exec host3 ncat -nklv 2000 &
-docker exec host4 ncat -e '/usr/bin/echo helloagainfromhost4' -nkluv 0.0.0.0 4000 &
+docker exec host4 ncat -e '/usr/bin/echo helloagainfromhost4' -nkluv 4000 &
-docker exec host2 ncat -e '/usr/bin/echo host2' -nkluv 0.0.0.0 3000 &
+docker exec host2 ncat -e '/usr/bin/echo host2' -nkluv 3000 &
-docker exec host3 ncat -e '/usr/bin/echo host3' -nkluv 0.0.0.0 3000 &
+docker exec host3 ncat -e '/usr/bin/echo host3' -nkluv 3000 &
 
 set +x
 echo
 echo " *** Testing ping from lighthouse1"
 echo
 set -x
-docker exec lighthouse1 ping -c1 192.168.100.2
+docker exec lighthouse1 ping -c1 $HOST2_NIP
-docker exec lighthouse1 ping -c1 192.168.100.3
+docker exec lighthouse1 ping -c1 $HOST3_NIP
 
 set +x
 echo
 echo " *** Testing ping from host2"
 echo
 set -x
-docker exec host2 ping -c1 192.168.100.1
+docker exec host2 ping -c1 $LIGHTHOUSE_NIP
 # Should fail because not allowed by host3 inbound firewall
-! docker exec host2 ping -c1 192.168.100.3 -w5 || exit 1
+! docker exec host2 ping -c1 $HOST3_NIP -w5 || exit 1
 
 set +x
 echo
@@ -109,34 +122,34 @@ echo " *** Testing ncat from host2"
 echo
 set -x
 # Should fail because not allowed by host3 inbound firewall
-! docker exec host2 ncat -nzv -w5 192.168.100.3 2000 || exit 1
+! docker exec host2 ncat -nzv -w5 $HOST3_NIP 2000 || exit 1
-! docker exec host2 ncat -nzuv -w5 192.168.100.3 3000 | grep -q host3 || exit 1
+! docker exec host2 ncat -nzuv -w5 $HOST3_NIP 3000 | grep -q host3 || exit 1
 
 set +x
 echo
 echo " *** Testing ping from host3"
 echo
 set -x
-docker exec host3 ping -c1 192.168.100.1
+docker exec host3 ping -c1 $LIGHTHOUSE_NIP
-docker exec host3 ping -c1 192.168.100.2
+docker exec host3 ping -c1 $HOST2_NIP
 
 set +x
 echo
 echo " *** Testing ncat from host3"
 echo
 set -x
-docker exec host3 ncat -nzv -w5 192.168.100.2 2000
+docker exec host3 ncat -nzv -w5 $HOST2_NIP 2000
-docker exec host3 ncat -nzuv -w5 192.168.100.2 3000 | grep -q host2
+docker exec host3 ncat -nzuv -w5 $HOST2_NIP 3000 | grep -q host2
 
 set +x
 echo
 echo " *** Testing ping from host4"
 echo
 set -x
-docker exec host4 ping -c1 192.168.100.1
+docker exec host4 ping -c1 $LIGHTHOUSE_NIP
 # Should fail because not allowed by host4 outbound firewall
-! docker exec host4 ping -c1 192.168.100.2 -w5 || exit 1
+! docker exec host4 ping -c1 $HOST2_NIP -w5 || exit 1
-! docker exec host4 ping -c1 192.168.100.3 -w5 || exit 1
+! docker exec host4 ping -c1 $HOST3_NIP -w5 || exit 1
 
 set +x
 echo
@@ -144,10 +157,10 @@ echo " *** Testing ncat from host4"
 echo
 set -x
 # Should fail because not allowed by host4 outbound firewall
-! docker exec host4 ncat -nzv -w5 192.168.100.2 2000 || exit 1
+! docker exec host4 ncat -nzv -w5 $HOST2_NIP 2000 || exit 1
-! docker exec host4 ncat -nzv -w5 192.168.100.3 2000 || exit 1
+! docker exec host4 ncat -nzv -w5 $HOST3_NIP 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 $HOST2_NIP 3000 | grep -q host2 || exit 1
-! docker exec host4 ncat -nzuv -w5 192.168.100.3 3000 | grep -q host3 || exit 1
+! docker exec host4 ncat -nzuv -w5 $HOST3_NIP 3000 | grep -q host3 || exit 1
 
 set +x
 echo
@@ -159,7 +172,7 @@ 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 192.168.100.4 4000" | grep -q helloagainfromhost4
+docker exec host2 sh -c "(/usr/bin/echo host2; sleep 2) | ncat -nuv $HOST4_NIP 4000" | grep -q helloagainfromhost4
 
 docker exec host4 sh -c 'kill 1'
 docker exec host3 sh -c 'kill 1'

.github/workflows/test.yml

@@ -13,20 +13,28 @@ on:
       - 'go.sum'
 jobs:
 
-  test-linux:
+  static:
-    name: Build all and test on ubuntu-linux
+    name: Static checks
     runs-on: ubuntu-latest
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
-    - name: Build
+    - name: Install goimports
-      run: make all
+      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: Vet
       run: make vet
@@ -34,99 +42,109 @@ jobs:
     - name: golangci-lint
       uses: golangci/golangci-lint-action@v9
       with:
-        version: v2.5
+        version: v2.12
-
-    - 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@v6
-
-    - uses: actions/setup-go@v6
-      with:
-        go-version: '1.25'
-        check-latest: true
-
-    - name: Build
-      run: make bin-boringcrypto
-
-    - name: Test
-      run: make test-boringcrypto
-
-    - name: End 2 end
-      run: make e2e GOEXPERIMENT=boringcrypto CGO_ENABLED=1 TEST_ENV="TEST_LOGS=1" TEST_FLAGS="-v -ldflags -checklinkname=0"
-
-  test-linux-pkcs11:
-    name: Build and test on linux with pkcs11
-    runs-on: ubuntu-latest
-    steps:
-
-    - uses: actions/checkout@v6
-
-    - uses: actions/setup-go@v6
-      with:
-        go-version: '1.25'
-        check-latest: true
-
-    - name: Build
-      run: make bin-pkcs11
-
-    - name: Test
-      run: make test-pkcs11
 
   test:
-    name: Build and test on ${{ matrix.os }}
+    name: Test ${{ matrix.name }}
     runs-on: ${{ matrix.os }}
     strategy:
+      fail-fast: false
       matrix:
-        os: [windows-latest, macos-latest]
+        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
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
-    - name: Build nebula
+    - name: Build
-      run: go build ./cmd/nebula
+      run: ${{ matrix.build-cmd }}
 
-    - name: Build nebula-cert
+    - name: Cross-build darwin-amd64
-      run: go build ./cmd/nebula-cert
+      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
-    - name: Vet
-      run: make vet
-
-    - name: golangci-lint
-      uses: golangci/golangci-lint-action@v9
-      with:
-        version: v2.5
 
     - name: Test
-      run: make test
+      run: ${{ matrix.test-cmd }}
 
     - name: End 2 end
-      run: make e2evv
+      if: matrix.e2e-cmd != ''
+      run: ${{ matrix.e2e-cmd }}
 
     - uses: actions/upload-artifact@v7
+      if: matrix.e2e-cmd != '' && always()
       with:
-        name: e2e packet flow ${{ matrix.os }}
+        name: e2e packet flow ${{ matrix.name }}
-        path: e2e/mermaid/${{ matrix.os }}
+        path: e2e/mermaid/
         if-no-files-found: warn
+
+  cross-build:
+    name: Cross-build ${{ matrix.name }}
+    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/setup-go@v7
+      with:
+        go-version: '1.26'
+        check-latest: true
+
+    - name: Build ${{ matrix.name }}
+      run: make -j"$(nproc)" ${{ matrix.make-target }}
+
+  finish:
+    name: CI status
+    if: always()
+    needs: [static, test, cross-build]
+    runs-on: ubuntu-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
+
+    - name: All upstream jobs passed
+      run: echo "ok"

CHANGELOG.md

@@ -7,6 +7,88 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.11.0] - 2026-07-23
+
+See the [v1.11.0](https://github.com/slackhq/nebula/milestone/25?closed=1) milestone for a complete list of changes.
+
+### Breaking
+
+- Logging has switched from logrus to Go's structured `slog`. Log output changes: levels are upper case
+  (`level=INFO`), trace prints as `level=DEBUG-4`, timestamps are always RFC3339Nano and `logging.timestamp_format`
+  is ignored, and some messages were reworded. Review any log parsing before upgrading. This is also an API break
+  for embedders, as constructors now take a `*slog.Logger`. (#1672, #1734, #1621)
+- `firewall.inbound_action` and `firewall.outbound_action` (used to set reject vs. drop policy) were each being
+  applied to the opposite direction, that is now corrected. This only affects how blocked packets are answered, not
+  which packets the firewall allows or denies. If you set either of these you are getting the behavior of the other
+  one today and likely want to swap them before upgrading. (#1798)
+- On Windows, Nebula now installs WFP PERMIT filters for the nebula adapter and the listener port by default. WFP
+  sits below Windows Defender Firewall, so any WDF inbound rules you rely on for either will no longer apply. Set
+  `tun.windows_bypass_wdf` and `listen.windows_bypass_wdf` to false to leave WDF in charge. (#1710)
+- On Windows, the nebula device is now set to the `private` network category instead of whatever Windows decided,
+  which is usually `Public`. This makes the host firewall less restrictive on the overlay. Set
+  `tun.network_category` to `unset` to keep the old behavior. (#1710)
+- Reject packets for non-TCP now use ICMP code 13, communication administratively prohibited, instead of code 3,
+  port unreachable. Anything keying off the old code needs updating. (#1766, #1768)
+- The SSH debug server's profiling commands are now confined to `sshd.sandbox_dir`, which defaults to
+  `$TMP/nebula-debug`. Relative paths resolve inside it and absolute paths outside it are rejected, so anything
+  scripting `start-cpu-profile`, `save-heap-profile`, or `save-mutex-profile` with a path elsewhere needs the
+  directory set. The directory is not created for you. (#1622)
+
+### Added
+
+- Sign the Windows release binaries. (#1718)
+- Generate IPv6 reject packets, matching the existing IPv4 behavior. (#1766, #1767, #1768)
+- Accept `-` in `nebula-cert` to read from stdin or write to stdout. (#1714)
+- Search for both `config.yml` and `config.yaml` in service and command line modes. (#1717)
+- Add version labels to the Docker/OCI images. (#1772)
+- Rebind the listener and re-query lighthouses on macOS when the underlay network changes, so devices moving
+  between wifi and wired or between networks recover without waiting for dead tunnel detection. Controlled by
+  `listen.rebind_on_network_change` (default `true`, not reloadable). (#1816)
+
+### Changed
+
+- Reload the firewall when the unsafe networks in the certificate change. (#1719)
+- Reconfigure, start, and stop the stats listener on a config reload instead of requiring a restart. (#1670)
+- Update a static host's addresses when they change on reload. (#1713)
+- Don't require a port on ICMP firewall rules. (#1609)
+- Connection track ICMP traffic. (#1602)
+- Return `NODATA` instead of `NXDOMAIN` from the DNS server for a name that exists but has no record of the
+  requested type, so clients that query `AAAA` first (busybox/Alpine) fall through to `A`. (#1668)
+- Record the local host's details in the DNS server. (#1716)
+- Install Windows unsafe routes as link routes. (#1709)
+- Reduce relay handshake log spam, and only log a handshake send error at error level when the remote list
+  changes. (#1733, #1765, #1810)
+- Start, stop, and reload subsystems (DNS, stats, conntrack, ssh, punchy) cleanly without leaking goroutines. (#1640, #1654, #1661, #1667, #1669, #1708, #1806, #1815)
+- `Control` is now safe to stop and wait on from any lifecycle state, and a new `Control.Wait` blocks until nebula
+  has fully stopped and returns the first fatal reader error. Failed starts release the udp sockets and tun fd
+  instead of leaking them. (#1794)
+- Trigger an immediate lighthouse update when reconnecting to or adding a lighthouse instead of waiting for the next update tick. (#1645)
+- Bring the Darwin and OpenBSD tun implementations in line with the other BSDs. (#1703)
+- Update to build against go v1.26. (#1818)
+- Various dependency updates. (#1586, #1587, #1604, #1617, #1618, #1627, #1628, #1629, #1652, #1664, #1665, #1697, #1721, #1732, #1742, #1743, #1750, #1763, #1771, #1782, #1800, #1807)
+
+### Fixed
+
+- Fix a data race on a host's remote address that could send packets to the wrong address during a roam. (#1773)
+- Fix tunnels that could permanently escape connection manager monitoring. (#1752)
+- Fix a crash when reloading the SSH server's trusted keys. (#1787)
+- Fix hostmap corruption when a host has multiple overlay addresses. Each address now gets its own list instead of
+  a single shared chain, which also fixes two latent bugs on the add and makePrimary paths. (#1788, #1790)
+- Apply `remote_allow_list` IPv4 rules to 4-in-6 mapped addresses. (#1786)
+- Don't panic in the DNS server on a short or empty query name. (#1635)
+- Advance the replay window on relayed packets so a relay drops replayed frames instead of re-forwarding them. (#1751)
+- Fix a race in relay state handling. (#1753)
+- Lock replay window updates so concurrent readers can't corrupt it. (#1802)
+- Reject malformed handshakes more reliably, including invalid ed25519 key lengths. (#1601, #1756)
+- Properly handle `closetunnel` packets. (#1638)
+- Fix an IPv6 extension-header length overflow that could make the firewall parse the wrong protocol and ports. (#1789)
+- Fix relay re-establishment when a handshake arrives over a relay entry that a one-sided teardown left
+  `Disestablished`, which silently dropped every send until dead tunnel detection forced a re-handshake. (#1805)
+- Don't build new relay state on a tunnel that was just discarded. (#1796)
+- Don't delete the wrong pending hostinfo in the handshake manager. (#1811)
+- Don't call the packet reader after a UDP error on Darwin. (#1755)
+- Open the FreeBSD tun device non blocking. (#1666)
+
 ## [1.10.3] - 2026-02-06
 
 ### Security

Makefile

@@ -60,6 +60,18 @@ 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
 
@@ -82,6 +94,35 @@ 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)
@@ -227,6 +268,9 @@ 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
@@ -236,5 +280,5 @@ smoke-vagrant/%: bin-docker build/%/nebula
 	cd .github/workflows/smoke/ && ./smoke-vagrant.sh $*
 
 .FORCE:
-.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/%
+.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/%
 .DEFAULT_GOAL := bin

cert/helper_test.go

@@ -13,6 +13,12 @@ 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
@@ -34,10 +40,10 @@ func NewTestCaCert(version Version, curve Curve, before, after time.Time, networ
 	}
 
 	if before.IsZero() {
-		before = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}
 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}
 
 	t := &TBSCertificate{
@@ -70,11 +76,11 @@ 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 = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}
 
 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}
 
 	if len(networks) == 0 {

cert/pem.go

@@ -148,6 +148,9 @@ func MarshalSigningPublicKeyToPEM(curve Curve, b []byte) []byte {
 	}
 }
 
+// UnmarshalPublicKeyFromPEM will try to unmarshal the first pem block in a byte array, returning any non
+// consumed data or an error on failure. Only key-agreement (ECDH) public key banners are accepted.
+// Use UnmarshalSigningPublicKeyFromPEM for Ed25519/ECDSA banners.
 func UnmarshalPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
 	k, r := pem.Decode(b)
 	if k == nil {
@@ -156,10 +159,10 @@ func UnmarshalPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
 	var expectedLen int
 	var curve Curve
 	switch k.Type {
-	case X25519PublicKeyBanner, Ed25519PublicKeyBanner:
+	case X25519PublicKeyBanner:
 		expectedLen = 32
 		curve = Curve_CURVE25519
-	case P256PublicKeyBanner, ECDSAP256PublicKeyBanner:
+	case P256PublicKeyBanner:
 		// Uncompressed
 		expectedLen = 65
 		curve = Curve_P256
@@ -172,6 +175,33 @@ func UnmarshalPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
 	return k.Bytes, r, curve, nil
 }
 
+// UnmarshalSigningPublicKeyFromPEM will try to unmarshal the first pem block in a byte array, returning any non
+// consumed data or an error on failure. Only Ed25519/ECDSA public key banners are accepted.
+// Use UnmarshalPublicKeyFromPEM for X25519/P256 (ECDH) banners.
+func UnmarshalSigningPublicKeyFromPEM(b []byte) ([]byte, []byte, Curve, error) {
+	k, r := pem.Decode(b)
+	if k == nil {
+		return nil, r, 0, fmt.Errorf("input did not contain a valid PEM encoded block")
+	}
+	var expectedLen int
+	var curve Curve
+	switch k.Type {
+	case Ed25519PublicKeyBanner:
+		expectedLen = 32
+		curve = Curve_CURVE25519
+	case ECDSAP256PublicKeyBanner:
+		// Uncompressed
+		expectedLen = 65
+		curve = Curve_P256
+	default:
+		return nil, r, 0, fmt.Errorf("bytes did not contain a proper Ed25519/ECDSA public key banner")
+	}
+	if len(k.Bytes) != expectedLen {
+		return nil, r, 0, fmt.Errorf("key was not %d bytes, is invalid %s public key", expectedLen, curve)
+	}
+	return k.Bytes, r, curve, nil
+}
+
 func MarshalPrivateKeyToPEM(curve Curve, b []byte) []byte {
 	switch curve {
 	case Curve_CURVE25519:

cert/pem_test.go

@@ -255,60 +255,6 @@ AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 func TestUnmarshalPublicKeyFromPEM(t *testing.T) {
 	t.Parallel()
 	pubKey := []byte(`# A good key
------BEGIN NEBULA ED25519 PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
------END NEBULA ED25519 PUBLIC KEY-----
-`)
-	shortKey := []byte(`# A short key
------BEGIN NEBULA ED25519 PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==
------END NEBULA ED25519 PUBLIC KEY-----
-`)
-	invalidBanner := []byte(`# Invalid banner
------BEGIN NOT A NEBULA PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
------END NOT A NEBULA PUBLIC KEY-----
-`)
-	invalidPem := []byte(`# Not a valid PEM format
--BEGIN NEBULA ED25519 PUBLIC KEY-----
-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
--END NEBULA ED25519 PUBLIC KEY-----`)
-
-	keyBundle := appendByteSlices(pubKey, shortKey, invalidBanner, invalidPem)
-
-	// Success test case
-	k, rest, curve, err := UnmarshalPublicKeyFromPEM(keyBundle)
-	assert.Len(t, k, 32)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	require.NoError(t, err)
-	assert.Equal(t, rest, appendByteSlices(shortKey, invalidBanner, invalidPem))
-
-	// Fail due to short key
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Nil(t, k)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	assert.Equal(t, rest, appendByteSlices(invalidBanner, invalidPem))
-	require.EqualError(t, err, "key was not 32 bytes, is invalid CURVE25519 public key")
-
-	// Fail due to invalid banner
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Nil(t, k)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	require.EqualError(t, err, "bytes did not contain a proper public key banner")
-	assert.Equal(t, rest, invalidPem)
-
-	// Fail due to invalid PEM format, because
-	// it's missing the requisite pre-encapsulation boundary.
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Nil(t, k)
-	assert.Equal(t, Curve_CURVE25519, curve)
-	assert.Equal(t, rest, invalidPem)
-	require.EqualError(t, err, "input did not contain a valid PEM encoded block")
-}
-
-func TestUnmarshalX25519PublicKey(t *testing.T) {
-	t.Parallel()
-	pubKey := []byte(`# A good key
 -----BEGIN NEBULA X25519 PUBLIC KEY-----
 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 -----END NEBULA X25519 PUBLIC KEY-----
@@ -319,7 +265,7 @@ AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
 AAAAAAAAAAAAAAAAAAAAAAA=
 -----END NEBULA P256 PUBLIC KEY-----
 `)
-	oldPubP256Key := []byte(`# A good key
+	signingKey := []byte(`# A signing key has the wrong scope for this function
 -----BEGIN NEBULA ECDSA P256 PUBLIC KEY-----
 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
 AAAAAAAAAAAAAAAAAAAAAAA=
@@ -340,44 +286,118 @@ AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
 -END NEBULA X25519 PUBLIC KEY-----`)
 
-	keyBundle := appendByteSlices(pubKey, pubP256Key, oldPubP256Key, shortKey, invalidBanner, invalidPem)
+	keyBundle := appendByteSlices(pubKey, pubP256Key, signingKey, shortKey, invalidBanner, invalidPem)
 
-	// Success test case
+	// X25519 key
 	k, rest, curve, err := UnmarshalPublicKeyFromPEM(keyBundle)
 	assert.Len(t, k, 32)
 	require.NoError(t, err)
-	assert.Equal(t, rest, appendByteSlices(pubP256Key, oldPubP256Key, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, rest, appendByteSlices(pubP256Key, signingKey, shortKey, invalidBanner, invalidPem))
 	assert.Equal(t, Curve_CURVE25519, curve)
 
-	// Success test case
+	// P256 key
 	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
 	assert.Len(t, k, 65)
 	require.NoError(t, err)
-	assert.Equal(t, rest, appendByteSlices(oldPubP256Key, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, rest, appendByteSlices(signingKey, shortKey, invalidBanner, invalidPem))
 	assert.Equal(t, Curve_P256, curve)
 
-	// Success test case
+	// Reject a signing public key (Ed25519/ECDSA banner)
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
-	assert.Len(t, k, 65)
+	assert.Nil(t, k)
-	require.NoError(t, err)
 	assert.Equal(t, rest, appendByteSlices(shortKey, invalidBanner, invalidPem))
-	assert.Equal(t, Curve_P256, curve)
+	require.EqualError(t, err, "bytes did not contain a proper public key banner")
 
 	// Fail due to short key
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
 	assert.Nil(t, k)
 	assert.Equal(t, rest, appendByteSlices(invalidBanner, invalidPem))
 	require.EqualError(t, err, "key was not 32 bytes, is invalid CURVE25519 public key")
 
 	// Fail due to invalid banner
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
 	assert.Nil(t, k)
 	require.EqualError(t, err, "bytes did not contain a proper public key banner")
 	assert.Equal(t, rest, invalidPem)
 
 	// Fail due to invalid PEM format, because
 	// it's missing the requisite pre-encapsulation boundary.
-	k, rest, curve, err = UnmarshalPublicKeyFromPEM(rest)
+	k, rest, _, err = UnmarshalPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	assert.Equal(t, rest, invalidPem)
+	require.EqualError(t, err, "input did not contain a valid PEM encoded block")
+}
+
+func TestUnmarshalSigningPublicKeyFromPEM(t *testing.T) {
+	t.Parallel()
+	pubKey := []byte(`# A good key
+-----BEGIN NEBULA ED25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-----END NEBULA ED25519 PUBLIC KEY-----
+`)
+	pubP256Key := []byte(`# A good key
+-----BEGIN NEBULA ECDSA P256 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
+AAAAAAAAAAAAAAAAAAAAAAA=
+-----END NEBULA ECDSA P256 PUBLIC KEY-----
+`)
+	ecdhKey := []byte(`# A key-agreement key has the wrong scope for this function
+-----BEGIN NEBULA X25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-----END NEBULA X25519 PUBLIC KEY-----
+`)
+	shortKey := []byte(`# A short key
+-----BEGIN NEBULA ED25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==
+-----END NEBULA ED25519 PUBLIC KEY-----
+`)
+	invalidBanner := []byte(`# Invalid banner
+-----BEGIN NOT A NEBULA PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-----END NOT A NEBULA PUBLIC KEY-----
+`)
+	invalidPem := []byte(`# Not a valid PEM format
+-BEGIN NEBULA ED25519 PUBLIC KEY-----
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=
+-END NEBULA ED25519 PUBLIC KEY-----`)
+
+	keyBundle := appendByteSlices(pubKey, pubP256Key, ecdhKey, shortKey, invalidBanner, invalidPem)
+
+	// Ed25519 key
+	k, rest, curve, err := UnmarshalSigningPublicKeyFromPEM(keyBundle)
+	assert.Len(t, k, 32)
+	require.NoError(t, err)
+	assert.Equal(t, rest, appendByteSlices(pubP256Key, ecdhKey, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, Curve_CURVE25519, curve)
+
+	// ECDSA P256 key
+	k, rest, curve, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Len(t, k, 65)
+	require.NoError(t, err)
+	assert.Equal(t, rest, appendByteSlices(ecdhKey, shortKey, invalidBanner, invalidPem))
+	assert.Equal(t, Curve_P256, curve)
+
+	// Reject a key-agreement public key (X25519/P256 banner)
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	assert.Equal(t, rest, appendByteSlices(shortKey, invalidBanner, invalidPem))
+	require.EqualError(t, err, "bytes did not contain a proper Ed25519/ECDSA public key banner")
+
+	// Fail due to short key
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	assert.Equal(t, rest, appendByteSlices(invalidBanner, invalidPem))
+	require.EqualError(t, err, "key was not 32 bytes, is invalid CURVE25519 public key")
+
+	// Fail due to invalid banner
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
+	assert.Nil(t, k)
+	require.EqualError(t, err, "bytes did not contain a proper Ed25519/ECDSA public key banner")
+	assert.Equal(t, rest, invalidPem)
+
+	// Fail due to invalid PEM format, because
+	// it's missing the requisite pre-encapsulation boundary.
+	k, rest, _, err = UnmarshalSigningPublicKeyFromPEM(rest)
 	assert.Nil(t, k)
 	assert.Equal(t, rest, invalidPem)
 	require.EqualError(t, err, "input did not contain a valid PEM encoded block")

cert_test/cert.go

@@ -14,6 +14,12 @@ 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
@@ -35,10 +41,10 @@ func NewTestCaCert(version cert.Version, curve cert.Curve, before, after time.Ti
 	}
 
 	if before.IsZero() {
-		before = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}
 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}
 
 	t := &cert.TBSCertificate{
@@ -71,11 +77,11 @@ 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 = time.Now().Add(time.Second * -60).Round(time.Second)
+		before = testCertNow.Add(time.Second * -60)
 	}
 
 	if after.IsZero() {
-		after = time.Now().Add(time.Second * 60).Round(time.Second)
+		after = testCertNow.Add(time.Second * 60)
 	}
 
 	var pub, priv []byte

cmd/nebula-cert/ca.go

@@ -97,6 +97,19 @@ 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
@@ -171,12 +184,21 @@ 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++ {
-				out.Write([]byte("Enter passphrase: "))
+				errOut.Write([]byte("Enter passphrase: "))
 				passphrase, err = pr.ReadPassword()
 
 				if err == ErrNoTerminal {
@@ -261,15 +283,17 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		Curve:          curve,
 	}
 
-	if !isP11 {
+	if !isP11 && !isStdio(*cf.outKeyPath) {
 		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)
 		}
+	}
 
 	var c cert.Certificate
 	var b []byte
@@ -294,7 +318,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 			b = cert.MarshalSigningPrivateKeyToPEM(curve, rawPriv)
 		}
 
-		err = os.WriteFile(*cf.outKeyPath, b, 0600)
+		err = writeOutput(*cf.outKeyPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
@@ -305,7 +329,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 		return fmt.Errorf("error while marshalling certificate: %s", err)
 	}
 
-	err = os.WriteFile(*cf.outCertPath, b, 0600)
+	err = writeOutput(*cf.outCertPath, b, 0600, out)
 	if err != nil {
 		return fmt.Errorf("error while writing out-crt: %s", err)
 	}
@@ -316,7 +340,7 @@ func ca(args []string, out io.Writer, errOut io.Writer, pr PasswordReader) error
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}
 
-		err = os.WriteFile(*cf.outQRPath, b, 0600)
+		err = writeOutput(*cf.outQRPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -332,6 +356,7 @@ 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

@@ -27,6 +27,7 @@ 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"+
@@ -84,7 +85,7 @@ func Test_ca(t *testing.T) {
 		err:      nil,
 	}
 
-	pwPromptOb := "Enter passphrase: "
+	pwPromptEB := "Enter passphrase: "
 
 	// required args
 	assertHelpError(t, ca(
@@ -168,8 +169,8 @@ 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.Equal(t, pwPromptOb, ob.String())
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Equal(t, pwPromptEB, eb.String())
 
 	// test encrypted key with passphrase environment variable
 	os.Remove(keyF.Name())
@@ -207,8 +208,8 @@ 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.Equal(t, pwPromptOb, ob.String())
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Equal(t, pwPromptEB, eb.String())
 
 	// test when user fails to enter a password
 	os.Remove(keyF.Name())
@@ -217,8 +218,8 @@ 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.Equal(t, strings.Repeat(pwPromptOb, 5), ob.String()) // prompts 5 times before giving up
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Equal(t, strings.Repeat(pwPromptEB, 5), eb.String()) // prompts 5 times before giving up
 
 	// create valid cert/key for overwrite tests
 	os.Remove(keyF.Name())
@@ -247,3 +248,67 @@ 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

@@ -42,6 +42,8 @@ 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
@@ -69,6 +71,14 @@ 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 {
@@ -82,12 +92,12 @@ func keygen(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while getting public key: %w", err)
 		}
 	} else {
-		err = os.WriteFile(*cf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600)
+		err = writeOutput(*cf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
 	}
-	err = os.WriteFile(*cf.outPubPath, cert.MarshalPublicKeyToPEM(curve, pub), 0600)
+	err = writeOutput(*cf.outPubPath, cert.MarshalPublicKeyToPEM(curve, pub), 0600, out)
 	if err != nil {
 		return fmt.Errorf("error while writing out-pub: %s", err)
 	}
@@ -102,6 +112,7 @@ 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

@@ -20,6 +20,7 @@ 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"+
@@ -93,3 +94,43 @@ 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

@@ -22,7 +22,9 @@ func (pr StdinPasswordReader) ReadPassword() ([]byte, error) {
 	}
 
 	password, err := term.ReadPassword(int(os.Stdin.Fd()))
-	fmt.Println()
+	// 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)
 
 	return password, err
 }

cmd/nebula-cert/print.go

@@ -40,11 +40,23 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}
 
-	rawCert, err := os.ReadFile(*pf.path)
+	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)
 	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
@@ -57,12 +69,14 @@ 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.outQRPath != "" {
 			b, err := c.MarshalPEM()
@@ -79,7 +93,7 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 		part++
 	}
 
-	if *pf.json {
+	if *pf.json && !qrToStdout {
 		b, _ := json.Marshal(jsonCerts)
 		_, _ = out.Write(b)
 		_, _ = out.Write([]byte("\n"))
@@ -91,7 +105,7 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}
 
-		err = os.WriteFile(*pf.outQRPath, b, 0600)
+		err = writeOutput(*pf.outQRPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -107,6 +121,7 @@ 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

@@ -25,6 +25,7 @@ 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"+
@@ -178,6 +179,44 @@ 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

@@ -85,6 +85,9 @@ 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
@@ -102,13 +105,35 @@ 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 := os.ReadFile(*sf.caKeyPath)
+		rawCAKey, err = readInput("ca-key", *sf.caKeyPath, &claims)
-
 		if err != nil {
 			return fmt.Errorf("error while reading ca-key: %s", err)
 		}
@@ -121,7 +146,7 @@ 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++ {
-					out.Write([]byte("Enter passphrase: "))
+					errOut.Write([]byte("Enter passphrase: "))
 					passphrase, err = pr.ReadPassword()
 
 					if errors.Is(err, ErrNoTerminal) {
@@ -147,7 +172,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		}
 	}
 
-	rawCACert, err := os.ReadFile(*sf.caCertPath)
+	rawCACert, err := readInput("ca-crt", *sf.caCertPath, &claims)
 	if err != nil {
 		return fmt.Errorf("error while reading ca-crt: %s", err)
 	}
@@ -245,7 +270,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 
 	if *sf.inPubPath != "" {
 		var pubCurve cert.Curve
-		rawPub, err := os.ReadFile(*sf.inPubPath)
+		rawPub, err := readInput("in-pub", *sf.inPubPath, &claims)
 		if err != nil {
 			return fmt.Errorf("error while reading in-pub: %s", err)
 		}
@@ -266,17 +291,11 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		pub, rawPriv = newKeypair(curve)
 	}
 
-	if *sf.outKeyPath == "" {
+	if !isStdio(*sf.outCertPath) {
-		*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
 
@@ -360,11 +379,13 @@ 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)
 			}
+		}
 
-		err = os.WriteFile(*sf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600)
+		err = writeOutput(*sf.outKeyPath, cert.MarshalPrivateKeyToPEM(curve, rawPriv), 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-key: %s", err)
 		}
@@ -379,7 +400,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 		b = append(b, sb...)
 	}
 
-	err = os.WriteFile(*sf.outCertPath, b, 0600)
+	err = writeOutput(*sf.outCertPath, b, 0600, out)
 	if err != nil {
 		return fmt.Errorf("error while writing out-crt: %s", err)
 	}
@@ -390,7 +411,7 @@ func signCert(args []string, out io.Writer, errOut io.Writer, pr PasswordReader)
 			return fmt.Errorf("error while generating qr code: %s", err)
 		}
 
-		err = os.WriteFile(*sf.outQRPath, b, 0600)
+		err = writeOutput(*sf.outQRPath, b, 0600, out)
 		if err != nil {
 			return fmt.Errorf("error while writing out-qr: %s", err)
 		}
@@ -440,6 +461,7 @@ 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

@@ -27,6 +27,7 @@ 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"+
@@ -376,15 +377,18 @@ 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.Equal(t, "Enter passphrase: ", ob.String())
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Equal(t, "Enter passphrase: ", 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", "")
 
@@ -395,8 +399,8 @@ 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.Equal(t, "Enter passphrase: ", ob.String())
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Equal(t, "Enter passphrase: ", eb.String())
 
 	// test with the wrong password in environment
 	ob.Reset()
@@ -416,8 +420,8 @@ 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.Equal(t, "Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: ", ob.String())
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.String())
+	assert.Equal(t, "Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: Enter passphrase: ", eb.String())
 
 	// test an error condition
 	ob.Reset()
@@ -425,6 +429,106 @@ 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.Equal(t, "Enter passphrase: ", ob.String())
+	assert.Empty(t, ob.String())
-	assert.Empty(t, eb.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())
 }

cmd/nebula-cert/stdio.go

@@ -0,0 +1,117 @@
+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

@@ -0,0 +1,167 @@
+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

@@ -39,18 +39,26 @@ func verify(args []string, out io.Writer, errOut io.Writer) error {
 		return err
 	}
 
-	caFile, err := os.Open(*vf.caPath)
+	var claims ioClaims
+	if err := reserveInputs(&claims,
+		"ca", *vf.caPath,
+		"crt", *vf.certPath,
+	); err != nil {
+		return err
+	}
+
+	caReader, err := openInput("ca", *vf.caPath, &claims)
 	if err != nil {
 		return fmt.Errorf("error while reading ca: %w", err)
 	}
-	defer caFile.Close()
+	defer caReader.Close()
 
-	caPool, err := cert.NewCAPoolFromPEMReader(caFile)
+	caPool, err := cert.NewCAPoolFromPEMReader(caReader)
 	if err != nil && !errors.Is(err, cert.ErrExpired) {
 		return fmt.Errorf("error while adding ca cert to pool: %w", err)
 	}
 
-	rawCert, err := os.ReadFile(*vf.certPath)
+	rawCert, err := readInput("crt", *vf.certPath, &claims)
 	if err != nil {
 		return fmt.Errorf("unable to read crt: %w", err)
 	}
@@ -85,6 +93,7 @@ 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

@@ -23,6 +23,7 @@ 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"+
@@ -122,3 +123,46 @@ 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

@@ -53,7 +53,12 @@ func main() {
 	l := logging.NewLogger(os.Stdout)
 
 	if *serviceFlag != "" {
-		if err := doService(configPath, configTest, Build, serviceFlag); err != nil {
+		if *configTest {
+			fmt.Println("-test is not supported with -service, run the config test without -service")
+			os.Exit(1)
+		}
+
+		if err := doService(configPath, Build, serviceFlag); err != nil {
 			l.Error("Service command failed", "error", err)
 			os.Exit(1)
 		}
@@ -61,10 +66,13 @@ func main() {
 	}
 
 	if *configPath == "" {
-		fmt.Println("-config flag must be set")
+		p, err := config.DefaultPath()
-		flag.Usage()
+		if err != nil {
+			fmt.Println(err)
 			os.Exit(1)
 		}
+		*configPath = p
+	}
 
 	c := config.NewC(l)
 	err := c.Load(*configPath)
@@ -90,15 +98,14 @@ func main() {
 	}
 
 	if !*configTest {
-		wait, err := ctrl.Start()
+		if err := ctrl.Start(); err != nil {
-		if err != nil {
 			util.LogWithContextIfNeeded("Error while running", err, l)
 			os.Exit(1)
 		}
 
 		go ctrl.ShutdownBlock()
 
-		if err := wait(); err != nil {
+		if err := ctrl.Wait(); err != nil {
 			l.Error("Nebula stopped due to fatal error", "error", err)
 			os.Exit(2)
 		}

cmd/nebula-service/service.go

@@ -4,7 +4,6 @@ import (
 	"fmt"
 	"log"
 	"os"
-	"path/filepath"
 
 	"github.com/kardianos/service"
 	"github.com/slackhq/nebula"
@@ -16,7 +15,6 @@ var logger service.Logger
 
 type program struct {
 	configPath *string
-	configTest *bool
 	build      string
 	control    *nebula.Control
 }
@@ -42,39 +40,47 @@ func (p *program) Start(s service.Service) error {
 		}
 	})
 
-	p.control, err = nebula.Main(c, *p.configTest, Build, l, nil)
+	p.control, err = nebula.Main(c, false, Build, l, nil)
 	if err != nil {
 		return err
 	}
 
-	p.control.Start()
+	if err := p.control.Start(); err != nil {
+		return err
+	}
+
+	// Nebula can stop itself on a fatal packet reader error, make sure to log it if it happens.
+	go func() {
+		if err := p.control.Wait(); err != nil {
+			logger.Error(fmt.Sprintf("Nebula stopped due to fatal error: %v", err))
+			os.Exit(2)
+		}
+	}()
+
 	return nil
 }
 
 func (p *program) Stop(s service.Service) error {
 	logger.Info("Nebula service stopping.")
+	if p.control == nil {
+		return nil
+	}
+
 	p.control.Stop()
+
+	// block until nebula has fully drained before reporting stopped.
+	// error logging is handled by Start.
+	_ = p.control.Wait()
 	return nil
 }
 
-func fileExists(filename string) bool {
+func doService(configPath *string, build string, serviceFlag *string) error {
-	_, 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 == "" {
-		ex, err := os.Executable()
+		p, err := config.DefaultPath()
 		if err != nil {
 			return err
 		}
-		*configPath = filepath.Dir(ex) + "/config.yaml"
+		*configPath = p
-		if !fileExists(*configPath) {
-			*configPath = filepath.Dir(ex) + "/config.yml"
-		}
 	}
 
 	svcConfig := &service.Config{
@@ -86,7 +92,6 @@ func doService(configPath *string, configTest *bool, build string, serviceFlag *
 
 	prg := &program{
 		configPath: configPath,
-		configTest: configTest,
 		build:      build,
 	}
 
@@ -118,8 +123,9 @@ func doService(configPath *string, configTest *bool, build string, serviceFlag *
 	switch *serviceFlag {
 	case "run":
 		if err := s.Run(); err != nil {
-			// Route any errors to the system logger
+			// Route any errors to the system logger and report the failure
 			logger.Error(err)
+			return err
 		}
 	default:
 		if err := service.Control(s, *serviceFlag); err != nil {

cmd/nebula/close_on_timer_test.go

@@ -0,0 +1,96 @@
+//go:build linux && !android && !e2e_testing
+
+package main
+
+import (
+	"fmt"
+	"net/netip"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	cert_test "github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/test"
+	"github.com/stretchr/testify/require"
+)
+
+// TestControlStopClosesOnTimer reproduces the dnclient lifecycle: nebula runs as
+// a library, and on a config update dnclient calls Stop() in-process to tear the
+// old instance down before starting a new one. This boots a real nebula (real
+// blocking UDP sockets, tun disabled), lets it run, then Stop()s it on a timer
+// and asserts it actually closes. If the reader goroutines parked in recvmmsg
+// don't wake on Close(), Wait() blocks forever and this fails with a goroutine
+// dump instead of relying on a process signal to unstick them.
+func TestControlStopClosesOnTimer(t *testing.T) {
+	l := test.NewLogger()
+	dir := t.TempDir()
+
+	before := time.Now().Add(-time.Hour)
+	after := time.Now().Add(time.Hour)
+	ca, _, caKey, caPEM := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, before, after, nil, nil, nil)
+	networks := []netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")}
+	_, _, keyPEM, certPEM := cert_test.NewTestCert(cert.Version2, cert.Curve_CURVE25519, ca, caKey, "close-on-timer", before, after, networks, nil, nil)
+
+	caPath := filepath.Join(dir, "ca.pem")
+	certPath := filepath.Join(dir, "cert.pem")
+	keyPath := filepath.Join(dir, "key.pem")
+	require.NoError(t, os.WriteFile(caPath, caPEM, 0o600))
+	require.NoError(t, os.WriteFile(certPath, certPEM, 0o600))
+	require.NoError(t, os.WriteFile(keyPath, keyPEM, 0o600))
+
+	// tun disabled so no device/root is needed; routines: 2 so we exercise the
+	// multi-socket (SO_REUSEPORT) teardown, which is where dnclient runs.
+	configBody := fmt.Sprintf(`
+pki:
+  ca: %s
+  cert: %s
+  key: %s
+listen:
+  host: 127.0.0.1
+  port: 0
+tun:
+  disabled: true
+firewall:
+  outbound:
+    - port: any
+      proto: any
+      host: any
+  inbound:
+    - port: any
+      proto: any
+      host: any
+routines: 2
+`, caPath, certPath, keyPath)
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "config.yml"), []byte(configBody), 0o600))
+
+	c := config.NewC(l)
+	require.NoError(t, c.Load(dir))
+
+	ctrl, err := nebula.Main(c, false, "close-on-timer", l, nil)
+	require.NoError(t, err)
+	require.NoError(t, ctrl.Start())
+
+	// Run like a live nebula, then close on a timer, exactly as dnclient does.
+	<-time.NewTimer(5 * time.Second).C
+
+	stopped := make(chan struct{})
+	go func() {
+		ctrl.Stop() // closes the udp sockets (shutdown(2)) and the tun
+		ctrl.Wait() // blocks until every reader goroutine has returned
+		close(stopped)
+	}()
+
+	select {
+	case <-stopped:
+		t.Log("nebula closed cleanly on timer")
+	case <-time.After(10 * time.Second):
+		buf := make([]byte, 1<<20)
+		n := runtime.Stack(buf, true)
+		t.Fatalf("nebula did NOT close within 10s of Stop(): a blocking reader never woke\n%s", buf[:n])
+	}
+}

cmd/nebula/main.go

@@ -50,10 +50,13 @@ func main() {
 	}
 
 	if *configPath == "" {
-		fmt.Println("-config flag must be set")
+		p, err := config.DefaultPath()
-		flag.Usage()
+		if err != nil {
+			fmt.Println(err)
 			os.Exit(1)
 		}
+		*configPath = p
+	}
 
 	l := logging.NewLogger(os.Stdout)
 
@@ -81,8 +84,7 @@ func main() {
 	}
 
 	if !*configTest {
-		wait, err := ctrl.Start()
+		if err := ctrl.Start(); err != nil {
-		if err != nil {
 			util.LogWithContextIfNeeded("Error while running", err, l)
 			os.Exit(1)
 		}
@@ -90,7 +92,7 @@ func main() {
 		go ctrl.ShutdownBlock()
 		notifyReady(l)
 
-		if err := wait(); err != nil {
+		if err := ctrl.Wait(); err != nil {
 			l.Error("Nebula stopped due to fatal error", "error", err)
 			os.Exit(2)
 		}

config/default.go

@@ -0,0 +1,29 @@
+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

@@ -0,0 +1,67 @@
+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

@@ -105,11 +105,12 @@ func (cm *connectionManager) getInactivityTimeout() time.Duration {
 }
 
 func (cm *connectionManager) In(h *HostInfo) {
-	h.in.Store(true)
+	h.markIn()
 }
 
-func (cm *connectionManager) Out(h *HostInfo) {
+// Out records outbound traffic and reports whether the local network changed since this tunnel last sent.
-	h.out.Store(true)
+func (cm *connectionManager) Out(h *HostInfo) bool {
+	return h.markOut(cm.intf.rebindEpoch.Load())
 }
 
 func (cm *connectionManager) RelayUsed(localIndex uint32) {
@@ -128,22 +129,13 @@ func (cm *connectionManager) RelayUsed(localIndex uint32) {
 // getAndResetTrafficCheck returns if there was any inbound or outbound traffic within the last tick and
 // resets the state for this local index
 func (cm *connectionManager) getAndResetTrafficCheck(h *HostInfo, now time.Time) (bool, bool) {
-	in := h.in.Swap(false)
+	in, out := h.takeTraffic()
-	out := h.out.Swap(false)
 	if in || out {
 		h.lastUsed = now
 	}
 	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()
@@ -306,8 +298,8 @@ 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", req.RelayFromAddr,
+				"relayFrom", relayFrom,
-				"relayTo", req.RelayToAddr,
+				"relayTo", relayTo,
 				"initiatorRelayIndex", req.InitiatorRelayIndex,
 				"responderRelayIndex", req.ResponderRelayIndex,
 				"vpnAddrs", newhostinfo.vpnAddrs,
@@ -348,7 +340,7 @@ func (cm *connectionManager) makeTrafficDecision(localIndex uint32, now time.Tim
 				"tunnelCheck", m{"state": "alive", "method": "passive"},
 			)
 		}
-		hostinfo.pendingDeletion.Store(false)
+		hostinfo.setPendingDeletion(false)
 
 		if mainHostInfo {
 			decision = tryRehandshake
@@ -371,7 +363,7 @@ func (cm *connectionManager) makeTrafficDecision(localIndex uint32, now time.Tim
 		return decision, hostinfo, primary
 	}
 
-	if hostinfo.pendingDeletion.Load() {
+	if hostinfo.isPendingDeletion() {
 		// We have already sent a test packet and nothing was returned, this hostinfo is dead
 		hostinfo.logger(cm.l).Info("Tunnel status",
 			"tunnelCheck", m{"state": "dead", "method": "active"},
@@ -422,7 +414,7 @@ func (cm *connectionManager) makeTrafficDecision(localIndex uint32, now time.Tim
 		}
 	}
 
-	hostinfo.pendingDeletion.Store(true)
+	hostinfo.setPendingDeletion(true)
 	cm.trafficTimer.Add(hostinfo.localIndexId, cm.pendingDeletionInterval)
 	return decision, hostinfo, nil
 }

connection_manager_test.go

@@ -25,6 +25,7 @@ func newTestLighthouse() *LightHouse {
 	lighthouses := []netip.Addr{}
 	staticList := map[netip.Addr]struct{}{}
 
+	lh.localAddrsFn = func(*LocalAllowList) []netip.Addr { return nil }
 	lh.lighthouses.Store(&lighthouses)
 	lh.staticList.Store(&staticList)
 
@@ -85,25 +86,25 @@ func Test_NewConnectionManagerTest(t *testing.T) {
 	// We saw traffic out to vpnIp
 	nc.Out(hostinfo)
 	nc.In(hostinfo)
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
-	assert.True(t, hostinfo.out.Load())
+	assert.True(t, hostinfo.sentSinceCheck())
-	assert.True(t, hostinfo.in.Load())
+	assert.True(t, (hostinfo.state.Load()&stateIn != 0))
 
 	// Do a traffic check tick, should not be pending deletion but should not have any in/out packets recorded
 	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 
 	// Do another traffic check tick, this host should be pending deletion now
 	nc.Out(hostinfo)
-	assert.True(t, hostinfo.out.Load())
+	assert.True(t, hostinfo.sentSinceCheck())
 	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
-	assert.True(t, hostinfo.pendingDeletion.Load())
+	assert.True(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 
@@ -167,33 +168,33 @@ func Test_NewConnectionManagerTest2(t *testing.T) {
 	// We saw traffic out to vpnIp
 	nc.Out(hostinfo)
 	nc.In(hostinfo)
-	assert.True(t, hostinfo.in.Load())
+	assert.True(t, (hostinfo.state.Load()&stateIn != 0))
-	assert.True(t, hostinfo.out.Load())
+	assert.True(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 
 	// Do a traffic check tick, should not be pending deletion but should not have any in/out packets recorded
 	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 
 	// Do another traffic check tick, this host should be pending deletion now
 	nc.Out(hostinfo)
 	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
-	assert.True(t, hostinfo.pendingDeletion.Load())
+	assert.True(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 
 	// We saw traffic, should no longer be pending deletion
 	nc.In(hostinfo)
 	nc.doTrafficCheck(hostinfo.localIndexId, p, nb, out, time.Now())
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 }
@@ -252,31 +253,31 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 	// Do a traffic check tick, in and out should be cleared but should not be pending deletion
 	nc.Out(hostinfo)
 	nc.In(hostinfo)
-	assert.True(t, hostinfo.out.Load())
+	assert.True(t, hostinfo.sentSinceCheck())
-	assert.True(t, hostinfo.in.Load())
+	assert.True(t, (hostinfo.state.Load()&stateIn != 0))
 
 	now := time.Now()
 	decision, _, _ := nc.makeTrafficDecision(hostinfo.localIndexId, now)
 	assert.Equal(t, tryRehandshake, decision)
 	assert.Equal(t, now, hostinfo.lastUsed)
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 
 	decision, _, _ = nc.makeTrafficDecision(hostinfo.localIndexId, now.Add(time.Second*5))
 	assert.Equal(t, doNothing, decision)
 	assert.Equal(t, now, hostinfo.lastUsed)
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 
 	// Do another traffic check tick, should still not be pending deletion
 	decision, _, _ = nc.makeTrafficDecision(hostinfo.localIndexId, now.Add(time.Second*10))
 	assert.Equal(t, doNothing, decision)
 	assert.Equal(t, now, hostinfo.lastUsed)
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 
@@ -284,9 +285,9 @@ func Test_NewConnectionManager_DisconnectInactive(t *testing.T) {
 	decision, _, _ = nc.makeTrafficDecision(hostinfo.localIndexId, now.Add(time.Minute*10))
 	assert.Equal(t, closeTunnel, decision)
 	assert.Equal(t, now, hostinfo.lastUsed)
-	assert.False(t, hostinfo.pendingDeletion.Load())
+	assert.False(t, hostinfo.isPendingDeletion())
-	assert.False(t, hostinfo.out.Load())
+	assert.False(t, hostinfo.sentSinceCheck())
-	assert.False(t, hostinfo.in.Load())
+	assert.False(t, (hostinfo.state.Load()&stateIn != 0))
 	assert.Contains(t, nc.hostMap.Indexes, hostinfo.localIndexId)
 	assert.Contains(t, nc.hostMap.Hosts, hostinfo.vpnAddrs[0])
 }

connection_state.go

@@ -2,15 +2,17 @@ package nebula
 
 import (
 	"encoding/json"
+	"log/slog"
 	"sync"
 	"sync/atomic"
 
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/handshake"
+	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/noiseutil"
 )
 
-const ReplayWindow = 8192
+const ReplayWindow = 1024
 
 type ConnectionState struct {
 	eKey           noiseutil.CipherState
@@ -20,6 +22,7 @@ type ConnectionState struct {
 	initiator      bool
 	messageCounter atomic.Uint64
 	window         *Bits
+	decryptLock    sync.Mutex
 	writeLock      sync.Mutex
 }
 
@@ -54,3 +57,52 @@ func (cs *ConnectionState) MarshalJSON() ([]byte, error) {
 func (cs *ConnectionState) Curve() cert.Curve {
 	return cs.myCert.Curve()
 }
+
+func (cs *ConnectionState) Decrypt(l *slog.Logger, messageCounter uint64, out []byte, packet []byte, nb []byte) ([]byte, error) {
+	var err error
+	cs.decryptLock.Lock()
+	result := cs.window.Check(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return nil, ErrAlreadySeen
+	}
+
+	out, err = cs.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], messageCounter, nb)
+	if err != nil {
+		return nil, err
+	}
+
+	cs.decryptLock.Lock()
+	result = cs.window.Update(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return nil, ErrAlreadySeen
+	}
+	return out, nil
+}
+
+// VerifyRelay verifies AEAD protected (but not encrypted) relay frames. packet must be length-checked by the caller.
+func (cs *ConnectionState) VerifyRelay(l *slog.Logger, messageCounter uint64, packet []byte, nb []byte) error {
+	cs.decryptLock.Lock()
+	result := cs.window.Check(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return ErrAlreadySeen
+	}
+
+	signedPayload := packet[:len(packet)-cs.dKey.Overhead()]
+	signatureValue := packet[len(packet)-cs.dKey.Overhead():]
+	_, err := cs.dKey.DecryptDanger(nil, signedPayload, signatureValue, messageCounter, nb)
+	if err != nil {
+		return err
+	}
+
+	cs.decryptLock.Lock()
+	result = cs.window.Update(l, messageCounter)
+	cs.decryptLock.Unlock()
+	if !result {
+		return ErrAlreadySeen
+	}
+
+	return nil
+}

control.go

@@ -53,6 +53,7 @@ type Control struct {
 	statsStart             func()
 	dnsStart               func()
 	lighthouseStart        func()
+	networkChangeStart     func(rebind func())
 	connectionManagerStart func(context.Context)
 }
 
@@ -69,29 +70,29 @@ type ControlHostInfo struct {
 }
 
 // Start actually runs nebula, this is a nonblocking call.
-// The returned function blocks until nebula has fully stopped and returns the
+// Use Wait to block until nebula has fully stopped and to learn whether a fatal reader error caused the shutdown.
-// first fatal reader error (if any). A nil error means nebula shut down
+func (c *Control) Start() error {
-// gracefully; a non-nil error means a reader hit an unexpected failure that
-// triggered the shutdown.
-func (c *Control) Start() (func() error, error) {
 	c.stateLock.Lock()
 	defer c.stateLock.Unlock()
 	switch c.state {
 	case StateReady:
 		//yay!
 	case StateStopped, StateStopping:
-		return nil, ErrAlreadyStopped
+		return ErrAlreadyStopped
 	case StateStarted:
-		return nil, ErrAlreadyStarted
+		return ErrAlreadyStarted
 	default:
-		return nil, ErrUnknownState
+		return ErrUnknownState
 	}
 
 	// Activate the interface
 	err := c.f.activate()
 	if err != nil {
+		// Cancel before Close so a caller returning from Wait always observes a dead Context
+		c.cancel()
+		_ = c.f.Close()
 		c.state = StateStopped
-		return nil, err
+		return err
 	}
 
 	// Call all the delayed funcs that waited patiently for the interface to be created.
@@ -104,6 +105,9 @@ func (c *Control) Start() (func() error, error) {
 	if c.dnsStart != nil {
 		go c.dnsStart()
 	}
+	if c.networkChangeStart != nil {
+		go c.networkChangeStart(c.RebindUDPServer)
+	}
 	if c.connectionManagerStart != nil {
 		go c.connectionManagerStart(c.ctx)
 	}
@@ -114,13 +118,9 @@ func (c *Control) Start() (func() error, error) {
 	c.f.triggerShutdown = c.Stop
 
 	// Start reading packets.
-	out, err := c.f.run()
+	c.f.run()
-	if err != nil {
-		c.state = StateStopped
-		return nil, err
-	}
 	c.state = StateStarted
-	return out, nil
+	return nil
 }
 
 func (c *Control) State() RunState {
@@ -133,10 +133,26 @@ func (c *Control) Context() context.Context {
 	return c.ctx
 }
 
-// Stop is a non-blocking call that signals nebula to close all tunnels and shut down
+// Stop tears nebula down, closing all tunnels and releasing everything it holds.
+// Use Wait to block until the shutdown has completed.
+// A Control that has been stopped cannot be started again, Start will return ErrAlreadyStopped.
 func (c *Control) Stop() {
 	c.stateLock.Lock()
-	if c.state != StateStarted {
+	switch c.state {
+	case StateStarted:
+		// Fall through to the full teardown below
+
+	case StateReady:
+		// Never started
+		c.cancel()
+		c.state = StateStopped
+		if err := c.f.Close(); err != nil {
+			c.l.Error("Close interface failed", "error", err)
+		}
+		c.stateLock.Unlock()
+		return
+
+	default:
 		c.stateLock.Unlock()
 		// We are stopping or stopped already
 		return
@@ -145,19 +161,26 @@ func (c *Control) Stop() {
 	c.state = StateStopping
 	c.stateLock.Unlock()
 
-	// Stop the handshakeManager (and other services), to prevent new tunnels from
+	// Closing tunnels can be slow with a large hostmap, don't hold the lock for it
-	// being created while we're shutting them all down.
 	c.cancel()
-
 	c.CloseAllTunnels(false)
+
+	c.stateLock.Lock()
+	c.state = StateStopped
 	if err := c.f.Close(); err != nil {
 		c.l.Error("Close interface failed", "error", err)
 	}
-	c.stateLock.Lock()
-	c.state = StateStopped
 	c.stateLock.Unlock()
 }
 
+// Wait blocks until nebula has fully stopped, either via Stop or an internal fatal error,
+// and returns the first fatal packet reader error if there was one.
+// It is safe to call from multiple goroutines and at any point in the lifecycle,
+// but a Wait on a Control that is never started and never stopped will block forever.
+func (c *Control) Wait() error {
+	return c.f.wait()
+}
+
 // ShutdownBlock will listen for and block on term and interrupt signals, calling Control.Stop() once signalled
 func (c *Control) ShutdownBlock() {
 	sigChan := make(chan os.Signal, 1)
@@ -170,15 +193,26 @@ func (c *Control) ShutdownBlock() {
 	c.Stop()
 }
 
-// RebindUDPServer asks the UDP listener to rebind it's listener. Mainly used on mobile clients when interfaces change
+// RebindUDPServer asks the UDP listener to rebind it's listener. Mainly used on mobile clients when interfaces change.
 func (c *Control) RebindUDPServer() {
-	_ = c.f.outside.Rebind()
+	c.stateLock.Lock()
+	defer c.stateLock.Unlock()
+
+	if c.state != StateStarted {
+		return
+	}
+
+	// A failure here means we are likely still pinned to the interface we came up on, so the rest of this is
+	// unlikely to help. Say so instead of silently carrying on as if we rebound.
+	if err := c.f.outside.Rebind(); err != nil {
+		c.l.Error("Failed to rebind udp socket", "error", err)
+	}
 
 	// Trigger a lighthouse update, useful for mobile clients that should have an update interval of 0
 	c.f.lightHouse.SendUpdate()
 
 	// Let the main interface know that we rebound so that underlying tunnels know to trigger punches from their remotes
-	c.f.rebindCount++
+	c.f.rebindEpoch.Add(1)
 }
 
 // ListHostmapHosts returns details about the actual or pending (handshaking) hostmap by vpn ip
@@ -305,7 +339,7 @@ func (c *Control) CloseAllTunnels(excludeLighthouses bool) (closed int) {
 
 		c.l.Debug("Sending close tunnel message",
 			"vpnAddrs", h.vpnAddrs,
-			"udpAddr", h.remote,
+			"udpAddr", h.GetRemote(),
 		)
 		closed++
 	}
@@ -350,7 +384,7 @@ func copyHostInfo(h *HostInfo, preferredRanges []netip.Prefix) ControlHostInfo {
 		RemoteAddrs:            h.remotes.CopyAddrs(preferredRanges),
 		CurrentRelaysToMe:      h.relayState.CopyRelayIps(),
 		CurrentRelaysThroughMe: h.relayState.CopyRelayForIps(),
-		CurrentRemote:          h.remote,
+		CurrentRemote:          h.GetRemote(),
 	}
 
 	for i, a := range h.vpnAddrs {

control_lifecycle_test.go

@@ -0,0 +1,292 @@
+package nebula
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/netip"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/gaissmai/bart"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/routing"
+	"github.com/slackhq/nebula/test"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+type fakeDevice struct {
+	closeOnce sync.Once
+	closedCh  chan struct{}
+	closed    bool
+}
+
+func newFakeDevice() *fakeDevice {
+	return &fakeDevice{closedCh: make(chan struct{})}
+}
+
+// Read blocks until Close like a real tun with no traffic, then reports EOF
+// the same way a closed device does
+func (d *fakeDevice) Read(p []byte) (int, error) {
+	<-d.closedCh
+	return 0, io.EOF
+}
+
+func (d *fakeDevice) Write(p []byte) (int, error) { return len(p), nil }
+
+func (d *fakeDevice) Close() error {
+	d.closeOnce.Do(func() {
+		d.closed = true
+		close(d.closedCh)
+	})
+	return nil
+}
+
+func (d *fakeDevice) Activate() error                       { return nil }
+func (d *fakeDevice) Networks() []netip.Prefix              { return nil }
+func (d *fakeDevice) Name() string                          { return "fake" }
+func (d *fakeDevice) RoutesFor(netip.Addr) routing.Gateways { return nil }
+func (d *fakeDevice) SupportsMultiqueue() bool              { return false }
+func (d *fakeDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, errors.New("unsupported")
+}
+
+// newReadyControl hand-builds the minimum Control that Main would have
+// produced right before Start, including the construction token NewInterface
+// takes so waiters block until Close releases the resources
+func newReadyControl(t *testing.T) (*Control, *fakeDevice, *fakeConn) {
+	l := test.NewLogger()
+	dev := newFakeDevice()
+	conn := &fakeConn{}
+	ctx, cancel := context.WithCancel(context.Background())
+
+	myVpnNet := netip.MustParsePrefix("10.128.0.1/16")
+	nt := new(bart.Lite)
+	nt.Insert(myVpnNet)
+	cs := &CertState{
+		myVpnNetworks:      []netip.Prefix{myVpnNet},
+		myVpnNetworksTable: nt,
+	}
+	lh, err := NewLightHouseFromConfig(ctx, l, config.NewC(l), cs, nil, nil)
+	require.NoError(t, err)
+
+	f := &Interface{
+		ctx:        ctx,
+		inside:     dev,
+		outside:    conn,
+		writers:    []udp.Conn{conn},
+		readers:    make([]io.ReadWriteCloser, 1),
+		routines:   1,
+		hostMap:    newHostMap(l),
+		lightHouse: lh,
+		l:          l,
+	}
+	f.wg.Add(1)
+
+	return &Control{
+		state:  StateReady,
+		f:      f,
+		l:      l,
+		ctx:    ctx,
+		cancel: cancel,
+	}, dev, conn
+}
+
+func TestControl_StopBeforeStart(t *testing.T) {
+	c, dev, conn := newReadyControl(t)
+
+	// A Stop on a never started control must release everything Main acquired
+	c.Stop()
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed, "the tun device should have been closed")
+	assert.True(t, conn.closed, "the udp socket should have been closed")
+	require.ErrorIs(t, c.ctx.Err(), context.Canceled, "the service context should have been cancelled")
+
+	// Wait must return promptly now that the resources are released
+	require.NoError(t, c.Wait())
+
+	// A stopped control can never be started
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
+
+	// A second Stop is a harmless no-op
+	c.Stop()
+	assert.Equal(t, StateStopped, c.State())
+	require.NoError(t, c.Wait())
+}
+
+func TestControl_WaitBlocksUntilStop(t *testing.T) {
+	c, _, _ := newReadyControl(t)
+
+	done := make(chan error, 1)
+	go func() { done <- c.Wait() }()
+
+	select {
+	case <-done:
+		t.Fatal("Wait returned before Stop")
+	case <-time.After(50 * time.Millisecond):
+	}
+
+	c.Stop()
+	select {
+	case err := <-done:
+		require.NoError(t, err)
+	case <-time.After(time.Second):
+		t.Fatal("Wait did not return after Stop")
+	}
+}
+
+type fakeConn struct {
+	closed  bool
+	rebinds int
+}
+
+func (c *fakeConn) Rebind() error                            { c.rebinds++; return nil }
+func (c *fakeConn) LocalAddr() (netip.AddrPort, error)       { return netip.AddrPort{}, nil }
+func (c *fakeConn) ListenOut(_ udp.EncReader) error          { return nil }
+func (c *fakeConn) WriteTo(_ []byte, _ netip.AddrPort) error { return nil }
+func (c *fakeConn) ReloadConfig(_ *config.C)                 {}
+func (c *fakeConn) SupportsMultipleReaders() bool            { return true }
+func (c *fakeConn) Close() error                             { c.closed = true; return nil }
+
+type multiqueueDevice struct {
+	*fakeDevice
+}
+
+func (d *multiqueueDevice) SupportsMultiqueue() bool { return true }
+
+func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
+	dev := &multiqueueDevice{fakeDevice: newFakeDevice()}
+	conn := &fakeConn{}
+	ctx, cancel := context.WithCancel(context.Background())
+	f := &Interface{
+		ctx:      ctx,
+		inside:   dev,
+		outside:  conn,
+		writers:  []udp.Conn{conn},
+		readers:  make([]io.ReadWriteCloser, 2),
+		routines: 2,
+		l:        test.NewLogger(),
+	}
+	f.wg.Add(1)
+
+	c := &Control{
+		state:  StateReady,
+		f:      f,
+		l:      test.NewLogger(),
+		ctx:    ctx,
+		cancel: cancel,
+	}
+
+	// The second reader fails to open, everything must be released
+	require.Error(t, c.Start())
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed, "the tun device should have been closed")
+	assert.True(t, conn.closed, "the udp socket should have been closed")
+	require.ErrorIs(t, c.ctx.Err(), context.Canceled)
+
+	// And Wait must not hang on the construction token
+	require.NoError(t, c.Wait())
+}
+
+func TestInterface_CloseIsIdempotent(t *testing.T) {
+	dev := newFakeDevice()
+	f := &Interface{
+		inside: dev,
+		l:      test.NewLogger(),
+	}
+	f.wg.Add(1)
+
+	require.NoError(t, f.Close())
+	assert.True(t, dev.closed)
+
+	// A second Close must not double release the wg token or the device
+	require.NoError(t, f.Close())
+	require.NoError(t, f.wait())
+}
+
+func TestControl_FatalErrorReportsThroughWait(t *testing.T) {
+	c, dev, conn := newReadyControl(t)
+
+	// Mirror what Start wires up, without needing real packet readers
+	c.f.triggerShutdown = c.Stop
+	c.state = StateStarted
+
+	boom := errors.New("boom")
+	c.f.onFatal(boom)
+
+	require.ErrorIs(t, c.Wait(), boom)
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed)
+	assert.True(t, conn.closed)
+
+	// A second fatal error must not fire the shutdown again or replace the first
+	c.f.onFatal(errors.New("later"))
+	require.ErrorIs(t, c.Wait(), boom)
+
+	// Wait stays factual, a Stop after the death does not mask the error
+	c.Stop()
+	require.ErrorIs(t, c.Wait(), boom)
+}
+
+func TestControl_ConcurrentStopAndStart(t *testing.T) {
+	c, _, _ := newReadyControl(t)
+
+	var wg sync.WaitGroup
+	for i := 0; i < 2; i++ {
+		wg.Go(func() { c.Stop() })
+	}
+	wg.Go(func() { _ = c.Start() })
+	wg.Go(func() {
+		_ = c.Wait()
+		// A returned Wait must always observe the final state, no matter how
+		// the race resolved
+		assert.Equal(t, StateStopped, c.State())
+	})
+	wg.Wait()
+
+	// However the race resolves, the control must end fully stopped with no
+	// panic and Wait must observe the final state
+	require.NoError(t, c.Wait())
+	assert.Equal(t, StateStopped, c.State())
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
+}
+
+func TestControl_StartStopLifecycle(t *testing.T) {
+	c, dev, conn := newReadyControl(t)
+
+	require.NoError(t, c.Start())
+	assert.Equal(t, StateStarted, c.State())
+	require.ErrorIs(t, c.Start(), ErrAlreadyStarted)
+
+	// Stop must unpark the reader blocked in the device and release everything
+	c.Stop()
+	assert.Equal(t, StateStopped, c.State())
+	assert.True(t, dev.closed, "the tun device should have been closed")
+	assert.True(t, conn.closed, "the udp socket should have been closed")
+	require.ErrorIs(t, c.ctx.Err(), context.Canceled)
+
+	// The reader drained off a closed device, that is not a fatal error
+	require.NoError(t, c.Wait())
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
+}
+
+func TestControl_RebindIsGatedByState(t *testing.T) {
+	c, _, conn := newReadyControl(t)
+
+	// A rebind before Start reaches nothing, the interface is not up
+	c.RebindUDPServer()
+	assert.Equal(t, 0, conn.rebinds, "rebind before start must be a no-op")
+
+	require.NoError(t, c.Start())
+	c.RebindUDPServer()
+	assert.Equal(t, 1, conn.rebinds, "rebind while started must reach the conn")
+
+	// A rebind racing a completed stop must not touch the closed conn
+	c.Stop()
+	require.NoError(t, c.Wait())
+	c.RebindUDPServer()
+	assert.Equal(t, 1, conn.rebinds, "rebind after stop must be a no-op")
+}

control_test.go

@@ -1,6 +1,8 @@
 package nebula
 
 import (
+	"bytes"
+	"log/slog"
 	"net"
 	"net/netip"
 	"reflect"
@@ -9,6 +11,7 @@ import (
 	"github.com/slackhq/nebula/cert"
 	"github.com/slackhq/nebula/test"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestControl_GetHostInfoByVpnIp(t *testing.T) {
@@ -42,8 +45,7 @@ func TestControl_GetHostInfoByVpnIp(t *testing.T) {
 	assert.True(t, ok)
 
 	crt := &dummyCert{}
-	hm.unlockedAddHostInfo(&HostInfo{
+	hi := &HostInfo{
-		remote:  remote1,
 		remotes: remotes,
 		ConnectionState: &ConnectionState{
 			peerCert: &cert.CachedCertificate{Certificate: crt},
@@ -56,13 +58,14 @@ func TestControl_GetHostInfoByVpnIp(t *testing.T) {
 			relayForByAddr: map[netip.Addr]*Relay{},
 			relayForByIdx:  map[uint32]*Relay{},
 		},
-	}, &Interface{})
+	}
+	hi.remote.Store(&remote1)
+	hm.unlockedAddHostInfo(hi, &Interface{})
 
 	vpnIp2, ok := netip.AddrFromSlice(ipNet2.IP)
 	assert.True(t, ok)
 
-	hm.unlockedAddHostInfo(&HostInfo{
+	hi2 := &HostInfo{
-		remote:  remote1,
 		remotes: remotes,
 		ConnectionState: &ConnectionState{
 			peerCert: nil,
@@ -75,7 +78,9 @@ func TestControl_GetHostInfoByVpnIp(t *testing.T) {
 			relayForByAddr: map[netip.Addr]*Relay{},
 			relayForByIdx:  map[uint32]*Relay{},
 		},
-	}, &Interface{})
+	}
+	hi2.remote.Store(&remote1)
+	hm.unlockedAddHostInfo(hi2, &Interface{})
 
 	c := Control{
 		state: StateReady,
@@ -119,3 +124,153 @@ func assertFields(t *testing.T, expected []string, actualStruct any) {
 
 	assert.Equal(t, expected, fields)
 }
+
+// alwaysAllowV4/V6 are check funcs that accept every entry (including nil pointers),
+// letting us inject a nil *V4AddrPort/*V6AddrPort into a RemoteList's reported cache
+// the same way a malformed proto message off the wire could.
+func alwaysAllowV4(netip.Addr, *V4AddrPort) bool { return true }
+func alwaysAllowV6(netip.Addr, *V6AddrPort) bool { return true }
+
+// TestGetRelays_SkipsNilRelayAddrs proves GetRelays tolerates nil entries in the
+// RelayVpnAddrs proto slice (which protoAddrToNetAddr would nil-deref on) and still
+// returns the valid relays, including the legacy OldRelayVpnAddrs.
+func TestGetRelays_SkipsNilRelayAddrs(t *testing.T) {
+	good := netip.MustParseAddr("10.0.0.9")
+
+	d := &NebulaMetaDetails{
+		OldRelayVpnAddrs: []uint32{0x0a000001}, // 10.0.0.1
+		RelayVpnAddrs: []*Addr{
+			nil,
+			netAddrToProtoAddr(good),
+			nil,
+		},
+	}
+
+	var relays []netip.Addr
+	require.NotPanics(t, func() { relays = d.GetRelays() })
+
+	assert.Equal(t, []netip.Addr{
+		netip.MustParseAddr("10.0.0.1"),
+		good,
+	}, relays)
+}
+
+// TestGetRelays_AllNil ensures an all-nil RelayVpnAddrs slice yields no relays and no panic.
+func TestGetRelays_AllNil(t *testing.T) {
+	d := &NebulaMetaDetails{RelayVpnAddrs: []*Addr{nil, nil}}
+	var relays []netip.Addr
+	require.NotPanics(t, func() { relays = d.GetRelays() })
+	assert.Empty(t, relays)
+}
+
+// TestRemoteList_CopyCache_SkipsNilReported proves CopyCache skips nil reported
+// pointers (v4 and v6) instead of nil-dereferencing them in protoV*AddrPortToNetAddrPort.
+func TestRemoteList_CopyCache_SkipsNilReported(t *testing.T) {
+	owner := netip.MustParseAddr("10.0.0.1")
+	rl := NewRemoteList([]netip.Addr{owner}, nil)
+
+	rl.unlockedSetV4(owner, owner, []*V4AddrPort{
+		nil,
+		newIp4AndPortFromString("1.2.3.4:5"),
+		nil,
+	}, alwaysAllowV4)
+
+	rl.unlockedSetV6(owner, owner, []*V6AddrPort{
+		nil,
+		newIp6AndPortFromString("[1::1]:6"),
+		nil,
+	}, alwaysAllowV6)
+
+	var cm *CacheMap
+	require.NotPanics(t, func() { cm = rl.CopyCache() })
+
+	c := (*cm)[owner.String()]
+	require.NotNil(t, c)
+	assert.ElementsMatch(t, []netip.AddrPort{
+		netip.MustParseAddrPort("1.2.3.4:5"),
+		netip.MustParseAddrPort("[1::1]:6"),
+	}, c.Reported)
+}
+
+// TestRemoteList_Rebuild_SkipsNilReported drives unlockedCollect (via Rebuild) with
+// nil reported entries and confirms only the valid addresses survive, with no panic.
+func TestRemoteList_Rebuild_SkipsNilReported(t *testing.T) {
+	owner := netip.MustParseAddr("10.0.0.1")
+	rl := NewRemoteList([]netip.Addr{owner}, nil)
+
+	rl.unlockedSetV4(owner, owner, []*V4AddrPort{
+		nil,
+		newIp4AndPortFromString("1.2.3.4:5"),
+	}, alwaysAllowV4)
+	rl.unlockedSetV6(owner, owner, []*V6AddrPort{
+		newIp6AndPortFromString("[1::1]:6"),
+		nil,
+	}, alwaysAllowV6)
+
+	require.NotPanics(t, func() { rl.Rebuild([]netip.Prefix{}) })
+
+	assert.ElementsMatch(t, []netip.AddrPort{
+		netip.MustParseAddrPort("1.2.3.4:5"),
+		netip.MustParseAddrPort("[1::1]:6"),
+	}, rl.addrs)
+}
+
+// newRelayControl marshals a NebulaControl the way it arrives on the wire so we can feed
+// it through HandleControlMsg's unmarshal + validate path.
+func newRelayControl(t *testing.T, typ NebulaControl_MessageType, from, to *Addr) []byte {
+	t.Helper()
+	msg := &NebulaControl{
+		Type:          typ,
+		RelayFromAddr: from,
+		RelayToAddr:   to,
+	}
+	b, err := msg.Marshal()
+	require.NoError(t, err)
+	return b
+}
+
+// TestRelayManager_HandleControlMsg_NilRelayAddrs verifies the validation block added to
+// HandleControlMsg: CreateRelay{Request,Response} carrying a nil RelayFromAddr or
+// RelayToAddr are dropped with a debug log rather than nil-dereferencing downstream.
+func TestRelayManager_HandleControlMsg_NilRelayAddrs(t *testing.T) {
+	good := netAddrToProtoAddr(netip.MustParseAddr("10.0.0.9"))
+
+	cases := []struct {
+		name    string
+		typ     NebulaControl_MessageType
+		from    *Addr
+		to      *Addr
+		wantLog string // debug substring expected, "" == expect no drop log
+	}{
+		{"request nil from", NebulaControl_CreateRelayRequest, nil, good, "nil RelayFromAddr"},
+		{"request nil to", NebulaControl_CreateRelayRequest, good, nil, "nil RelayToAddr"},
+		{"request both nil", NebulaControl_CreateRelayRequest, nil, nil, "nil RelayFromAddr"},
+		{"response nil from", NebulaControl_CreateRelayResponse, nil, good, "nil RelayFromAddr"},
+		{"response nil to", NebulaControl_CreateRelayResponse, good, nil, "nil RelayToAddr"},
+		// A non-relay control type is not subject to the relay-addr validation and must
+		// pass through it untouched (the final switch simply no-ops on it).
+		{"unrelated type nil addrs", NebulaControl_None, nil, nil, ""},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			var buf bytes.Buffer
+			l := test.NewLoggerWithOutputAndLevel(&buf, slog.LevelDebug)
+			rm := &relayManager{l: l, hostmap: newHostMap(l)}
+			rm.useRelays.Store(true)
+
+			f := &Interface{l: l}
+			h := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("10.0.0.2")}, localIndexId: 1}
+
+			d := newRelayControl(t, tc.typ, tc.from, tc.to)
+
+			require.NotPanics(t, func() { rm.HandleControlMsg(h, d, f) })
+
+			if tc.wantLog == "" {
+				assert.NotContains(t, buf.String(), "nil Relay")
+			} else {
+				assert.Contains(t, buf.String(), tc.wantLog)
+			}
+		})
+	}
+}

control_tester.go

@@ -108,7 +108,19 @@ func (c *Control) GetVpnAddrs() []netip.Addr {
 }
 
 func (c *Control) GetUDPAddr() netip.AddrPort {
-	return c.f.outside.(*udp.TesterConn).Addr
+	return c.f.outside.(*udp.TesterConn).GetAddr()
+}
+
+// SetUDPAddr moves this node to a new underlay address, standing in for a laptop waking up on a different
+// network. Register the new address with the router as well or nothing will route back.
+func (c *Control) SetUDPAddr(addr netip.AddrPort) {
+	c.f.outside.(*udp.TesterConn).SetAddr(addr)
+}
+
+// SetLocalAddrsFn replaces underlay address discovery so a test can advertise its simulated address instead of
+// whatever this machine's NICs happen to be. Call it before Start, SendUpdate reads it from the update worker.
+func (c *Control) SetLocalAddrsFn(fn func(*LocalAllowList) []netip.Addr) {
+	c.f.lightHouse.localAddrsFn = fn
 }
 
 func (c *Control) KillPendingTunnel(vpnIp netip.Addr) bool {
@@ -125,6 +137,14 @@ func (c *Control) GetHostmap() *HostMap {
 	return c.f.hostMap
 }
 
+// GetHostmapIndexCount returns the number of entries in the main hostmap Indexes table, holding
+// the hostmap read lock so tests can poll it while connection manager churns tunnels.
+func (c *Control) GetHostmapIndexCount() int {
+	c.f.hostMap.RLock()
+	defer c.f.hostMap.RUnlock()
+	return len(c.f.hostMap.Indexes)
+}
+
 func (c *Control) GetF() *Interface {
 	return c.f
 }

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 = bit32.band(tvbuf:range(0,1):uint(), 0x0F)
+    local nebula_type = bit.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

@@ -11,7 +11,6 @@ import (
 	"sync"
 	"sync/atomic"
 
-	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
 	"github.com/slackhq/nebula/config"
 )
@@ -23,7 +22,10 @@ type dnsServer struct {
 	dnsMap4 map[string]netip.Addr
 	dnsMap6 map[string]netip.Addr
 	hostMap *HostMap
-	myVpnAddrsTable *bart.Lite
+	pki     *PKI
+
+	// selfHost is the cached FQDN we last seeded for ourselves
+	selfHost string
 
 	mux *dns.ServeMux
 
@@ -55,14 +57,14 @@ 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, cs *CertState, hostMap *HostMap, c *config.C) (*dnsServer, error) {
+func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, pki *PKI, 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,
-		myVpnAddrsTable: cs.myVpnAddrsTable,
+		pki:     pki,
 	}
 	ds.mux = dns.NewServeMux()
 	ds.mux.HandleFunc(".", ds.handleDnsRequest)
@@ -76,6 +78,7 @@ func newDnsServerFromConfig(ctx context.Context, l *slog.Logger, cs *CertState,
 	if err := ds.reload(c, true); err != nil {
 		return ds, err
 	}
+	ds.seedSelf()
 	return ds, nil
 }
 
@@ -94,8 +97,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	newAddr := getDnsServerAddr(c)
 
 	d.serverMu.Lock()
-	running := d.server
+	running := d.server != nil
-	runningStarted := d.started
 	sameAddr := d.addr == newAddr
 	d.addr = newAddr
 	d.enabled.Store(enabled)
@@ -109,29 +111,26 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	}
 
 	if !enabled {
-		if running != nil {
+		if running {
 			d.Stop()
 		}
 		// Drop any records that accumulated while enabled; a later re-enable
-		// will repopulate from fresh handshakes.
+		// will repopulate from fresh handshakes and a fresh seedSelf.
 		d.clearRecords()
 		return nil
 	}
 
-	if running == nil {
+	if !running {
 		// Was disabled (or never started); bring it up now.
 		go d.Start()
-		return nil
+	} else if !sameAddr {
-	}
+		// Stop clears the slot before shutting down, otherwise the Start below can find the dying server and refuse
-
+		d.Stop()
-	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()
+	}
+
+	// Refresh the self entry every enabled reload so cert renewals that change our name or VPN addresses are picked up.
+	d.seedSelf()
 	return nil
 }
 
@@ -162,7 +161,9 @@ func (d *dnsServer) Start() {
 
 	started := make(chan struct{})
 	d.serverMu.Lock()
-	if d.ctx.Err() != nil {
+	// Re-check enabled under the lock, a disable that raced our check above snapshots the slot under it too.
+	// Two reloads in quick succession can both spawn a Start, the loser would orphan the live listener past Stop
+	if d.ctx.Err() != nil || d.server != nil || !d.enabled.Load() {
 		d.serverMu.Unlock()
 		return
 	}
@@ -200,6 +201,14 @@ func (d *dnsServer) Start() {
 		close(started)
 	}
 
+	// Release our slot, unless a reload already replaced us, so a dead listener can't block a future Start
+	d.serverMu.Lock()
+	if d.server == server {
+		d.server = nil
+		d.started = nil
+	}
+	d.serverMu.Unlock()
+
 	if err != nil {
 		d.l.Warn("Failed to run the DNS responder", "error", err)
 	}
@@ -249,6 +258,20 @@ 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 ""
@@ -266,12 +289,60 @@ func (d *dnsServer) QueryCert(data string) string {
 	return string(b)
 }
 
-// clearRecords drops all DNS records.
+// clearRecords drops all DNS records, including the self entry.
 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`
@@ -309,8 +380,12 @@ 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 d.myVpnAddrsTable.Contains(b)
+	return cs.myVpnAddrsTable.Contains(b)
 }
 
 func (d *dnsServer) parseQuery(m *dns.Msg, w dns.ResponseWriter) {

dns_server_test.go

@@ -9,7 +9,10 @@ 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"
@@ -191,14 +194,51 @@ func TestDnsServer_reload_initial_serveDnsWithoutLighthouse(t *testing.T) {
 }
 
 func TestDnsServer_reload_sameAddr_noOp(t *testing.T) {
+	port := freeUDPPort(t)
 	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
-
 	require.NoError(t, ds.reload(c, true))
-	// No server running yet, no addr change. Reload should not spawn anything.
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	before := ds.server
+	ds.serverMu.Unlock()
+	require.NotNil(t, before)
+
+	// Same address, so the running listener must be left alone rather than rebuilt under live queries
 	require.NoError(t, ds.reload(c, false))
 	assert.True(t, ds.enabled.Load())
-	assert.Nil(t, ds.server)
+
+	ds.serverMu.Lock()
+	after := ds.server
+	ds.serverMu.Unlock()
+	assert.Same(t, before, after, "a same-address reload must not restart the listener")
+
+	ds.Stop()
+}
+
+// The branch the old sameAddr test was accidentally hitting: enabled with nothing running means reload starts it.
+func TestDnsServer_reload_whenNotRunning_starts(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+
+	// initial only records config, it never starts anything
+	require.NoError(t, ds.reload(c, true))
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "the initial reload must not start a listener")
+	ds.serverMu.Unlock()
+
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	assert.NotNil(t, ds.server, "a reload with nothing running should bring DNS up")
+	ds.serverMu.Unlock()
+
+	ds.Stop()
 }
 
 func TestDnsServer_StartStop_lifecycle(t *testing.T) {
@@ -276,6 +316,92 @@ 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)
@@ -338,3 +464,168 @@ func waitFor(t *testing.T, cond func() bool) {
 	}
 	t.Fatal("timed out waiting for condition")
 }
+
+// Two reloads in quick succession, or a HUP before Control.Start, can race two Starts at the same listener.
+func TestDnsServer_Start_isIdempotent(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	first := ds.server
+	ds.serverMu.Unlock()
+	require.NotNil(t, first)
+
+	// If the second Start replaces the tracked server, Stop kills the wrong one and the port leaks
+	done := make(chan struct{})
+	go func() {
+		ds.Start()
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(time.Second * 5):
+		t.Fatal("second Start never returned")
+	}
+
+	ds.serverMu.Lock()
+	second := ds.server
+	ds.serverMu.Unlock()
+	assert.Same(t, first, second, "a second Start must not replace the running server")
+
+	// The real proof, after Stop the port must actually be free
+	ds.Stop()
+	waitFor(t, func() bool {
+		pc, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+		if err != nil {
+			return false
+		}
+		_ = pc.Close()
+		return true
+	})
+}
+
+// An address change must actually end up listening on the new port. Start's guard refuses when a server is already
+// installed, so reload has to clear the slot before shutting the old one down.
+func TestDnsServer_reload_addrChange_restarts(t *testing.T) {
+	first := freeUDPPort(t)
+	second := freeUDPPort(t)
+
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", first, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	// Cycle a few times, the failure this guards against depends on which goroutine wins serverMu
+	for i := range 8 {
+		want := second
+		if i%2 == 1 {
+			want = first
+		}
+		setDnsConfig(c, "127.0.0.1", want, true, true)
+		require.NoError(t, ds.reload(c, false))
+		waitForBind(t, ds)
+
+		ds.serverMu.Lock()
+		srv := ds.server
+		ds.serverMu.Unlock()
+		require.NotNil(t, srv, "reload left DNS down instead of restarting it")
+		require.Equal(t, "127.0.0.1:"+want, srv.Addr, "reload should be serving the new address")
+	}
+
+	// Land back on second so the port assertions below are meaningful
+	setDnsConfig(c, "127.0.0.1", second, true, true)
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	// The old port must be released and the new one actually held
+	waitFor(t, func() bool {
+		pc, err := net.ListenPacket("udp", "127.0.0.1:"+first)
+		if err != nil {
+			return false
+		}
+		_ = pc.Close()
+		return true
+	})
+	_, err := net.ListenPacket("udp", "127.0.0.1:"+second)
+	require.Error(t, err, "the new address should be bound by the DNS responder")
+
+	ds.Stop()
+}
+
+// A listener that dies on its own must release the slot, or a later same-addr reload sees it as running and no-ops.
+func TestDnsServer_Start_bindFailure_releasesSlot(t *testing.T) {
+	port := freeUDPPort(t)
+	blocker, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+	require.NoError(t, err)
+
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	ds.Start() // returns once the bind fails
+
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "a listener that failed to bind must not stay parked in the slot")
+	ds.serverMu.Unlock()
+
+	// With the slot released, a reload can retry once the port frees up
+	require.NoError(t, blocker.Close())
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	assert.NotNil(t, ds.server, "a same-addr reload should retry after a failed bind")
+	ds.serverMu.Unlock()
+
+	ds.Stop()
+}
+
+// A disable that lands while Start is between its unlocked check and the guard must not leave a listener behind.
+func TestDnsServer_Start_refusesWhenDisabledUnderLock(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+	require.True(t, ds.enabled.Load())
+
+	// Holding serverMu parks Start on the lock, the only way to land the disable in that window on purpose
+	ds.serverMu.Lock()
+
+	done := make(chan struct{})
+	go func() {
+		ds.Start()
+		close(done)
+	}()
+
+	select {
+	case <-done:
+		ds.serverMu.Unlock()
+		t.Fatal("Start returned early, the test never exercised the window")
+	case <-time.After(time.Millisecond * 100):
+	}
+
+	// The disable reload's critical section. It sees nothing running, so it never calls Stop.
+	ds.enabled.Store(false)
+	ds.serverMu.Unlock()
+
+	select {
+	case <-done:
+	case <-time.After(time.Second * 5):
+		t.Fatal("Start never returned")
+	}
+
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "Start must not install a listener a disable already cancelled")
+	ds.serverMu.Unlock()
+
+	pc, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+	require.NoError(t, err, "an orphaned listener is still holding the port")
+	_ = pc.Close()
+}

docker/Dockerfile

@@ -1,6 +1,16 @@
 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/echo_test.go

@@ -0,0 +1,85 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func assertTestRequestEchoed(t *testing.T, cipher string) {
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	over := m{"cipher": cipher}
+	a, aNet, aUdp, _ := newSimpleServer(cert.Version1, ca, caKey, "a", "10.128.0.1/24", over)
+	b, bNet, bUdp, _ := newSimpleServer(cert.Version1, ca, caKey, "b", "10.128.0.2/24", over)
+
+	a.InjectLightHouseAddr(bNet[0].Addr(), bUdp)
+	b.InjectLightHouseAddr(aNet[0].Addr(), aUdp)
+	a.Start()
+	b.Start()
+	t.Cleanup(func() { a.Stop(); b.Stop() })
+	r := router.NewR(t, a, b)
+	defer r.RenderFlow()
+
+	assertTunnel(t, aNet[0].Addr(), bNet[0].Addr(), a, b, r)
+	drainUDPTx(a)
+	drainUDPTx(b)
+
+	payload := []byte("a test payload well over sixteen bytes long, wow it's so very long long long!")
+	require.Greater(t, len(payload), header.Len)
+	a.GetF().SendMessageToVpnAddr(header.Test, header.TestRequest, bNet[0].Addr(), payload, make([]byte, 12, 12), make([]byte, udp.MTU))
+
+	// Deliver A's request to B; B must echo a reply back
+	b.InjectUDPPacket(a.GetFromUDP(true))
+	reply := nextUDPTxOfType(t, b, header.Test, header.TestReply, 2*time.Second)
+
+	assert.Equal(t, aUdp, reply.To, "the reply must go back to the requester")
+	// header + echoed payload + 16-byte AEAD tag: proves the whole payload
+	// round-tripped rather than being dropped or truncated.
+	assert.Equal(t, header.Len+len(payload)+16, len(reply.Data), "the full payload must be echoed back")
+}
+
+func TestTestRequestEchoesLongPayloadAES(t *testing.T) {
+	assertTestRequestEchoed(t, "aes")
+}
+
+func TestTestRequestEchoesLongPayloadChaChaPoly(t *testing.T) {
+	assertTestRequestEchoed(t, "chachapoly")
+}
+
+// drainUDPTx empties a control's UDP tx queue without blocking.
+func drainUDPTx(c *nebula.Control) {
+	for c.GetFromUDP(false) != nil {
+	}
+}
+
+// nextUDPTxOfType returns the next packet a control transmits whose nebula
+// header matches (wantType, wantSub), skipping unrelated packets.
+// It fails the test if none arrives within the timeout.
+func nextUDPTxOfType(t *testing.T, c *nebula.Control, wantType header.MessageType, wantSub header.MessageSubType, within time.Duration) *udp.Packet {
+	t.Helper()
+	ch := c.GetUDPTxChan()
+	timeout := time.After(within)
+	for {
+		select {
+		case p := <-ch:
+			var h header.H
+			if err := h.Parse(p.Data); err == nil && h.Type == wantType && h.Subtype == wantSub {
+				return p
+			}
+		case <-timeout:
+			t.Fatalf("timed out waiting for a %v/%v packet on the udp tx queue", wantType, wantSub)
+			return nil
+		}
+	}
+}

e2e/handshakes_test.go

@@ -405,7 +405,7 @@ func TestStage1Race(t *testing.T) {
 
 	r.Log("Spin until connection manager tears down a tunnel")
 
-	for len(myControl.GetHostmap().Indexes)+len(theirControl.GetHostmap().Indexes) > 2 {
+	for myControl.GetHostmapIndexCount()+theirControl.GetHostmapIndexCount() > 2 {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
 		time.Sleep(time.Second)
@@ -453,9 +453,11 @@ func TestUncleanShutdownRaceLoser(t *testing.T) {
 
 	r.Log("Nuke my hostmap")
 	myHostmap := myControl.GetHostmap()
+	myHostmap.Lock()
 	myHostmap.Hosts = map[netip.Addr]*nebula.HostInfo{}
 	myHostmap.Indexes = map[uint32]*nebula.HostInfo{}
 	myHostmap.RemoteIndexes = map[uint32]*nebula.HostInfo{}
+	myHostmap.Unlock()
 
 	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me again")))
 	p = r.RouteForAllUntilTxTun(theirControl)
@@ -465,10 +467,10 @@ func TestUncleanShutdownRaceLoser(t *testing.T) {
 	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 
 	r.Log("Wait for the dead index to go away")
-	start := len(theirControl.GetHostmap().Indexes)
+	start := theirControl.GetHostmapIndexCount()
 	for {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		if len(theirControl.GetHostmap().Indexes) < start {
+		if theirControl.GetHostmapIndexCount() < start {
 			break
 		}
 		time.Sleep(time.Second)
@@ -504,9 +506,11 @@ func TestUncleanShutdownRaceWinner(t *testing.T) {
 
 	r.Log("Nuke my hostmap")
 	theirHostmap := theirControl.GetHostmap()
+	theirHostmap.Lock()
 	theirHostmap.Hosts = map[netip.Addr]*nebula.HostInfo{}
 	theirHostmap.Indexes = map[uint32]*nebula.HostInfo{}
 	theirHostmap.RemoteIndexes = map[uint32]*nebula.HostInfo{}
+	theirHostmap.Unlock()
 
 	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirVpnIpNet[0].Addr(), 80, []byte("Hi from them again")))
 	p = r.RouteForAllUntilTxTun(myControl)
@@ -517,10 +521,10 @@ func TestUncleanShutdownRaceWinner(t *testing.T) {
 	assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 
 	r.Log("Wait for the dead index to go away")
-	start := len(myControl.GetHostmap().Indexes)
+	start := myControl.GetHostmapIndexCount()
 	for {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
-		if len(myControl.GetHostmap().Indexes) < start {
+		if myControl.GetHostmapIndexCount() < start {
 			break
 		}
 		time.Sleep(time.Second)
@@ -628,10 +632,10 @@ func TestReestablishRelays(t *testing.T) {
 	r.Log("Close the tunnel")
 	relayControl.CloseTunnel(theirVpnIpNet[0].Addr(), true)
 
-	start := len(myControl.GetHostmap().Indexes)
+	start := myControl.GetHostmapIndexCount()
-	curIndexes := len(myControl.GetHostmap().Indexes)
+	curIndexes := myControl.GetHostmapIndexCount()
 	for curIndexes >= start {
-		curIndexes = len(myControl.GetHostmap().Indexes)
+		curIndexes = myControl.GetHostmapIndexCount()
 		r.Logf("Wait for the dead index to go away:start=%v indexes, current=%v indexes", start, curIndexes)
 		myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me should fail")))
 
@@ -721,6 +725,70 @@ func TestReestablishRelays(t *testing.T) {
 
 }
 
+func TestRelayHandshakeOverDisestablishedEntry(t *testing.T) {
+	t.Parallel()
+	// If them tears down the tunnel while me keeps Established relay state, me's next
+	// handshake flows through the relay with no fresh CreateRelayRequest and lands on
+	// them's Disestablished terminal relay entry. them must re-establish that entry, or
+	// its first transmit deletes its only relay and the tunnel is born transmit-dead:
+	// them can receive but every send is silently dropped.
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
+	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "relay  ", "10.128.0.128/24", m{"relay": m{"am_relay": true}})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them   ", "10.128.0.2/24", m{"relay": m{"use_relays": true}})
+
+	// Teach my how to get to the relay and that their can be reached via the relay
+	myControl.InjectLightHouseAddr(relayVpnIpNet[0].Addr(), relayUdpAddr)
+	myControl.InjectRelays(theirVpnIpNet[0].Addr(), []netip.Addr{relayVpnIpNet[0].Addr()})
+	relayControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	// Build a router so we don't have to reason who gets which packet
+	r := router.NewR(t, myControl, relayControl, theirControl)
+	defer r.RenderFlow()
+
+	// Start the servers
+	myControl.Start()
+	relayControl.Start()
+	theirControl.Start()
+
+	t.Log("Trigger a handshake from me to them via the relay")
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me")))
+
+	p := r.RouteForAllUntilTxTun(theirControl)
+	assertUdpPacket(t, []byte("Hi from me"), p, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), 80, 80)
+	oldIdx := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false).LocalIndex
+
+	t.Log("Close the tunnel on them only, marking their relay entry Disestablished")
+	theirControl.CloseTunnel(myVpnIpNet[0].Addr(), true)
+
+	t.Log("Re-handshake from me, riding the still-Established relay state")
+	myControl.ReHandshake(theirVpnIpNet[0].Addr())
+	for {
+		h := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
+		if h != nil && h.LocalIndex != oldIdx && h.RemoteIndex != 0 {
+			break
+		}
+		r.RouteForAllExitFunc(func(*udp.Packet, *nebula.Control) router.ExitType {
+			return router.RouteAndExit
+		})
+	}
+
+	hAtThem := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	require.NotNil(t, hAtThem, "them should have completed the relayed handshake")
+	require.Equal(t, []netip.Addr{relayVpnIpNet[0].Addr()}, hAtThem.CurrentRelaysToMe, "them should know a relay for the new tunnel")
+
+	t.Log("Send from them to me; their only relay entry must survive the transmit")
+	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirVpnIpNet[0].Addr(), 80, []byte("Hi from them")))
+	require.Never(t, func() bool {
+		h := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+		return h == nil || len(h.CurrentRelaysToMe) == 0
+	}, time.Second, 10*time.Millisecond, "them deleted its only relay entry; the tunnel is permanently transmit-dead")
+
+	p = r.RouteForAllUntilTxTun(myControl)
+	assertUdpPacket(t, []byte("Hi from them"), p, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), 80, 80)
+	r.RenderHostmaps("Final hostmaps", myControl, relayControl, theirControl)
+}
+
 func TestStage1RaceRelays(t *testing.T) {
 	t.Parallel()
 	//NOTE: this is a race between me and relay resulting in a full tunnel from me to them via relay
@@ -819,18 +887,18 @@ func TestStage1RaceRelays2(t *testing.T) {
 
 	t.Log("Wait until we remove extra tunnels")
 	t.Logf("Waiting for hostinfos to be removed... myControl=%d theirControl=%d relayControl=%d",
-		len(myControl.GetHostmap().Indexes),
+		myControl.GetHostmapIndexCount(),
-		len(theirControl.GetHostmap().Indexes),
+		theirControl.GetHostmapIndexCount(),
-		len(relayControl.GetHostmap().Indexes),
+		relayControl.GetHostmapIndexCount(),
 	)
-	hostInfos := len(myControl.GetHostmap().Indexes) + len(theirControl.GetHostmap().Indexes) + len(relayControl.GetHostmap().Indexes)
+	hostInfos := myControl.GetHostmapIndexCount() + theirControl.GetHostmapIndexCount() + relayControl.GetHostmapIndexCount()
 	retries := 60
 	for hostInfos > 6 && retries > 0 {
-		hostInfos = len(myControl.GetHostmap().Indexes) + len(theirControl.GetHostmap().Indexes) + len(relayControl.GetHostmap().Indexes)
+		hostInfos = myControl.GetHostmapIndexCount() + theirControl.GetHostmapIndexCount() + relayControl.GetHostmapIndexCount()
 		t.Logf("Waiting for hostinfos to be removed... myControl=%d theirControl=%d relayControl=%d",
-			len(myControl.GetHostmap().Indexes),
+			myControl.GetHostmapIndexCount(),
-			len(theirControl.GetHostmap().Indexes),
+			theirControl.GetHostmapIndexCount(),
-			len(relayControl.GetHostmap().Indexes),
+			relayControl.GetHostmapIndexCount(),
 		)
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
@@ -924,24 +992,24 @@ func TestRehandshakingRelays(t *testing.T) {
 	assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 	r.RenderHostmaps("working hostmaps", myControl, relayControl, theirControl)
 	// We should have two hostinfos on all sides
-	for len(myControl.GetHostmap().Indexes) != 2 {
+	for myControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(myControl.GetHostmap().Indexes))
+		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", myControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("myControl hostinfos got cleaned up!")
-	for len(theirControl.GetHostmap().Indexes) != 2 {
+	for theirControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(theirControl.GetHostmap().Indexes))
+		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", theirControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("theirControl hostinfos got cleaned up!")
-	for len(relayControl.GetHostmap().Indexes) != 2 {
+	for relayControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(relayControl.GetHostmap().Indexes))
+		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", relayControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
@@ -1029,24 +1097,24 @@ func TestRehandshakingRelaysPrimary(t *testing.T) {
 	assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 	r.RenderHostmaps("working hostmaps", myControl, relayControl, theirControl)
 	// We should have two hostinfos on all sides
-	for len(myControl.GetHostmap().Indexes) != 2 {
+	for myControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(myControl.GetHostmap().Indexes))
+		t.Logf("Waiting for myControl hostinfos (%v != 2) to get cleaned up from lack of use...", myControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("myControl hostinfos got cleaned up!")
-	for len(theirControl.GetHostmap().Indexes) != 2 {
+	for theirControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(theirControl.GetHostmap().Indexes))
+		t.Logf("Waiting for theirControl hostinfos (%v != 2) to get cleaned up from lack of use...", theirControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
 		time.Sleep(time.Second)
 	}
 	t.Logf("theirControl hostinfos got cleaned up!")
-	for len(relayControl.GetHostmap().Indexes) != 2 {
+	for relayControl.GetHostmapIndexCount() != 2 {
-		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", len(relayControl.GetHostmap().Indexes))
+		t.Logf("Waiting for relayControl hostinfos (%v != 2) to get cleaned up from lack of use...", relayControl.GetHostmapIndexCount())
 		r.Log("Assert the relay tunnel still works")
 		assertTunnel(t, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), theirControl, myControl, r)
 		r.Log("yupitdoes")
@@ -1123,7 +1191,7 @@ func TestRehandshaking(t *testing.T) {
 	theirConfig.ReloadConfigString(string(rc))
 
 	r.Log("Spin until there is only 1 tunnel")
-	for len(myControl.GetHostmap().Indexes)+len(theirControl.GetHostmap().Indexes) > 2 {
+	for myControl.GetHostmapIndexCount()+theirControl.GetHostmapIndexCount() > 2 {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
 		time.Sleep(time.Second)
@@ -1223,7 +1291,7 @@ func TestRehandshakingLoser(t *testing.T) {
 	myConfig.ReloadConfigString(string(rc))
 
 	r.Log("Spin until there is only 1 tunnel")
-	for len(myControl.GetHostmap().Indexes)+len(theirControl.GetHostmap().Indexes) > 2 {
+	for myControl.GetHostmapIndexCount()+theirControl.GetHostmapIndexCount() > 2 {
 		assertTunnel(t, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), myControl, theirControl, r)
 		t.Log("Connection manager hasn't ticked yet")
 		time.Sleep(time.Second)
@@ -1535,3 +1603,78 @@ func TestGoodHandshakeUnsafeDest(t *testing.T) {
 	myControl.Stop()
 	theirControl.Stop()
 }
+
+func TestMultiVpnAddrDeletePrimaryKeepsSecondAddr(t *testing.T) {
+	t.Parallel()
+	// Regression for the hostmap multi-vpnAddr delete bug. A dual-stack (v4+v6) V2-cert peer that
+	// handshakes twice at once ends up with two hostinfos linked in the shared next/prev chain, with the
+	// primary owning both addresses. Deleting that primary (e.g. connection manager dropping it, a
+	// CloseTunnel, a collision) must promote the surviving sibling for EVERY address. The pre-fix code
+	// unlinked the chain once per address, so it promoted the sibling for the first address and orphaned
+	// the second: the peer stayed reachable at its v4 addr but not its v6 addr despite a live tunnel.
+
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me  ", "10.128.0.1/24,fd00::1/64", nil)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "10.128.0.2/24,fd00::2/64", nil)
+
+	// This bug only exists for peers carrying more than one vpn address
+	require.Len(t, theirVpnIpNet, 2)
+	theirV4 := theirVpnIpNet[0].Addr()
+	theirV6 := theirVpnIpNet[1].Addr()
+
+	// Put their info in our lighthouse and vice versa
+	myControl.InjectLightHouseAddr(theirV4, theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	// Build a router so we don't have to reason who gets which packet
+	r := router.NewR(t, myControl, theirControl)
+	defer r.RenderFlow()
+
+	myControl.Start()
+	theirControl.Start()
+
+	// Race a handshake so both of us build a hostinfo for the other, leaving my hostmap with a single
+	// host (them) backed by two linked hostinfos, just like TestStage1Race.
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirV4, 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me")))
+	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirV4, 80, []byte("Hi from them")))
+
+	myHsForThem := myControl.GetFromUDP(true)
+	theirHsForMe := theirControl.GetFromUDP(true)
+
+	r.InjectUDPPacket(theirControl, myControl, theirHsForMe)
+	r.InjectUDPPacket(myControl, theirControl, myHsForThem)
+
+	r.RouteForAllUntilTxTun(theirControl)
+	r.RouteForAllUntilTxTun(myControl)
+
+	r.RenderHostmaps("Racing hostmaps", myControl, theirControl)
+
+	// Two hostinfos for them means the shared next/prev chain has a sibling to promote. The Hosts map has
+	// one entry per vpn address (two, for dual stack), so the index count is what tells us there are two
+	// hostinfos.
+	require.Len(t, myControl.ListHostmapIndexes(false), 2)
+
+	// The primary owns both of their addresses
+	primaryV4 := myControl.GetHostInfoByVpnAddr(theirV4, false)
+	primaryV6 := myControl.GetHostInfoByVpnAddr(theirV6, false)
+	require.NotNil(t, primaryV4)
+	require.NotNil(t, primaryV6)
+	require.Equal(t, primaryV4.LocalIndex, primaryV6.LocalIndex, "both addrs should point at the same primary")
+
+	// Delete the primary tunnel. localOnly so we don't perturb their side, we only care about my hostmap.
+	require.True(t, myControl.CloseTunnel(theirV4, true))
+
+	// The surviving sibling must still serve BOTH addresses.
+	survivorV4 := myControl.GetHostInfoByVpnAddr(theirV4, false)
+	survivorV6 := myControl.GetHostInfoByVpnAddr(theirV6, false)
+	require.NotNil(t, survivorV4, "v4 addr should still resolve to the surviving tunnel")
+	// Pre-fix this is nil: the second address was orphaned when the primary was deleted.
+	require.NotNil(t, survivorV6, "v6 addr was orphaned after deleting the primary (multi-vpnAddr delete bug)")
+	assert.Equal(t, survivorV4.LocalIndex, survivorV6.LocalIndex, "both addrs should promote to the same survivor")
+	assert.NotEqual(t, primaryV4.LocalIndex, survivorV4.LocalIndex, "a different hostinfo should now be primary")
+
+	r.RenderHostmaps("Final hostmaps", myControl, theirControl)
+
+	myControl.Stop()
+	theirControl.Stop()
+}

e2e/helpers_test.go

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

e2e/rebind_test.go

@@ -0,0 +1,225 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// reportedAddrs is what the lighthouse would hand a peer asking where vpnAddr is.
+func reportedAddrs(t *testing.T, lh *nebula.Control, vpnAddr netip.Addr) []netip.AddrPort {
+	t.Helper()
+	cm := lh.QueryLighthouse(vpnAddr)
+	if cm == nil {
+		return nil
+	}
+	var out []netip.AddrPort
+	for _, c := range *cm {
+		out = append(out, c.Reported...)
+		out = append(out, c.Learned...)
+	}
+	return out
+}
+
+// waitForLighthouseMsg routes until a lighthouse message lands on lh, or gives up. Reports whether one arrived.
+func waitForLighthouseMsg(t *testing.T, r *router.R, lh *nebula.Control, wait time.Duration) bool {
+	t.Helper()
+	h := &header.H{}
+	return r.RouteForAllExitFuncOrTimeout(wait, func(p *udp.Packet, c *nebula.Control) router.ExitType {
+		if c != lh {
+			return router.KeepRouting
+		}
+		// Punches are a single byte and never parse, they are just not what we are after
+		if err := h.Parse(p.Data); err != nil {
+			return router.KeepRouting
+		}
+		if h.Type == header.LightHouse {
+			return router.RouteAndExit
+		}
+		return router.KeepRouting
+	})
+}
+
+// A laptop that changes networks has to tell the lighthouse promptly, otherwise the lighthouse keeps handing peers
+// the old address and their punches land nowhere. On a long lighthouse interval the only thing that closes that
+// window is the rebind, which on darwin the network change monitor drives. The e2e build compiles the monitor out,
+// so we call RebindUDPServer directly, which is the same thing the monitor does.
+func TestRebindSendsLighthouseUpdate(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{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	// 600s interval, so nothing scheduled can send an update during this test. A rebind is the only thing that can.
+	myControl, _, _, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	})
+
+	r := router.NewR(t, lhControl, myControl)
+	defer r.RenderFlow()
+
+	lhControl.Start()
+	myControl.Start()
+
+	// Let the startup registration finish, then clear everything it left behind
+	require.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5), "expected an initial registration")
+	r.RouteFor(time.Millisecond * 400)
+
+	// Nothing should be talking to the lighthouse on its own now
+	require.False(t, waitForLighthouseMsg(t, r, lhControl, time.Millisecond*200),
+		"nothing should reach the lighthouse before the rebind")
+
+	myControl.RebindUDPServer()
+
+	assert.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5),
+		"a rebind should push an update to the lighthouse rather than waiting out the interval")
+
+	lhControl.Stop()
+	myControl.Stop()
+}
+
+// The other half of a rebind: every live tunnel requeries the lighthouse on its next send. That query is what makes
+// the lighthouse tell the peer to punch toward our new address, which is the part that actually revives a tunnel
+// whose remote NAT state died while we were on a different network.
+func TestRebindRequeriesPeersOnNextSend(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{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	lhCfg := m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+			// Without this the peers advertise this machine's real addresses and then try to punch at them,
+			// which the router has no route for.
+			"local_allow_list": m{
+				"10.0.0.0/24": true,
+				"::/0":        false,
+			},
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	}
+
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", lhCfg)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "10.128.0.3/24", lhCfg)
+
+	r := router.NewR(t, lhControl, myControl, theirControl)
+	defer r.RenderFlow()
+
+	lhControl.Start()
+	myControl.Start()
+	theirControl.Start()
+	r.RouteFor(time.Millisecond * 500)
+
+	// Point the peers at each other directly, this test is about the rebind and not about lighthouse discovery
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("initial")))
+	r.RouteFor(time.Second)
+	require.NotNil(t, myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false), "expected a tunnel to them")
+	r.RouteFor(time.Millisecond * 300)
+
+	// Assert on what the peer sees rather than on lighthouse traffic. A query for them makes the lighthouse send
+	// them a punch notification, which is the whole point. Our own update to the lighthouse sends them nothing,
+	// so this cannot be satisfied by the update the rebind itself pushes.
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("quiet")))
+	require.False(t, waitForLighthouseMsg(t, r, theirControl, time.Millisecond*300),
+		"an ordinary send should not requery the lighthouse")
+
+	myControl.RebindUDPServer()
+	r.RouteFor(time.Millisecond * 300) // let the update the rebind itself sends pass by
+
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("after rebind")))
+	assert.True(t, waitForLighthouseMsg(t, r, theirControl, time.Second*5),
+		"the first send after a rebind should requery the lighthouse, which then tells the peer to punch at us")
+
+	lhControl.Stop()
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+// The scenario this whole thing exists for: a laptop sleeps at the office and wakes up at home on a new address.
+// Until it tells the lighthouse, the lighthouse keeps handing peers the office address, so their punches land
+// nowhere and the tunnel stays dead. On a long interval the rebind is the only thing that closes that window.
+func TestRebindAdvertisesNewAddressAfterMove(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{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	})
+
+	// Advertise wherever we currently are rather than this machine's real NICs, read fresh each time so a move
+	// is picked up.
+	myControl.SetLocalAddrsFn(func(*nebula.LocalAllowList) []netip.Addr {
+		return []netip.Addr{myControl.GetUDPAddr().Addr()}
+	})
+
+	r := router.NewR(t, lhControl, myControl)
+	defer r.RenderFlow()
+
+	lhControl.Start()
+	myControl.Start()
+
+	require.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5), "expected an initial registration")
+	r.RouteFor(time.Millisecond * 400)
+
+	require.Contains(t, reportedAddrs(t, lhControl, myVpnIpNet[0].Addr()), myUdpAddr,
+		"the lighthouse should know the address we started on")
+
+	// Wake up somewhere else
+	newAddr := netip.MustParseAddrPort("10.0.0.99:4242")
+	myControl.SetUDPAddr(newAddr)
+	r.AddRoute(newAddr.Addr(), newAddr.Port(), myControl)
+
+	// Nothing has told the lighthouse, and with interval 600 nothing scheduled will
+	r.RouteFor(time.Millisecond * 400)
+	require.NotContains(t, reportedAddrs(t, lhControl, myVpnIpNet[0].Addr()), newAddr,
+		"the lighthouse should still be handing out the old address before the rebind")
+
+	myControl.RebindUDPServer()
+	require.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5), "expected an update after the rebind")
+	r.RouteFor(time.Millisecond * 400)
+
+	assert.Contains(t, reportedAddrs(t, lhControl, myVpnIpNet[0].Addr()), newAddr,
+		"after the rebind the lighthouse should hand peers our new address")
+
+	lhControl.Stop()
+	myControl.Stop()
+}

e2e/recovery_timing_test.go

@@ -0,0 +1,136 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/udp"
+)
+
+// TestRecoveryTiming measures how long a tunnel takes to come back after the peer stops accepting our traffic,
+// which is what a laptop waking on a new network looks like from the peer's side: its NAT has no state for where
+// we are now, so everything we send disappears.
+//
+// It is a measurement, not a pass/fail assertion. Recovery is timed to the moment the peer punches back at us,
+// since that is when its NAT opens and the tunnel is usable again.
+//
+//	go test -tags e2e_testing -v -run TestRecoveryTiming ./e2e/
+func TestRecoveryTiming(t *testing.T) {
+	for _, tc := range []struct {
+		name   string
+		rebind bool
+	}{
+		{"no trigger", false},
+		{"rebind counter", true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			d, lost := measureRecovery(t, tc.rebind)
+			t.Logf("RESULT %-16s recovered in %-9v (%d packets lost)", tc.name, d.Round(time.Millisecond), lost)
+		})
+	}
+}
+
+// measureRecovery returns how long until the peer punched back, and how many of our packets died meanwhile. When
+// rebind is true we call RebindUDPServer once the tunnel goes dark, which is what the darwin network change
+// monitor does and what iOS has always done. When false, nothing tells nebula anything is wrong.
+func measureRecovery(t *testing.T, rebind bool) (time.Duration, int) {
+	t.Helper()
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	peerCfg := m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+			"local_allow_list": m{
+				"10.0.0.0/24": true,
+				"::/0":        false,
+			},
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	}
+
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", peerCfg)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "10.128.0.3/24", peerCfg)
+
+	r := router.NewR(t, lhControl, myControl, theirControl)
+	defer r.RenderFlow()
+	defer func() {
+		lhControl.Stop()
+		myControl.Stop()
+		theirControl.Stop()
+	}()
+
+	lhControl.Start()
+	myControl.Start()
+	theirControl.Start()
+	r.RouteFor(time.Millisecond * 500)
+
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("establish")))
+	r.RouteFor(time.Second)
+	if myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false) == nil {
+		t.Fatal("failed to establish the tunnel we are measuring")
+	}
+	r.RouteFor(time.Millisecond * 500)
+
+	// From here the peer's NAT has no state for us, everything we send it disappears
+	start := time.Now()
+	blackholed := 0
+	var recovered time.Duration
+
+	if rebind {
+		myControl.RebindUDPServer()
+	}
+
+	// Keep the tun busy the way someone retrying a stalled connection would
+	stop := make(chan struct{})
+	defer close(stop)
+	go func() {
+		tick := time.NewTicker(time.Millisecond * 200)
+		defer tick.Stop()
+		for {
+			select {
+			case <-stop:
+				return
+			case <-tick.C:
+				myControl.InjectTunPacket(BuildTunUDPPacket(
+					theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("retry")))
+			}
+		}
+	}()
+
+	r.RouteForAllExitFuncOrTimeout(time.Second*30, func(p *udp.Packet, c *nebula.Control) router.ExitType {
+		if c == theirControl && p.From == myControl.GetUDPAddr() {
+			blackholed++
+			return router.Drop
+		}
+
+		// The peer reaching us directly is the moment its NAT opened, whether that is a punch or a handshake
+		if c == myControl && p.From == theirUdpAddr {
+			recovered = time.Since(start)
+			return router.RouteAndExit
+		}
+
+		return router.KeepRouting
+	})
+
+	if recovered == 0 {
+		t.Fatalf("no recovery within 30s (%d packets blackholed)", blackholed)
+	}
+	return recovered, blackholed
+}

e2e/router/router.go

@@ -114,6 +114,28 @@ type packet struct {
 	packet *udp.Packet
 	tun    bool // a packet pulled off a tun device
 	rx     bool // the packet was received by a udp device
+
+	// h is the nebula header, parsed once when the packet is recorded. parseErr says why there isn't one, which
+	// the flow log reports rather than hiding. Punchy sends a single byte, so an unparseable packet is normal.
+	h        header.H
+	parseErr error
+}
+
+// fromAddr and toAddr are the addresses this packet actually travelled between. Reading them off the control
+// instead would misreport the whole history once a test moves a node. Tun packets are synthesized without
+// addresses, so they fall back to the control.
+func (p *packet) fromAddr() netip.AddrPort {
+	if p.tun || !p.packet.From.IsValid() {
+		return p.from.GetUDPAddr()
+	}
+	return p.packet.From
+}
+
+func (p *packet) toAddr() netip.AddrPort {
+	if p.tun || !p.packet.To.IsValid() {
+		return p.to.GetUDPAddr()
+	}
+	return p.packet.To
 }
 
 func (p *packet) WasReceived() {
@@ -131,6 +153,9 @@ const (
 	ExitNow ExitType = 1
 	// RouteAndExit routes this packet and exits immediately afterwards
 	RouteAndExit ExitType = 2
+	// Drop discards this packet without delivering it and keeps routing. Use it to simulate a blackhole, such as
+	// a restrictive NAT refusing traffic from an address it has not seen.
+	Drop ExitType = 3
 )
 
 type ExitFunc func(packet *udp.Packet, receiver *nebula.Control) ExitType
@@ -141,7 +166,9 @@ type ExitFunc func(packet *udp.Packet, receiver *nebula.Control) ExitType
 func NewR(t testing.TB, controls ...*nebula.Control) *R {
 	ctx, cancel := context.WithCancel(context.Background())
 
-	if err := os.MkdirAll("mermaid", 0755); err != nil {
+	// t.Name() contains a slash for subtests, so the flow log can land in a nested directory
+	fn := filepath.Join("mermaid", fmt.Sprintf("%s.md", t.Name()))
+	if err := os.MkdirAll(filepath.Dir(fn), 0755); err != nil {
 		panic(err)
 	}
 
@@ -152,7 +179,7 @@ func NewR(t testing.TB, controls ...*nebula.Control) *R {
 		outNat:       make(map[outNatKey]netip.AddrPort),
 		flow:         []flowEntry{},
 		ignoreFlows:  []ignoreFlow{},
-		fn:           filepath.Join("mermaid", fmt.Sprintf("%s.md", t.Name())),
+		fn:           fn,
 		t:            t,
 		cancelRender: cancel,
 	}
@@ -249,7 +276,7 @@ func (r *R) renderFlow() {
 			continue
 		}
 
-		addr := e.packet.from.GetUDPAddr()
+		addr := e.packet.fromAddr()
 		if _, ok := participants[addr]; ok {
 			continue
 		}
@@ -268,7 +295,6 @@ func (r *R) renderFlow() {
 	}
 
 	// Print packets
-	h := &header.H{}
 	for _, e := range r.flow {
 		if e.packet == nil {
 			//fmt.Fprintf(f, "    note over %s: %s\n", strings.Join(participantsVals, ", "), e.note)
@@ -280,21 +306,22 @@ func (r *R) renderFlow() {
 			fmt.Fprintln(f, r.formatUdpPacket(p))
 
 		} else {
-			if err := h.Parse(p.packet.Data); err != nil {
-				panic(err)
-			}
-
 			line := "--x"
 			if p.rx {
 				line = "->>"
 			}
 
-			fmt.Fprintf(f,
+			detail := fmt.Sprintf("%s(%s), index %v, counter: %v",
-				"    %s%s%s: %s(%s), index %v, counter: %v\n",
+				p.h.TypeName(), p.h.SubTypeName(), p.h.RemoteIndex, p.h.MessageCounter)
-				normalizeName(p.from.GetUDPAddr().String()),
+			if p.parseErr != nil {
+				detail = fmt.Sprintf("unparsed, %v (%d bytes)", p.parseErr, len(p.packet.Data))
+			}
+
+			fmt.Fprintf(f, "    %s%s%s: %s\n",
+				normalizeName(p.fromAddr().String()),
 				line,
-				normalizeName(p.to.GetUDPAddr().String()),
+				normalizeName(p.toAddr().String()),
-				h.TypeName(), h.SubTypeName(), h.RemoteIndex, h.MessageCounter,
+				detail,
 			)
 		}
 	}
@@ -408,21 +435,24 @@ func (r *R) unlockedInjectFlow(from, to *nebula.Control, p *udp.Packet, tun bool
 
 	r.renderHostmaps(fmt.Sprintf("Packet %v", len(r.flow)))
 
-	if len(r.ignoreFlows) > 0 {
 	var h header.H
-		err := h.Parse(p.Data)
+	var parseErr error
-		if err != nil {
+	if !tun {
-			panic(err)
+		parseErr = h.Parse(p.Data)
 	}
 
+	// Decide before copying, the copy comes from a freelist and an ignored packet would never be released
 	for _, i := range r.ignoreFlows {
-			if !tun {
+		if tun {
-				if i.messageType == h.Type && i.subType == h.Subtype {
+			if i.tun.HasValue && i.tun.IsTrue {
 				return nil
 			}
-			} else if i.tun.HasValue && i.tun.IsTrue {
+			continue
-				return nil
 		}
+
+		// A packet we could not parse has no type to match against, so no rule can ignore it
+		if parseErr == nil && i.messageType == h.Type && i.subType == h.Subtype {
+			return nil
 		}
 	}
 
@@ -431,6 +461,8 @@ func (r *R) unlockedInjectFlow(from, to *nebula.Control, p *udp.Packet, tun bool
 		to:       to,
 		packet:   p.Copy(),
 		tun:      tun,
+		h:        h,
+		parseErr: parseErr,
 	}
 
 	r.flow = append(r.flow, flowEntry{packet: fp})
@@ -660,6 +692,10 @@ func (r *R) RouteExitFunc(sender *nebula.Control, whatDo ExitFunc) {
 			p.Release()
 			return
 
+		case Drop:
+			// Record it so the flow log shows the attempt, but never hand it to the receiver
+			r.unlockedInjectFlow(sender, receiver, p, false)
+
 		case KeepRouting:
 			fp := r.unlockedInjectFlow(sender, receiver, p, false)
 			receiver.InjectUDPPacket(p)
@@ -690,6 +726,85 @@ func (r *R) RouteUntilAfterMsgType(sender *nebula.Control, msgType header.Messag
 	})
 }
 
+// RouteFor routes everything that shows up for the given duration and then returns. Use it to let a test settle
+// deterministically rather than sleeping and hoping: a single FlushAll races a completing handshake, which queues
+// more packets right behind it.
+func (r *R) RouteFor(d time.Duration) {
+	r.RouteForAllExitFuncOrTimeout(d, func(*udp.Packet, *nebula.Control) ExitType {
+		return KeepRouting
+	})
+}
+
+// RouteForAllExitFuncOrTimeout is RouteForAllExitFunc with a deadline, reporting whether whatDo asked to exit
+// before time ran out. The unbounded version blocks forever on a quiet network, so this is what a test needs to
+// assert that something does NOT happen, or to route for a fixed settling period.
+func (r *R) RouteForAllExitFuncOrTimeout(timeout time.Duration, whatDo ExitFunc) bool {
+	sc := make([]reflect.SelectCase, 0, len(r.controls)+1)
+	cm := make([]*nebula.Control, 0, len(r.controls))
+
+	for _, c := range r.controls {
+		sc = append(sc, reflect.SelectCase{
+			Dir:  reflect.SelectRecv,
+			Chan: reflect.ValueOf(c.GetUDPTxChan()),
+			Send: reflect.Value{},
+		})
+		cm = append(cm, c)
+	}
+
+	timer := time.NewTimer(timeout)
+	defer timer.Stop()
+	sc = append(sc, reflect.SelectCase{
+		Dir:  reflect.SelectRecv,
+		Chan: reflect.ValueOf(timer.C),
+		Send: reflect.Value{},
+	})
+
+	for {
+		x, rx, _ := reflect.Select(sc)
+		if x == len(cm) {
+			return false
+		}
+
+		r.Lock()
+		p := rx.Interface().(*udp.Packet)
+		receiver := r.getControl(cm[x].GetUDPAddr(), p.To, p)
+		if receiver == nil {
+			r.Unlock()
+			panic("Can't RouteForAllExitFuncOrTimeout for host: " + p.To.String())
+		}
+
+		e := whatDo(p, receiver)
+		switch e {
+		case ExitNow:
+			r.Unlock()
+			p.Release()
+			return true
+
+		case RouteAndExit:
+			fp := r.unlockedInjectFlow(cm[x], receiver, p, false)
+			receiver.InjectUDPPacket(p)
+			fp.WasReceived()
+			r.Unlock()
+			p.Release()
+			return true
+
+		case Drop:
+			// Record it so the flow log shows the attempt, but never hand it to the receiver
+			r.unlockedInjectFlow(cm[x], receiver, p, false)
+
+		case KeepRouting:
+			fp := r.unlockedInjectFlow(cm[x], receiver, p, false)
+			receiver.InjectUDPPacket(p)
+			fp.WasReceived()
+
+		default:
+			panic(fmt.Sprintf("Unknown exitFunc return: %v", e))
+		}
+		r.Unlock()
+		p.Release()
+	}
+}
+
 func (r *R) RouteForAllUntilAfterMsgTypeTo(receiver *nebula.Control, msgType header.MessageType, subType header.MessageSubType) {
 	h := &header.H{}
 	r.RouteForAllExitFunc(func(p *udp.Packet, r *nebula.Control) ExitType {
@@ -782,6 +897,10 @@ func (r *R) RouteForAllExitFunc(whatDo ExitFunc) {
 			p.Release()
 			return
 
+		case Drop:
+			// Record it so the flow log shows the attempt, but never hand it to the receiver
+			r.unlockedInjectFlow(cm[x], receiver, p, false)
+
 		case KeepRouting:
 			fp := r.unlockedInjectFlow(cm[x], receiver, p, false)
 			receiver.InjectUDPPacket(p)

e2e/tunnels_test.go

@@ -15,6 +15,7 @@ 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"
 )
 
@@ -42,8 +43,8 @@ func TestDropInactiveTunnels(t *testing.T) {
 	r.Log("Go inactive and wait for the tunnels to get dropped")
 	waitStart := time.Now()
 	for {
-		myIndexes := len(myControl.GetHostmap().Indexes)
+		myIndexes := myControl.GetHostmapIndexCount()
-		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		theirIndexes := theirControl.GetHostmapIndexCount()
 		if myIndexes == 0 && theirIndexes == 0 {
 			break
 		}
@@ -373,6 +374,100 @@ 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{})
@@ -398,8 +493,8 @@ func TestCloseTunnelAuthenticated(t *testing.T) {
 
 	waitStart := time.Now()
 	for {
-		myIndexes := len(myControl.GetHostmap().Indexes)
+		myIndexes := myControl.GetHostmapIndexCount()
-		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		theirIndexes := theirControl.GetHostmapIndexCount()
 		if myIndexes == 0 && theirIndexes == 0 {
 			break
 		}
@@ -453,8 +548,8 @@ func TestCloseTunnelAuthenticated(t *testing.T) {
 	r.Log("Injected bogus close tunnel. Let's see!")
 	waitStart = time.Now()
 	for {
-		myIndexes := len(myControl.GetHostmap().Indexes)
+		myIndexes := myControl.GetHostmapIndexCount()
-		theirIndexes := len(theirControl.GetHostmap().Indexes)
+		theirIndexes := theirControl.GetHostmapIndexCount()
 		if myIndexes == 0 {
 			t.Fatal("myIndexes should not be 0")
 		}

ecn_inner_test.go

@@ -1,124 +0,0 @@
-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

@@ -146,6 +146,14 @@ listen:
   # Default true; set to false to leave WDF in charge of inbound decisions on the listener port. Not reloadable.
   #windows_bypass_wdf: true
 
+  # On macOS only
+  # macOS scopes the udp socket to the interface it was created on, so moving between networks (wifi to wired,
+  # office to home) leaves Nebula sending out an interface that no longer has a route. When true, Nebula watches
+  # the routing socket and rebinds the listener once the change settles.
+  # iOS does not use this, the host app drives the same rebind itself.
+  # Default true. Not reloadable.
+  #rebind_on_network_change: true
+
   # By default, Nebula replies to packets it has no tunnel for with a "recv_error" packet. This packet helps speed up reconnection
   # in the case that Nebula on either side did not shut down cleanly. This response can be abused as a way to discover if Nebula is running
   # on a host though. This option lets you configure if you want to send "recv_error" packets always, never, or only to private network remotes.
@@ -397,7 +405,7 @@ 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 port unreachable packet.
+  #     - For other protocols, this will be an ICMP "Destination unreachable: Communication administratively prohibited" packet.
   outbound_action: drop
   inbound_action: drop
 

examples/service_scripts/nebula.service

@@ -8,6 +8,15 @@ Before=sshd.service
 Type=notify
 NotifyAccess=main
 SyslogIdentifier=nebula
+
+# Uncomment to run as an unprivileged user with only CAP_NET_ADMIN. Requires a
+# nebula user that owns the config directory. Add CAP_NET_BIND_SERVICE to both
+# lines if any listener (lighthouse DNS, listen.port, stats, sshd) binds <1024.
+#User=nebula
+#Group=nebula
+#CapabilityBoundingSet=CAP_NET_ADMIN
+#AmbientCapabilities=CAP_NET_ADMIN
+
 ExecReload=/bin/kill -HUP $MAINPID
 ExecStart=/usr/local/bin/nebula -config /etc/nebula/config.yml
 Restart=always

firewall.go

@@ -44,8 +44,8 @@ type Firewall struct {
 	InRules  *FirewallTable
 	OutRules *FirewallTable
 
-	InSendReject  bool
+	InboundSendReject  bool
-	OutSendReject bool
+	OutboundSendReject bool
 
 	//TODO: we should have many more options for TCP, an option for ICMP, and mimic the kernel a bit better
 	// https://www.kernel.org/doc/Documentation/networking/nf_conntrack-sysctl.txt
@@ -59,7 +59,8 @@ type Firewall struct {
 
 	// assignedNetworks is a list of vpn networks assigned to us in the certificate.
 	assignedNetworks []netip.Prefix
-	hasUnsafeNetworks bool
+	// unsafeNetworks is the list of unsafe networks issued to us in the certificate
+	unsafeNetworks []netip.Prefix
 
 	rules        string
 	rulesVersion uint16
@@ -158,10 +159,9 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 		assignedNetworks = append(assignedNetworks, network)
 	}
 
-	hasUnsafeNetworks := false
+	unsafeNetworks := c.UnsafeNetworks()
-	for _, n := range c.UnsafeNetworks() {
+	for _, n := range unsafeNetworks {
 		routableNetworks.Insert(n)
-		hasUnsafeNetworks = true
 	}
 
 	return &Firewall{
@@ -176,7 +176,7 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 		DefaultTimeout:   defaultTimeout,
 		routableNetworks: routableNetworks,
 		assignedNetworks: assignedNetworks,
-		hasUnsafeNetworks: hasUnsafeNetworks,
+		unsafeNetworks:   unsafeNetworks,
 		l:                l,
 
 		incomingMetrics: firewallMetrics{
@@ -216,23 +216,23 @@ func NewFirewallFromConfig(l *slog.Logger, cs *CertState, c *config.C) (*Firewal
 	inboundAction := c.GetString("firewall.inbound_action", "drop")
 	switch inboundAction {
 	case "reject":
-		fw.InSendReject = true
+		fw.InboundSendReject = true
 	case "drop":
-		fw.InSendReject = false
+		fw.InboundSendReject = false
 	default:
 		l.Warn("invalid firewall.inbound_action, defaulting to `drop`", "action", inboundAction)
-		fw.InSendReject = false
+		fw.InboundSendReject = false
 	}
 
 	outboundAction := c.GetString("firewall.outbound_action", "drop")
 	switch outboundAction {
 	case "reject":
-		fw.OutSendReject = true
+		fw.OutboundSendReject = true
 	case "drop":
-		fw.OutSendReject = false
+		fw.OutboundSendReject = false
 	default:
 		l.Warn("invalid firewall.outbound_action, defaulting to `drop`", "action", outboundAction)
-		fw.OutSendReject = false
+		fw.OutboundSendReject = false
 	}
 
 	err := AddFirewallRulesFromConfig(l, false, c, fw)
@@ -423,11 +423,6 @@ var ErrNoMatchingRule = errors.New("no matching rule in firewall table")
 // Drop returns an error if the packet should be dropped, explaining why. It
 // returns nil if the packet should not be dropped.
 func (f *Firewall) Drop(fp firewall.Packet, incoming bool, h *HostInfo, caPool *cert.CAPool, localCache firewall.ConntrackCache) error {
-	// Check if we spoke to this tuple, if we did then allow this packet
-	if f.inConns(fp, h, caPool, localCache) {
-		return nil
-	}
-
 	// Make sure remote address matches nebula certificate, and determine how to treat it
 	if h.networks == nil {
 		// Simple case: Certificate has one address and no unsafe networks
@@ -461,6 +456,11 @@ func (f *Firewall) Drop(fp firewall.Packet, incoming bool, h *HostInfo, caPool *
 		return ErrInvalidLocalIP
 	}
 
+	// Check if we spoke to this tuple, if we did then allow this packet
+	if f.inConns(fp, h, caPool, localCache) {
+		return nil
+	}
+
 	table := f.OutRules
 	if incoming {
 		table = f.InRules
@@ -897,7 +897,7 @@ func (flc *firewallLocalCIDR) addRule(f *Firewall, localCidr string) error {
 	}
 
 	if localCidr == "" {
-		if !f.hasUnsafeNetworks || f.defaultLocalCIDRAny {
+		if len(f.unsafeNetworks) == 0 || f.defaultLocalCIDRAny {
 			flc.Any = true
 			return nil
 		}
@@ -1055,7 +1055,6 @@ 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
@@ -1064,11 +1063,11 @@ func parsePort(s string) (int32, int32, error) {
 		return firewall.PortFragment, firewall.PortFragment, nil
 	}
 	if !strings.Contains(s, `-`) {
-		rPort, err := strconv.Atoi(s)
+		rPort, err := parsePortValue("", s)
 		if err != nil {
-			return notAPort, notAPort, fmt.Errorf("was not a number; `%s`", s)
+			return notAPort, notAPort, err
 		}
-		return int32(rPort), int32(rPort), nil
+		return rPort, rPort, nil
 	}
 
 	sPorts := strings.SplitN(s, `-`, 2)
@@ -1079,22 +1078,40 @@ func parsePort(s string) (int32, int32, error) {
 		return notAPort, notAPort, fmt.Errorf("appears to be a range but could not be parsed; `%s`", s)
 	}
 
-	rStartPort, err := strconv.Atoi(sPorts[0])
+	startPort, err := parsePortValue("beginning range ", sPorts[0])
 	if err != nil {
-		return notAPort, notAPort, fmt.Errorf("beginning range was not a number; `%s`", sPorts[0])
+		return notAPort, notAPort, err
 	}
 
-	rEndPort, err := strconv.Atoi(sPorts[1])
+	endPort, err := parsePortValue("ending range ", sPorts[1])
 	if err != nil {
-		return notAPort, notAPort, fmt.Errorf("ending range was not a number; `%s`", sPorts[1])
+		return notAPort, notAPort, err
 	}
 
-	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

@@ -5,8 +5,6 @@ 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
@@ -58,8 +56,8 @@ 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(), logging.LevelTrace) {
+			if c.l.Enabled(context.Background(), slog.LevelDebug) {
-				c.l.Log(context.Background(), logging.LevelTrace, "resetting conntrack cache", "len", ll)
+				c.l.Debug("resetting conntrack cache", "len", ll)
 			}
 			c.cache = make(ConntrackCache, ll)
 		}

firewall/cache_test.go

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

firewall_test.go

@@ -916,6 +916,159 @@ func TestFirewall_DropIPSpoofing(t *testing.T) {
 	assert.Equal(t, fw.Drop(p, true, &h1, cp, nil), ErrInvalidRemoteIP)
 }
 
+func TestFirewall_ConntrackSourceSpoofingAcrossPeers(t *testing.T) {
+	l := test.NewLoggerWithOutput(&bytes.Buffer{})
+
+	myVpnNetworksTable := new(bart.Lite)
+	myVpnNetworksTable.Insert(netip.MustParsePrefix("192.0.2.1/24"))
+
+	owner := &dummyCert{
+		name:     "owner",
+		networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.1/24")},
+	}
+
+	victim := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "victim",
+			networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.2/24")},
+		},
+	}
+	victimHI := HostInfo{
+		ConnectionState: &ConnectionState{peerCert: victim},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.2")},
+	}
+	victimHI.buildNetworks(myVpnNetworksTable, victim.Certificate)
+
+	attacker := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "attacker",
+			networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.3/24")},
+		},
+	}
+	attackerHI := HostInfo{
+		ConnectionState: &ConnectionState{peerCert: attacker},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.3")},
+	}
+	attackerHI.buildNetworks(myVpnNetworksTable, attacker.Certificate)
+
+	fw := NewFirewall(l, time.Second, time.Minute, time.Hour, owner)
+	// Allow any inbound traffic that passes the cert / source-IP checks.
+	require.NoError(t, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
+	cp := cert.NewCAPool()
+
+	flow := firewall.Packet{
+		LocalAddr:  netip.MustParseAddr("192.0.2.1"),
+		RemoteAddr: netip.MustParseAddr("192.0.2.2"),
+		LocalPort:  443,
+		RemotePort: 55000,
+		Protocol:   firewall.ProtoUDP,
+	}
+
+	require.NoError(t, fw.Drop(flow, true, &victimHI, cp, nil),
+		"victim's own traffic from its own overlay IP must be allowed")
+
+	unseen := flow
+	unseen.RemotePort = 55001
+	assert.Equal(t, ErrInvalidRemoteIP, fw.Drop(unseen, true, &attackerHI, cp, nil),
+		"sanity: attacker forging victim's source IP must be rejected when no conntrack entry exists")
+
+	got := fw.Drop(flow, true, &attackerHI, cp, nil)
+	t.Logf("attacker replaying victim's 4-tuple: Drop returned %v (nil == packet ALLOWED == spoof succeeded)", got)
+	assert.Equal(t, ErrInvalidRemoteIP, got,
+		"SECURITY: attacker spoofed victim's overlay source IP (192.0.2.2) by reusing an existing conntrack 4-tuple; Drop returned %v instead of rejecting", got)
+}
+
+// BenchmarkFirewallDropConntrackHit measures Drop on an already-established flow
+// (a conntrack hit). This is the fast path that the source-IP<->cert binding
+// reordering adds work to, so it quantifies the cost of moving the address checks
+// ahead of the conntrack lookup. Cases:
+//   - simple:  peer cert has one address, no unsafe networks (h.networks == nil),
+//     so the remote-address check is a single netip.Addr compare.
+//   - complex: peer cert has unsafe networks (h.networks populated), so the
+//     remote-address check is a BART lookup.
+//   - noCache/localCache: whether a per-batch ConntrackCache is supplied, which in
+//     the original code let the fast path skip straight past the address checks.
+func BenchmarkFirewallDropConntrackHit(b *testing.B) {
+	l := test.NewLoggerWithOutput(&bytes.Buffer{})
+
+	myVpnNetworksTable := new(bart.Lite)
+	myVpnNetworksTable.Insert(netip.MustParsePrefix("192.0.2.1/24"))
+
+	owner := &dummyCert{
+		name:     "owner",
+		networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.1/24")},
+	}
+
+	simpleCert := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:     "simple",
+			networks: []netip.Prefix{netip.MustParsePrefix("192.0.2.2/24")},
+		},
+	}
+	simpleHost := &HostInfo{
+		ConnectionState: &ConnectionState{peerCert: simpleCert},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.2")},
+	}
+	simpleHost.buildNetworks(myVpnNetworksTable, simpleCert.Certificate)
+
+	complexCert := &cert.CachedCertificate{
+		Certificate: &dummyCert{
+			name:           "complex",
+			networks:       []netip.Prefix{netip.MustParsePrefix("192.0.2.2/24")},
+			unsafeNetworks: []netip.Prefix{netip.MustParsePrefix("198.51.100.0/24")},
+		},
+	}
+	complexHost := &HostInfo{
+		ConnectionState: &ConnectionState{peerCert: complexCert},
+		vpnAddrs:        []netip.Addr{netip.MustParseAddr("192.0.2.2")},
+	}
+	complexHost.buildNetworks(myVpnNetworksTable, complexCert.Certificate)
+
+	cp := cert.NewCAPool()
+
+	flow := firewall.Packet{
+		LocalAddr:  netip.MustParseAddr("192.0.2.1"),
+		RemoteAddr: netip.MustParseAddr("192.0.2.2"),
+		LocalPort:  443,
+		RemotePort: 55000,
+		Protocol:   firewall.ProtoUDP,
+	}
+
+	cases := []struct {
+		name     string
+		host     *HostInfo
+		useCache bool
+	}{
+		{"simple/noCache", simpleHost, false},
+		{"simple/localCache", simpleHost, true},
+		{"complex/noCache", complexHost, false},
+		{"complex/localCache", complexHost, true},
+	}
+
+	for _, tc := range cases {
+		b.Run(tc.name, func(b *testing.B) {
+			fw := NewFirewall(l, time.Second, time.Minute, time.Hour, owner)
+			require.NoError(b, fw.AddRule(true, firewall.ProtoAny, 0, 0, []string{"any"}, "", "", "", "", ""))
+
+			// Establish the conntrack entry so every benchmarked Drop is a hit.
+			require.NoError(b, fw.Drop(flow, true, tc.host, cp, nil))
+
+			var cache firewall.ConntrackCache
+			if tc.useCache {
+				cache = firewall.ConntrackCache{}
+			}
+
+			b.ReportAllocs()
+			b.ResetTimer()
+			for i := 0; i < b.N; i++ {
+				if err := fw.Drop(flow, true, tc.host, cp, cache); err != nil {
+					b.Fatal(err)
+				}
+			}
+		})
+	}
+}
+
 func BenchmarkLookup(b *testing.B) {
 	ml := func(m map[string]struct{}, a [][]string) {
 		for n := 0; n < b.N; n++ {
@@ -1029,6 +1182,75 @@ 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

@@ -1,6 +1,6 @@
 module github.com/slackhq/nebula
 
-go 1.25.0
+go 1.26.0
 
 require (
 	dario.cat/mergo v1.0.2
@@ -9,10 +9,10 @@ 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 v0.26.1
+	github.com/gaissmai/bart v0.28.0
 	github.com/gogo/protobuf v1.3.2
 	github.com/google/gopacket v1.1.19
-	github.com/kardianos/service v1.2.4
+	github.com/kardianos/service v1.3.0
 	github.com/miekg/dns v1.1.72
 	github.com/miekg/pkcs11 v1.1.2
 	github.com/nbrownus/go-metrics-prometheus v0.0.0-20210712211119-974a6260965f
@@ -24,15 +24,15 @@ 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 v0.50.0
+	golang.org/x/crypto v0.54.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.53.0
+	golang.org/x/net v0.57.0
-	golang.org/x/sync v0.20.0
+	golang.org/x/sync v0.22.0
-	golang.org/x/sys v0.43.0
+	golang.org/x/sys v0.47.0
-	golang.org/x/term v0.42.0
+	golang.org/x/term v0.45.0
 	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
+	golang.zx2c4.com/wireguard/windows v1.0.1
 	google.golang.org/protobuf v1.36.11
 	gopkg.in/yaml.v3 v3.0.1
 	gvisor.dev/gvisor v0.0.0-20240423190808-9d7a357edefe
@@ -50,7 +50,7 @@ require (
 	github.com/prometheus/procfs v0.16.1 // indirect
 	github.com/vishvananda/netns v0.0.5 // indirect
 	go.yaml.in/yaml/v2 v2.4.2 // indirect
-	golang.org/x/mod v0.34.0 // indirect
+	golang.org/x/mod v0.36.0 // indirect
 	golang.org/x/time v0.5.0 // indirect
-	golang.org/x/tools v0.43.0 // indirect
+	golang.org/x/tools v0.45.0 // indirect
 )

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 v0.26.1 h1:+w4rnLGNlA2GDVn382Tfe3jOsK5vOr5n4KmigJ9lbTo=
+github.com/gaissmai/bart v0.28.0 h1:89yZLo8NmyqD0RYgJ3QO9HhqqGGw+oWhf90cZm69Lko=
-github.com/gaissmai/bart v0.26.1/go.mod h1:GREWQfTLRWz/c5FTOsIw+KkscuFkIV5t8Rp7Nd1Td5c=
+github.com/gaissmai/bart v0.28.0/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=
@@ -66,8 +66,8 @@ github.com/json-iterator/go v1.1.10/go.mod h1:KdQUCv79m/52Kvf8AW2vK1V8akMuk1QjK/
 github.com/json-iterator/go v1.1.11/go.mod h1:KdQUCv79m/52Kvf8AW2vK1V8akMuk1QjK/uOdHXbAo4=
 github.com/julienschmidt/httprouter v1.2.0/go.mod h1:SYymIcj16QtmaHHD7aYtjjsJG7VTCxuUUipMqKk8s4w=
 github.com/julienschmidt/httprouter v1.3.0/go.mod h1:JR6WtHb+2LUe8TCKY3cZOxFyyO8IZAc4RVcycCCAKdM=
-github.com/kardianos/service v1.2.4 h1:XNlGtZOYNx2u91urOdg/Kfmc+gfmuIo1Dd3rEi2OgBk=
+github.com/kardianos/service v1.3.0 h1:/LGy+xPP2TM+GLTiCZ2di7cy0Jd/qrawlTUfqKYFdTI=
-github.com/kardianos/service v1.2.4/go.mod h1:E4V9ufUuY82F7Ztlu1eN9VXWIQxg8NoLQlmFe0MtrXc=
+github.com/kardianos/service v1.3.0/go.mod h1:E4V9ufUuY82F7Ztlu1eN9VXWIQxg8NoLQlmFe0MtrXc=
 github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8=
 github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+oQHNcck=
 github.com/klauspost/compress v1.18.0 h1:c/Cqfb0r+Yi+JtIEq73FWXVkRonBlf0CRNYc8Zttxdo=
@@ -162,16 +162,16 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
 golang.org/x/crypto v0.0.0-20210322153248-0c34fe9e7dc2/go.mod h1:T9bdIzuCu7OtxOm1hfPfRQxPLYneinmdGuTeoZ9dtd4=
-golang.org/x/crypto v0.50.0 h1:zO47/JPrL6vsNkINmLoo/PH1gcxpls50DNogFvB5ZGI=
+golang.org/x/crypto v0.54.0 h1:YLIA59K4fiNzHzjnZt2tUJQjQtUWfWbeHBqKtk3eScw=
-golang.org/x/crypto v0.50.0/go.mod h1:3muZ7vA7PBCE6xgPX7nkzzjiUq87kRItoJQM1Yo8S+Q=
+golang.org/x/crypto v0.54.0/go.mod h1:KWL8ny2AZdGR2cWmzeHrp2azQPGogOv+HeQaVEXC2dk=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090 h1:Di6/M8l0O2lCLc6VVRWhgCiApHV8MnQurBnFSHsQtNY=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090/go.mod h1:FXUEEKJgO7OQYeo8N01OfiKP8RXMtf6e8aTskBGqWdc=
 golang.org/x/lint v0.0.0-20200302205851-738671d3881b/go.mod h1:3xt1FjdF8hUf6vQPIChWIBhFzV8gjjsPE/fR3IyQdNY=
 golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzBzNggFXnrqF1CaUcvgkdR5Ot7KZg=
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
-golang.org/x/mod v0.34.0 h1:xIHgNUUnW6sYkcM5Jleh05DvLOtwc6RitGHbDk4akRI=
+golang.org/x/mod v0.36.0 h1:JJjpVx6myfUsUdAzZuOSTTmRE0PfZeNWzzvKrP7amb4=
-golang.org/x/mod v0.34.0/go.mod h1:ykgH52iCZe79kzLLMhyCUzhMci+nQj+0XkbXpNYtVjY=
+golang.org/x/mod v0.36.0/go.mod h1:moc6ELqsWcOw5Ef3xVprK5ul/MvtVvkIXLziUOICjUQ=
 golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20181114220301-adae6a3d119a/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20190108225652-1e06a53dbb7e/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
@@ -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 v0.53.0 h1:d+qAbo5L0orcWAr0a9JweQpjXF19LMXJE8Ey7hwOdUA=
+golang.org/x/net v0.57.0 h1:K5+3DljvIuDG9/Jv9rvyMywYNFCQ9RSUY6OOTTkT+tE=
-golang.org/x/net v0.53.0/go.mod h1:JvMuJH7rrdiCfbeHoo3fCQU24Lf5JJwT9W3sJFulfgs=
+golang.org/x/net v0.57.0/go.mod h1:KpXc8iv+r3XplLAG/f7Jsf9RPszJzdR0f58q9vGOuEU=
 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 v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
-golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sync v0.22.0/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 v0.43.0 h1:Rlag2XtaFTxp19wS8MXlJwTvoh8ArU6ezoyFsMyCTNI=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
-golang.org/x/sys v0.43.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.42.0 h1:UiKe+zDFmJobeJ5ggPwOshJIVt6/Ft0rcfrXZDLWAWY=
+golang.org/x/term v0.45.0 h1:NwWyBmoJCbfTHpxrWoZ9C6/VxOf7ic219I8xZZFdrf0=
-golang.org/x/term v0.42.0/go.mod h1:Dq/D+snpsbazcBG5+F9Q1n2rXV8Ma+71xEjTRufARgY=
+golang.org/x/term v0.45.0/go.mod h1:9aqxs0blBcrm/n0L9QW0aRVD+ktan8ssZromtqJC43w=
 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=
@@ -223,8 +223,8 @@ golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtn
 golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapKDTa4BR/hXlZSLoq2Wpct/0txZ28=
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
-golang.org/x/tools v0.43.0 h1:12BdW9CeB3Z+J/I/wj34VMl8X+fEXBxVR90JeMX5E7s=
+golang.org/x/tools v0.45.0 h1:18qN3FAooORvApf5XjCXgsuayZOEtXf6JK18I3+ONa8=
-golang.org/x/tools v0.43.0/go.mod h1:uHkMso649BX2cZK6+RpuIPXS3ho2hZo4FVwfoy1vIk0=
+golang.org/x/tools v0.45.0/go.mod h1:LuUGqqaXcXMEFEruIVJVm5mgDD8vww/z/SR1gQ4uE/0=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
@@ -233,8 +233,8 @@ golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 h1:B82qJJgjvYKsXS9jeu
 golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2/go.mod h1:deeaetjYA+DHMHg+sMSMI58GrEteJUUzzw7en6TJQcI=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b h1:J1CaxgLerRR5lgx3wnr6L04cJFbWoceSK9JWBdglINo=
 golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b/go.mod h1:tqur9LnfstdR9ep2LaJT4lFUl0EjlHtge+gAjmsHUG4=
-golang.zx2c4.com/wireguard/windows v0.6.1 h1:XMaKojH1Hs/raMrmnir4n35nTvzvWj7NmSYzHn2F4qU=
+golang.zx2c4.com/wireguard/windows v1.0.1 h1:eOxiDVbywPC+ZQqvdCK7x+ZwWXKbYv50TtH8ysFIbw8=
-golang.zx2c4.com/wireguard/windows v0.6.1/go.mod h1:04aqInu5GYuTFvMuDw/rKBAF7mHrltW/3rekpfbbZDM=
+golang.zx2c4.com/wireguard/windows v1.0.1/go.mod h1:+fbT3FFdX4zzYDLwJh5+HPEcNN/3HyNdzhNSVsQM+zs=
 google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
 google.golang.org/protobuf v0.0.0-20200109180630-ec00e32a8dfd/go.mod h1:DFci5gLYBciE7Vtevhsrf46CRTquxDuWsQurQQe4oz8=
 google.golang.org/protobuf v0.0.0-20200221191635-4d8936d0db64/go.mod h1:kwYJMbMJ01Woi6D6+Kah6886xMZcty6N08ah7+eCXa0=

handshake/errors.go

@@ -13,6 +13,7 @@ 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

@@ -312,11 +312,19 @@ func (m *Machine) processPayload(msg []byte, flags msgFlags) error {
 
 	// Process payload
 	if flags.expectsPayload {
+		var remoteIndex uint32
 		if m.result.Initiator {
-			m.result.RemoteIndex = payload.ResponderIndex
+			remoteIndex = payload.ResponderIndex
 		} else {
-			m.result.RemoteIndex = payload.InitiatorIndex
+			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

@@ -229,6 +229,24 @@ 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

@@ -83,6 +83,7 @@ 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
@@ -217,7 +218,6 @@ 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
@@ -295,7 +295,13 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		hm.messageMetrics.Tx(header.Handshake, hh.machine.Subtype(), 1)
 		err := hm.outside.WriteTo(stage0, addr)
 		if err != nil {
-			hostinfo.logger(hm.l).Error("Failed to send handshake message",
+			// These repeat every attempt, so match the success log below and only shout when the remotes changed
+			level := slog.LevelDebug
+			if remotesHaveChanged {
+				level = slog.LevelError
+			}
+
+			hostinfo.logger(hm.l).Log(context.Background(), level, "Failed to send handshake message",
 				"udpAddr", addr,
 				"initiatorIndex", hostinfo.localIndexId,
 				"handshake", hsFields,
@@ -323,7 +329,7 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		)
 	}
 
-	hm.f.relayManager.StartRelays(hm.f, vpnIp, hostinfo, stage0)
+	hm.f.relayManager.StartRelays(hm.f, vpnIp, hh, stage0)
 
 	// If a lighthouse triggered this attempt then we are still in the timer wheel and do not need to re-add
 	if !lighthouseTriggered {
@@ -430,14 +436,11 @@ func (hm *HandshakeManager) CheckAndComplete(hostinfo *HostInfo, handshakePacket
 	// Check if we already have a tunnel with this vpn ip
 	existingHostInfo, found := hm.mainHostMap.Hosts[hostinfo.vpnAddrs[0]]
 	if found && existingHostInfo != nil {
-		testHostInfo := existingHostInfo
+		// Is it just a delayed handshake packet? Check every hostinfo we hold for this address.
-		for testHostInfo != nil {
+		for _, testHostInfo := range hm.mainHostMap.unlockedGetHostList(hostinfo.vpnAddrs[0]) {
-			// Is it just a delayed handshake packet?
 			if bytes.Equal(hostinfo.HandshakePacket[handshakePacket], testHostInfo.HandshakePacket[handshakePacket]) {
 				return testHostInfo, ErrAlreadySeen
 			}
-
-			testHostInfo = testHostInfo.next
 		}
 
 		// Is this a newer handshake?
@@ -465,7 +468,6 @@ 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,
 		)
 	}
@@ -488,7 +490,6 @@ 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,
 		)
 	}
@@ -534,8 +535,10 @@ func (hm *HandshakeManager) DeleteHostInfo(hostinfo *HostInfo) {
 
 func (hm *HandshakeManager) unlockedDeleteHostInfo(hostinfo *HostInfo) {
 	for _, addr := range hostinfo.vpnAddrs {
+		if cur, ok := hm.vpnIps[addr]; ok && cur.hostinfo == hostinfo {
 			delete(hm.vpnIps, addr)
 		}
+	}
 
 	if len(hm.vpnIps) == 0 {
 		hm.vpnIps = map[netip.Addr]*HandshakeHostInfo{}
@@ -798,7 +801,6 @@ 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
@@ -965,7 +967,6 @@ 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) {
@@ -974,7 +975,6 @@ 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)))
@@ -1085,7 +1085,7 @@ func (hm *HandshakeManager) sendHandshakeResponse(via ViaSender, msg []byte, hos
 		hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
 		// We received a valid handshake on this relay, so make sure the relay
 		// state reflects that, in case it had been marked Disestablished.
-		via.relayHI.relayState.UpdateRelayForByIdxState(via.remoteIdx, Established)
+		via.relayHI.relayState.UpdateRelayForByIdxState(via.relay.LocalIndex, Established)
 		f.SendVia(via.relayHI, via.relay, msg, make([]byte, 12), make([]byte, mtu), false)
 		f.l.Info("Handshake message sent", append(logFields, "relay", via.relayHI.vpnAddrs[0])...)
 	}

hostmap.go

@@ -60,7 +60,16 @@ type HostMap struct {
 	Indexes       map[uint32]*HostInfo
 	Relays        map[uint32]*HostInfo // Maps a Relay IDX to a Relay HostInfo object
 	RemoteIndexes map[uint32]*HostInfo
+	// Hosts maps a vpn address to its primary hostinfo, one entry per address we hold a tunnel
+	// for. moreHosts only has an entry while an address is held by 2 or more hostinfos and stores
+	// the full most-recent-first list; moreHosts[a][0] is always the same hostinfo as Hosts[a].
+	// Each address gets its own independent list, so a hostinfo owning multiple addresses can
+	// never corrupt another address's ordering the way the old shared next/prev chain could.
+	// Entries in moreHosts are only ever written by unlockedSetHostsForAddr; Hosts is written
+	// directly only in the single-hostinfo fast paths where moreHosts is known to have no entry,
+	// and unlockedDeleteHostInfo swaps either map for a fresh one when it fully drains.
 	Hosts           map[netip.Addr]*HostInfo
+	moreHosts       map[netip.Addr][]*HostInfo
 	preferredRanges atomic.Pointer[[]netip.Prefix]
 	l               *slog.Logger
 }
@@ -138,9 +147,9 @@ 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
 }
@@ -229,18 +238,30 @@ const (
 )
 
 type HostInfo struct {
-	remote          netip.AddrPort
+	// The first cache line is everything the packet paths touch. Grouping them here means a send or receive
-	remotes         *RemoteList
+	// pulls in one line instead of two, which is what the layout looked like when state lived at the end.
-	promoteCounter  atomic.Uint32
+
+	remote          atomic.Pointer[netip.AddrPort]
 	ConnectionState *ConnectionState
+
+	// state holds everything the hot paths need to touch per packet, in one word: whether we have seen traffic
+	// each way since the connection manager last looked, whether it has given up on us, and the
+	// Interface.rebindEpoch this tunnel last sent under. Keeping the epoch here means it survives the traffic
+	// bits being cleared, so a tunnel that has not sent since a rebind still notices when it does.
+	state atomic.Uint32
+
+	promoteCounter atomic.Uint32
 	remoteIndexId  uint32
 	localIndexId   uint32
+	remotes        *RemoteList
 
 	// vpnAddrs is a list of vpn addresses assigned to this host that are within our own vpn networks
 	// The host may have other vpn addresses that are outside our
 	// vpn networks but were removed because they are not usable
 	vpnAddrs []netip.Addr
 
+	// Everything below is off the packet path: handshakes, relays, roaming and the connection manager.
+
 	// networks is a combination of specific vpn addresses (not prefixes!) and full unsafe networks assigned to this host.
 	networks   *bart.Table[NetworkType]
 	relayState RelayState
@@ -253,11 +274,6 @@ type HostInfo struct {
 	// This is used to limit lighthouse re-queries in chatty clients
 	nextLHQuery atomic.Int64
 
-	// lastRebindCount is the other side of Interface.rebindCount, if these values don't match then we need to ask LH
-	// for a punch from the remote end of this tunnel. The goal being to prime their conntrack for our traffic just like
-	// with a handshake
-	lastRebindCount int8
-
 	// lastHandshakeTime records the time the remote side told us about at the stage when the handshake was completed locally
 	// Stage 1 packet will contain it if I am a responder, stage 2 packet if I am an initiator
 	// This is used to avoid an attack where a handshake packet is replayed after some time
@@ -266,13 +282,6 @@ type HostInfo struct {
 	lastRoam       time.Time
 	lastRoamRemote netip.AddrPort
 
-	// Used to track other hostinfos for this vpn ip since only 1 can be primary
-	// Synchronised via hostmap lock and not the hostinfo lock.
-	next, prev *HostInfo
-
-	//TODO: in, out, and others might benefit from being an atomic.Int32. We could collapse connectionManager pendingDeletion, relayUsed, and in/out into this 1 thing
-	in, out, pendingDeletion atomic.Bool
-
 	// lastUsed tracks the last time ConnectionManager checked the tunnel and it was in use.
 	// This value will be behind against actual tunnel utilization in the hot path.
 	// This should only be used by the ConnectionManagers ticker routine.
@@ -282,7 +291,6 @@ type HostInfo struct {
 type ViaSender struct {
 	UdpAddr   netip.AddrPort
 	relayHI   *HostInfo // relayHI is the host info object of the relay
-	remoteIdx uint32    // remoteIdx is the index included in the header of the received packet
 	relay     *Relay    // relay contains the rest of the relay information, including the PeerIP of the host trying to communicate with us.
 	IsRelayed bool      // IsRelayed is true if the packet was sent through a relay
 }
@@ -334,6 +342,7 @@ func newHostMap(l *slog.Logger) *HostMap {
 		Relays:        map[uint32]*HostInfo{},
 		RemoteIndexes: map[uint32]*HostInfo{},
 		Hosts:         map[netip.Addr]*HostInfo{},
+		moreHosts:     map[netip.Addr][]*HostInfo{},
 		l:             l,
 	}
 }
@@ -382,13 +391,55 @@ func (hm *HostMap) EmitStats() {
 	metrics.GetOrRegisterGauge("hostmap.main.relayIndexes", nil).Update(int64(relaysLen))
 }
 
-// DeleteHostInfo will fully unlink the hostinfo and return true if it was the final hostinfo for this vpn ip
+// unlockedSetHostsForAddr stores the per-address hostinfo list (list[0] is the primary). An empty
+// list removes the address. This is the one place Hosts and moreHosts are written together, keep
+// it that way. Callers must hold the write lock.
+func (hm *HostMap) unlockedSetHostsForAddr(addr netip.Addr, list []*HostInfo) {
+	if len(list) == 0 {
+		delete(hm.Hosts, addr)
+		delete(hm.moreHosts, addr)
+		return
+	}
+	hm.Hosts[addr] = list[0]
+	if len(list) > 1 {
+		hm.moreHosts[addr] = list
+	} else {
+		delete(hm.moreHosts, addr)
+	}
+}
+
+// unlockedGetHostList returns every hostinfo holding addr, primary first, or nil if we have no
+// tunnel for addr. The common single-hostinfo case builds a fresh one element list, so keep this
+// off the packet hot path; the primary is a direct Hosts read. Callers must hold the lock (read
+// or write).
+func (hm *HostMap) unlockedGetHostList(addr netip.Addr) []*HostInfo {
+	if list, ok := hm.moreHosts[addr]; ok {
+		return list
+	}
+	if h, ok := hm.Hosts[addr]; ok {
+		return []*HostInfo{h}
+	}
+	return nil
+}
+
+// removeHostInfo returns list with hi removed (order preserved), or list unchanged if hi is
+// absent. It deletes in place: every mutator holds the hostmap write lock and no reader ever
+// retains a slice across a mutation (readers iterate under RLock), so there is no snapshot to
+// invalidate.
+func removeHostInfo(list []*HostInfo, hi *HostInfo) []*HostInfo {
+	idx := slices.Index(list, hi)
+	if idx < 0 {
+		return list
+	}
+	return slices.Delete(list, idx, idx+1)
+}
+
+// DeleteHostInfo will fully unlink the hostinfo and return true if no other hostinfo still holds
+// any of its vpn addrs, meaning we no longer have a tunnel to the peer
 func (hm *HostMap) DeleteHostInfo(hostinfo *HostInfo) bool {
 	// Delete the host itself, ensuring it's not modified anymore
 	hm.Lock()
-	// If we have a previous or next hostinfo then we are not the last one for this vpn ip
+	final := hm.unlockedDeleteHostInfo(hostinfo)
-	final := (hostinfo.next == nil && hostinfo.prev == nil)
-	hm.unlockedDeleteHostInfo(hostinfo)
 	hm.Unlock()
 
 	return final
@@ -400,86 +451,67 @@ func (hm *HostMap) MakePrimary(hostinfo *HostInfo) {
 	hm.unlockedMakePrimary(hostinfo)
 }
 
-func (hm *HostMap) unlockedMakePrimary(hostinfo *HostInfo) {
+// unlockedMakePrimary reports whether hostinfo is (now) the primary for each of its addresses,
-	// Get the current primary, if it exists
+// false only when it is no longer in the hostmap at all.
-	oldHostinfo := hm.Hosts[hostinfo.vpnAddrs[0]]
+func (hm *HostMap) unlockedMakePrimary(hostinfo *HostInfo) bool {
-
+	// A hostinfo that is no longer in the hostmap must not be re-inserted here. Callers can race
-	// Every address in the hostinfo gets elevated to primary
+	// tunnel teardown, deciding to promote under the read lock and only taking the write lock
-	for _, vpnAddr := range hostinfo.vpnAddrs {
+	// after a delete fully unlinked the hostinfo (connection manager swapPrimary, AddRelay). Every
-		//NOTE: It is possible that we leave a dangling hostinfo here but connection manager works on
+	// live hostinfo is registered in Indexes by unlockedAddHostInfo, so this is a membership test.
-		// indexes so it should be fine.
+	if hm.Indexes[hostinfo.localIndexId] != hostinfo {
-		hm.Hosts[vpnAddr] = hostinfo
+		return false
 	}
 
-	// If we are already primary then we won't bother re-linking
+	// Move hostinfo to the front (primary) of each of its address lists. The lists are
-	if oldHostinfo == hostinfo {
+	// independent per address, so this can never leave a dangling entry the way promoting
-		return
+	// against a single shared chain could.
-	}
-
-	// Unlink this hostinfo
-	if hostinfo.prev != nil {
-		hostinfo.prev.next = hostinfo.next
-	}
-	if hostinfo.next != nil {
-		hostinfo.next.prev = hostinfo.prev
-	}
-
-	// If there wasn't a previous primary then clear out any links
-	if oldHostinfo == nil {
-		hostinfo.next = nil
-		hostinfo.prev = nil
-		return
-	}
-
-	// Relink the hostinfo as primary
-	hostinfo.next = oldHostinfo
-	oldHostinfo.prev = hostinfo
-	hostinfo.prev = nil
-}
-
-func (hm *HostMap) unlockedDeleteHostInfo(hostinfo *HostInfo) {
 	for _, addr := range hostinfo.vpnAddrs {
-		h := hm.Hosts[addr]
+		if hm.Hosts[addr] == hostinfo {
-		for h != nil {
+			// Already primary for this address, the list is already in the right order
-			if h == hostinfo {
+			continue
-				hm.unlockedInnerDeleteHostInfo(h, addr)
-			}
-			h = h.next
 		}
+		list := removeHostInfo(hm.unlockedGetHostList(addr), hostinfo)
+		list = append([]*HostInfo{hostinfo}, list...)
+		hm.unlockedSetHostsForAddr(addr, list)
 	}
+	return true
 }
 
-func (hm *HostMap) unlockedInnerDeleteHostInfo(hostinfo *HostInfo, addr netip.Addr) {
+// unlockedDeleteHostInfo removes hostinfo from every one of its address lists and from the index
-	primary, ok := hm.Hosts[addr]
+// maps. It returns true if this was the last hostinfo for all of its addresses (we no longer have
-	isLastHostinfo := hostinfo.next == nil && hostinfo.prev == nil
+// any tunnel to the peer), which the caller uses to decide whether to clear learned lighthouse
-	if ok && primary == hostinfo {
+// state and disestablish relays.
-		// The vpn addr pointer points to the same hostinfo as the local index id, we can remove it
+func (hm *HostMap) unlockedDeleteHostInfo(hostinfo *HostInfo) bool {
+	// Remove this hostinfo from each of its address lists. The lists are independent, so a
+	// sibling is never promoted to an address it does not own and no other list is touched.
+	final := true
+	for _, addr := range hostinfo.vpnAddrs {
+		if list, ok := hm.moreHosts[addr]; ok {
+			list = removeHostInfo(list, hostinfo)
+			hm.unlockedSetHostsForAddr(addr, list)
+			if len(list) > 0 {
+				final = false
+			}
+		} else if existing, ok := hm.Hosts[addr]; ok {
+			if existing == hostinfo {
+				// Common case, the only hostinfo for this address. moreHosts has no entry to clean up.
 				delete(hm.Hosts, addr)
+			} else {
+				// We don't hold this address but another hostinfo does, we still have a tunnel to the peer
+				final = false
+			}
+		}
+	}
+
+	// Go maps never shrink their buckets, replace fully drained maps so a node that churned
+	// through a large peer count gives the memory back. Same idiom as the index maps below.
 	if len(hm.Hosts) == 0 {
 		hm.Hosts = map[netip.Addr]*HostInfo{}
 	}
-
+	if len(hm.moreHosts) == 0 {
-		if hostinfo.next != nil {
+		hm.moreHosts = map[netip.Addr][]*HostInfo{}
-			// We had more than 1 hostinfo at this vpn addr, promote the next in the list to primary
-			hm.Hosts[addr] = hostinfo.next
-			// It is primary, there is no previous hostinfo now
-			hostinfo.next.prev = nil
 	}
 
-	} else {
-		// Relink if we were in the middle of multiple hostinfos for this vpn addr
-		if hostinfo.prev != nil {
-			hostinfo.prev.next = hostinfo.next
-		}
-
-		if hostinfo.next != nil {
-			hostinfo.next.prev = hostinfo.prev
-		}
-	}
-
-	hostinfo.next = nil
-	hostinfo.prev = nil
-
 	// The remote index uses index ids outside our control so lets make sure we are only removing
 	// the remote index pointer here if it points to the hostinfo we are deleting
 	hostinfo2, ok := hm.RemoteIndexes[hostinfo.remoteIndexId]
@@ -502,7 +534,7 @@ func (hm *HostMap) unlockedInnerDeleteHostInfo(hostinfo *HostInfo, addr netip.Ad
 		)
 	}
 
-	if isLastHostinfo {
+	if final {
 		// I have lost connectivity to my peers. My relay tunnel is likely broken. Mark the next
 		// hops as 'Requested' so that new relay tunnels are created in the future.
 		hm.unlockedDisestablishVpnAddrRelayFor(hostinfo)
@@ -511,6 +543,8 @@ func (hm *HostMap) unlockedInnerDeleteHostInfo(hostinfo *HostInfo, addr netip.Ad
 	for _, localRelayIdx := range hostinfo.relayState.CopyRelayForIdxs() {
 		delete(hm.Relays, localRelayIdx)
 	}
+
+	return final
 }
 
 func (hm *HostMap) QueryIndex(index uint32) *HostInfo {
@@ -554,19 +588,30 @@ func (hm *HostMap) QueryVpnAddrsRelayFor(targetIps []netip.Addr, relayHostIp net
 	hm.RLock()
 	defer hm.RUnlock()
 
+	// This runs per relayed packet, so check the primary with a single map probe and only consult
+	// moreHosts when the primary can't relay for us.
 	h, ok := hm.Hosts[relayHostIp]
 	if !ok {
 		return nil, nil, errors.New("unable to find host")
 	}
 
-	for h != nil {
 	for _, targetIp := range targetIps {
 		r, ok := h.relayState.QueryRelayForByIp(targetIp)
 		if ok && r.State == Established {
 			return h, r, nil
 		}
 	}
-		h = h.next
+
+	if list, ok := hm.moreHosts[relayHostIp]; ok {
+		// list[0] is the primary we already checked
+		for _, h := range list[1:] {
+			for _, targetIp := range targetIps {
+				r, ok := h.relayState.QueryRelayForByIp(targetIp)
+				if ok && r.State == Established {
+					return h, r, nil
+				}
+			}
+		}
 	}
 
 	return nil, nil, errors.New("unable to find host with relay")
@@ -574,20 +619,14 @@ func (hm *HostMap) QueryVpnAddrsRelayFor(targetIps []netip.Addr, relayHostIp net
 
 func (hm *HostMap) unlockedDisestablishVpnAddrRelayFor(hi *HostInfo) {
 	for _, relayHostIp := range hi.relayState.CopyRelayIps() {
-		if h, ok := hm.Hosts[relayHostIp]; ok {
+		for _, h := range hm.unlockedGetHostList(relayHostIp) {
-			for h != nil {
 			h.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-				h = h.next
-			}
 		}
 	}
 	for _, rs := range hi.relayState.CopyAllRelayFor() {
 		if rs.Type == ForwardingType {
-			if h, ok := hm.Hosts[rs.PeerAddr]; ok {
+			for _, h := range hm.unlockedGetHostList(rs.PeerAddr) {
-				for h != nil {
 				h.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-					h = h.next
-				}
 			}
 		}
 	}
@@ -623,6 +662,11 @@ func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
 	hm.Indexes[hostinfo.localIndexId] = hostinfo
 	hm.RemoteIndexes[hostinfo.remoteIndexId] = hostinfo
 
+	hostinfo.markOut(f.rebindEpoch.Load())
+	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),
@@ -632,22 +676,27 @@ func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
 }
 
 func (hm *HostMap) unlockedInnerAddHostInfo(vpnAddr netip.Addr, hostinfo *HostInfo, f *Interface) {
-	existing := hm.Hosts[vpnAddr]
+	existing, ok := hm.Hosts[vpnAddr]
+	if !ok {
+		// Common case, the first hostinfo for this address. moreHosts stays empty.
 		hm.Hosts[vpnAddr] = hostinfo
-
+		return
-	if existing != nil && existing != hostinfo {
-		hostinfo.next = existing
-		existing.prev = hostinfo
 	}
 
-	i := 1
+	// The new hostinfo becomes the primary for this address. Remove any stale copy of it first so
-	check := hostinfo
+	// we never hold a duplicate, then prepend.
-	for check != nil {
+	list, ok := hm.moreHosts[vpnAddr]
-		if i > MaxHostInfosPerVpnIp {
+	if !ok {
-			hm.unlockedDeleteHostInfo(check)
+		list = []*HostInfo{existing}
 	}
-		check = check.next
+	list = removeHostInfo(list, hostinfo)
-		i++
+	list = append([]*HostInfo{hostinfo}, list...)
+	hm.unlockedSetHostsForAddr(vpnAddr, list)
+
+	// Enforce the per-address cap by fully retiring the oldest hostinfo once we exceed it.
+	// Deleting it removes it from all of its addresses and the index maps, matching prior behavior.
+	if len(list) > MaxHostInfosPerVpnIp {
+		hm.unlockedDeleteHostInfo(list[len(list)-1])
 	}
 }
 
@@ -679,7 +728,7 @@ func (hm *HostMap) ForEachIndex(f controlEach) {
 func (i *HostInfo) TryPromoteBest(preferredRanges []netip.Prefix, ifce *Interface) {
 	c := i.promoteCounter.Add(1)
 	if c%ifce.tryPromoteEvery.Load() == 0 {
-		remote := i.remote
+		remote := i.GetRemote()
 
 		// return early if we are already on a preferred remote
 		if remote.IsValid() {
@@ -714,6 +763,64 @@ func (i *HostInfo) TryPromoteBest(preferredRanges []netip.Prefix, ifce *Interfac
 	}
 }
 
+// Bits within HostInfo.state. Everything above stateEpochShift is the rebind epoch.
+const (
+	stateIn uint32 = 1 << iota
+	stateOut
+	statePendingDeletion
+
+	stateFlags      = stateIn | stateOut | statePendingDeletion
+	stateEpochShift = 3
+)
+
+// markIn records inbound traffic. Reading first keeps the cache line shared on the common path, where the bit
+// is already set.
+func (i *HostInfo) markIn() {
+	if i.state.Load()&stateIn == 0 {
+		i.state.Or(stateIn)
+	}
+}
+
+// markOut records that we sent on this tunnel under the given rebind epoch. It reports whether the epoch moved
+// since our last send, which means the local network changed and we want the far side to punch at us again.
+// The common path is a single load that matches and returns.
+func (i *HostInfo) markOut(epoch uint32) bool {
+	e := epoch << stateEpochShift
+	for {
+		old := i.state.Load()
+		if old&stateOut != 0 && old&^stateFlags == e {
+			return false
+		}
+
+		if i.state.CompareAndSwap(old, old&stateFlags|stateOut|e) {
+			return old&^stateFlags != e
+		}
+	}
+}
+
+// sentSinceCheck reports whether anything has been sent since the connection manager last looked.
+func (i *HostInfo) sentSinceCheck() bool {
+	return i.state.Load()&stateOut != 0
+}
+
+// takeTraffic clears both traffic bits, leaving the epoch alone, and reports what they were.
+func (i *HostInfo) takeTraffic() (in bool, out bool) {
+	old := i.state.And(^(stateIn | stateOut))
+	return old&stateIn != 0, old&stateOut != 0
+}
+
+func (i *HostInfo) setPendingDeletion(v bool) {
+	if v {
+		i.state.Or(statePendingDeletion)
+	} else {
+		i.state.And(^statePendingDeletion)
+	}
+}
+
+func (i *HostInfo) isPendingDeletion() bool {
+	return i.state.Load()&statePendingDeletion != 0
+}
+
 func (i *HostInfo) GetCert() *cert.CachedCertificate {
 	if i.ConnectionState != nil {
 		return i.ConnectionState.peerCert
@@ -721,11 +828,18 @@ func (i *HostInfo) GetCert() *cert.CachedCertificate {
 	return nil
 }
 
+func (i *HostInfo) GetRemote() netip.AddrPort {
+	if p := i.remote.Load(); p != nil {
+		return *p
+	}
+	return netip.AddrPort{}
+}
+
 // TODO: Maybe use ViaSender here?
 func (i *HostInfo) SetRemote(remote netip.AddrPort) {
 	// We copy here because we likely got this remote from a source that reuses the object
-	if i.remote != remote {
+	if i.GetRemote() != remote {
-		i.remote = remote
+		i.remote.Store(&remote)
 		i.remotes.LearnRemote(i.vpnAddrs[0], remote)
 	}
 }
@@ -737,7 +851,7 @@ func (i *HostInfo) SetRemoteIfPreferred(hm *HostMap, via ViaSender) bool {
 		return false
 	}
 
-	currentRemote := i.remote
+	currentRemote := i.GetRemote()
 	if !currentRemote.IsValid() {
 		i.SetRemote(via.UdpAddr)
 		return true

hostmap_test.go

@@ -2,6 +2,7 @@ package nebula
 
 import (
 	"net/netip"
+	"slices"
 	"testing"
 
 	"github.com/slackhq/nebula/config"
@@ -10,78 +11,84 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+// chainIds returns the localIndexIds of the hostinfos holding addr, primary (index 0) first. It
+// also validates the Hosts/moreHosts sync contract on every call so a mutation that broke it
+// fails fast.
+func chainIds(t *testing.T, hm *HostMap, addr netip.Addr) []uint32 {
+	t.Helper()
+	assertHostMapInvariants(t, hm)
+	list := hm.unlockedGetHostList(addr)
+	ids := make([]uint32, len(list))
+	for i, h := range list {
+		ids[i] = h.localIndexId
+	}
+	return ids
+}
+
+// assertHostMapInvariants checks the Hosts/moreHosts contract: moreHosts only holds addresses
+// with 2 or more hostinfos, its first entry is always the primary in Hosts, lists never hold
+// duplicates, every hostinfo in a list owns the address and is registered in Indexes, and every
+// indexed hostinfo is reachable through each of its addresses.
+func assertHostMapInvariants(t *testing.T, hm *HostMap) {
+	t.Helper()
+	for addr, list := range hm.moreHosts {
+		require.GreaterOrEqualf(t, len(list), 2, "moreHosts[%s] must hold at least 2 hostinfos", addr)
+		require.Samef(t, hm.Hosts[addr], list[0], "moreHosts[%s][0] must match the primary in Hosts", addr)
+		seen := map[*HostInfo]bool{}
+		for _, h := range list {
+			require.NotNilf(t, h, "moreHosts[%s] must never hold a nil hostinfo", addr)
+			require.Falsef(t, seen[h], "moreHosts[%s] holds hostinfo %d twice", addr, h.localIndexId)
+			seen[h] = true
+			require.Samef(t, hm.Indexes[h.localIndexId], h, "moreHosts[%s] member %d is not registered in Indexes", addr, h.localIndexId)
+			require.Truef(t, slices.Contains(h.vpnAddrs, addr), "moreHosts[%s] member %d does not own the address", addr, h.localIndexId)
+		}
+	}
+	for addr, h := range hm.Hosts {
+		require.NotNilf(t, h, "Hosts[%s] must never be nil", addr)
+		require.Samef(t, hm.Indexes[h.localIndexId], h, "Hosts[%s] primary %d is not registered in Indexes", addr, h.localIndexId)
+		require.Truef(t, slices.Contains(h.vpnAddrs, addr), "Hosts[%s] primary (index %d) does not own the address", addr, h.localIndexId)
+	}
+	for idx, h := range hm.Indexes {
+		require.Equalf(t, idx, h.localIndexId, "Indexes[%d] holds hostinfo with localIndexId %d", idx, h.localIndexId)
+		for _, va := range h.vpnAddrs {
+			require.Truef(t, slices.Contains(hm.unlockedGetHostList(va), h), "indexed hostinfo %d is missing from the list for %s", idx, va)
+		}
+	}
+}
+
 func TestHostMap_MakePrimary(t *testing.T) {
 	l := test.NewLogger()
 	hm := newHostMap(l)
 
 	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
 
-	h1 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 1}
+	h1 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
-	h2 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 2}
+	h2 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 2}
-	h3 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 3}
+	h3 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 3}
-	h4 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 4}
+	h4 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 4}
 
 	hm.unlockedAddHostInfo(h4, f)
 	hm.unlockedAddHostInfo(h3, f)
 	hm.unlockedAddHostInfo(h2, f)
 	hm.unlockedAddHostInfo(h1, f)
 
-	// Make sure we go h1 -> h2 -> h3 -> h4
+	// Most-recently-added is primary: h1, h2, h3, h4
-	prim := hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.Equal(t, []uint32{1, 2, 3, 4}, chainIds(t, hm, a))
-	assert.Equal(t, h1.localIndexId, prim.localIndexId)
+	assert.Equal(t, h1, hm.QueryVpnAddr(a))
-	assert.Equal(t, h2.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Equal(t, h3.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h4.prev.localIndexId)
-	assert.Nil(t, h4.next)
 
-	// Swap h3/middle to primary
+	// Swap the middle to primary: h3, h1, h2, h4
 	hm.MakePrimary(h3)
+	assert.Equal(t, []uint32{3, 1, 2, 4}, chainIds(t, hm, a))
+	assert.Equal(t, h3, hm.QueryVpnAddr(a))
 
-	// Make sure we go h3 -> h1 -> h2 -> h4
+	// Swap the tail to primary: h4, h3, h1, h2
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
-	assert.Equal(t, h3.localIndexId, prim.localIndexId)
-	assert.Equal(t, h1.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h2.localIndexId, h1.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h1.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Equal(t, h2.localIndexId, h4.prev.localIndexId)
-	assert.Nil(t, h4.next)
-
-	// Swap h4/tail to primary
 	hm.MakePrimary(h4)
+	assert.Equal(t, []uint32{4, 3, 1, 2}, chainIds(t, hm, a))
 
-	// Make sure we go h4 -> h3 -> h1 -> h2
+	// Swapping the current primary again is a no-op
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
-	assert.Equal(t, h4.localIndexId, prim.localIndexId)
-	assert.Equal(t, h3.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h2.localIndexId, h1.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h1.prev.localIndexId)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Nil(t, h2.next)
-
-	// Swap h4 again should be no-op
 	hm.MakePrimary(h4)
-
+	assert.Equal(t, []uint32{4, 3, 1, 2}, chainIds(t, hm, a))
-	// Make sure we go h4 -> h3 -> h1 -> h2
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
-	assert.Equal(t, h4.localIndexId, prim.localIndexId)
-	assert.Equal(t, h3.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h2.localIndexId, h1.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h1.prev.localIndexId)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Nil(t, h2.next)
 }
 
 func TestHostMap_DeleteHostInfo(t *testing.T) {
@@ -89,13 +96,14 @@ func TestHostMap_DeleteHostInfo(t *testing.T) {
 	hm := newHostMap(l)
 
 	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
 
-	h1 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 1}
+	h1 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
-	h2 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 2}
+	h2 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 2}
-	h3 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 3}
+	h3 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 3}
-	h4 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 4}
+	h4 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 4}
-	h5 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 5}
+	h5 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 5}
-	h6 := &HostInfo{vpnAddrs: []netip.Addr{netip.MustParseAddr("0.0.0.1")}, localIndexId: 6}
+	h6 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 6}
 
 	hm.unlockedAddHostInfo(h6, f)
 	hm.unlockedAddHostInfo(h5, f)
@@ -104,94 +112,243 @@ func TestHostMap_DeleteHostInfo(t *testing.T) {
 	hm.unlockedAddHostInfo(h2, f)
 	hm.unlockedAddHostInfo(h1, f)
 
-	// h6 should be deleted
+	// h6 is evicted by the MaxHostInfosPerVpnIp cap; the rest are newest-first.
-	assert.Nil(t, h6.next)
+	assert.Nil(t, hm.QueryIndex(h6.localIndexId))
-	assert.Nil(t, h6.prev)
+	assert.Equal(t, []uint32{1, 2, 3, 4, 5}, chainIds(t, hm, a))
-	h := hm.QueryIndex(h6.localIndexId)
-	assert.Nil(t, h)
 
-	// Make sure we go h1 -> h2 -> h3 -> h4 -> h5
+	// Delete primary; not final since siblings remain.
-	prim := hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.False(t, hm.DeleteHostInfo(h1))
-	assert.Equal(t, h1.localIndexId, prim.localIndexId)
+	assert.Equal(t, []uint32{2, 3, 4, 5}, chainIds(t, hm, a))
-	assert.Equal(t, h2.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h1.localIndexId, h2.prev.localIndexId)
-	assert.Equal(t, h3.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h4.prev.localIndexId)
-	assert.Equal(t, h5.localIndexId, h4.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h5.prev.localIndexId)
-	assert.Nil(t, h5.next)
 
-	// Delete primary
+	// Deleting the same hostinfo again must not report final while siblings remain and must not
-	hm.DeleteHostInfo(h1)
+	// disturb the list. The old chain code got this wrong: the first delete nil'd next/prev, so a
-	assert.Nil(t, h1.prev)
+	// second delete looked final and wiped lighthouse state out from under the live sibling.
-	assert.Nil(t, h1.next)
+	assert.False(t, hm.DeleteHostInfo(h1))
+	assert.Equal(t, []uint32{2, 3, 4, 5}, chainIds(t, hm, a))
 
-	// Make sure we go h2 -> h3 -> h4 -> h5
+	// Delete a middle node.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.False(t, hm.DeleteHostInfo(h3))
-	assert.Equal(t, h2.localIndexId, prim.localIndexId)
+	assert.Equal(t, []uint32{2, 4, 5}, chainIds(t, hm, a))
-	assert.Equal(t, h3.localIndexId, prim.next.localIndexId)
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h3.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h3.prev.localIndexId)
-	assert.Equal(t, h4.localIndexId, h3.next.localIndexId)
-	assert.Equal(t, h3.localIndexId, h4.prev.localIndexId)
-	assert.Equal(t, h5.localIndexId, h4.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h5.prev.localIndexId)
-	assert.Nil(t, h5.next)
 
-	// Delete in the middle
+	// Delete the tail.
-	hm.DeleteHostInfo(h3)
+	assert.False(t, hm.DeleteHostInfo(h5))
-	assert.Nil(t, h3.prev)
+	assert.Equal(t, []uint32{2, 4}, chainIds(t, hm, a))
-	assert.Nil(t, h3.next)
 
-	// Make sure we go h2 -> h4 -> h5
+	// Delete the head; h4 remains and becomes primary.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.False(t, hm.DeleteHostInfo(h2))
-	assert.Equal(t, h2.localIndexId, prim.localIndexId)
+	assert.Equal(t, []uint32{4}, chainIds(t, hm, a))
-	assert.Equal(t, h4.localIndexId, prim.next.localIndexId)
+	assert.Equal(t, h4, hm.QueryVpnAddr(a))
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h4.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h4.prev.localIndexId)
-	assert.Equal(t, h5.localIndexId, h4.next.localIndexId)
-	assert.Equal(t, h4.localIndexId, h5.prev.localIndexId)
-	assert.Nil(t, h5.next)
 
-	// Delete the tail
+	// Delete the only remaining item; final is true and the address is gone.
-	hm.DeleteHostInfo(h5)
+	assert.True(t, hm.DeleteHostInfo(h4))
-	assert.Nil(t, h5.prev)
+	assert.Empty(t, chainIds(t, hm, a))
-	assert.Nil(t, h5.next)
+	assert.Nil(t, hm.QueryVpnAddr(a))
 
-	// Make sure we go h2 -> h4
+	// Deleting an already-gone hostinfo is still final; nothing holds the address anymore.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	assert.True(t, hm.DeleteHostInfo(h4))
-	assert.Equal(t, h2.localIndexId, prim.localIndexId)
+	assert.Empty(t, chainIds(t, hm, a))
-	assert.Equal(t, h4.localIndexId, prim.next.localIndexId)
+}
-	assert.Nil(t, prim.prev)
-	assert.Equal(t, h4.localIndexId, h2.next.localIndexId)
-	assert.Equal(t, h2.localIndexId, h4.prev.localIndexId)
-	assert.Nil(t, h4.next)
 
-	// Delete the head
+// TestHostMap_MakePrimary_DeletedHostInfo covers promoting a hostinfo that lost a race with
-	hm.DeleteHostInfo(h2)
+// tunnel teardown: swapPrimary and AddRelay decide to promote while holding a stale pointer and
-	assert.Nil(t, h2.prev)
+// only take the write lock after a delete fully unlinked the hostinfo. MakePrimary must be a
-	assert.Nil(t, h2.next)
+// no-op, not a resurrection that installs an unmanaged primary.
+func TestHostMap_MakePrimary_DeletedHostInfo(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
 
-	// Make sure we only have h4
+	h1 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	h2 := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 2}
-	assert.Equal(t, h4.localIndexId, prim.localIndexId)
+	hm.unlockedAddHostInfo(h1, f)
-	assert.Nil(t, prim.prev)
+	hm.unlockedAddHostInfo(h2, f)
-	assert.Nil(t, prim.next)
-	assert.Nil(t, h4.next)
 
-	// Delete the only item
+	// h1 is fully deleted while another goroutine still holds a pointer to it.
-	hm.DeleteHostInfo(h4)
+	assert.False(t, hm.DeleteHostInfo(h1))
-	assert.Nil(t, h4.prev)
+	assert.Equal(t, []uint32{2}, chainIds(t, hm, a))
-	assert.Nil(t, h4.next)
 
-	// Make sure we have nil
+	// The stale promote must not bring it back.
-	prim = hm.QueryVpnAddr(netip.MustParseAddr("0.0.0.1"))
+	hm.MakePrimary(h1)
-	assert.Nil(t, prim)
+	assert.Equal(t, []uint32{2}, chainIds(t, hm, a))
+	assert.Equal(t, h2, hm.QueryVpnAddr(a))
+	assert.Nil(t, hm.QueryIndex(h1.localIndexId))
+}
+
+// TestHostMap_QueryVpnAddrsRelayFor_NonPrimary makes sure a relay established on an older
+// hostinfo is still found after a newer tunnel without relay state takes primary for the same
+// address. The lookup checks the primary first and falls back to the rest of the list.
+func TestHostMap_QueryVpnAddrsRelayFor_NonPrimary(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	relayAddr := netip.MustParseAddr("0.0.0.9")
+	target := netip.MustParseAddr("0.0.0.1")
+
+	older := &HostInfo{
+		vpnAddrs:     []netip.Addr{relayAddr},
+		localIndexId: 1,
+		relayState: RelayState{
+			relayForByAddr: map[netip.Addr]*Relay{},
+			relayForByIdx:  map[uint32]*Relay{},
+		},
+	}
+	older.relayState.InsertRelay(target, 100, &Relay{Type: ForwardingType, State: Established, LocalIndex: 100, PeerAddr: target})
+	hm.unlockedAddHostInfo(older, f)
+
+	// The relay is found on the primary.
+	h, r, err := hm.QueryVpnAddrsRelayFor([]netip.Addr{target}, relayAddr)
+	require.NoError(t, err)
+	assert.Equal(t, older, h)
+	assert.Equal(t, uint32(100), r.LocalIndex)
+
+	// A re-handshake with no relay state takes primary; the established relay on the older
+	// hostinfo must still be found through the fallback.
+	newer := &HostInfo{vpnAddrs: []netip.Addr{relayAddr}, localIndexId: 2}
+	hm.unlockedAddHostInfo(newer, f)
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, relayAddr))
+
+	h, r, err = hm.QueryVpnAddrsRelayFor([]netip.Addr{target}, relayAddr)
+	require.NoError(t, err)
+	assert.Equal(t, older, h)
+	assert.Equal(t, uint32(100), r.LocalIndex)
+
+	// No hostinfo at all is a plain miss.
+	_, _, err = hm.QueryVpnAddrsRelayFor([]netip.Addr{target}, netip.MustParseAddr("0.0.0.42"))
+	require.Error(t, err)
+}
+
+// TestHostMap_DeleteHostInfo_MultipleVpnAddrs exercises the case where a hostinfo carries more than one
+// vpnAddr and shares its next/prev chain with a live sibling. Deleting the head must not corrupt the
+// sibling: every address the sibling owns has to keep pointing at it. The pre-fix code unlinked the shared
+// chain once per vpnAddr, so on the first address it nil'd next/prev, and on the second address the node
+// looked already-detached: it dropped the map entry instead of promoting the sibling (and tripped the
+// isLastHostinfo relay teardown). See unlockedDeleteHostInfo.
+func TestHostMap_DeleteHostInfo_MultipleVpnAddrs(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+
+	f := &Interface{}
+
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	// Two tunnels for the same peer, each reachable at both a and b.
+	other := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 1}
+	head := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 2}
+
+	hm.unlockedAddHostInfo(other, f)
+	hm.unlockedAddHostInfo(head, f)
+
+	// head is primary for both addresses, other is next in each address's list.
+	assert.Equal(t, head, hm.QueryVpnAddr(a))
+	assert.Equal(t, head, hm.QueryVpnAddr(b))
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, b))
+
+	// Delete the head. other is still live, so it must become primary for BOTH addresses.
+	assert.False(t, hm.DeleteHostInfo(head))
+	assert.Equal(t, other, hm.QueryVpnAddr(a))
+	assert.Equal(t, other, hm.QueryVpnAddr(b))
+	assert.Equal(t, []uint32{1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{1}, chainIds(t, hm, b))
+
+	// head is fully removed from the index map.
+	assert.Nil(t, hm.QueryIndex(head.localIndexId))
+}
+
+// TestHostMap_DeleteHostInfo_DivergentVpnAddrs covers chained hostinfos for the same peer whose
+// vpnAddrs sets differ (a re-handshake cert added a second address). Deleting the superset node
+// must not promote a sibling to an address it does not own.
+func TestHostMap_DeleteHostInfo_DivergentVpnAddrs(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	// sub owns only a; super (a newer handshake) owns a and b.
+	sub := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
+	super := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 2}
+	hm.unlockedAddHostInfo(sub, f)
+	hm.unlockedAddHostInfo(super, f)
+
+	assert.Equal(t, []uint32{2, 1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{2}, chainIds(t, hm, b))
+
+	// Delete super: a promotes to sub (which owns it); b has no remaining owner and must be
+	// removed, not dangled at sub (which does not own b).
+	assert.False(t, hm.DeleteHostInfo(super))
+	assert.Equal(t, []uint32{1}, chainIds(t, hm, a))
+	assert.Empty(t, chainIds(t, hm, b))
+	assert.Equal(t, sub, hm.QueryVpnAddr(a))
+	assert.Nil(t, hm.QueryVpnAddr(b))
+	assert.Nil(t, hm.QueryIndex(super.localIndexId))
+
+	// Deleting sub cleans up fully.
+	assert.True(t, hm.DeleteHostInfo(sub))
+	assert.Nil(t, hm.QueryVpnAddr(a))
+	assertHostMapInvariants(t, hm)
+}
+
+// TestHostMap_AddDivergentOverlap covers a new hostinfo claiming addresses currently owned by two
+// DIFFERENT hostinfos. The old single shared next/prev chain overwrote a pointer and orphaned one
+// of them (in Indexes but unreachable via its address); independent per-address lists cannot.
+func TestHostMap_AddDivergentOverlap(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+	f := &Interface{}
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	hiA := &HostInfo{vpnAddrs: []netip.Addr{a}, localIndexId: 1}
+	hiP := &HostInfo{vpnAddrs: []netip.Addr{b}, localIndexId: 2}
+	hm.unlockedAddHostInfo(hiA, f)
+	hm.unlockedAddHostInfo(hiP, f)
+
+	hiB := &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: 3}
+	hm.unlockedAddHostInfo(hiB, f)
+
+	assert.Equal(t, []uint32{3, 1}, chainIds(t, hm, a))
+	assert.Equal(t, []uint32{3, 2}, chainIds(t, hm, b))
+	// hiA is still reachable via its address (not orphaned) and still indexed.
+	assert.Contains(t, chainIds(t, hm, a), hiA.localIndexId)
+	assert.NotNil(t, hm.QueryIndex(hiA.localIndexId))
+}
+
+// TestHostMap_MaxHostInfosPerVpnIp_MultipleVpnAddrs verifies the MaxHostInfosPerVpnIp overflow prune
+// (unlockedInnerAddHostInfo calls unlockedDeleteHostInfo on the oldest node once the chain is too long)
+// still behaves when hostinfos carry more than one vpnAddr. The pruned node is always the tail, so it is
+// primary for none of the addresses, and both address chains must stay consistent afterwards.
+func TestHostMap_MaxHostInfosPerVpnIp_MultipleVpnAddrs(t *testing.T) {
+	l := test.NewLogger()
+	hm := newHostMap(l)
+
+	f := &Interface{}
+
+	a := netip.MustParseAddr("0.0.0.1")
+	b := netip.MustParseAddr("0.0.0.2")
+
+	// Add one more than the cap, newest last so it becomes head. Every hostinfo owns both a and b.
+	hostinfos := make([]*HostInfo, 0, MaxHostInfosPerVpnIp+1)
+	for i := 0; i <= MaxHostInfosPerVpnIp; i++ {
+		hostinfos = append(hostinfos, &HostInfo{vpnAddrs: []netip.Addr{a, b}, localIndexId: uint32(i + 1)})
+	}
+	// Add oldest first (highest index in our slice) so the very first one added is the overflow victim.
+	for i := len(hostinfos) - 1; i >= 0; i-- {
+		hm.unlockedAddHostInfo(hostinfos[i], f)
+	}
+
+	oldest := hostinfos[len(hostinfos)-1]
+
+	// The oldest hostinfo was pruned from both lists and the index map.
+	assert.Nil(t, hm.QueryIndex(oldest.localIndexId))
+
+	// Both addresses hold exactly MaxHostInfosPerVpnIp survivors in the same order; oldest is absent.
+	require.Len(t, chainIds(t, hm, a), MaxHostInfosPerVpnIp)
+	assert.Equal(t, chainIds(t, hm, a), chainIds(t, hm, b), "both addresses must list the same survivors in the same order")
+	assert.NotContains(t, chainIds(t, hm, a), oldest.localIndexId)
+	assert.Equal(t, hm.QueryVpnAddr(a), hm.QueryVpnAddr(b))
 }
 
 func TestHostMap_reload(t *testing.T) {

inside.go

@@ -2,7 +2,6 @@ package nebula
 
 import (
 	"context"
-	"io"
 	"log/slog"
 	"net/netip"
 
@@ -10,17 +9,10 @@ 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(pkt wire.TunPacket, fwPacket *firewall.Packet, nb []byte, sendBatch *batch.SendBatch, rejectBuf []byte, q int, localCache firewall.ConntrackCache) {
+func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []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) {
@@ -45,10 +37,7 @@ func (f *Interface) consumeInsidePacket(pkt wire.TunPacket, fwPacket *firewall.P
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
 		// TUN device.
 		if immediatelyForwardToSelf {
-			err := pkt.PerSegment(func(seg []byte) error {
+			_, err := f.readers[q].Write(packet)
-				_, werr := f.readers[q].Write(seg)
-				return werr
-			})
 			if err != nil {
 				f.l.Error("Failed to forward to tun", "error", err)
 			}
@@ -64,23 +53,11 @@ func (f *Interface) consumeInsidePacket(pkt wire.TunPacket, fwPacket *firewall.P
 	}
 
 	hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
-		// borrowed: PerSegment builds each segment in the kernel-supplied pkt
+		hh.cachePacket(f.l, header.Message, 0, packet, f.sendMessageNow, f.cachedPacketMetrics)
-		// 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, rejectBuf, q)
+		f.rejectInside(packet, out, 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,
@@ -96,9 +73,10 @@ func (f *Interface) consumeInsidePacket(pkt wire.TunPacket, fwPacket *firewall.P
 
 	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendInsideMessage(hostinfo, pkt, nb, sendBatch)
+		f.sendNoMetrics(header.Message, 0, hostinfo.ConnectionState, hostinfo, netip.AddrPort{}, packet, nb, out, q)
+
 	} else {
-		f.rejectInside(packet, rejectBuf, q)
+		f.rejectInside(packet, out, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("dropping outbound packet",
 				"fwPacket", fwPacket,
@@ -108,151 +86,8 @@ func (f *Interface) consumeInsidePacket(pkt wire.TunPacket, fwPacket *firewall.P
 	}
 }
 
-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 {
+	if !f.firewall.OutboundSendReject {
 		return
 	}
 
@@ -268,7 +103,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 }
 
 func (f *Interface) rejectOutside(packet []byte, ci *ConnectionState, hostinfo *HostInfo, nb, out []byte, q int) {
-	if !f.firewall.OutSendReject {
+	if !f.firewall.InboundSendReject {
 		return
 	}
 
@@ -440,13 +275,21 @@ func (f *Interface) sendTo(t header.MessageType, st header.MessageSubType, ci *C
 	f.sendNoMetrics(t, st, ci, hostinfo, remote, p, nb, out, 0)
 }
 
-func (f *Interface) prepareSendVia(via *HostInfo,
+// 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,
 	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()
@@ -468,7 +311,7 @@ func (f *Interface) prepareSendVia(via *HostInfo,
 			"headerLen", len(out),
 			"cipherOverhead", via.ConnectionState.eKey.Overhead(),
 		)
-		return nil, io.ErrShortBuffer
+		return
 	}
 
 	// The header bytes are written to the 'out' slice; Grow the slice to hold the header and associated data payload.
@@ -488,44 +331,20 @@ func (f *Interface) prepareSendVia(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(toSend, via.remote)
+	err = f.writers[0].WriteTo(out, via.GetRemote())
 	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) {
 	if ci.eKey == nil {
 		return
 	}
-	useRelay := !remote.IsValid() && !hostinfo.remote.IsValid()
+	useRelay := !remote.IsValid() && !hostinfo.GetRemote().IsValid()
 	fullOut := out
 
 	if useRelay {
@@ -546,17 +365,12 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 
 	//l.WithField("trace", string(debug.Stack())).Error("out Header ", &Header{Version, t, st, 0, hostinfo.remoteIndexId, c}, p)
 	out = header.Encode(out, header.Version, t, st, hostinfo.remoteIndexId, c)
-	f.connectionManager.Out(hostinfo)
+	// We rebound since this tunnel last sent, so the local network moved. Ask the lighthouse to have the far side
-
+	// punch at where we are now, which primes their conntrack the same way a handshake would.
-	// Query our LH if we haven't since the last time we've been rebound, this will cause the remote to punch against
+	if f.connectionManager.Out(hostinfo) && t != header.CloseTunnel {
-	// all our addrs and enable a faster roaming.
-	if t != header.CloseTunnel && 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) {
-			f.l.Debug("Lighthouse update triggered for punch due to rebind counter",
+			f.l.Debug("Lighthouse update triggered for punch due to rebind epoch",
 				"vpnAddrs", hostinfo.vpnAddrs,
 			)
 		}
@@ -572,7 +386,6 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 			"error", err,
 			"udpAddr", remote,
 			"counter", c,
-			"attemptedCounter", c,
 		)
 		return
 	}
@@ -585,12 +398,12 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 				"udpAddr", remote,
 			)
 		}
-	} else if hostinfo.remote.IsValid() {
+	} else if hr := hostinfo.GetRemote(); hr.IsValid() {
-		err = f.writers[q].WriteTo(out, hostinfo.remote)
+		err = f.writers[q].WriteTo(out, hr)
 		if err != nil {
 			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
 				"error", err,
-				"udpAddr", remote,
+				"udpAddr", hr,
 			)
 		}
 	} else {

interface.go

@@ -4,24 +4,22 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"io"
 	"log/slog"
 	"net/netip"
-	"runtime"
+	"slices"
 	"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"
 )
 
@@ -41,7 +39,6 @@ type InterfaceConfig struct {
 	DropLocalBroadcast bool
 	DropMulticast      bool
 	routines           int
-	batchSize          int
 	MessageMetrics     *MessageMetrics
 	version            string
 	relayManager       *relayManager
@@ -52,13 +49,6 @@ type InterfaceConfig struct {
 	reQueryWait     time.Duration
 
 	ConntrackCacheTimeout time.Duration
-
-	// 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
 }
 
@@ -81,18 +71,8 @@ type Interface struct {
 	dropLocalBroadcast    bool
 	dropMulticast         bool
 	routines              int
-	batchSize             int
 	disconnectInvalid     atomic.Bool
 	closed                atomic.Bool
-	// 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
@@ -102,19 +82,17 @@ type Interface struct {
 	sendRecvErrorConfig   recvErrorConfig
 	acceptRecvErrorConfig recvErrorConfig
 
-	// rebindCount is used to decide if an active tunnel should trigger a punch notification through a lighthouse
+	// rebindEpoch bumps every time the udp listener is rebound, which means the local network moved. Tunnels
-	rebindCount int8
+	// compare it against their own copy to decide they need a punch from the far side. Read on every send, only
+	// written on a rebind, so the cache line stays shared across the routines.
+	rebindEpoch atomic.Uint32
 	version     string
 
 	conntrackCacheTimeout time.Duration
 
 	ctx     context.Context
 	writers []udp.Conn
-	readers []tio.Queue
+	readers []io.ReadWriteCloser
-	// 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.
@@ -211,11 +189,9 @@ 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([]tio.Queue, c.routines),
+		readers:               make([]io.ReadWriteCloser, c.routines),
-		batchers:              make([]batch.RxBatcher, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -224,7 +200,6 @@ 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,
@@ -242,6 +217,9 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 
 	ifce.connectionManager.intf = ifce
 
+	// Held until Close so waiting on the interface blocks until the resources are actually released
+	ifce.wg.Add(1)
+
 	return ifce, nil
 }
 
@@ -274,40 +252,27 @@ 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 {
-			if err = f.inside.NewMultiQueueReader(); err != nil {
+			reader, err = f.inside.NewMultiQueueReader()
+			if 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
+	// On error the caller owns the cleanup, Control.Start cancels the service context
+	// before releasing our resources so a waiter never observes a live context
 	if err = f.inside.Activate(); err != nil {
-		f.wg.Done()
-		f.inside.Close()
 		return err
 	}
 
 	return nil
 }
 
-func (f *Interface) run() (func() error, error) {
+func (f *Interface) run() {
 	// Launch n queues to read packets from udp
 	for i := 0; i < f.routines; i++ {
 		f.wg.Go(func() {
@@ -322,13 +287,14 @@ func (f *Interface) run() (func() error, error) {
 		})
 	}
 
-	return func() error {
+}
+
+func (f *Interface) wait() error {
 	f.wg.Wait()
 	if e := f.fatalErr.Load(); e != nil {
 		return *e
 	}
 	return nil
-	}, nil
 }
 
 // onFatal stores the first fatal reader error, and calls triggerShutdown if it was the first one
@@ -352,23 +318,19 @@ 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)
 
-	listener := func(fromUdpAddr netip.AddrPort, payload []byte, meta udp.RxMeta) {
+	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
-		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, payload, h, fwPacket, lhh, nb, i, ctCache.Get(), meta)
+		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get())
-	}
+	})
 
-	flusher := func() {
+	// An error after teardown began is shutdown noise, the closed flag covers resources
-		if err := f.batchers[i].Flush(); err != nil {
+	// Close releases itself and the cancelled ctx covers ones torn down by their owners
-			f.l.Error("Failed to flush tun coalescer", "error", err)
+	// reacting to it, like the user device pipes
-		}
+	if err != nil && !f.closed.Load() && f.ctx.Err() == nil {
-	}
-
-	err := li.ListenOut(listener, flusher)
-
-	if err != nil && !f.closed.Load() {
 		f.l.Error("Error while reading inbound packet, closing", "error", err)
 		f.onFatal(err)
 	}
@@ -376,57 +338,29 @@ func (f *Interface) listenOut(i int) {
 	f.l.Debug("underlay reader is done", "reader", i)
 }
 
-func (f *Interface) listenIn(reader tio.Queue, q int) {
+func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
-	// Pinning this thread (and goroutine) to a single CPU keeps every sendmmsg from this goroutine going through the
+	packet := make([]byte, mtu)
-	// same TX ring on the nic, so the wire sees per-flow order.
+	out := make([]byte, mtu)
-	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(tunPackets, packetMem)
+		n, err := reader.Read(packet)
 		if err != nil {
-			if !f.closed.Load() {
+			// Same shutdown noise handling as listenOut
-				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", q)
+			if !f.closed.Load() && f.ctx.Err() == nil {
+				f.l.Error("Error while reading outbound packet, closing", "error", err, "reader", i)
 				f.onFatal(err)
 			}
 			break
 		}
 
-		ctCache := conntrackCache.Get()
+		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, 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", q)
+	f.l.Debug("overlay reader is done", "reader", i)
 }
 
 func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
@@ -435,7 +369,6 @@ 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)
@@ -453,13 +386,22 @@ func (f *Interface) reloadDisconnectInvalid(c *config.C) {
 }
 
 func (f *Interface) reloadFirewall(c *config.C) {
-	//TODO: need to trigger/detect if the certificate changed too
+	cs := f.pki.getCertState()
-	if c.HasChanged("firewall") == false {
+	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 {
 		f.l.Debug("No firewall config change detected")
 		return
 	}
 
-	fw, err := NewFirewallFromConfig(f.l, f.pki.getCertState(), c)
+	fw, err := NewFirewallFromConfig(f.l, cs, c)
 	if err != nil {
 		f.l.Error("Error while creating firewall during reload", "error", err)
 		return
@@ -559,20 +501,6 @@ 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()
@@ -583,11 +511,7 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 	certInitiatingVersion := metrics.GetOrRegisterGauge("certificate.initiating_version", nil)
 	certMaxVersion := metrics.GetOrRegisterGauge("certificate.max_version", nil)
 
-	for {
+	emit := func() {
-		select {
-		case <-ctx.Done():
-			return
-		case <-ticker.C:
 		f.firewall.EmitStats()
 		f.handshakeManager.EmitStats()
 		udpStats()
@@ -604,6 +528,18 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 			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()
+		}
 	}
 }
 
@@ -615,9 +551,15 @@ func (f *Interface) GetCertState() *CertState {
 	return f.pki.getCertState()
 }
 
+// Close releases the interface's resources: the udp sockets and the tun device.
+// It is idempotent and safe to call at any point in the lifecycle, including on an interface that never activated,
+// calls after the first return nil without doing anything.
 func (f *Interface) Close() error {
+	if !f.closed.CompareAndSwap(false, true) {
+		return nil
+	}
+
 	var errs []error
-	f.closed.Store(true)
 
 	// Release the udp readers
 	for i, u := range f.writers {
@@ -633,6 +575,8 @@ func (f *Interface) Close() error {
 	if closeErr != nil {
 		errs = append(errs, closeErr)
 	}
+
+	// Release the construction token so waiters know the resources are gone
 	f.wg.Done()
 	return errors.Join(errs...)
 }

interface_emit_test.go

@@ -0,0 +1,73 @@
+//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

@@ -0,0 +1,120 @@
+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

@@ -4,27 +4,55 @@ import (
 	"encoding/binary"
 
 	"golang.org/x/net/ipv4"
+	"golang.org/x/net/ipv6"
 )
 
 const (
-	// Need 96 bytes for the largest reject packet:
+	// MaxIPv4RejectPacketSize is the largest IPv4 reject packet:
 	// - 20 byte ipv4 header
 	// - 8 byte icmpv4 header
 	// - 68 byte body (60 byte max orig ipv4 header + 8 byte orig icmpv4 header)
-	MaxRejectPacketSize = ipv4.HeaderLen + 8 + 60 + 8
+	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
 )
 
 func CreateRejectPacket(packet []byte, out []byte) []byte {
-	if len(packet) < ipv4.HeaderLen || int(packet[0]>>4) != ipv4.Version {
+	if len(packet) < 1 {
 		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)
+	default:
+		return nil
+	}
 }
 
 func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
@@ -35,11 +63,16 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
 		return nil
 	}
 
-	// ICMP reply includes original header and first 8 bytes of the packet
+	// Do not generate ICMP errors in response to ICMP error packets
-	packetLen := len(packet)
+	if packet[9] == 1 && len(packet) > ihl {
-	if packetLen > ihl+8 {
+		icmpType := packet[ihl]
-		packetLen = ihl + 8
+		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)
 
 	outLen := ipv4.HeaderLen + 8 + packetLen
 	if outLen > cap(out) {
@@ -72,7 +105,7 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
 	// ICMP Destination Unreachable
 	icmpOut := out[ipv4.HeaderLen:]
 	icmpOut[0] = 3  // type (Destination unreachable)
-	icmpOut[1] = 3 // code (Port unreachable error)
+	icmpOut[1] = 13 // code (Communication administratively prohibited)
 	icmpOut[2] = 0  // checksum
 	icmpOut[3] = 0  //  .
 	icmpOut[4] = 0  // unused
@@ -165,7 +198,193 @@ 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) {
@@ -199,6 +418,43 @@ func CreateICMPEchoResponse(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.
 //
@@ -236,3 +492,18 @@ 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

@@ -1,11 +1,13 @@
 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) {
@@ -43,7 +45,7 @@ func Test_CreateRejectPacket(t *testing.T) {
 	}
 	b = append(b, []byte{0, 3, 0, 4, 0, 0, 0, 0}...)
 
-	expectedLen = MaxRejectPacketSize
+	expectedLen = maxIPv4RejectPacketSize
 	out = make([]byte, MaxRejectPacketSize)
 	rejectPacket = CreateRejectPacket(b, out)
 	assert.NotNil(t, rejectPacket)
@@ -71,3 +73,404 @@ 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

@@ -36,6 +36,10 @@ type LightHouse struct {
 	myVpnNetworksTable *bart.Lite
 	punchy             *Punchy
 
+	// localAddrsFn enumerates the underlay addresses we advertise. It is a field so tests can supply simulated
+	// addresses rather than whatever this machine's NICs happen to be. Set it before Start.
+	localAddrsFn func(*LocalAllowList) []netip.Addr
+
 	// Local cache of answers from light houses
 	// map of vpn addr to answers
 	addrMap map[netip.Addr]*RemoteList
@@ -107,6 +111,10 @@ func NewLightHouseFromConfig(ctx context.Context, l *slog.Logger, c *config.C, c
 		queryChan:          make(chan netip.Addr, c.GetUint32("handshakes.query_buffer", 64)),
 		l:                  l,
 	}
+	h.localAddrsFn = func(al *LocalAllowList) []netip.Addr {
+		return localAddrs(h.l, al)
+	}
+
 	lighthouses := make([]netip.Addr, 0)
 	h.lighthouses.Store(&lighthouses)
 	staticList := make(map[netip.Addr]struct{})
@@ -272,16 +280,18 @@ 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.
-		// Entries no longer present must have their (possible) background DNS goroutines stopped.
+		ourselves := lh.myVpnNetworks[0].Addr()
-		if existingStaticList := lh.staticList.Load(); existingStaticList != nil {
+		oldStaticList := lh.staticList.Load()
+		if oldStaticList != nil {
 			lh.RLock()
-			for staticVpnAddr := range *existingStaticList {
+			for staticVpnAddr := range *oldStaticList {
 				if am, ok := lh.addrMap[staticVpnAddr]; ok && am != nil {
-					am.hr.Cancel()
+					am.ResetForOwner(ourselves)
 				}
 			}
 			lh.RUnlock()
 		}
+
 		// Build a new list based on current config.
 		staticList := make(map[netip.Addr]struct{})
 		err := lh.loadStaticMap(c, staticList)
@@ -289,6 +299,21 @@ 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") {
@@ -901,7 +926,7 @@ func (lh *LightHouse) SendUpdate() {
 	}
 
 	lal := lh.GetLocalAllowList()
-	for _, e := range localAddrs(lh.l, lal) {
+	for _, e := range lh.localAddrsFn(lal) {
 		if lh.myVpnNetworksTable.Contains(e) {
 			continue
 		}
@@ -1401,6 +1426,9 @@ func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpn
 
 	remoteAllowList := lhh.lh.GetRemoteAllowList()
 	for _, a := range n.Details.V4AddrPorts {
+		if a == nil {
+			continue
+		}
 		b := protoV4AddrPortToNetAddrPort(a)
 		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
 			lhh.lh.punchy.Schedule(b, detailsVpnAddr)
@@ -1408,6 +1436,9 @@ func (lhh *LightHouseHandler) handleHostPunchNotification(n *NebulaMeta, fromVpn
 	}
 
 	for _, a := range n.Details.V6AddrPorts {
+		if a == nil {
+			continue
+		}
 		b := protoV6AddrPortToNetAddrPort(a)
 		if remoteAllowList.Allow(detailsVpnAddr, b.Addr()) {
 			lhh.lh.punchy.Schedule(b, detailsVpnAddr)
@@ -1437,7 +1468,7 @@ func protoV6AddrPortToNetAddrPort(ap *V6AddrPort) netip.AddrPort {
 	b := [16]byte{}
 	binary.BigEndian.PutUint64(b[:8], ap.Hi)
 	binary.BigEndian.PutUint64(b[8:], ap.Lo)
-	return netip.AddrPortFrom(netip.AddrFrom16(b), uint16(ap.Port))
+	return netip.AddrPortFrom(netip.AddrFrom16(b).Unmap(), uint16(ap.Port))
 }
 
 func netAddrToProtoAddr(addr netip.Addr) *Addr {
@@ -1477,9 +1508,11 @@ func (d *NebulaMetaDetails) GetRelays() []netip.Addr {
 
 	if len(d.RelayVpnAddrs) > 0 {
 		for _, r := range d.RelayVpnAddrs {
+			if r != nil {
 				relays = append(relays, protoAddrToNetAddr(r))
 			}
 		}
+	}
 	return relays
 }
 

lighthouse_test.go

@@ -303,6 +303,132 @@ 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

@@ -5,10 +5,7 @@ import (
 	"fmt"
 	"log/slog"
 	"net"
-	"net/http"
-	_ "net/http/pprof"
 	"net/netip"
-	"runtime"
 	"runtime/debug"
 	"strings"
 	"time"
@@ -36,9 +33,6 @@ 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)
@@ -136,6 +130,17 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	udpConns := make([]udp.Conn, routines)
 	port := c.GetInt("listen.port", 0)
 
+	// Callers get no handle to these until the Control is returned, release them on any error.
+	defer func() {
+		if reterr != nil {
+			for _, u := range udpConns {
+				if u != nil {
+					_ = u.Close()
+				}
+			}
+		}
+	}()
+
 	if !configTest {
 		rawListenHost := c.GetString("listen.host", "0.0.0.0")
 		var listenHost netip.Addr
@@ -200,7 +205,7 @@ 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.getCertState(), hostMap, c)
+	ds, err := newDnsServerFromConfig(ctx, l, pki, hostMap, c)
 	if err != nil {
 		l.Warn("Failed to start DNS responder", "error", err)
 	}
@@ -221,13 +226,11 @@ 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,
 	}
 
@@ -245,7 +248,6 @@ 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)
@@ -266,6 +268,8 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 
 	attachCommands(l, c, ssh, ifce)
 
+	networkChanges := udp.NewNetworkChangeMonitor(ctx, l, c)
+
 	return &Control{
 		state:                  StateReady,
 		f:                      ifce,
@@ -276,57 +280,11 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		statsStart:             stats.Start,
 		dnsStart:               ds.Start,
 		lighthouseStart:        lightHouse.StartUpdateWorker,
+		networkChangeStart:     networkChanges.Start,
 		connectionManagerStart: connManager.Start,
 	}, 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

@@ -147,62 +147,6 @@ 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

@@ -13,7 +13,6 @@ import (
 
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
-	"github.com/slackhq/nebula/udp"
 	"golang.org/x/net/ipv4"
 )
 
@@ -23,7 +22,7 @@ const (
 
 var ErrOutOfWindow = errors.New("out of window packet")
 
-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) {
+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) {
 	err := h.Parse(packet)
 	if err != nil {
 		// Hole punch packets are 0 or 1 byte big, so lets ignore printing those errors
@@ -103,27 +102,31 @@ func (f *Interface) readOutsidePackets(via ViaSender, packet []byte, h *header.H
 		return
 	}
 
+	if len(packet) < header.Len+hostinfo.ConnectionState.dKey.Overhead() {
+		f.messageMetrics.RxInvalid(1)
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			f.l.Debug("packet too small", "from", via, "length", len(packet))
+		}
+		return
+	}
+
 	// All remaining packets are encrypted
-	ci := hostinfo.ConnectionState
-	if !ci.window.Check(f.l, h.MessageCounter) {
-		return
-	}
-
-	// Relay packets are special
 	if isMessageRelay {
-		f.handleOutsideRelayPacket(hostinfo, via, packet, h, fwPacket, lhf, nb, q, localCache, meta)
+		// Relay packets are special, this branch should always early-return
+		if err = hostinfo.ConnectionState.VerifyRelay(f.l, h.MessageCounter, packet, nb); err != nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				hostinfo.logger(f.l).Debug("Failed to verify relay packet", "error", err, "from", via, "header", h)
+			}
+			return
+		}
+		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache)
 		return
 	}
 
-	out := f.batchers[q].Reserve(len(packet))[:0]
+	out, err = hostinfo.ConnectionState.Decrypt(f.l, h.MessageCounter, out, packet, nb)
-	out, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("Failed to decrypt packet",
+			hostinfo.logger(f.l).Debug("Failed to decrypt packet", "error", err, "from", via, "header", h)
-				"error", err,
-				"from", via,
-				"header", h,
-			)
 		}
 		return
 	}
@@ -136,7 +139,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, packet []byte, h *header.H
 	case header.Message:
 		switch h.Subtype {
 		case header.MessageNone:
-			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache, meta)
+			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected message subtype seen", "from", via, "header", h)
 			return
@@ -151,7 +154,8 @@ func (f *Interface) readOutsidePackets(via ViaSender, packet []byte, h *header.H
 		case header.TestReply:
 			// No-op, useful for the Roaming and connectionManager side-effects above
 		case header.TestRequest:
-			f.send(header.Test, header.TestReply, ci, hostinfo, out, nb, out)
+			//recycle the input packet ciphertext as our output buffer
+			f.send(header.Test, header.TestReply, hostinfo.ConnectionState, hostinfo, out, nb, packet)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected test subtype seen", "from", via, "header", h)
 			return
@@ -169,22 +173,9 @@ func (f *Interface) readOutsidePackets(via ViaSender, packet []byte, h *header.H
 	}
 }
 
-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) {
+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) {
-	// The entire body is sent as AD, not encrypted.
+	// Successfully validated the thing. Get rid of the Relay header and the AEAD tag
-	// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
+	signedPayload := packet[header.Len : len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
-	// 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
-	// otherwise malformed (meaning, there is no trailing 16 byte AEAD value), then this will result in at worst a 0-length slice
-	// which will gracefully fail in the DecryptDanger call.
-	signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
-	signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
-	// 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.
-	signedPayload = signedPayload[header.Len:]
 	// Pull the Roaming parts up here, and return in all call paths.
 	f.handleHostRoaming(hostinfo, via)
 	// Track usage of both the HostInfo and the Relay for the received & authenticated packet
@@ -196,8 +187,7 @@ 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",
-			"vpnAddrs", hostinfo.vpnAddrs,
+			"relayRemoteIndex", h.RemoteIndex,
-			"remoteIndex", h.RemoteIndex,
 		)
 		return
 	}
@@ -209,19 +199,18 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		via = ViaSender{
 			UdpAddr:   via.UdpAddr,
 			relayHI:   hostinfo,
-			remoteIdx: relay.RemoteIndex,
 			relay:     relay,
 			IsRelayed: true,
 		}
-		f.readOutsidePackets(via, signedPayload, h, fwPacket, lhf, nb, q, localCache, meta)
+		f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache)
 	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
 		}
@@ -230,10 +219,10 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		if targetRelay.State == Established {
 			switch targetRelay.Type {
 			case ForwardingType:
-				// Forward this packet through the relay tunnel
+				// Forward this packet through the relay tunnel, rebuilding it in place.
-				// Find the target HostInfo //todo it would potentially be nice to batch these
+				// Encode overwrites the old outer header, and the new AEAD tag lands where the old one was
-				out := f.batchers[q].Reserve(len(packet) + header.Len + hostinfo.ConnectionState.dKey.Overhead())[:0]
+				fwdBuf := packet[:0:len(packet)] // Cap to len(packet) to protect memory from a larger parent buffer
-				f.SendVia(targetHI, targetRelay, signedPayload, nb, out, false)
+				f.SendVia(targetHI, targetRelay, signedPayload, nb, fwdBuf, true)
 			case TerminalType:
 				hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
 				return
@@ -273,7 +262,8 @@ func (f *Interface) sendCloseTunnel(h *HostInfo) {
 }
 
 func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
-	if !via.IsRelayed && hostinfo.remote != via.UdpAddr {
+	curRemote := hostinfo.GetRemote()
+	if !via.IsRelayed && curRemote != via.UdpAddr {
 		if !f.lightHouse.GetRemoteAllowList().AllowAll(hostinfo.vpnAddrs, via.UdpAddr.Addr()) {
 			if f.l.Enabled(context.Background(), slog.LevelDebug) {
 				hostinfo.logger(f.l).Debug("lighthouse.remote_allow_list denied roaming", "newAddr", via.UdpAddr)
@@ -285,7 +275,7 @@ func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
 			if f.l.Enabled(context.Background(), slog.LevelDebug) {
 				hostinfo.logger(f.l).Debug("Suppressing roam back to previous remote",
 					"suppressSeconds", RoamingSuppressSeconds,
-					"udpAddr", hostinfo.remote,
+					"udpAddr", curRemote,
 					"newAddr", via.UdpAddr,
 				)
 			}
@@ -293,11 +283,11 @@ func (f *Interface) handleHostRoaming(hostinfo *HostInfo, via ViaSender) {
 		}
 
 		hostinfo.logger(f.l).Info("Host roamed to new udp ip/port.",
-			"udpAddr", hostinfo.remote,
+			"udpAddr", curRemote,
 			"newAddr", via.UdpAddr,
 		)
 		hostinfo.lastRoam = time.Now()
-		hostinfo.lastRoamRemote = hostinfo.remote
+		hostinfo.lastRoamRemote = curRemote
 		hostinfo.SetRemote(via.UdpAddr)
 	}
 
@@ -417,16 +407,14 @@ func parseV6(data []byte, incoming bool, fp *firewall.Packet) error {
 			if dataLen <= offset+1 {
 				break
 			}
-
+			next = (int(data[offset+1]) + 2) << 2
-			next = int(data[offset+1]+2) << 2
 
 		default:
 			// Normal ipv6 header length processing
 			if dataLen <= offset+1 {
 				break
 			}
-
+			next = (int(data[offset+1]) + 1) << 3
-			next = int(data[offset+1]+1) << 3
 		}
 
 		if next <= 0 {
@@ -501,75 +489,7 @@ func parseV4(data []byte, incoming bool, fp *firewall.Packet) error {
 	return nil
 }
 
-func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, out []byte, packet []byte, h *header.H, nb []byte) ([]byte, error) {
+func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache) {
-	var err error
-	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], mc, nb)
-	if err != nil {
-		return nil, err
-	}
-
-	if !hostinfo.ConnectionState.window.Update(f.l, mc) {
-		return nil, ErrOutOfWindow
-	}
-
-	return out, nil
-}
-
-// 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",
@@ -593,7 +513,7 @@ func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, p
 		return
 	}
 
-	err = f.batchers[q].Commit(out)
+	_, err = f.readers[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}
@@ -640,10 +560,11 @@ func (f *Interface) handleRecvError(addr netip.AddrPort, h *header.H) {
 		return
 	}
 
-	if hostinfo.remote.IsValid() && hostinfo.remote != addr {
+	hr := hostinfo.GetRemote()
+	if hr.IsValid() && hr != addr {
 		f.l.Info("Someone spoofing recv_errors?",
 			"addr", addr,
-			"hostinfoRemote", hostinfo.remote,
+			"hostinfoRemote", hr,
 		)
 		return
 	}

outside_test.go

@@ -640,3 +640,38 @@ func serializeAH(ah *layers.IPSecAH) []byte {
 
 	return buf.Bytes()
 }
+
+// Test_newPacket_v6ExtHeaderOverflow is a regression test for the IPv6 extension-header
+// length uint8 overflow in parseV6. A Destination-Options header with HdrExtLen=255 spans
+// (255+1)*8 = 2048 bytes, so the real transport header sits at offset 2088. Before the fix
+// the advance was computed in uint8 and wrapped to 0 (then clamped to 8), so the firewall
+// read the transport header ~2KB too early from attacker-controlled option bytes while the
+// host OS parses the real header, a firewall port/proto bypass. The fix makes parseV6 land
+// on the same offset the host does.
+func Test_newPacket_v6ExtHeaderOverflow(t *testing.T) {
+	p := &firewall.Packet{}
+
+	const (
+		hdrLen      = 40              // IPv6 header
+		extLen      = 2048            // (255+1)*8, the true Destination-Options header size
+		realTCPAt   = hdrLen + extLen // 2088, where the host reads the transport header
+		forgedTCPAt = hdrLen + 8      // 48, where the pre-fix wrapped+clamped walk landed
+	)
+
+	pkt := make([]byte, realTCPAt+4)
+	pkt[0] = 0x60                                   // version 6
+	pkt[6] = byte(layers.IPProtocolIPv6Destination) // NextHeader -> Destination Options
+	pkt[40] = byte(firewall.ProtoTCP)               // Dest-Options NextHeader -> TCP
+	pkt[41] = 255                                   // HdrExtLen = 255
+
+	// Forged transport header at the pre-fix (wrong) offset: dst port 443.
+	binary.BigEndian.PutUint16(pkt[forgedTCPAt+2:forgedTCPAt+4], 443)
+	// Real transport header at the offset the host actually uses: dst port 22.
+	binary.BigEndian.PutUint16(pkt[realTCPAt+2:realTCPAt+4], 22)
+
+	require.NoError(t, newPacket(pkt, true, p))
+	assert.Equal(t, uint8(firewall.ProtoTCP), p.Protocol)
+	// LocalPort is the destination port for incoming traffic. It must be the real port (22)
+	// the host delivers to, not the forged 443 at the overflowed offset.
+	assert.Equal(t, uint16(22), p.LocalPort, "firewall must parse the real transport header, not the overflowed offset")
+}

overlay/batch/coalesce_core.go

@@ -1,149 +0,0 @@
-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

@@ -1,136 +0,0 @@
-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

@@ -1,97 +0,0 @@
-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

@@ -1,47 +0,0 @@
-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

@@ -1,12 +0,0 @@
-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

@@ -1,743 +0,0 @@
-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

@@ -1,244 +0,0 @@
-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

@@ -1,60 +0,0 @@
-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

@@ -1,126 +0,0 @@
-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

@@ -1,339 +0,0 @@
-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

@@ -1,385 +0,0 @@
-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

@@ -1,23 +0,0 @@
-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

@@ -1,157 +0,0 @@
-#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

@@ -1,12 +0,0 @@
-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

@@ -1,143 +0,0 @@
-#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

@@ -1,10 +0,0 @@
-//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

@@ -1,190 +0,0 @@
-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

@@ -4,21 +4,15 @@ 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.Closer
+	io.ReadWriteCloser
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
 	SupportsMultiqueue() bool
-	NewMultiQueueReader() error
+	NewMultiQueueReader() (io.ReadWriteCloser, error)
-	Readers() []tio.Queue
 }

overlay/overlaytest/noop.go

@@ -4,11 +4,10 @@ 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.
@@ -16,10 +15,6 @@ 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{}
 }
@@ -36,7 +31,7 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+func (NoopTun) Read([]byte) (int, error) {
 	return 0, nil
 }
 
@@ -48,12 +43,8 @@ func (NoopTun) SupportsMultiqueue() bool {
 	return false
 }
 
-func (NoopTun) NewMultiQueueReader() error {
+func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return errors.New("unsupported")
+	return nil, errors.New("unsupported")
-}
-
-func (NoopTun) Readers() []tio.Queue {
-	return []tio.Queue{NoopTun{}}
 }
 
 func (NoopTun) Close() error {

overlay/tio/queueset_gso_linux.go

@@ -1,90 +0,0 @@
-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

@@ -1,80 +0,0 @@
-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...)
-}