pin tun reader threads to CPUs so per-flow packets keep wire order

Each listenIn goroutine locks its OS thread and pins it to one CPU
(sched_setaffinity), so every UDP send from that goroutine leaves
through the same XPS-selected NIC TX ring instead of being sprayed
across rings and reordered. On by default via tun.pin_threads; queue i
pins to the i-th entry of the process's allowed CPU set (respecting
cpuset/taskset masks, whose IDs are often not 0..NumCPU-1), or to an
explicit tun.cpu_affinity list, validated against that same allowed
set. Linux only; pinning is a no-op elsewhere.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.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

@@ -10,7 +10,7 @@ jobs:
     name: Build Linux/BSD All
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
       - uses: actions/setup-go@v6
         with:
@@ -36,7 +36,7 @@ jobs:
       id-token: write
       contents: read
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
       - uses: actions/setup-go@v6
         with:
@@ -76,7 +76,7 @@ jobs:
       HAS_SIGNING_CREDS: ${{ secrets.AC_USERNAME != '' }}
     runs-on: macos-latest
     steps:
-      - uses: actions/checkout@v6
+      - uses: actions/checkout@v7
 
       - uses: actions/setup-go@v6
         with:
@@ -134,7 +134,7 @@ jobs:
       # be overwritten
       - name: Checkout code
         if: ${{ env.HAS_DOCKER_CREDS == 'true' }}
-        uses: actions/checkout@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

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

.github/workflows/smoke.yml

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
 
-    - uses: actions/checkout@v6
+    - uses: actions/checkout@v7
 
     - uses: actions/setup-go@v6
       with:
@@ -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
       with:
         go-version: '1.25'
         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
@@ -36,97 +44,107 @@ jobs:
       with:
         version: v2.5
 
-    - 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
       with:
         go-version: '1.25'
         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@v6
+      with:
+        go-version: '1.25'
+        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"

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 {
@@ -261,14 +283,16 @@ 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 _, err := os.Stat(*cf.outCertPath); err == nil {
+	if !isStdio(*cf.outCertPath) {
-		return fmt.Errorf("refusing to overwrite existing CA cert: %s", *cf.outCertPath)
+		if _, err := os.Stat(*cf.outCertPath); err == nil {
+			return fmt.Errorf("refusing to overwrite existing CA cert: %s", *cf.outCertPath)
+		}
 	}
 
 	var c cert.Certificate
@@ -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
@@ -57,11 +69,13 @@ func printCert(args []string, out io.Writer, errOut io.Writer) error {
 			return fmt.Errorf("error while unmarshaling cert: %s", err)
 		}
 
-		if *pf.json {
+		if !qrToStdout {
-			jsonCerts = append(jsonCerts, c)
+			if *pf.json {
-		} else {
+				jsonCerts = append(jsonCerts, c)
-			_, _ = out.Write([]byte(c.String()))
+			} else {
-			_, _ = out.Write([]byte("\n"))
+				_, _ = out.Write([]byte(c.String()))
+				_, _ = out.Write([]byte("\n"))
+			}
 		}
 
 		if *pf.outQRPath != "" {
@@ -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)
 		}
@@ -266,16 +291,10 @@ 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 _, err := os.Stat(*sf.outCertPath); err == nil {
-	}
+			return fmt.Errorf("refusing to overwrite existing cert: %s", *sf.outCertPath)
-
+		}
-	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 _, err := os.Stat(*sf.outKeyPath); err == nil {
+		if !isStdio(*sf.outKeyPath) {
-			return fmt.Errorf("refusing to overwrite existing key: %s", *sf.outKeyPath)
+			if _, err := os.Stat(*sf.outKeyPath); err == nil {
+				return fmt.Errorf("refusing to overwrite existing key: %s", *sf.outKeyPath)
+			}
 		}
 
-		err = 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)
 		}
@@ -61,9 +66,12 @@ func main() {
 	}
 
 	if *configPath == "" {
-		fmt.Println("-config flag must be set")
+		p, err := config.DefaultPath()
-		flag.Usage()
+		if err != nil {
-		os.Exit(1)
+			fmt.Println(err)
+			os.Exit(1)
+		}
+		*configPath = p
 	}
 
 	c := config.NewC(l)
@@ -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/main.go

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

@@ -136,14 +136,6 @@ func (cm *connectionManager) getAndResetTrafficCheck(h *HostInfo, now time.Time)
 	return in, out
 }
 
-// AddTrafficWatch must be called for every new HostInfo.
-// We will continue to monitor the HostInfo until the tunnel is dropped.
-func (cm *connectionManager) AddTrafficWatch(h *HostInfo) {
-	if h.out.Swap(true) == false {
-		cm.trafficTimer.Add(h.localIndexId, cm.checkInterval)
-	}
-}
-
 func (cm *connectionManager) Start(ctx context.Context) {
 	clockSource := time.NewTicker(cm.trafficTimer.t.tickDuration)
 	defer clockSource.Stop()
@@ -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,

control.go

@@ -69,29 +69,29 @@ 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.
@@ -114,13 +114,9 @@ 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 +129,26 @@ 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 +157,26 @@ 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)
@@ -170,8 +189,15 @@ 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.stateLock.Lock()
+	defer c.stateLock.Unlock()
+
+	if c.state != StateStarted {
+		return
+	}
+
 	_ = c.f.outside.Rebind()
 
 	// Trigger a lighthouse update, useful for mobile clients that should have an update interval of 0
@@ -305,7 +331,7 @@ 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 +376,7 @@ 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 +1,296 @@
+package nebula
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/netip"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/gaissmai/bart"
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
+	"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() ([]tio.Packet, error) {
+	<-d.closedCh
+	return nil, 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) Queues(int) ([]tio.Queue, error) { return []tio.Queue{d}, nil }
+
+// 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},
+		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
+}
+
+// Queues claims multiqueue support but fails to open the second queue,
+// exercising the activation error path.
+func (d *multiqueueDevice) Queues(n int) ([]tio.Queue, error) {
+	if n > 1 {
+		return nil, errors.New("second queue failed to open")
+	}
+	return d.fakeDevice.Queues(n)
+}
+
+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},
+		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

@@ -125,6 +125,14 @@ 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

@@ -11,19 +11,21 @@ import (
 	"sync"
 	"sync/atomic"
 
-	"github.com/gaissmai/bart"
 	"github.com/miekg/dns"
 	"github.com/slackhq/nebula/config"
 )
 
 type dnsServer struct {
 	sync.RWMutex
-	l               *slog.Logger
+	l       *slog.Logger
-	ctx             context.Context
+	ctx     context.Context
-	dnsMap4         map[string]netip.Addr
+	dnsMap4 map[string]netip.Addr
-	dnsMap6         map[string]netip.Addr
+	dnsMap6 map[string]netip.Addr
-	hostMap         *HostMap
+	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,
+		l:       l,
-		ctx:             ctx,
+		ctx:     ctx,
-		dnsMap4:         make(map[string]netip.Addr),
+		dnsMap4: make(map[string]netip.Addr),
-		dnsMap6:         make(map[string]netip.Addr),
+		dnsMap6: make(map[string]netip.Addr),
-		hostMap:         hostMap,
+		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
 }
 
@@ -113,7 +116,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 			d.Stop()
 		}
 		// Drop any records that accumulated while enabled; a later re-enable
-		// will repopulate from fresh handshakes.
+		// will repopulate from fresh handshakes and a fresh seedSelf.
 		d.clearRecords()
 		return nil
 	}
@@ -121,17 +124,14 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	if running == nil {
 		// Was disabled (or never started); bring it up now.
 		go d.Start()
-		return nil
+	} else if !sameAddr {
+		d.shutdownServer(running, runningStarted, "reload")
+		// Old Start goroutine has now exited; bring up a fresh listener on the new address.
+		go d.Start()
 	}
 
-	if sameAddr {
+	// Refresh the self entry every enabled reload so cert renewals that change our name or VPN addresses are picked up.
-		return nil
+	d.seedSelf()
-	}
-
-	d.shutdownServer(running, runningStarted, "reload")
-	// Old Start goroutine has now exited; bring up a fresh listener on the
-	// new address.
-	go d.Start()
 	return nil
 }
 
@@ -249,6 +249,20 @@ 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 +280,60 @@ 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 +371,12 @@ 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"
@@ -276,6 +279,92 @@ 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)

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")))
 
@@ -819,18 +823,18 @@ 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 +928,24 @@ 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 +1033,24 @@ 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 +1127,7 @@ 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 +1227,7 @@ 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 +1539,78 @@ 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/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")
 		}

examples/config.yml

@@ -254,6 +254,20 @@ tun:
   # Default MTU for every packet, safe setting is (and the default) 1300 for internet based traffic
   mtu: 1300
 
+  # Linux only. pin_threads pins each tun reader/encrypt OS thread to a single CPU. This keeps every goroutine's
+  # sends flowing through one XPS-selected NIC TX ring, so packets within a flow stay ordered on the wire
+  # instead of being sprayed across multiple TX rings and reordered. Not reloadable.
+  #pin_threads: true
+
+  # Linux only. cpu_affinity overrides which CPUs the tun reader threads pin to: a list of CPU IDs, one per routine
+  # (see the top-level `routines` setting). Lists shorter than `routines` are modulo-cycled across the queues; extra
+  # entries are ignored. IDs must be within the process's allowed CPU set, so this respects taskset / cgroup cpusets;
+  # a non-integer or not-allowed entry disables the override and falls back to spreading queues across the allowed
+  # CPUs. Only meaningful while pin_threads is true. Not reloadable.
+  #cpu_affinity:
+  #  - 2
+  #  - 4
+
   # Route based MTU overrides, you have known vpn ip paths that can support larger MTUs you can increase/decrease them here
   routes:
     #- mtu: 8800
@@ -397,7 +411,7 @@ 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
 

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
@@ -58,8 +58,9 @@ type Firewall struct {
 	routableNetworks *bart.Lite
 
 	// assignedNetworks is a list of vpn networks assigned to us in the certificate.
-	assignedNetworks  []netip.Prefix
+	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{
@@ -169,15 +169,15 @@ func NewFirewall(l *slog.Logger, tcpTimeout, UDPTimeout, defaultTimeout time.Dur
 			Conns:      make(map[firewall.Packet]*conn),
 			TimerWheel: NewTimerWheel[firewall.Packet](tmin, tmax),
 		},
-		InRules:           newFirewallTable(),
+		InRules:          newFirewallTable(),
-		OutRules:          newFirewallTable(),
+		OutRules:         newFirewallTable(),
-		TCPTimeout:        tcpTimeout,
+		TCPTimeout:       tcpTimeout,
-		UDPTimeout:        UDPTimeout,
+		UDPTimeout:       UDPTimeout,
-		DefaultTimeout:    defaultTimeout,
+		DefaultTimeout:   defaultTimeout,
-		routableNetworks:  routableNetworks,
+		routableNetworks: routableNetworks,
-		assignedNetworks:  assignedNetworks,
+		assignedNetworks: assignedNetworks,
-		hasUnsafeNetworks: hasUnsafeNetworks,
+		unsafeNetworks:   unsafeNetworks,
-		l:                 l,
+		l:                l,
 
 		incomingMetrics: firewallMetrics{
 			droppedLocalAddr:  metrics.GetOrRegisterCounter("firewall.incoming.dropped.local_addr", nil),
@@ -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_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

@@ -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 v0.53.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.53.0
+	golang.org/x/net v0.56.0
-	golang.org/x/sync v0.20.0
+	golang.org/x/sync v0.21.0
-	golang.org/x/sys v0.43.0
+	golang.org/x/sys v0.46.0
-	golang.org/x/term v0.42.0
+	golang.org/x/term v0.44.0
 	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 v0.53.0 h1:QZ4Muo8THX6CizN2vPPd5fBGHyogrdK9fG4wLPFUsto=
-golang.org/x/crypto v0.50.0/go.mod h1:3muZ7vA7PBCE6xgPX7nkzzjiUq87kRItoJQM1Yo8S+Q=
+golang.org/x/crypto v0.53.0/go.mod h1:DNLU434OwVakk9PzuwV8w62mAJpRJL3vsgcfp4Qnsio=
 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 v0.56.0 h1:Rw8j/hFzGvJUZwNBXnAtf5sVDVt+65SK2C7IxCxZt5o=
-golang.org/x/net v0.53.0/go.mod h1:JvMuJH7rrdiCfbeHoo3fCQU24Lf5JJwT9W3sJFulfgs=
+golang.org/x/net v0.56.0/go.mod h1:D3Ku6r+V6JROoZK144D2XfMHFcMq/0zSfLelVTCFKec=
 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 v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sync v0.21.0/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 v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.43.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.46.0/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 v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=
-golang.org/x/term v0.42.0/go.mod h1:Dq/D+snpsbazcBG5+F9Q1n2rXV8Ma+71xEjTRufARgY=
+golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=
 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
@@ -323,7 +323,7 @@ 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 +430,11 @@ 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 +462,6 @@ 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 +484,6 @@ 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,
 		)
 	}
@@ -798,7 +793,6 @@ 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 +959,6 @@ 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 +967,6 @@ 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)))

hostmap.go

@@ -56,11 +56,20 @@ type Relay struct {
 }
 
 type HostMap struct {
-	sync.RWMutex    //Because we concurrently read and write to our maps
+	sync.RWMutex  //Because we concurrently read and write to our maps
-	Indexes         map[uint32]*HostInfo
+	Indexes       map[uint32]*HostInfo
-	Relays          map[uint32]*HostInfo // Maps a Relay IDX to a Relay HostInfo object
+	Relays        map[uint32]*HostInfo // Maps a Relay IDX to a Relay HostInfo object
-	RemoteIndexes   map[uint32]*HostInfo
+	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
 }
@@ -229,7 +238,7 @@ const (
 )
 
 type HostInfo struct {
-	remote          netip.AddrPort
+	remote          atomic.Pointer[netip.AddrPort]
 	remotes         *RemoteList
 	promoteCounter  atomic.Uint32
 	ConnectionState *ConnectionState
@@ -266,10 +275,6 @@ 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
 
@@ -334,6 +339,7 @@ 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 +388,55 @@ 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
@@ -400,85 +448,66 @@ 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 {
-		delete(hm.Hosts, addr)
+	// Remove this hostinfo from each of its address lists. The lists are independent, so a
-		if len(hm.Hosts) == 0 {
+	// sibling is never promoted to an address it does not own and no other list is touched.
-			hm.Hosts = map[netip.Addr]*HostInfo{}
+	final := true
-		}
+	for _, addr := range hostinfo.vpnAddrs {
-
+		if list, ok := hm.moreHosts[addr]; ok {
-		if hostinfo.next != nil {
+			list = removeHostInfo(list, hostinfo)
-			// We had more than 1 hostinfo at this vpn addr, promote the next in the list to primary
+			hm.unlockedSetHostsForAddr(addr, list)
-			hm.Hosts[addr] = hostinfo.next
+			if len(list) > 0 {
-			// It is primary, there is no previous hostinfo now
+				final = false
-			hostinfo.next.prev = nil
+			}
-		}
+		} else if existing, ok := hm.Hosts[addr]; ok {
-
+			if existing == hostinfo {
-	} else {
+				// Common case, the only hostinfo for this address. moreHosts has no entry to clean up.
-		// Relink if we were in the middle of multiple hostinfos for this vpn addr
+				delete(hm.Hosts, addr)
-		if hostinfo.prev != nil {
+			} else {
-			hostinfo.prev.next = hostinfo.next
+				// We don't hold this address but another hostinfo does, we still have a tunnel to the peer
-		}
+				final = false
-
+			}
-		if hostinfo.next != nil {
-			hostinfo.next.prev = hostinfo.prev
 		}
 	}
 
-	hostinfo.next = nil
+	// Go maps never shrink their buckets, replace fully drained maps so a node that churned
-	hostinfo.prev = nil
+	// 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 {
+		hm.moreHosts = map[netip.Addr][]*HostInfo{}
+	}
 
 	// 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
@@ -502,7 +531,7 @@ 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 +540,8 @@ 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 +585,30 @@ 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 {
-		for _, targetIp := range targetIps {
+		r, ok := h.relayState.QueryRelayForByIp(targetIp)
-			r, ok := h.relayState.QueryRelayForByIp(targetIp)
+		if ok && r.State == Established {
-			if ok && r.State == Established {
+			return h, r, nil
-				return h, r, nil
+		}
+	}
+
+	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
+				}
 			}
 		}
-		h = h.next
 	}
 
 	return nil, nil, errors.New("unable to find host with relay")
@@ -574,20 +616,14 @@ 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.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.relayState.UpdateRelayForByIpState(hi.vpnAddrs[0], Disestablished)
-					h = h.next
-				}
 			}
 		}
 	}
@@ -623,6 +659,11 @@ func (hm *HostMap) unlockedAddHostInfo(hostinfo *HostInfo, f *Interface) {
 	hm.Indexes[hostinfo.localIndexId] = hostinfo
 	hm.RemoteIndexes[hostinfo.remoteIndexId] = hostinfo
 
+	hostinfo.out.Store(true)
+	if f.connectionManager != nil { // f.connectionManager is only nil in some unit tests
+		f.connectionManager.trafficTimer.Add(hostinfo.localIndexId, f.connectionManager.checkInterval)
+	}
+
 	if hm.l.Enabled(context.Background(), slog.LevelDebug) {
 		hm.l.Debug("Hostmap vpnIp added",
 			"hostMap", m{"vpnAddrs": hostinfo.vpnAddrs, "mapTotalSize": len(hm.Hosts),
@@ -632,22 +673,27 @@ 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]
-	hm.Hosts[vpnAddr] = hostinfo
+	if !ok {
-
+		// Common case, the first hostinfo for this address. moreHosts stays empty.
-	if existing != nil && existing != hostinfo {
+		hm.Hosts[vpnAddr] = hostinfo
-		hostinfo.next = existing
+		return
-		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 +725,7 @@ 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() {
@@ -721,11 +767,18 @@ 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 +790,7 @@ 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"
 
@@ -10,16 +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) {
-	// borrowed: pkt.Bytes is owned by the originating tio.Queue and is
-	// only valid until the next Read on that queue. If you must keep
-	// the packet, use pkt.Clone() to detach it
-	packet := pkt.Bytes
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -44,10 +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.queues[q].Write(packet)
-				_, werr := f.readers[q].Write(seg)
-				return werr
-			})
 			if err != nil {
 				f.l.Error("Failed to forward to tun", "error", err)
 			}
@@ -63,23 +53,11 @@ func (f *Interface) consumeInsidePacket(pkt wire.TunPacket, fwPacket *firewall.P
 	}
 
 	hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
-		// borrowed: SegmentSuperpacket 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,
@@ -95,9 +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,
@@ -107,126 +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
-	}
-
-	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
-			}
-
-			sendBatch.Commit(toSend, relayHostInfo.remote, 0)
-			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
-		}
-
-		sendBatch.Commit(out, hostinfo.remote, 0)
-		return nil
-	})
-	if err != nil {
-		hostinfo.logger(f.l).Error("Failed to segment superpacket for send",
-			"error", err,
-		)
-	}
-}
-
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
-	if !f.firewall.InSendReject {
+	if !f.firewall.OutboundSendReject {
 		return
 	}
 
@@ -235,14 +96,14 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 		return
 	}
 
-	_, err := f.readers[q].Write(out)
+	_, err := f.queues[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}
 }
 
 func (f *Interface) rejectOutside(packet []byte, ci *ConnectionState, hostinfo *HostInfo, nb, out []byte, q int) {
-	if !f.firewall.OutSendReject {
+	if !f.firewall.InboundSendReject {
 		return
 	}
 
@@ -414,13 +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()
@@ -442,7 +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.
@@ -462,43 +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
-// q indicates which writer to use to send the packet.
-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 {
@@ -545,7 +391,6 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 			"error", err,
 			"udpAddr", remote,
 			"counter", c,
-			"attemptedCounter", c,
 		)
 		return
 	}
@@ -558,8 +403,8 @@ 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,

interface.go

@@ -6,22 +6,23 @@ import (
 	"fmt"
 	"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"
+	"github.com/slackhq/nebula/util"
 )
 
 const mtu = 9001
@@ -40,7 +41,6 @@ type InterfaceConfig struct {
 	DropLocalBroadcast bool
 	DropMulticast      bool
 	routines           int
-	batchSize          int
 	MessageMetrics     *MessageMetrics
 	version            string
 	relayManager       *relayManager
@@ -51,7 +51,19 @@ type InterfaceConfig struct {
 	reQueryWait     time.Duration
 
 	ConntrackCacheTimeout time.Duration
-	l                     *slog.Logger
+
+	// CpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
+	// should pin to. Queue i pins to CpuAffinity[i % len(CpuAffinity)] —
+	// shorter lists than `routines` cycle. Empty list keeps the default
+	// pin-to-(i % NumCPU) behavior. Only consulted when PinThreads is true.
+	CpuAffinity []int
+	// PinThreads controls whether each TUN reader OS thread is pinned to a
+	// single CPU (via tun.pin_threads, default true). Pinning keeps each
+	// goroutine's UDP sends on one XPS-selected NIC TX ring so per-flow
+	// packets stay ordered on the wire.
+	PinThreads bool
+
+	l *slog.Logger
 }
 
 type Interface struct {
@@ -73,10 +85,18 @@ type Interface struct {
 	dropLocalBroadcast    bool
 	dropMulticast         bool
 	routines              int
-	batchSize             int
 	disconnectInvalid     atomic.Bool
 	closed                atomic.Bool
-	relayManager          *relayManager
+	// cpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
+	// should pin to. Queue i pins to cpuAffinity[i % len(cpuAffinity)].
+	// Empty falls back to the default pin-to-(allowed CPU) behavior.
+	// Only consulted when pinThreads is true.
+	cpuAffinity []int
+	// pinThreads controls whether listenIn pins each TUN reader OS thread to
+	// a CPU at all (tun.pin_threads, default true). When false, threads are
+	// left free to migrate as on stock nebula.
+	pinThreads   bool
+	relayManager *relayManager
 
 	tryPromoteEvery atomic.Uint32
 	reQueryEvery    atomic.Uint32
@@ -93,12 +113,8 @@ type Interface struct {
 
 	ctx     context.Context
 	writers []udp.Conn
-	readers []tio.Queue
+	queues  []tio.Queue
-	// batchers is one per tun queue, wrapping readers[i].
+	wg      sync.WaitGroup
-	// decryptToTun sends plaintext into the batch.RxBatcher;
-	// listenOut calls its Flush at the end of each UDP recvmmsg batch.
-	batchers []batch.RxBatcher
-	wg       sync.WaitGroup
 
 	// fatalErr holds the first unexpected reader error that caused shutdown.
 	// nil means "no fatal error" (yet)
@@ -194,11 +210,8 @@ 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),
-		batchers:              make([]batch.RxBatcher, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -207,6 +220,8 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		relayManager:          c.relayManager,
 		connectionManager:     c.connectionManager,
 		conntrackCacheTimeout: c.ConntrackCacheTimeout,
+		cpuAffinity:           c.CpuAffinity,
+		pinThreads:            c.PinThreads,
 
 		metricHandshakes: metrics.GetOrRegisterHistogram("handshakes", nil, metrics.NewExpDecaySample(1028, 0.015)),
 		messageMetrics:   c.MessageMetrics,
@@ -224,6 +239,9 @@ 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
 }
 
@@ -246,40 +264,37 @@ func (f *Interface) activate() error {
 		"boringcrypto", boringEnabled(),
 	)
 
-	if f.routines > 1 {
+	if f.routines > 1 && !f.outside.SupportsMultipleReaders() {
-		if !f.inside.SupportsMultiqueue() || !f.outside.SupportsMultipleReaders() {
+		f.routines = 1
-			f.routines = 1
+		f.l.Warn("multiple udp readers are not supported on this platform, falling back to a single routine")
-			f.l.Warn("routines is not supported on this platform, falling back to a single routine")
-		}
 	}
 
+	// Prepare the tun queues. A device that can't open that many hands back
+	// fewer (a single queue on platforms without multiqueue support) and we
+	// size the reader routines to what we actually got.
+	queues, err := f.inside.Queues(f.routines)
+	if err != nil {
+		return err
+	}
+	if len(queues) < f.routines {
+		f.l.Warn("tun multiqueue is not supported on this platform, falling back to fewer routines",
+			"requested", f.routines, "opened", len(queues))
+		f.routines = len(queues)
+	}
+	f.queues = queues
+
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
-	// Prepare n tun queues
+	// On error the caller owns the cleanup, Control.Start cancels the service context
-	for i := 0; i < f.routines; i++ {
+	// before releasing our resources so a waiter never observes a live context
-		if i > 0 {
-			if err = f.inside.NewMultiQueueReader(); err != nil {
-				return err
-			}
-		}
-	}
-	f.readers = f.inside.Readers()
-	for i := range f.readers {
-		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
 	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() {
@@ -290,17 +305,18 @@ func (f *Interface) run() (func() error, error) {
 	// Launch n queues to read packets from tun dev
 	for i := 0; i < f.routines; i++ {
 		f.wg.Go(func() {
-			f.listenIn(f.readers[i], i)
+			f.listenIn(f.queues[i], i)
 		})
 	}
 
-	return func() error {
+}
-		f.wg.Wait()
+
-		if e := f.fatalErr.Load(); e != nil {
+func (f *Interface) wait() error {
-			return *e
+	f.wg.Wait()
-		}
+	if e := f.fatalErr.Load(); e != nil {
-		return nil
+		return *e
-	}, nil
+	}
+	return nil
 }
 
 // onFatal stores the first fatal reader error, and calls triggerShutdown if it was the first one
@@ -324,23 +340,19 @@ 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())
+		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)
 	}
@@ -348,38 +360,54 @@ 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(queue tio.Queue, i int) {
-	packetMem := make([]byte, mtu+16) //MTU + some leading slack space for platforms that return "bonus info"
+	// Pinning this thread (and goroutine) to a single CPU keeps every UDP send from this goroutine going through
-	// TODO get the amount of bonus info from the reader
+	// the same TX ring on the nic (XPS selects the ring by CPU), so the wire sees per-flow order. Skip entirely
-	packets := make([]wire.TunPacket, 1)
+	// when tun.pin_threads is false.
-	rejectBuf := make([]byte, mtu)
+	if f.pinThreads {
-	arenaSize := batch.SendBatchCap * (udp.MTU + 32)
+		var cpu int
-	sb := batch.NewSendBatch(f.writers[q], batch.SendBatchCap, util.NewArena(arenaSize))
+		if n := len(f.cpuAffinity); n > 0 {
+			// Explicit tun.cpu_affinity list wins; parseCpuAffinity already
+			// validated the entries against the allowed CPU set.
+			cpu = f.cpuAffinity[i%n]
+		} else if allowed, err := util.AllowedCPUs(); err == nil && len(allowed) > 0 {
+			// Default: spread queues across the CPUs we're actually allowed to
+			// run on. Under a cpuset/taskset mask these aren't 0..NumCPU-1, so
+			// i % NumCPU would pick unrunnable IDs and every pin would fail.
+			cpu = allowed[i%len(allowed)]
+		} else {
+			cpu = i % runtime.NumCPU()
+		}
+		if err := util.PinThreadToCPU(cpu); err != nil {
+			f.l.Warn("failed to pin tun reader to CPU", "queue", i, "cpu", cpu, "err", err)
+		}
+	}
+
+	out := make([]byte, mtu)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
 	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.Read(packets, packetMem)
+		pkts, err := queue.Read()
 		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()
+		for _, pkt := range pkts {
-		for i := range n {
+			// borrowed: pkt.Bytes is owned by the queue and only valid until
-			f.consumeInsidePacket(packets[i], fwPacket, nb, sb, rejectBuf, q, ctCache)
+			// the next Read; consumeInsidePacket reads it synchronously.
-		}
+			f.consumeInsidePacket(pkt.Bytes, fwPacket, nb, out, i, conntrackCache.Get())
-		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) {
@@ -405,13 +433,22 @@ 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
@@ -521,26 +558,34 @@ func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 	certInitiatingVersion := metrics.GetOrRegisterGauge("certificate.initiating_version", nil)
 	certMaxVersion := metrics.GetOrRegisterGauge("certificate.max_version", nil)
 
+	emit := func() {
+		f.firewall.EmitStats()
+		f.handshakeManager.EmitStats()
+		udpStats()
+
+		certState := f.pki.getCertState()
+		defaultCrt := certState.GetDefaultCertificate()
+		certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
+		certInitiatingVersion.Update(int64(defaultCrt.Version()))
+
+		// Report the max certificate version we are capable of using
+		if certState.v2Cert != nil {
+			certMaxVersion.Update(int64(certState.v2Cert.Version()))
+		} else {
+			certMaxVersion.Update(int64(certState.v1Cert.Version()))
+		}
+	}
+
+	// Prime gauges so a Prometheus scrape that lands before the first tick
+	// sees real values instead of the zero defaults (issue #907).
+	emit()
+
 	for {
 		select {
 		case <-ctx.Done():
 			return
 		case <-ticker.C:
-			f.firewall.EmitStats()
+			emit()
-			f.handshakeManager.EmitStats()
-			udpStats()
-
-			certState := f.pki.getCertState()
-			defaultCrt := certState.GetDefaultCertificate()
-			certExpirationGauge.Update(int64(defaultCrt.NotAfter().Sub(time.Now()) / time.Second))
-			certInitiatingVersion.Update(int64(defaultCrt.Version()))
-
-			// Report the max certificate version we are capable of using
-			if certState.v2Cert != nil {
-				certMaxVersion.Update(int64(certState.v2Cert.Version()))
-			} else {
-				certMaxVersion.Update(int64(certState.v1Cert.Version()))
-			}
 		}
 	}
 }
@@ -553,9 +598,15 @@ 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 {
@@ -571,6 +622,8 @@ 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

@@ -4,26 +4,54 @@ 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
 	}
 
-	switch packet[9] {
+	version := int(packet[0] >> 4)
-	case 6: // tcp
+	switch version {
-		return ipv4CreateRejectTCPPacket(packet, out)
+	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 ipv4CreateRejectICMPPacket(packet, out)
+		return nil
 	}
 }
 
@@ -35,12 +63,17 @@ 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) {
 		return nil
@@ -71,14 +104,14 @@ func ipv4CreateRejectICMPPacket(packet []byte, out []byte) []byte {
 
 	// ICMP Destination Unreachable
 	icmpOut := out[ipv4.HeaderLen:]
-	icmpOut[0] = 3 // type (Destination unreachable)
+	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[2] = 0  // checksum
-	icmpOut[3] = 0 //  .
+	icmpOut[3] = 0  //  .
-	icmpOut[4] = 0 // unused
+	icmpOut[4] = 0  // unused
-	icmpOut[5] = 0 //  .
+	icmpOut[5] = 0  //  .
-	icmpOut[6] = 0 //  .
+	icmpOut[6] = 0  //  .
-	icmpOut[7] = 0 //  .
+	icmpOut[7] = 0  //  .
 
 	// Copy original IP header and first 8 bytes as body
 	copy(icmpOut[8:], packet[:packetLen])
@@ -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

@@ -272,16 +272,18 @@ 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 +291,21 @@ 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") {
@@ -1401,6 +1418,9 @@ 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 +1428,9 @@ 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 +1460,7 @@ 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 {
@@ -1477,7 +1500,9 @@ func (d *NebulaMetaDetails) GetRelays() []netip.Addr {
 
 	if len(d.RelayVpnAddrs) > 0 {
 		for _, r := range d.RelayVpnAddrs {
-			relays = append(relays, protoAddrToNetAddr(r))
+			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

@@ -7,6 +7,7 @@ import (
 	"net"
 	"net/netip"
 	"runtime/debug"
+	"slices"
 	"strings"
 	"time"
 
@@ -130,6 +131,17 @@ 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
@@ -194,7 +206,7 @@ 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)
 	}
@@ -215,12 +227,13 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		DropLocalBroadcast:    c.GetBool("tun.drop_local_broadcast", false),
 		DropMulticast:         c.GetBool("tun.drop_multicast", false),
 		routines:              routines,
-		batchSize:             c.GetInt("listen.batch", 64),
 		MessageMetrics:        messageMetrics,
 		version:               buildVersion,
 		relayManager:          NewRelayManager(ctx, l, hostMap, c),
 		punchy:                punchy,
 		ConntrackCacheTimeout: conntrackCacheTimeout,
+		CpuAffinity:           parseCpuAffinity(c, l, routines),
+		PinThreads:            c.GetBool("tun.pin_threads", true),
 		l:                     l,
 	}
 
@@ -272,6 +285,70 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	}, nil
 }
 
+// parseCpuAffinity reads `tun.cpu_affinity` from the config — a list of
+// integer CPU IDs, one per TUN reader goroutine. Empty / unset returns nil
+// (listenIn falls back to spreading queues across the allowed CPU set).
+// 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, or a CPU ID we're not allowed to run on) are also a
+// warning and disable the override entirely so we don't silently pin to the
+// wrong CPU. Entries are validated against the process's current affinity
+// mask (util.AllowedCPUs) rather than 0..NumCPU-1: under a cgroup cpuset or
+// taskset the runnable IDs are frequently not that contiguous range, and
+// pinning to an unrunnable ID always fails. If the allowed set can't be
+// determined we fall back to a plain non-negative check.
+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
+	}
+	// allowed is the set of CPU IDs we're actually permitted to run on. A nil
+	// slice (unsupported platform or lookup error) means "can't tell", so we
+	// only apply the weaker non-negative check in that case.
+	allowed, err := util.AllowedCPUs()
+	if err != nil {
+		l.Warn("could not determine allowed CPUs; validating tun.cpu_affinity against non-negative only", "error", err)
+		allowed = nil
+	}
+	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 {
+			l.Warn("tun.cpu_affinity entry out of range; ignoring affinity",
+				"index", i, "cpu", cpu)
+			return nil
+		}
+		if len(allowed) > 0 && !slices.Contains(allowed, cpu) {
+			l.Warn("tun.cpu_affinity entry not in allowed CPU set; ignoring affinity",
+				"index", i, "cpu", cpu, "allowed", allowed)
+			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 {

main_test.go

@@ -0,0 +1,51 @@
+package nebula
+
+import (
+	"testing"
+
+	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/test"
+	"github.com/slackhq/nebula/util"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestParseCpuAffinity(t *testing.T) {
+	l := test.NewLogger()
+
+	// newConfig returns a config.C with tun.cpu_affinity set to v. A nil v
+	// leaves the key unset.
+	newConfig := func(v any) *config.C {
+		c := config.NewC(l)
+		if v != nil {
+			c.Settings["tun"] = map[string]any{"cpu_affinity": v}
+		}
+		return c
+	}
+
+	// unset -> nil (listenIn falls back to spreading across the allowed set)
+	assert.Nil(t, parseCpuAffinity(newConfig(nil), l, 1))
+
+	// Pick a CPU we're actually allowed to run on so a valid list survives
+	// validation regardless of the host's affinity mask.
+	allowed, _ := util.AllowedCPUs()
+	validCPU := 0
+	if len(allowed) > 0 {
+		validCPU = allowed[0]
+	}
+
+	// valid list -> parsed through unchanged
+	assert.Equal(t, []int{validCPU, validCPU}, parseCpuAffinity(newConfig([]any{validCPU, validCPU}), l, 2))
+
+	// a negative entry is out of range on every platform -> disables the override
+	assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU, -1}), l, 2))
+
+	// a non-integer entry -> disables the override
+	assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU, "not-a-cpu"}), l, 2))
+
+	// a CPU id outside the allowed set -> disables the override. Only assertable
+	// where we can enumerate the allowed set (e.g. linux); 1<<20 is far beyond
+	// any representable CPU id so it can never be in the mask.
+	if len(allowed) > 0 {
+		assert.Nil(t, parseCpuAffinity(newConfig([]any{1 << 20}), l, 1))
+	}
+}

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

@@ -22,7 +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) {
+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
@@ -110,11 +110,11 @@ func (f *Interface) readOutsidePackets(via ViaSender, packet []byte, h *header.H
 
 	// Relay packets are special
 	if isMessageRelay {
-		f.handleOutsideRelayPacket(hostinfo, via, packet, h, fwPacket, lhf, nb, q, localCache)
+		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache)
+
 		return
 	}
 
-	out := f.batchers[q].Reserve(len(packet))[:0]
 	out, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -150,7 +150,8 @@ 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, ci, hostinfo, out, nb, packet)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected test subtype seen", "from", via, "header", h)
 			return
@@ -168,7 +169,7 @@ 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) {
+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.
 	// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
 	// The packet is guaranteed to be at least 16 bytes at this point, b/c it got past the h.Parse() call above. If it's
@@ -176,10 +177,16 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 	// which will gracefully fail in the DecryptDanger call.
 	signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
 	signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
-	// The decrypted output is empty (relay packets carry their payload as AD) and unused.
+	var err error
-	// The recursive readOutsidePackets call below operates on signedPayload. Passing
+	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, signedPayload, signatureValue, h.MessageCounter, nb)
-	// nil avoids reserving an arena slot.
+	if err != nil {
-	if _, err := hostinfo.ConnectionState.dKey.DecryptDanger(nil, signedPayload, signatureValue, h.MessageCounter, nb); err != nil {
+		return
+	}
+	// Advance the replay window now that the frame is authenticated
+	if !hostinfo.ConnectionState.window.Update(f.l, h.MessageCounter) {
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			hostinfo.logger(f.l).Debug("dropping out of window relay packet", "header", h)
+		}
 		return
 	}
 	// Successfully validated the thing. Get rid of the Relay header.
@@ -195,8 +202,7 @@ 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
 	}
@@ -212,15 +218,15 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 			relay:     relay,
 			IsRelayed: true,
 		}
-		f.readOutsidePackets(via, signedPayload, h, fwPacket, lhf, nb, q, localCache)
+		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
 		}
@@ -230,8 +236,7 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 			switch targetRelay.Type {
 			case ForwardingType:
 				// Forward this packet through the relay tunnel
-				// Find the target HostInfo //todo it would potentially be nice to batch these
+				// Find the target HostInfo
-				out := f.batchers[q].Reserve(len(packet) + header.Len + hostinfo.ConnectionState.dKey.Overhead())[:0]
 				f.SendVia(targetHI, targetRelay, signedPayload, nb, out, false)
 			case TerminalType:
 				hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
@@ -272,7 +277,8 @@ 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)
@@ -284,7 +290,7 @@ 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,
 				)
 			}
@@ -292,11 +298,11 @@ 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)
 	}
 
@@ -416,16 +422,14 @@ 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 {
@@ -538,7 +542,7 @@ func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, p
 		return
 	}
 
-	err = f.batchers[q].Commit(out)
+	_, err = f.queues[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}
@@ -585,10 +589,11 @@ 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

@@ -1,42 +0,0 @@
-package batch
-
-// Arena is an injectable byte-slab that hands out non-overlapping borrowed
-// slices via Reserve and releases them in bulk via Reset. Coalescers take
-// an *Arena at construction so the caller controls the slab lifetime and
-// can share one slab across multiple coalescers (MultiCoalescer hands the
-// same *Arena to every lane so the lanes don't carry their own backings).
-//
-// Reserve borrows; the slice is valid until the next Reset. The slab grows
-// (by allocating a fresh, larger backing array) if a Reserve doesn't fit;
-// pre-size the arena via NewArena to avoid that path on the hot path.
-type Arena struct {
-	buf []byte
-}
-
-// NewArena returns an Arena with a pre-allocated backing of the given
-// capacity. Pass 0 if you don't intend to call Reserve (e.g. a test that
-// only feeds the coalescer pre-made []byte packets via Commit).
-func NewArena(capacity int) *Arena {
-	return &Arena{buf: make([]byte, 0, capacity)}
-}
-
-// Reserve hands out a non-overlapping sz-byte slice from the arena. If the
-// request doesn't fit the current backing, a fresh, larger backing is
-// allocated; already-borrowed slices reference the old backing and remain
-// valid until Reset.
-func (a *Arena) Reserve(sz int) []byte {
-	if len(a.buf)+sz > cap(a.buf) {
-		newCap := max(cap(a.buf)*2, sz)
-		a.buf = make([]byte, 0, newCap)
-	}
-	start := len(a.buf)
-	a.buf = a.buf[:start+sz]
-	return a.buf[start : start+sz : start+sz]
-}
-
-// Reset releases every slice handed out since the last Reset. Callers must
-// not use any previously-borrowed slice after this returns. The underlying
-// backing array is retained so subsequent Reserves don't re-allocate.
-func (a *Arena) Reset() {
-	a.buf = a.buf[:0]
-}

overlay/batch/passthrough.go

@@ -1,46 +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  *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/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/device.go

@@ -8,8 +8,8 @@ import (
 	"github.com/slackhq/nebula/routing"
 )
 
-// defaultBatchBufSize is the per-Queue scratch size for Read on backends
+// defaultBatchBufSize is the per-Queue scratch size for Read. 65535 covers
-// that don't do TSO segmentation. 65535 covers any single IP packet.
+// any single IP packet.
 const defaultBatchBufSize = 65535
 
 type Device interface {
@@ -18,7 +18,11 @@ type Device interface {
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
-	SupportsMultiqueue() bool
+	// Queues returns the device's packet queues, opening additional ones as
-	NewMultiQueueReader() error
+	// needed until there are n. Platforms without multiqueue support return
-	Readers() []tio.Queue
+	// their single queue regardless of n, so callers must size reader loops
+	// to len(result), not n; implementations never return more than n. An
+	// error means a queue that should have opened could not; the caller owns
+	// cleanup via Close. Called once, during interface activation.
+	Queues(n int) ([]tio.Queue, error)
 }

overlay/overlaytest/noop.go

@@ -3,12 +3,10 @@
 package overlaytest
 
 import (
-	"errors"
 	"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 +14,6 @@ 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{}
 }
@@ -36,24 +30,16 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+func (NoopTun) Read() ([]tio.Packet, error) {
-	return 0, nil
+	return nil, nil
 }
 
 func (NoopTun) Write([]byte) (int, error) {
 	return 0, nil
 }
 
-func (NoopTun) SupportsMultiqueue() bool {
+func (NoopTun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{NoopTun{}}, nil
-}
-
-func (NoopTun) NewMultiQueueReader() error {
-	return errors.New("unsupported")
-}
-
-func (NoopTun) Readers() []tio.Queue {
-	return []tio.Queue{NoopTun{}}
 }
 
 func (NoopTun) Close() error {

overlay/tio/blockon_linux.go

@@ -0,0 +1,45 @@
+//go:build linux && !android
+// +build linux,!android
+
+package tio
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// blockOn parks the calling goroutine until fd is ready (events is POLLIN for
+// reads, POLLOUT for writes) or shutdownFd signals teardown. It builds the
+// pollfd array on the stack every call, so concurrent callers on the same
+// Queue never share Revents storage.
+//
+// Returns os.ErrClosed when shutdown was signaled (POLLIN on shutdownFd)
+// or either fd reported a problem condition (POLLHUP|POLLNVAL|POLLERR).
+func blockOn(fd, shutdownFd int32, events int16) error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	pfds := [2]unix.PollFd{
+		{Fd: fd, Events: events},
+		{Fd: shutdownFd, Events: unix.POLLIN},
+	}
+	var err error
+	for {
+		_, err = unix.Poll(pfds[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	tunEvents := pfds[0].Revents
+	shutdownEvents := pfds[1].Revents
+	// Check err before trusting the potentially bogus bits we just got.
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	}
+	if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}

overlay/tio/queueset_poll_linux.go

@@ -1,9 +1,13 @@
+//go:build linux && !android
+// +build linux,!android
+
 package tio
 
 import (
 	"encoding/binary"
 	"errors"
 	"fmt"
+	"sync/atomic"
 
 	"golang.org/x/sys/unix"
 )
@@ -13,6 +17,7 @@ type pollQueueSet struct {
 	// pqi is exactly the same as pq, but stored as the interface type
 	pqi        []Queue
 	shutdownFd int
+	closed     atomic.Bool
 }
 
 func NewPollQueueSet() (QueueSet, error) {
@@ -48,29 +53,34 @@ func (c *pollQueueSet) Add(fd int) error {
 func (c *pollQueueSet) wakeForShutdown() error {
 	var buf [8]byte
 	binary.NativeEndian.PutUint64(buf[:], 1)
-	_, err := unix.Write(c.shutdownFd, buf[:])
+	_, err := unix.Write(int(c.shutdownFd), buf[:])
 	return err
 }
 
 func (c *pollQueueSet) Close() error {
-	if c.shutdownFd < 0 {
+	if c.closed.Swap(true) {
 		return nil
 	}
 
 	errs := []error{}
 
+	// Wake any reader blocked in poll so it observes POLLIN on the shutdown
+	// eventfd and returns os.ErrClosed.
 	if err := c.wakeForShutdown(); err != nil {
 		errs = append(errs, err)
 	}
 
+	// Close the per-queue tun fds; this also unblocks any in-flight reads.
+	// The per-queue Close deliberately leaves shutdownFd alone - it belongs
+	// to this container.
 	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
+	// Close the shutdown eventfd last: every reader's pollfd set references
-	// only after every Poll.Close has returned.
+	// it, so it must outlive the wake + per-queue teardown above.
 	if err := unix.Close(c.shutdownFd); err != nil {
 		errs = append(errs, err)
 	}

overlay/tio/single.go

@@ -0,0 +1,50 @@
+package tio
+
+import "io"
+
+// singleQueue adapts a legacy one-datagram-per-Read source into a Queue.
+// Read fills a private scratch buffer and returns exactly one Packet whose
+// Bytes borrow from that buffer, valid only until the next Read, per the
+// Queue contract. Single-reader like every Queue; Write is exactly as safe
+// for concurrent use as the underlying source's Write.
+type singleQueue struct {
+	rw     io.ReadWriter
+	closer io.Closer // nil: Close is a no-op (the source is shared and owned elsewhere)
+	buf    []byte
+	ret    [1]Packet
+}
+
+// NewSingleQueue wraps a one-datagram-per-Read ReadWriteCloser (a legacy tun
+// device) into a Queue. bufSize is the per-queue read scratch size and must
+// be at least the largest datagram the source can return. Close closes rwc.
+func NewSingleQueue(rwc io.ReadWriteCloser, bufSize int) Queue {
+	return &singleQueue{rw: rwc, closer: rwc, buf: make([]byte, bufSize)}
+}
+
+// NewSingleQueueNoClose is NewSingleQueue for a source owned by someone else,
+// e.g. several queues sharing one device. Close on the returned Queue is a
+// no-op so one queue can't tear the shared source out from under its
+// siblings; the owner remains responsible for closing the source itself.
+func NewSingleQueueNoClose(rw io.ReadWriter, bufSize int) Queue {
+	return &singleQueue{rw: rw, buf: make([]byte, bufSize)}
+}
+
+func (q *singleQueue) Read() ([]Packet, error) {
+	n, err := q.rw.Read(q.buf)
+	if err != nil {
+		return nil, err
+	}
+	q.ret[0] = Packet{Bytes: q.buf[:n]}
+	return q.ret[:], nil
+}
+
+func (q *singleQueue) Write(p []byte) (int, error) {
+	return q.rw.Write(p)
+}
+
+func (q *singleQueue) Close() error {
+	if q.closer == nil {
+		return nil
+	}
+	return q.closer.Close()
+}

overlay/tio/tio.go

@@ -2,8 +2,6 @@ package tio
 
 import (
 	"io"
-
-	"github.com/slackhq/nebula/wire"
 )
 
 // QueueSet holds one or many Queue objects and helps close them in an orderly way.
@@ -15,108 +13,40 @@ type QueueSet interface {
 	Add(fd int) error
 }
 
-// Capabilities advertises which kernel offload features a Queue successfully negotiated.
+// Queue is a readable/writable packet queue. Concurrency contract: a single
-// Callers consult this to decide which coalescers to wire onto the write path.
+// read goroutine drives Read; plain Write is safe for concurrent callers.
-type Capabilities struct {
-	// TSO means the FD was opened with IFF_VNET_HDR and the kernel agreed
-	// to TUN_F_TSO4|TSO6 — i.e. WriteGSO with GSOProtoTCP is safe.
-	TSO bool
-	// USO means the kernel additionally agreed to TUN_F_USO4|USO6, so
-	// WriteGSO with GSOProtoUDP is safe. Linux ≥ 6.2.
-	USO bool
-}
-
-// Queue is a readable/writable Poll queue. One Queue is driven by a single
-// read goroutine plus a single writer (see Write below).
 type Queue interface {
 	io.Closer
 
-	// Read will read at least 1 packet from the tun (up to len(p)).
+	// Read returns one or more packets. The returned Packet.Bytes slices
-	// mem will be used to provide the backing for each of p[n].Bytes.
+	// are borrowed from the Queue's internal buffer and are only valid
-	// Callers should size mem and p to avoid exhausting mem before p.
+	// until the next Read or Close on this Queue - callers must encrypt
-	// Returns the number of packets actually read, or error.
+	// or copy each slice before the next call. Single-reader only: not
-	Read(p []wire.TunPacket, mem []byte) (int, error)
+	// safe for concurrent Reads (it reuses per-queue rx scratch each call).
+	Read() ([]Packet, error)
 
 	// Write emits a single packet on the plaintext (outside→inside)
-	// delivery path.
+	// delivery path. Safe for concurrent use.
 	Write(p []byte) (int, error)
-
-	// Capabilities returns the Queue's negotiated offload capabilities,
-	// or the zero value when q does not advertise any.
-	Capabilities() Capabilities
 }
 
-// GSOInfo describes a kernel-supplied superpacket sitting in Packet.Bytes.
+// Packet is the unit Queue.Read returns. Bytes points into the queue's
-// The zero value means "not a superpacket" — Bytes is one regular IP
+// internal buffer and is only valid until the next Read or Close on the
-// datagram and no segmentation is required.
+// queue that produced it.
-type GSOInfo struct {
+type Packet struct {
-	// Size is the GSO segment size: max payload bytes per segment
+	Bytes []byte
-	// (== TCP MSS for TSO, == UDP payload chunk for USO). Zero means
-	// not a superpacket.
-	Size uint16
-	// HdrLen is the total L3+L4 header length within Bytes (already
-	// corrected via correctHdrLen, so safe to slice on).
-	HdrLen uint16
-	// CsumStart is the L4 header offset inside Bytes (== L3 header
-	// length).
-	CsumStart uint16
-	// Proto picks the L4 protocol (TCP or UDP) so the segmenter knows
-	// which checksum/header layout to apply.
-	Proto GSOProto
 }
 
-// GSOProto selects the L4 protocol for a GSO superpacket. Determines which
+// Clone returns a Packet whose Bytes is a freshly allocated copy of p.Bytes,
-// VIRTIO_NET_HDR_GSO_* type the writer stamps and which checksum offset
+// safe to retain past the next Read or Close on the originating Queue.
-// inside the transport header virtio NEEDS_CSUM expects.
+// Use this only when a caller genuinely needs to outlive the borrowed-slice
-type GSOProto uint8
+// contract — the hot path reads should continue to consume the borrow
-
+// synchronously to avoid the allocation.
-const (
+func (p Packet) Clone() Packet {
-	GSOProtoNone GSOProto = iota
+	if p.Bytes == nil {
-	GSOProtoTCP
+		return p
-	GSOProtoUDP
-)
-
-// GSOWriter is implemented by Queues that can emit a TCP or UDP superpacket
-// assembled from a header prefix plus one or more borrowed payload
-// fragments, in a single vectored write (writev with a leading
-// virtio_net_hdr). This lets the coalescer avoid copying payload bytes
-// between the caller's decrypt buffer and the TUN. Backends without GSO
-// support do not implement this interface and coalescing is skipped.
-//
-// hdr contains the IPv4/IPv6 header prefix (mutable - callers will have
-// filled in total length and IP csum). transportHdr is the TCP or UDP
-// header (mutable - the L4 checksum field must hold the pseudo-header
-// partial, single-fold not inverted, per virtio NEEDS_CSUM semantics).
-// pays are non-overlapping payload fragments whose concatenation is the
-// full superpacket payload; they are read-only from the writer's
-// perspective and must remain valid until the call returns. Every segment
-// in pays except possibly the last is exactly the same size. proto picks
-// the L4 protocol so the writer knows which GSOType / CsumOffset to set.
-//
-// Callers should also consult CapsProvider (via SupportsGSO or
-// QueueCapabilities) for the per-protocol negotiated capability; an
-// implementation of GSOWriter is necessary but not sufficient since USO
-// may not have been negotiated even when TSO was.
-type GSOWriter interface {
-	WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte, proto GSOProto) error
-}
-
-// SupportsGSO reports whether w implements GSOWriter and the underlying
-// queue advertises the negotiated capability for `want`. A writer that
-// implements GSOWriter but not CapsProvider is treated as permissive
-// (used by tests and fakes that don't negotiate).
-func SupportsGSO(w Queue, want GSOProto) (GSOWriter, bool) {
-	gw, ok := w.(GSOWriter)
-	if !ok {
-		return nil, false
-	}
-	caps := w.Capabilities()
-	switch want {
-	case GSOProtoTCP:
-		return gw, caps.TSO
-	case GSOProtoUDP:
-		return gw, caps.USO
-	default:
-		return gw, false
 	}
+	cp := make([]byte, len(p.Bytes))
+	copy(cp, p.Bytes)
+	return Packet{Bytes: cp}
 }

overlay/tio/tio_poll_linux.go

@@ -1,41 +1,42 @@
+//go:build linux && !android
+// +build linux,!android
+
 package tio
 
 import (
 	"fmt"
 	"os"
-	"sync"
 	"sync/atomic"
 
-	"github.com/slackhq/nebula/wire"
 	"golang.org/x/sys/unix"
 )
 
-type Poll struct {
+// Maximum size we accept for a single read from a TUN. 65535 covers any
-	fd int
+// single IP packet.
+const tunReadBufSize = 65535
 
-	readPoll  [2]unix.PollFd
+type Poll struct {
-	writePoll [2]unix.PollFd
+	fd         int
-	writeLock sync.Mutex
+	shutdownFd int
-	closed    atomic.Bool
+	closed     atomic.Bool
+
+	readBuf  []byte
+	batchRet [1]Packet
 }
 
+// newPoll wraps an existing tun fd. On failure it does NOT close fd: the
+// caller owns fd and is the sole closer (see pollQueueSet.Add callers in
+// overlay/tun_linux.go, which unix.Close on Add error). This keeps closes
+// at exactly one on every path.
 func newPoll(fd int, shutdownFd int) (*Poll, error) {
 	if err := unix.SetNonblock(fd, true); err != nil {
-		_ = unix.Close(fd)
 		return nil, fmt.Errorf("failed to set Poll device as nonblocking: %w", err)
 	}
 
 	out := &Poll{
-		fd: fd,
+		fd:         fd,
-		readPoll: [2]unix.PollFd{
+		shutdownFd: shutdownFd,
-			{Fd: int32(fd), Events: unix.POLLIN},
+		readBuf:    make([]byte, tunReadBufSize),
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writeLock: sync.Mutex{},
 	}
 	return out, nil
 }
@@ -43,68 +44,20 @@ func newPoll(fd int, shutdownFd int) (*Poll, error) {
 // blockOnRead waits until the Poll fd is readable or shutdown has been signaled.
 // Returns os.ErrClosed if Close was called.
 func (t *Poll) blockOnRead() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	return blockOn(int32(t.fd), int32(t.shutdownFd), unix.POLLIN)
-	var err error
-	for {
-		_, err = unix.Poll(t.readPoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	tunEvents := t.readPoll[0].Revents
-	shutdownEvents := t.readPoll[1].Revents
-	t.readPoll[0].Revents = 0
-	t.readPoll[1].Revents = 0
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	}
-	if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
 }
 
 func (t *Poll) blockOnWrite() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	return blockOn(int32(t.fd), int32(t.shutdownFd), unix.POLLOUT)
-	var err error
-	for {
-		_, err = unix.Poll(t.writePoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	t.writeLock.Lock()
-	tunEvents := t.writePoll[0].Revents
-	shutdownEvents := t.writePoll[1].Revents
-	t.writePoll[0].Revents = 0
-	t.writePoll[1].Revents = 0
-	t.writeLock.Unlock()
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	}
-	if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
 }
 
-func (t *Poll) Read(p []wire.TunPacket, mem []byte) (int, error) {
+func (t *Poll) Read() ([]Packet, error) {
-	if len(p) == 0 || len(mem) == 0 {
+	n, err := t.readOne(t.readBuf)
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.readOne(mem)
 	if err != nil {
-		return 0, err
+		return nil, err
 	}
-	p[0].Bytes = mem[:n]
+	t.batchRet[0] = Packet{Bytes: t.readBuf[:n]}
-	return 1, nil
+	return t.batchRet[:], nil
 }
 
 func (t *Poll) readOne(to []byte) (int, error) {
@@ -128,6 +81,7 @@ func (t *Poll) readOne(to []byte) (int, error) {
 	}
 }
 
+// Write is safe for concurrent use
 func (t *Poll) Write(from []byte) (int, error) {
 	for {
 		n, errno := unix.Write(t.fd, from)
@@ -154,15 +108,9 @@ func (t *Poll) Close() error {
 		return nil
 	}
 	//shutdownFd is owned by the container, so we should not close it
-	var err error
+	// Close the underlying fd but do NOT null t.fd: a reader may still be
-	if t.fd >= 0 {
+	// loading it in readOne, and mutating the field would race that load.
-		err = unix.Close(t.fd)
+	// It gets EBADF -> os.ErrClosed (or wakes via the shutdown eventfd's
-		t.fd = -1
+	// ppoll first). closed.Swap already guarantees we only close once.
-	}
+	return unix.Close(t.fd)
-
-	return err
-}
-
-func (t *Poll) Capabilities() Capabilities {
-	return Capabilities{}
 }

overlay/tio/tun_file_linux_test.go

@@ -10,7 +10,6 @@ import (
 	"testing"
 	"time"
 
-	"github.com/slackhq/nebula/wire"
 	"github.com/stretchr/testify/require"
 	"golang.org/x/sys/unix"
 )
@@ -28,13 +27,16 @@ func newReadPipe(t *testing.T) int {
 }
 
 func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
+	pipe1 := newReadPipe(t)
+	pipe2 := newReadPipe(t)
 	parent, err := NewPollQueueSet()
 	require.NoError(t, err)
-	require.NoError(t, parent.Add(newReadPipe(t)))
+	require.NoError(t, parent.Add(pipe1))
-	require.NoError(t, parent.Add(newReadPipe(t)))
+	require.NoError(t, parent.Add(pipe2))
-	// QueueSet.Close owns the read fds we Added — don't register a separate
+	t.Cleanup(func() {
-	// Cleanup to close them or we'll double-close whatever fd the kernel
+		_ = unix.Close(pipe1)
-	// has since reused.
+		_ = unix.Close(pipe2)
+	})
 
 	readers := parent.Queues()
 	errs := make([]error, len(readers))
@@ -43,8 +45,7 @@ func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
 		wg.Add(1)
 		go func(i int, r Queue) {
 			defer wg.Done()
-			pkts := make([]wire.TunPacket, 1)
+			_, errs[i] = r.Read()
-			_, errs[i] = r.Read(pkts, make([]byte, 64))
 		}(i, r)
 	}
 
@@ -69,12 +70,99 @@ func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
 	}
 }
 
-func TestPoll_Close_Idempotent(t *testing.T) {
+// TestPoll_ConcurrentWrite_NoRace hammers a single Poll queue from two writer
+// goroutines while a reader drains the other end of the pipe. The writers
+// overflow the pipe buffer, so both repeatedly park in blockOnWrite at the same
+// time — the exact scenario that raced on the old shared writePoll member
+// array. Run under -race; a shared-array regression trips the detector here.
+func TestPoll_ConcurrentWrite_NoRace(t *testing.T) {
+	var fds [2]int
+	require.NoError(t, unix.Pipe2(fds[:], unix.O_CLOEXEC))
+	readFd, writeFd := fds[0], fds[1]
+
 	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = unix.Close(shutdownFd) })
 
-	tf, err := newPoll(newReadPipe(t), shutdownFd)
+	p, err := newPoll(writeFd, shutdownFd)
+	require.NoError(t, err)
+
+	const writers = 2
+	const perWriter = 4000
+	payload := make([]byte, 100)
+	total := writers * perWriter * len(payload)
+
+	// Reader: drain the read end (blocking) until every writer's bytes are
+	// consumed, so the writers keep making progress rather than wedging on a
+	// permanently full pipe.
+	readDone := make(chan struct{})
+	go func() {
+		defer close(readDone)
+		buf := make([]byte, 4096)
+		got := 0
+		for got < total {
+			n, rerr := unix.Read(readFd, buf)
+			got += n
+			if rerr != nil {
+				if rerr == unix.EINTR {
+					continue
+				}
+				return
+			}
+			if n == 0 { // EOF
+				return
+			}
+		}
+	}()
+
+	var wg sync.WaitGroup
+	for w := 0; w < writers; w++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for i := 0; i < perWriter; i++ {
+				if _, werr := p.Write(payload); werr != nil {
+					t.Errorf("write: %v", werr)
+					return
+				}
+			}
+		}()
+	}
+	wg.Wait()
+
+	select {
+	case <-readDone:
+	case <-time.After(10 * time.Second):
+		t.Fatal("reader did not drain")
+	}
+
+	require.NoError(t, p.Close())
+	_ = unix.Close(readFd)
+}
+
+// TestPoll_NewPoll_DoesNotCloseFdOnFailure pins the ownership rule: when
+// newPoll fails, it must leave fd open so the caller (pollQueueSet.Add's
+// callers in tun_linux.go) is the sole closer. If newPoll also closed fd,
+// the poll path would double-close on Add error. We force the failure with
+// an O_PATH descriptor: fcntl(F_SETFL) — which SetNonblock performs — is not
+// permitted on O_PATH fds and fails with EBADF, while the fd itself stays
+// open so we can observe that newPoll left it alone.
+func TestPoll_NewPoll_DoesNotCloseFdOnFailure(t *testing.T) {
+	fd, err := unix.Open("/", unix.O_PATH|unix.O_CLOEXEC, 0)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = unix.Close(fd) })
+
+	p, err := newPoll(fd, 1)
+	require.Error(t, err, "SetNonblock on an O_PATH fd should fail")
+	require.Nil(t, p)
+
+	// If newPoll had closed fd, F_GETFD would report it closed. It staying
+	// open proves newPoll left the fd for the caller to close exactly once.
+	require.True(t, fdOpen(t, fd), "newPoll must not close fd on failure; caller is the sole closer")
+}
+
+func TestPoll_Close_Idempotent(t *testing.T) {
+	tf, err := newPoll(newReadPipe(t), 1)
 	require.NoError(t, err)
 	if err := tf.Close(); err != nil {
 		t.Fatalf("first Close: %v", err)
@@ -84,23 +172,37 @@ func TestPoll_Close_Idempotent(t *testing.T) {
 	}
 }
 
-func TestPollQueueSet_Close_ClosesEventfd(t *testing.T) {
+// fdOpen reports whether fd currently refers to an open file description.
+// A closed (or never-allocated) fd makes F_GETFD fail with EBADF.
+func fdOpen(t *testing.T, fd int) bool {
+	t.Helper()
+	_, err := unix.FcntlInt(uintptr(fd), unix.F_GETFD, 0)
+	if err == nil {
+		return true
+	}
+	if errors.Is(err, unix.EBADF) {
+		return false
+	}
+	t.Fatalf("unexpected fcntl(F_GETFD) error on fd %d: %v", fd, err)
+	return false
+}
+
+// TestPollQueueSet_Close_ClosesShutdownFd is the regression test for the
+// leaked shutdown eventfd: the container that owns shutdownFd must close it in
+// Close, and a second Close must be a safe no-op.
+func TestPollQueueSet_Close_ClosesShutdownFd(t *testing.T) {
 	qs, err := NewPollQueueSet()
 	require.NoError(t, err)
+	c, ok := qs.(*pollQueueSet)
+	require.True(t, ok)
 	require.NoError(t, qs.Add(newReadPipe(t)))
 
-	fd := qs.(*pollQueueSet).shutdownFd
+	shutdownFd := c.shutdownFd
+	require.True(t, fdOpen(t, shutdownFd), "shutdown eventfd should be open before Close")
+
 	require.NoError(t, qs.Close())
+	require.False(t, fdOpen(t, shutdownFd), "shutdown eventfd should be closed after Close")
 
-	// Closing the eventfd again should fail with EBADF, proving Close
+	// Second Close must not touch fds (shutdownFd is now -1) and must return nil.
-	// actually released it.
+	require.NoError(t, qs.Close())
-	if err := unix.Close(fd); err == nil {
-		t.Fatalf("eventfd %d still open after QueueSet.Close", fd)
-	}
-
-	// Second Close must be a no-op (and must not double-close the eventfd
-	// in case the kernel handed it out to another caller in the meantime).
-	if err := qs.Close(); err != nil {
-		t.Fatalf("second Close: %v", err)
-	}
 }

overlay/tun_android.go

@@ -16,11 +16,10 @@ import (
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
-	"github.com/slackhq/nebula/wire"
 )
 
 type tun struct {
-	rwc         io.ReadWriteCloser
+	io.ReadWriteCloser
 	fd          int
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
@@ -28,41 +27,21 @@ type tun struct {
 	l           *slog.Logger
 }
 
-func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.rwc.Read(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
-}
-
-func (t *tun) Write(p []byte) (int, error) {
-	return t.rwc.Write(p)
-}
-
-func (t *tun) Close() error {
-	return t.rwc.Close()
-}
-
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
 	// XXX Android returns an fd in non-blocking mode which is necessary for shutdown to work properly.
 	// Be sure not to call file.Fd() as it will set the fd to blocking mode.
 	file := os.NewFile(uintptr(deviceFd), "/dev/net/tun")
 
 	t := &tun{
-		rwc:         file,
+		ReadWriteCloser: file,
-		fd:          deviceFd,
+		fd:              deviceFd,
-		vpnNetworks: vpnNetworks,
+		vpnNetworks:     vpnNetworks,
-		l:           l,
+		l:               l,
 	}
 
 	err := t.reload(c, true)
 	if err != nil {
+		_ = file.Close()
 		return nil, err
 	}
 
@@ -118,18 +97,6 @@ func (t *tun) Name() string {
 	return "android"
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *tun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for android")
-}
-
-func (t *tun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *tun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
 }

overlay/tun_darwin.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -19,13 +18,12 @@ import (
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
-	"github.com/slackhq/nebula/wire"
 	netroute "golang.org/x/net/route"
 	"golang.org/x/sys/unix"
 )
 
 type tun struct {
-	rwc         io.ReadWriteCloser
+	f           *os.File
 	Device      string
 	vpnNetworks []netip.Prefix
 	DefaultMTU  int
@@ -33,9 +31,6 @@ type tun struct {
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	linkAddr    *netroute.LinkAddr
 	l           *slog.Logger
-
-	// cache out buffer since we need to prepend 4 bytes for tun metadata
-	out []byte
 }
 
 type ifReq struct {
@@ -126,7 +121,7 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, _ bool) (*t
 	}
 
 	t := &tun{
-		rwc:         os.NewFile(uintptr(fd), ""),
+		f:           os.NewFile(uintptr(fd), ""),
 		Device:      name,
 		vpnNetworks: vpnNetworks,
 		DefaultMTU:  c.GetInt("tun.mtu", DefaultMTU),
@@ -160,8 +155,8 @@ func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (*tun, e
 }
 
 func (t *tun) Close() error {
-	if t.rwc != nil {
+	if t.f != nil {
-		return t.rwc.Close()
+		return t.f.Close()
 	}
 	return nil
 }
@@ -504,46 +499,103 @@ func delRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 	return nil
 }
 
-func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+// tunWritev and tunReadv are linkname'd to x/sys/unix's libc-routed writev/readv stubs so the
-	if len(p) == 0 || len(mem) <= 4 {
+// calls go through libSystem's pinned trampoline. A raw syscall.Syscall(SYS_WRITEV/SYS_READV, ...)
-		return 0, nil //todo should this be an err?
+// on darwin/arm64 emits an SVC #0x80 trap (see $GOROOT/src/syscall/asm_darwin_arm64.s), the path
-	}
+// Apple keeps warning they will eventually disallow. We pull the low-level stubs instead of calling
-	p[0].Meta = struct{}{}
+// unix.Writev/unix.Readv because those take [][]byte and rebuild the []Iovec every call, which
-	n, err := t.rwc.Read(mem)
+// heap-allocates the header; linkname'ing the stubs lets us hand them our own stack-allocated
+// iovecs. See golang/go#78049.
+
+//go:linkname tunWritev golang.org/x/sys/unix.writev
+//go:noescape
+func tunWritev(fd int, iovecs []unix.Iovec) (n int, err error)
+
+//go:linkname tunReadv golang.org/x/sys/unix.readv
+//go:noescape
+func tunReadv(fd int, iovecs []unix.Iovec) (n int, err error)
+
+// Read pulls one IP packet off the utun device, scattering the 4 byte protocol header away from
+// the packet so the payload lands directly in to.
+func (t *tun) Read(to []byte) (int, error) {
+	var head [4]byte
+
+	rc, err := t.f.SyscallConn()
 	if err != nil {
 		return 0, err
 	}
-	p[0].Bytes = mem[4:n]
+
-	return 1, nil
+	var n int
+	var callErr error
+	err = rc.Read(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &to[0], Len: uint64(len(to))},
+		}
+		n, callErr = tunReadv(int(fd), iovecs)
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+	if n < 4 {
+		return 0, nil
+	}
+	return n - 4, nil
 }
 
-// Write is only valid for single threaded use
+// Write pushes one IP packet onto the utun device.
 func (t *tun) Write(from []byte) (int, error) {
-	buf := t.out
-	if cap(buf) < len(from)+4 {
-		buf = make([]byte, len(from)+4)
-		t.out = buf
-	}
-	buf = buf[:len(from)+4]
-
 	if len(from) == 0 {
 		return 0, syscall.EIO
 	}
 
-	// Determine the IP Family for the NULL L2 Header
 	ipVer := from[0] >> 4
-	if ipVer == 4 {
+	var head [4]byte
-		buf[3] = syscall.AF_INET
+	switch ipVer {
-	} else if ipVer == 6 {
+	case 4:
-		buf[3] = syscall.AF_INET6
+		head[3] = syscall.AF_INET
-	} else {
+	case 6:
+		head[3] = syscall.AF_INET6
+	default:
 		return 0, fmt.Errorf("unable to determine IP version from packet")
 	}
 
-	copy(buf[4:], from)
+	// Grab rc as a local so the compiler can devirtualize the call and keep the closure on the stack.
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
 
-	n, err := t.rwc.Write(buf)
+	var n int
-	return n - 4, err
+	var callErr error
+	err = rc.Write(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &from[0], Len: uint64(len(from))},
+		}
+		n, callErr = tunWritev(int(fd), iovecs)
+		// Type-assert to syscall.Errno so the EAGAIN/EWOULDBLOCK/EINTR check doesn't box the errno
+		// constants into error interfaces on every call.
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+
+	return n - 4, nil
 }
 
 func (t *tun) Networks() []netip.Prefix {
@@ -554,18 +606,6 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *tun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for darwin")
-}
-
-func (t *tun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *tun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
 }

overlay/tun_disabled.go

@@ -12,7 +12,6 @@ import (
 	"github.com/slackhq/nebula/iputil"
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
-	"github.com/slackhq/nebula/wire"
 )
 
 type disabledTun struct {
@@ -20,10 +19,26 @@ type disabledTun struct {
 	vpnNetworks []netip.Prefix
 
 	// Track these metrics since we don't have the tun device to do it for us
-	tx         metrics.Counter
+	tx metrics.Counter
-	rx         metrics.Counter
+	rx metrics.Counter
-	numReaders int
+	l  *slog.Logger
-	l          *slog.Logger
+}
+
+// Read hands the next queued packet to a reader, copying it into b. Reads
+// from concurrent queues are safe: the channel receive serializes them and
+// each queue copies into its own private scratch buffer.
+func (t *disabledTun) Read(b []byte) (int, error) {
+	r, ok := <-t.read
+	if !ok {
+		return 0, io.EOF
+	}
+
+	t.tx.Inc(1)
+	if t.l.Enabled(context.Background(), slog.LevelDebug) {
+		t.l.Debug("Write payload", "raw", prettyPacket(r))
+	}
+
+	return copy(b, r), nil
 }
 
 func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled bool, l *slog.Logger) *disabledTun {
@@ -31,7 +46,6 @@ func newDisabledTun(vpnNetworks []netip.Prefix, queueLen int, metricsEnabled boo
 		vpnNetworks: vpnNetworks,
 		read:        make(chan []byte, queueLen),
 		l:           l,
-		numReaders:  1,
 	}
 
 	if metricsEnabled {
@@ -61,37 +75,6 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
-func (t *disabledTun) readOne(b []byte) (int, error) {
-	r, ok := <-t.read
-	if !ok {
-		return 0, io.EOF
-	}
-
-	if len(r) > len(b) {
-		return 0, fmt.Errorf("packet larger than mtu: %d > %d bytes", len(r), len(b))
-	}
-
-	t.tx.Inc(1)
-	if t.l.Enabled(context.Background(), slog.LevelDebug) {
-		t.l.Debug("Write payload", "raw", prettyPacket(r))
-	}
-
-	return copy(b, r), nil
-}
-
-func (t *disabledTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.readOne(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
-}
-
 func (t *disabledTun) handleICMPEchoRequest(b []byte) bool {
 	out := make([]byte, len(b))
 	out = iputil.CreateICMPEchoResponse(b, out)
@@ -123,25 +106,14 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
-func (t *disabledTun) SupportsMultiqueue() bool {
+func (t *disabledTun) Queues(n int) ([]tio.Queue, error) {
-	return true
+	out := make([]tio.Queue, n)
-}
+	for i := range out {
-
+		// NoClose: the shared channel and metrics are owned by the
-func (t *disabledTun) NewMultiQueueReader() error {
+		// disabledTun; Close on the device tears them down once for everybody.
-	t.numReaders++
+		out[i] = tio.NewSingleQueueNoClose(t, defaultBatchBufSize)
-	return nil
-}
-
-func (t *disabledTun) Readers() []tio.Queue {
-	out := make([]tio.Queue, t.numReaders)
-	for i := range t.numReaders {
-		out[i] = t
 	}
-	return out
+	return out, nil
-}
-
-func (t *disabledTun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
 }
 
 func (t *disabledTun) Close() error {

overlay/tun_freebsd.go

@@ -17,7 +17,6 @@ import (
 	"unsafe"
 
 	"github.com/gaissmai/bart"
-	"github.com/slackhq/nebula/wire"
 
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/overlay/tio"
@@ -157,20 +156,7 @@ func (t *tun) blockOnWrite() error {
 	return nil
 }
 
-func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+func (t *tun) Read(to []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.readOne(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
-}
-
-func (t *tun) readOne(to []byte) (int, error) {
 	// first 4 bytes is protocol family, in network byte order
 	var head [4]byte
 	iovecs := [2]syscall.Iovec{
@@ -574,12 +560,8 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *tun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {
@@ -606,14 +588,6 @@ func (t *tun) addRoutes(logErrors bool) error {
 	return nil
 }
 
-func (t *tun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *tun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
-}
-
 func (t *tun) removeRoutes(routes []Route) error {
 	for _, r := range routes {
 		if !r.Install {
@@ -680,7 +654,6 @@ func addRoute(prefix netip.Prefix, gateway netroute.Addr) error {
 				return fmt.Errorf("failed to create route.RouteMessage for change: %w", err)
 			}
 			_, err = unix.Write(sock, data[:])
-			fmt.Println("DOING CHANGE")
 			return err
 		}
 		return fmt.Errorf("failed to write route.RouteMessage to socket: %w", err)

overlay/tun_ios.go

@@ -19,52 +19,38 @@ import (
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
-	"github.com/slackhq/nebula/wire"
+	"golang.org/x/sys/unix"
 )
 
 type tun struct {
-	rwc         io.ReadWriteCloser
+	io.ReadWriteCloser
 	vpnNetworks []netip.Prefix
 	Routes      atomic.Pointer[[]Route]
 	routeTree   atomic.Pointer[bart.Table[routing.Gateways]]
 	l           *slog.Logger
 }
 
-func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
-	if len(p) == 0 || len(mem) <= 4 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.rwc.Read(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[4:n]
-	return 1, nil
-}
-
-func (t *tun) Write(p []byte) (int, error) {
-	return t.rwc.Write(p)
-}
-
-func (t *tun) Close() error {
-	return t.rwc.Close()
-}
-
 func newTun(_ *config.C, _ *slog.Logger, _ []netip.Prefix, _ bool) (*tun, error) {
 	return nil, fmt.Errorf("newTun not supported in iOS")
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
+	if err := unix.SetNonblock(deviceFd, true); err != nil {
+		// We own the fd from the moment it is handed to us, same as the reload error path below
+		_ = unix.Close(deviceFd)
+		return nil, fmt.Errorf("failed to set the tun fd to non-blocking mode: %w", err)
+	}
+
 	file := os.NewFile(uintptr(deviceFd), "/dev/tun")
 	t := &tun{
-		vpnNetworks: vpnNetworks,
+		vpnNetworks:     vpnNetworks,
-		rwc:         &tunReadCloser{f: file},
+		ReadWriteCloser: &tunReadCloser{f: file},
-		l:           l,
+		l:               l,
 	}
 
 	err := t.reload(c, true)
 	if err != nil {
+		_ = file.Close()
 		return nil, err
 	}
 
@@ -119,9 +105,18 @@ type tunReadCloser struct {
 	wBuf []byte
 }
 
-// Read returns a packet with the BSD 4-byte header, watch out!
 func (tr *tunReadCloser) Read(to []byte) (int, error) {
-	return tr.f.Read(to)
+	tr.rMu.Lock()
+	defer tr.rMu.Unlock()
+
+	if cap(tr.rBuf) < len(to)+4 {
+		tr.rBuf = make([]byte, len(to)+4)
+	}
+	tr.rBuf = tr.rBuf[:len(to)+4]
+
+	n, err := tr.f.Read(tr.rBuf)
+	copy(to, tr.rBuf[4:])
+	return n - 4, err
 }
 
 func (tr *tunReadCloser) Write(from []byte) (int, error) {
@@ -165,18 +160,6 @@ func (t *tun) Name() string {
 	return "iOS"
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *tun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for ios")
-}
-
-func (t *tun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *tun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
 }

overlay/tun_linux.go

@@ -72,9 +72,7 @@ type ifreqQLEN struct {
 }
 
 func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip.Prefix) (*tun, error) {
-	// We don't know what flags the caller opened this fd with and can't turn
+	t, err := newTunGeneric(c, l, deviceFd, vpnNetworks)
-	// on IFF_VNET_HDR after TUNSETIFF, so skip offload on inherited fds.
-	t, err := newTunGeneric(c, l, deviceFd, false, false, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -130,14 +128,13 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 	if err != nil {
 		return nil, err
 	}
-
 	name, err := tunSetIff(fd, nameStr, baseFlags)
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, &NameError{Name: nameStr, Underlying: err}
 	}
 
-	t, err := newTunGeneric(c, l, fd, false, false, vpnNetworks)
+	t, err := newTunGeneric(c, l, fd, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -147,10 +144,10 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 	return t, nil
 }
 
-// newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
+// newTunGeneric does all the stuff common to different tun initialization
-func newTunGeneric(c *config.C, l *slog.Logger, fd int, vnetHdr, usoEnabled bool, vpnNetworks []netip.Prefix) (*tun, error) {
+// paths. It will close your files on error.
+func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
 	qs, err := tio.NewPollQueueSet()
-
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, err
@@ -261,11 +258,20 @@ func (t *tun) reload(c *config.C, initial bool) error {
 	return nil
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+// Queues opens additional kernel multiqueue fds until the device has n
-	return true
+// queues, then returns them all. The first queue was opened by newTun.
+func (t *tun) Queues(n int) ([]tio.Queue, error) {
+	for len(t.readers.Queues()) < n {
+		if err := t.addQueue(); err != nil {
+			return nil, err
+		}
+	}
+	return t.readers.Queues(), nil
 }
 
-func (t *tun) NewMultiQueueReader() error {
+// addQueue opens one more IFF_MULTI_QUEUE fd on the device and adds it to
+// the queue set.
+func (t *tun) addQueue() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
 
@@ -275,7 +281,6 @@ func (t *tun) NewMultiQueueReader() error {
 	}
 
 	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
-
 	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
 		_ = unix.Close(fd)
 		return err
@@ -455,15 +460,6 @@ func (t *tun) setDefaultRoute(cidr netip.Prefix) error {
 		Table:     unix.RT_TABLE_MAIN,
 		Type:      unix.RTN_UNICAST,
 	}
-	// Match the metric the kernel uses for its auto-installed connected
-	// route, so RouteReplace overwrites it in place instead of adding a
-	// second route at a worse metric. IPv6 connected routes are installed
-	// at metric 256 (IP6_RT_PRIO_KERN); IPv4 uses 0. Without this, the
-	// kernel route wins lookups and our MTU / AdvMSS / Features never
-	// apply on v6.
-	if cidr.Addr().Is6() {
-		nr.Priority = 256
-	}
 	err := netlink.RouteReplace(&nr)
 	if err != nil {
 		t.l.Warn("Failed to set default route MTU, retrying", "error", err, "cidr", cidr)
@@ -730,10 +726,6 @@ func (t *tun) updateRoutes(r netlink.RouteUpdate) {
 	t.routeTree.Store(newTree)
 }
 
-func (t *tun) Readers() []tio.Queue {
-	return t.readers.Queues()
-}
-
 func (t *tun) Close() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()

overlay/tun_linux_test.go

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

overlay/tun_netbsd.go

@@ -19,7 +19,6 @@ import (
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
-	"github.com/slackhq/nebula/wire"
 	netroute "golang.org/x/net/route"
 	"golang.org/x/sys/unix"
 )
@@ -69,27 +68,6 @@ type tun struct {
 	fd          int
 }
 
-func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.readOne(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
-}
-
-func (t *tun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *tun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
-}
-
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
 
 func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (*tun, error) {
@@ -163,7 +141,7 @@ func (t *tun) Close() error {
 	return nil
 }
 
-func (t *tun) readOne(to []byte) (int, error) {
+func (t *tun) Read(to []byte) (int, error) {
 	rc, err := t.f.SyscallConn()
 	if err != nil {
 		return 0, fmt.Errorf("failed to get syscall conn for tun: %w", err)
@@ -412,12 +390,8 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *tun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_openbsd.go

@@ -19,7 +19,6 @@ import (
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
-	"github.com/slackhq/nebula/wire"
 	netroute "golang.org/x/net/route"
 	"golang.org/x/sys/unix"
 )
@@ -58,21 +57,6 @@ type tun struct {
 	l           *slog.Logger
 	f           *os.File
 	fd          int
-	// cache out buffer since we need to prepend 4 bytes for tun metadata
-	out []byte
-}
-
-func (t *tun) Read(p []wire.TunPacket, mem []byte) (int, error) {
-	if len(p) == 0 || len(mem) <= 4 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.f.Read(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[4:n]
-	return 1, nil
 }
 
 var deviceNameRE = regexp.MustCompile(`^tun[0-9]+$`)
@@ -138,33 +122,103 @@ func (t *tun) Close() error {
 	return nil
 }
 
-// Write is only valid for single threaded use
+// tunWritev and tunReadv are linkname'd to x/sys/unix's libc-routed writev/readv stubs so the
-func (t *tun) Write(from []byte) (int, error) {
+// calls go through libc's pinned trampoline. OpenBSD's pinsyscall protection rejects a raw
-	buf := t.out
+// syscall.Syscall(SYS_WRITEV/SYS_READV, ...) because it doesn't originate from a libc-pinned
-	if cap(buf) < len(from)+4 {
+// address, so we can't use the syscall.Syscall pattern that freebsd / netbsd use. We pull the
-		buf = make([]byte, len(from)+4)
+// low-level stubs instead of calling unix.Writev/unix.Readv because those take [][]byte and rebuild
-		t.out = buf
+// the []Iovec every call, which heap-allocates the header; linkname'ing the stubs lets us hand them
-	}
+// our own stack-allocated iovecs. See golang/go#78049.
-	buf = buf[:len(from)+4]
 
+//go:linkname tunWritev golang.org/x/sys/unix.writev
+//go:noescape
+func tunWritev(fd int, iovecs []unix.Iovec) (n int, err error)
+
+//go:linkname tunReadv golang.org/x/sys/unix.readv
+//go:noescape
+func tunReadv(fd int, iovecs []unix.Iovec) (n int, err error)
+
+// Read pulls one IP packet off the tun device, scattering the 4 byte protocol header away from
+// the packet so the payload lands directly in to.
+func (t *tun) Read(to []byte) (int, error) {
+	var head [4]byte
+
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
+
+	var n int
+	var callErr error
+	err = rc.Read(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &to[0], Len: uint64(len(to))},
+		}
+		n, callErr = tunReadv(int(fd), iovecs)
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+	if n < 4 {
+		return 0, nil
+	}
+	return n - 4, nil
+}
+
+// Write pushes one IP packet onto the tun device.
+func (t *tun) Write(from []byte) (int, error) {
 	if len(from) == 0 {
 		return 0, syscall.EIO
 	}
 
-	// Determine the IP Family for the NULL L2 Header
 	ipVer := from[0] >> 4
-	if ipVer == 4 {
+	var head [4]byte
-		buf[3] = syscall.AF_INET
+	switch ipVer {
-	} else if ipVer == 6 {
+	case 4:
-		buf[3] = syscall.AF_INET6
+		head[3] = syscall.AF_INET
-	} else {
+	case 6:
+		head[3] = syscall.AF_INET6
+	default:
 		return 0, fmt.Errorf("unable to determine IP version from packet")
 	}
 
-	copy(buf[4:], from)
+	// Grab rc as a local so the compiler can devirtualize the call and keep the closure on the stack.
+	rc, err := t.f.SyscallConn()
+	if err != nil {
+		return 0, err
+	}
 
-	n, err := t.f.Write(buf)
+	var n int
-	return n - 4, err
+	var callErr error
+	err = rc.Write(func(fd uintptr) bool {
+		iovecs := []unix.Iovec{
+			{Base: &head[0], Len: 4},
+			{Base: &from[0], Len: uint64(len(from))},
+		}
+		n, callErr = tunWritev(int(fd), iovecs)
+		// Type-assert to syscall.Errno so the EAGAIN/EWOULDBLOCK/EINTR check doesn't box the errno
+		// constants into error interfaces on every call.
+		if errno, ok := callErr.(syscall.Errno); ok && errno.Temporary() {
+			return false
+		}
+		return true
+	})
+	if err != nil {
+		return 0, err
+	}
+	if callErr != nil {
+		return 0, callErr
+	}
+
+	return n - 4, nil
 }
 
 func (t *tun) addIp(cidr netip.Prefix) error {
@@ -315,12 +369,8 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsMultiqueue() bool {
+func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *tun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for openbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {
@@ -371,14 +421,6 @@ func (t *tun) deviceBytes() (o [16]byte) {
 	return
 }
 
-func (t *tun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *tun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
-}
-
 func addRoute(prefix netip.Prefix, gateways []netip.Prefix) error {
 	sock, err := unix.Socket(unix.AF_ROUTE, unix.SOCK_RAW, unix.AF_UNSPEC)
 	if err != nil {

overlay/tun_tester.go

@@ -17,7 +17,6 @@ import (
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/udp"
-	"github.com/slackhq/nebula/wire"
 )
 
 type TestTun struct {
@@ -164,20 +163,7 @@ func (t *TestTun) Close() error {
 	return nil
 }
 
-func (t *TestTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+func (t *TestTun) Read(b []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.read(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
-}
-
-func (t *TestTun) read(b []byte) (int, error) {
 	p, ok := <-t.rxPackets
 	if !ok {
 		return 0, os.ErrClosed
@@ -192,18 +178,6 @@ func (t *TestTun) read(b []byte) (int, error) {
 	return n, nil
 }
 
-func (t *TestTun) Readers() []tio.Queue {
+func (t *TestTun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{t}
+	return []tio.Queue{tio.NewSingleQueue(t, udp.MTU)}, nil
-}
-
-func (t *TestTun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
-}
-
-func (t *TestTun) SupportsMultiqueue() bool {
-	return false
-}
-
-func (t *TestTun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented")
 }

overlay/tun_windows.go

@@ -21,7 +21,6 @@ import (
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/slackhq/nebula/wintun"
-	"github.com/slackhq/nebula/wire"
 	"golang.org/x/sys/windows"
 	"golang.zx2c4.com/wireguard/windows/tunnel/winipcfg"
 )
@@ -48,17 +47,8 @@ type winTun struct {
 	tun *wintun.NativeTun
 }
 
-func (t *winTun) Read(p []wire.TunPacket, mem []byte) (int, error) {
+func (t *winTun) Read(b []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
+	return t.tun.Read(b, 0)
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := t.tun.Read(mem, 0)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
 }
 
 func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (Device, error) {
@@ -273,20 +263,8 @@ func (t *winTun) Write(b []byte) (int, error) {
 	return t.tun.Write(b, 0)
 }
 
-func (t *winTun) SupportsMultiqueue() bool {
+func (t *winTun) Queues(int) ([]tio.Queue, error) {
-	return false
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
-}
-
-func (t *winTun) NewMultiQueueReader() error {
-	return fmt.Errorf("TODO: multiqueue not implemented for windows")
-}
-
-func (t *winTun) Readers() []tio.Queue {
-	return []tio.Queue{t}
-}
-
-func (t *winTun) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
 }
 
 func (t *winTun) Close() error {

overlay/user.go

@@ -8,7 +8,6 @@ import (
 	"github.com/slackhq/nebula/config"
 	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
-	"github.com/slackhq/nebula/wire"
 )
 
 func NewUserDeviceFromConfig(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, routines int) (Device, error) {
@@ -25,13 +24,11 @@ func NewUserDevice(vpnNetworks []netip.Prefix) (Device, error) {
 		outboundWriter: ow,
 		inboundReader:  ir,
 		inboundWriter:  iw,
-		numReaders:     1,
 	}, nil
 }
 
 type UserDevice struct {
 	vpnNetworks []netip.Prefix
-	numReaders  int
 
 	outboundReader *io.PipeReader
 	outboundWriter *io.PipeWriter
@@ -40,23 +37,6 @@ type UserDevice struct {
 	inboundWriter *io.PipeWriter
 }
 
-func (d *UserDevice) Capabilities() tio.Capabilities {
-	return tio.Capabilities{}
-}
-
-func (d *UserDevice) Read(p []wire.TunPacket, mem []byte) (int, error) {
-	if len(p) == 0 || len(mem) == 0 {
-		return 0, nil //todo should this be an err?
-	}
-	p[0].Meta = struct{}{}
-	n, err := d.outboundReader.Read(mem)
-	if err != nil {
-		return 0, err
-	}
-	p[0].Bytes = mem[:n]
-	return 1, nil
-}
-
 func (d *UserDevice) Activate() error {
 	return nil
 }
@@ -67,27 +47,26 @@ func (d *UserDevice) RoutesFor(ip netip.Addr) routing.Gateways {
 	return routing.Gateways{routing.NewGateway(ip, 1)}
 }
 
-func (d *UserDevice) SupportsMultiqueue() bool {
+func (d *UserDevice) Queues(n int) ([]tio.Queue, error) {
-	return true
+	out := make([]tio.Queue, n)
-}
+	for i := range out {
-
+		// All queues share the underlying pipes (the io.Pipe serializes
-func (d *UserDevice) NewMultiQueueReader() error {
+		// concurrent callers) but each owns a private scratch buffer so
-	d.numReaders++
+		// concurrent Reads across queues never alias. NoClose: the pipes are
-	return nil
+		// owned by the UserDevice and torn down once by UserDevice.Close.
-}
+		out[i] = tio.NewSingleQueueNoClose(d, defaultBatchBufSize)
-
-func (d *UserDevice) Readers() []tio.Queue {
-	out := make([]tio.Queue, d.numReaders)
-	for i := range d.numReaders {
-		out[i] = d
 	}
-	return out
+	return out, nil
 }
 
 func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
 	return d.inboundReader, d.outboundWriter
 }
 
+func (d *UserDevice) Read(p []byte) (n int, err error) {
+	return d.outboundReader.Read(p)
+}
+
 func (d *UserDevice) Write(p []byte) (n int, err error) {
 	return d.inboundWriter.Write(p)
 }

overlay/user_test.go

@@ -0,0 +1,163 @@
+package overlay
+
+import (
+	"fmt"
+	"net/netip"
+	"sync"
+	"testing"
+
+	"github.com/slackhq/nebula/overlay/tio"
+)
+
+// newTestUserDevice returns the concrete *UserDevice so tests can reach Pipe()
+// and the internal queue plumbing.
+func newTestUserDevice(t *testing.T) *UserDevice {
+	t.Helper()
+	dev, err := NewUserDevice([]netip.Prefix{netip.MustParsePrefix("10.0.0.1/24")})
+	if err != nil {
+		t.Fatalf("NewUserDevice: %v", err)
+	}
+	ud, ok := dev.(*UserDevice)
+	if !ok {
+		t.Fatalf("NewUserDevice returned %T, want *UserDevice", dev)
+	}
+	return ud
+}
+
+// TestUserDeviceReadersDistinctBuffers ensures each Queue is actually different
+func TestUserDeviceReadersDistinctBuffers(t *testing.T) {
+	d := newTestUserDevice(t)
+
+	readers, err := d.Queues(2)
+	if err != nil {
+		t.Fatalf("Queues: %v", err)
+	}
+	if len(readers) != 2 {
+		t.Fatalf("Queues(2) returned %d queues, want 2", len(readers))
+	}
+
+	// Distinct queue objects.
+	if readers[0] == readers[1] {
+		t.Fatal("Queues(2) returned the same queue object twice")
+	}
+
+	// Drive one packet through each queue and confirm the borrowed bytes from
+	// the first read are NOT clobbered by the second read. With a shared
+	// buffer, reading pkt1 into q1 would corrupt q0's still-borrowed slice.
+	_, ow := d.Pipe()
+
+	pkt0 := []byte("packet-zero-aaaaaaaa")
+	pkt1 := []byte("packet-one-bbbbbbbbb")
+
+	// The pipe is unbuffered, so writes block until a reader consumes them.
+	// Serialize: write pkt0 (read on q0), then write pkt1 (read on q1).
+	go func() {
+		if _, err := ow.Write(pkt0); err != nil {
+			t.Errorf("write pkt0: %v", err)
+		}
+		if _, err := ow.Write(pkt1); err != nil {
+			t.Errorf("write pkt1: %v", err)
+		}
+	}()
+
+	got0, err := readers[0].Read()
+	if err != nil {
+		t.Fatalf("q0.Read: %v", err)
+	}
+	if len(got0) != 1 || string(got0[0].Bytes) != string(pkt0) {
+		t.Fatalf("q0 first read = %q, want %q", firstBytes(got0), pkt0)
+	}
+	// Hold onto q0's borrowed slice across q1's read.
+	borrowed := got0[0].Bytes
+
+	got1, err := readers[1].Read()
+	if err != nil {
+		t.Fatalf("q1.Read: %v", err)
+	}
+	if len(got1) != 1 || string(got1[0].Bytes) != string(pkt1) {
+		t.Fatalf("q1 read = %q, want %q", firstBytes(got1), pkt1)
+	}
+
+	// q0's borrowed bytes must still hold pkt0 - a shared buffer would now
+	// show pkt1's contents.
+	if string(borrowed) != string(pkt0) {
+		t.Fatalf("q0 borrowed bytes were clobbered by q1's read: got %q, want %q", borrowed, pkt0)
+	}
+}
+
+// TestUserDeviceReadersConcurrentRace exercises two queues reading distinct
+// packets concurrently. Run it under `go test -race`: with the old
+// shared-buffer implementation the concurrent Reads raced on readBuf/batchRet
+// and corrupted each other's returned slices.
+func TestUserDeviceReadersConcurrentRace(t *testing.T) {
+	d := newTestUserDevice(t)
+	readers, err := d.Queues(2)
+	if err != nil {
+		t.Fatalf("Queues: %v", err)
+	}
+	_, ow := d.Pipe()
+
+	const iterations = 200
+
+	errs := make(chan error, 3)
+
+	// Each reader parks in Read on the shared outboundReader; io.Pipe hands
+	// each write to whichever reader is currently waiting. We only care that
+	// concurrent Reads into distinct buffers are race-free, so any parked
+	// reader may serve any write.
+	var wg sync.WaitGroup
+	run := func(idx int) {
+		defer wg.Done()
+		for i := 0; i < iterations; i++ {
+			pkts, err := readers[idx].Read()
+			if err != nil {
+				errs <- err
+				return
+			}
+			if len(pkts) != 1 {
+				errs <- fmt.Errorf("reader %d: got %d packets, want 1", idx, len(pkts))
+				return
+			}
+			// Touch every byte of the borrowed slice while the other reader
+			// may be mid-Read; a shared buffer would race here.
+			total := 0
+			for _, c := range pkts[0].Bytes {
+				total += int(c)
+			}
+			_ = total
+		}
+	}
+
+	wg.Add(2)
+	go run(0)
+	go run(1)
+
+	// Feed 2*iterations packets. io.Pipe copies each write straight into the
+	// waiting reader's private buffer, so reusing buf between writes is safe.
+	go func() {
+		buf := make([]byte, 32)
+		for i := 0; i < 2*iterations; i++ {
+			for j := range buf {
+				buf[j] = byte(i + j)
+			}
+			if _, err := ow.Write(buf); err != nil {
+				errs <- err
+				return
+			}
+		}
+	}()
+
+	wg.Wait()
+	select {
+	case err := <-errs:
+		t.Fatalf("concurrent reader failed: %v", err)
+	default:
+	}
+}
+
+func firstBytes(p []tio.Packet) []byte {
+	if len(p) == 0 {
+		return nil
+	}
+	return p[0].Bytes
+}

punchy.go

@@ -174,9 +174,9 @@ func (p *Punchy) SendPunch(hostinfo *HostInfo) {
 
 	if p.punchEverything.Load() {
 		p.sendPunchToAllRemotes(hostinfo)
-	} else if hostinfo.remote.IsValid() {
+	} else if hr := hostinfo.GetRemote(); hr.IsValid() {
 		p.metricPunchyTx.Inc(1)
-		p.punchConn.WriteTo([]byte{1}, hostinfo.remote)
+		p.punchConn.WriteTo([]byte{1}, hr)
 	}
 }
 

relay_manager.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"log/slog"
 	"net/netip"
+	"slices"
 	"sync/atomic"
 
 	"github.com/slackhq/nebula/cert"
@@ -57,14 +58,25 @@ func (rm *relayManager) GetUseRelays() bool {
 // For each candidate relay it either kicks off a handshake to the relay, sends a CreateRelayRequest, retransmits
 // one that may have been lost, or, once the relay is Established, forwards the in-progress
 // stage 0 handshake packet for vpnIp through it.
-func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *HostInfo, stage0 []byte) {
+func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hh *HandshakeHostInfo, stage0 []byte) {
+	hostinfo := hh.hostinfo
 	if !rm.GetUseRelays() || len(hostinfo.remotes.relays) == 0 {
+		hh.lastRelays = nil
 		return
 	}
 
-	hostinfo.logger(rm.l).Info("Attempt to relay through hosts", "relays", hostinfo.remotes.relays)
+	relays := hostinfo.remotes.relays
+	listLevel := slog.LevelDebug
+	prior := hh.lastRelays
+	if !slices.Equal(relays, prior) {
+		listLevel = slog.LevelInfo
+		hh.lastRelays = slices.Clone(relays)
+	}
+	hl := hostinfo.logger(rm.l)
+	hl.Log(context.Background(), listLevel, "Attempt to relay through hosts", "relays", relays)
+
 	// Send a RelayRequest to all known Relay IP's
-	for _, relay := range hostinfo.remotes.relays {
+	for _, relay := range relays {
 		// Don't relay through the host I'm trying to connect to
 		if relay == vpnIp {
 			continue
@@ -75,20 +87,30 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 			continue
 		}
 
+		// Each relay's per-attempt log fires at Info on the first time we hit it and Debug after that.
+		level := slog.LevelInfo
+		if slices.Contains(prior, relay) {
+			level = slog.LevelDebug
+		}
+
 		relayHostInfo := rm.hostmap.QueryVpnAddr(relay)
-		if relayHostInfo == nil || !relayHostInfo.remote.IsValid() {
+		if relayHostInfo == nil || !relayHostInfo.GetRemote().IsValid() {
-			hostinfo.logger(rm.l).Info("Establish tunnel to relay target", "relay", relay.String())
+			hl.Log(context.Background(), level, "Establish tunnel to relay target", "relay", relay.String())
 			f.Handshake(relay)
 			continue
 		}
+
 		// Check the relay HostInfo to see if we already established a relay through
 		existingRelay, ok := relayHostInfo.relayState.QueryRelayForByIp(vpnIp)
 		if !ok {
 			// No relays exist or requested yet.
-			if relayHostInfo.remote.IsValid() {
+			if relayHostInfo.GetRemote().IsValid() {
 				idx, err := AddRelay(rm.l, relayHostInfo, rm.hostmap, vpnIp, nil, TerminalType, Requested)
 				if err != nil {
-					hostinfo.logger(rm.l).Info("Failed to add relay to hostmap", "relay", relay.String(), "error", err)
+					// No local relay state was installed, so a CreateRelayRequest would hand the
+					// peer an index we could never resolve. Skip it.
+					hl.Info("Failed to add relay to hostmap", "relay", relay.String(), "error", err)
+					continue
 				}
 
 				m := NebulaControl{
@@ -99,12 +121,12 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 				switch relayHostInfo.GetCert().Certificate.Version() {
 				case cert.Version1:
 					if !f.myVpnAddrs[0].Is4() {
-						hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
+						hl.Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
 						continue
 					}
 
 					if !vpnIp.Is4() {
-						hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
+						hl.Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
 						continue
 					}
 
@@ -116,16 +138,16 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 					m.RelayFromAddr = netAddrToProtoAddr(f.myVpnAddrs[0])
 					m.RelayToAddr = netAddrToProtoAddr(vpnIp)
 				default:
-					hostinfo.logger(rm.l).Error("Unknown certificate version found while creating relay")
+					hl.Error("Unknown certificate version found while creating relay")
 					continue
 				}
 
 				msg, err := m.Marshal()
 				if err != nil {
-					hostinfo.logger(rm.l).Error("Failed to marshal Control message to create relay", "error", err)
+					hl.Error("Failed to marshal Control message to create relay", "error", err)
 				} else {
 					f.SendMessageToHostInfo(header.Control, 0, relayHostInfo, msg, make([]byte, 12), make([]byte, mtu))
-					rm.l.Info("send CreateRelayRequest",
+					rm.l.Log(context.Background(), level, "send CreateRelayRequest",
 						"relayFrom", f.myVpnAddrs[0],
 						"relayTo", vpnIp,
 						"initiatorRelayIndex", idx,
@@ -138,14 +160,14 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 
 		switch existingRelay.State {
 		case Established:
-			hostinfo.logger(rm.l).Info("Send handshake via relay", "relay", relay.String())
+			hl.Log(context.Background(), level, "Send handshake via relay", "relay", relay.String())
 			f.SendVia(relayHostInfo, existingRelay, stage0, make([]byte, 12), make([]byte, mtu), false)
 		case Disestablished:
 			// Mark this relay as 'requested'
 			relayHostInfo.relayState.UpdateRelayForByIpState(vpnIp, Requested)
 			fallthrough
 		case Requested:
-			hostinfo.logger(rm.l).Info("Re-send CreateRelay request", "relay", relay.String())
+			hl.Log(context.Background(), level, "Re-send CreateRelay request", "relay", relay.String())
 			// Re-send the CreateRelay request, in case the previous one was lost.
 			m := NebulaControl{
 				Type:                NebulaControl_CreateRelayRequest,
@@ -155,12 +177,12 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 			switch relayHostInfo.GetCert().Certificate.Version() {
 			case cert.Version1:
 				if !f.myVpnAddrs[0].Is4() {
-					hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
+					hl.Error("can not establish v1 relay with a v6 network because the relay is not running a current nebula version")
 					continue
 				}
 
 				if !vpnIp.Is4() {
-					hostinfo.logger(rm.l).Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
+					hl.Error("can not establish v1 relay with a v6 remote network because the relay is not running a current nebula version")
 					continue
 				}
 
@@ -172,16 +194,16 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 				m.RelayFromAddr = netAddrToProtoAddr(f.myVpnAddrs[0])
 				m.RelayToAddr = netAddrToProtoAddr(vpnIp)
 			default:
-				hostinfo.logger(rm.l).Error("Unknown certificate version found while creating relay")
+				hl.Error("Unknown certificate version found while creating relay")
 				continue
 			}
 			msg, err := m.Marshal()
 			if err != nil {
-				hostinfo.logger(rm.l).Error("Failed to marshal Control message to create relay", "error", err)
+				hl.Error("Failed to marshal Control message to create relay", "error", err)
 			} else {
 				// This must send over the hostinfo, not over hm.Hosts[ip]
 				f.SendMessageToHostInfo(header.Control, 0, relayHostInfo, msg, make([]byte, 12), make([]byte, mtu))
-				rm.l.Info("send CreateRelayRequest",
+				rm.l.Log(context.Background(), level, "send CreateRelayRequest",
 					"relayFrom", f.myVpnAddrs[0],
 					"relayTo", vpnIp,
 					"initiatorRelayIndex", existingRelay.LocalIndex,
@@ -192,7 +214,7 @@ func (rm *relayManager) StartRelays(f *Interface, vpnIp netip.Addr, hostinfo *Ho
 			// PeerRequested only occurs in Forwarding relays, not Terminal relays, and this is a Terminal relay case.
 			fallthrough
 		default:
-			hostinfo.logger(rm.l).Error("Relay unexpected state",
+			hl.Error("Relay unexpected state",
 				"vpnIp", vpnIp,
 				"state", existingRelay.State,
 				"relay", relay,
@@ -218,7 +240,12 @@ func AddRelay(l *slog.Logger, relayHostInfo *HostInfo, hm *HostMap, vpnIp netip.
 			// Avoid standing up a relay that can't be used since only the primary hostinfo
 			// will be pointed to by the relay logic
 			//TODO: if there was an existing primary and it had relay state, should we merge?
-			hm.unlockedMakePrimary(relayHostInfo)
+			if !hm.unlockedMakePrimary(relayHostInfo) {
+				// The tunnel was torn down after the caller grabbed relayHostInfo. A relay standing
+				// on an unlinked hostinfo would never carry traffic, and its Relays entry could
+				// never be reclaimed since the delete-time cleanup has already run.
+				return 0, errors.New("relay hostinfo is no longer in the hostmap")
+			}
 
 			hm.Relays[index] = relayHostInfo
 			newRelay := Relay{
@@ -290,6 +317,22 @@ func (rm *relayManager) HandleControlMsg(h *HostInfo, d []byte, f *Interface) {
 		v = cert.Version2
 	}
 
+	// validate:
+	switch msg.Type {
+	case NebulaControl_CreateRelayRequest, NebulaControl_CreateRelayResponse:
+		if msg.RelayFromAddr == nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				h.logger(f.l).Debug("Control message received with nil RelayFromAddr", "type", msg.Type)
+			}
+			return
+		} else if msg.RelayToAddr == nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				h.logger(f.l).Debug("Control message received with nil RelayToAddr", "type", msg.Type)
+			}
+			return
+		}
+	}
+
 	switch msg.Type {
 	case NebulaControl_CreateRelayRequest:
 		rm.handleCreateRelayRequest(v, h, f, msg)
@@ -299,17 +342,17 @@ func (rm *relayManager) HandleControlMsg(h *HostInfo, d []byte, f *Interface) {
 }
 
 func (rm *relayManager) handleCreateRelayResponse(v cert.Version, h *HostInfo, f *Interface, m *NebulaControl) {
+	//nil-checks for protoAddrToNetAddr handled by caller
+	relayFrom := protoAddrToNetAddr(m.RelayFromAddr)
+	relayTo := protoAddrToNetAddr(m.RelayToAddr)
 	rm.l.Info("handleCreateRelayResponse",
-		"relayFrom", protoAddrToNetAddr(m.RelayFromAddr),
+		"relayFrom", relayFrom,
-		"relayTo", protoAddrToNetAddr(m.RelayToAddr),
+		"relayTo", relayTo,
 		"initiatorRelayIndex", m.InitiatorRelayIndex,
 		"responderRelayIndex", m.ResponderRelayIndex,
 		"vpnAddrs", h.vpnAddrs,
 	)
 
-	target := m.RelayToAddr
-	targetAddr := protoAddrToNetAddr(target)
-
 	relay, err := rm.EstablishRelay(h, m)
 	if err != nil {
 		rm.l.Error("Failed to update relay for relayTo", "error", err)
@@ -325,7 +368,7 @@ func (rm *relayManager) handleCreateRelayResponse(v cert.Version, h *HostInfo, f
 		rm.l.Error("Can't find a HostInfo for peer", "relayTo", relay.PeerAddr)
 		return
 	}
-	peerRelay, ok := peerHostInfo.relayState.QueryRelayForByIp(targetAddr)
+	peerRelay, ok := peerHostInfo.relayState.QueryRelayForByIp(relayTo)
 	if !ok {
 		rm.l.Error("peerRelay does not have Relay state for relayTo", "relayTo", peerHostInfo.vpnAddrs[0])
 		return
@@ -335,19 +378,19 @@ func (rm *relayManager) handleCreateRelayResponse(v cert.Version, h *HostInfo, f
 		// I initiated the request to this peer, but haven't heard back from the peer yet. I must wait for this peer
 		// to respond to complete the connection.
 	case PeerRequested, Disestablished, Established:
-		peerHostInfo.relayState.UpdateRelayForByIpState(targetAddr, Established)
+		peerHostInfo.relayState.UpdateRelayForByIpState(relayTo, Established)
 		resp := NebulaControl{
 			Type:                NebulaControl_CreateRelayResponse,
 			ResponderRelayIndex: peerRelay.LocalIndex,
 			InitiatorRelayIndex: peerRelay.RemoteIndex,
 		}
 
+		peer := peerHostInfo.vpnAddrs[0]
 		if v == cert.Version1 {
-			peer := peerHostInfo.vpnAddrs[0]
 			if !peer.Is4() {
 				rm.l.Error("Refusing to CreateRelayResponse for a v1 relay with an ipv6 address",
 					"relayFrom", peer,
-					"relayTo", target,
+					"relayTo", relayTo,
 					"initiatorRelayIndex", resp.InitiatorRelayIndex,
 					"responderRelayIndex", resp.ResponderRelayIndex,
 					"vpnAddrs", peerHostInfo.vpnAddrs,
@@ -357,30 +400,31 @@ func (rm *relayManager) handleCreateRelayResponse(v cert.Version, h *HostInfo, f
 
 			b := peer.As4()
 			resp.OldRelayFromAddr = binary.BigEndian.Uint32(b[:])
-			b = targetAddr.As4()
+			b = relayTo.As4()
 			resp.OldRelayToAddr = binary.BigEndian.Uint32(b[:])
 		} else {
-			resp.RelayFromAddr = netAddrToProtoAddr(peerHostInfo.vpnAddrs[0])
+			resp.RelayFromAddr = netAddrToProtoAddr(peer)
-			resp.RelayToAddr = target
+			resp.RelayToAddr = m.RelayToAddr
 		}
 
 		msg, err := resp.Marshal()
 		if err != nil {
 			rm.l.Error("relayManager Failed to marshal Control CreateRelayResponse message to create relay", "error", err)
-		} else {
+			return
-			f.SendMessageToHostInfo(header.Control, 0, peerHostInfo, msg, make([]byte, 12), make([]byte, mtu))
-			rm.l.Info("send CreateRelayResponse",
-				"relayFrom", resp.RelayFromAddr,
-				"relayTo", resp.RelayToAddr,
-				"initiatorRelayIndex", resp.InitiatorRelayIndex,
-				"responderRelayIndex", resp.ResponderRelayIndex,
-				"vpnAddrs", peerHostInfo.vpnAddrs,
-			)
 		}
+		f.SendMessageToHostInfo(header.Control, 0, peerHostInfo, msg, make([]byte, 12), make([]byte, mtu))
+		rm.l.Info("send CreateRelayResponse",
+			"relayFrom", peer,
+			"relayTo", relayTo,
+			"initiatorRelayIndex", resp.InitiatorRelayIndex,
+			"responderRelayIndex", resp.ResponderRelayIndex,
+			"vpnAddrs", peerHostInfo.vpnAddrs,
+		)
 	}
 }
 
 func (rm *relayManager) handleCreateRelayRequest(v cert.Version, h *HostInfo, f *Interface, m *NebulaControl) {
+	//nil-checks for protoAddrToNetAddr handled by caller
 	from := protoAddrToNetAddr(m.RelayFromAddr)
 	target := protoAddrToNetAddr(m.RelayToAddr)
 
@@ -490,7 +534,7 @@ func (rm *relayManager) handleCreateRelayRequest(v cert.Version, h *HostInfo, f
 			f.Handshake(target)
 			return
 		}
-		if !peer.remote.IsValid() {
+		if !peer.GetRemote().IsValid() {
 			// Only create relays to peers for whom I have a direct connection
 			return
 		}

relay_manager_test.go

@@ -0,0 +1,97 @@
+package nebula
+
+import (
+	"bytes"
+	"log/slog"
+	"net/netip"
+	"testing"
+
+	"github.com/gaissmai/bart"
+	"github.com/slackhq/nebula/test"
+	"github.com/stretchr/testify/assert"
+)
+
+// TestStartRelaysLogDedupe verifies that repeated attempts with the same relay set drop the log
+// chatter to Debug, mirroring how the normal handshake retry loop quiets down once it's already
+// announced its targets.
+func TestStartRelaysLogDedupe(t *testing.T) {
+	vpnIp := netip.MustParseAddr("100.64.99.4")
+	otherRelay := netip.MustParseAddr("100.64.99.5")
+
+	newHH := func() *HandshakeHostInfo {
+		// Use the target's own vpnIp as the "relay" so the loop body skips it without
+		// touching any sender-side state. That isolates the test to the level-selection
+		// behavior of the top-level "Attempt to relay through hosts" log.
+		hostinfo := &HostInfo{
+			vpnAddrs:     []netip.Addr{vpnIp},
+			localIndexId: 1,
+			remotes:      NewRemoteList([]netip.Addr{vpnIp}, nil),
+		}
+		hostinfo.remotes.relays = []netip.Addr{vpnIp}
+		return &HandshakeHostInfo{hostinfo: hostinfo}
+	}
+
+	// Park any extra relay addresses we'll introduce mid-test in myVpnAddrsTable so the loop
+	// body always skips before touching f.Handshake (which would need a real handshakeManager).
+	addrTable := new(bart.Lite)
+	addrTable.Insert(netip.PrefixFrom(otherRelay, otherRelay.BitLen()))
+	f := &Interface{myVpnAddrsTable: addrTable}
+
+	newRM := func(buf *bytes.Buffer) *relayManager {
+		l := test.NewLoggerWithOutputAndLevel(buf, slog.LevelDebug)
+		rm := &relayManager{l: l, hostmap: newHostMap(l)}
+		rm.useRelays.Store(true)
+		return rm
+	}
+
+	const msg = `msg="Attempt to relay through hosts"`
+
+	t.Run("first attempt logs at Info", func(t *testing.T) {
+		var buf bytes.Buffer
+		rm := newRM(&buf)
+		hh := newHH()
+		rm.StartRelays(f, vpnIp, hh, nil)
+		assert.Equal(t, []netip.Addr{vpnIp}, hh.lastRelays, "lastRelays should record the relay set we just attempted")
+		assert.Contains(t, buf.String(), "level=INFO "+msg, "expected Info level on first attempt")
+	})
+
+	t.Run("repeat attempt with same relays drops to Debug", func(t *testing.T) {
+		var buf bytes.Buffer
+		rm := newRM(&buf)
+		hh := newHH()
+		rm.StartRelays(f, vpnIp, hh, nil)
+		first := append([]netip.Addr(nil), hh.lastRelays...)
+		buf.Reset()
+		rm.StartRelays(f, vpnIp, hh, nil)
+		assert.Equal(t, first, hh.lastRelays)
+		assert.Contains(t, buf.String(), "level=DEBUG "+msg, "expected Debug level on identical retry")
+		assert.NotContains(t, buf.String(), "level=INFO "+msg, "Info should not fire on identical retry")
+	})
+
+	t.Run("changed relay list bumps back to Info", func(t *testing.T) {
+		var buf bytes.Buffer
+		rm := newRM(&buf)
+		hh := newHH()
+		rm.StartRelays(f, vpnIp, hh, nil)
+		buf.Reset()
+
+		// The lighthouse handed us a new set this round.
+		hh.hostinfo.remotes.relays = []netip.Addr{vpnIp, otherRelay}
+
+		rm.StartRelays(f, vpnIp, hh, nil)
+		assert.Equal(t, []netip.Addr{vpnIp, otherRelay}, hh.lastRelays)
+		assert.Contains(t, buf.String(), "level=INFO "+msg, "expected Info when the relay list changes")
+	})
+
+	t.Run("disabled relays clears lastRelays and emits no Attempt log", func(t *testing.T) {
+		var buf bytes.Buffer
+		rm := newRM(&buf)
+		rm.useRelays.Store(false)
+		hh := newHH()
+		hh.lastRelays = []netip.Addr{vpnIp}
+
+		rm.StartRelays(f, vpnIp, hh, nil)
+		assert.Nil(t, hh.lastRelays, "with relays disabled lastRelays should be cleared")
+		assert.NotContains(t, buf.String(), msg, "should not log when we shortcut out")
+	})
+}

remote_list.go

@@ -239,6 +239,31 @@ func (r *RemoteList) unlockedSetHostnamesResults(hr *hostnamesResults) {
 	r.hr = hr
 }
 
+// ResetForOwner zeros the reported address slices for the given owner and marks the addrs list dirty.
+// Any pending hostname resolution will be canceled.
+func (r *RemoteList) ResetForOwner(ownerVpnAddr netip.Addr) {
+	r.Lock()
+	defer r.Unlock()
+	r.hr.Cancel()
+	if c, ok := r.cache[ownerVpnAddr]; ok {
+		if c.v4 != nil {
+			c.v4.reported = c.v4.reported[:0]
+		}
+		if c.v6 != nil {
+			c.v6.reported = c.v6.reported[:0]
+		}
+	}
+	r.shouldRebuild = true
+}
+
+// ClearHostnameResults cancels the in-flight DNS resolver goroutine (if any) and drops the resolved IP cache.
+func (r *RemoteList) ClearHostnameResults() {
+	r.Lock()
+	defer r.Unlock()
+	r.unlockedSetHostnamesResults(nil)
+	r.shouldRebuild = true
+}
+
 // Len locks and reports the size of the deduplicated address list
 // The deduplication work may need to occur here, so you must pass preferredRanges
 func (r *RemoteList) Len(preferredRanges []netip.Prefix) int {
@@ -319,6 +344,9 @@ func (r *RemoteList) CopyCache() *CacheMap {
 			}
 
 			for _, a := range mc.v4.reported {
+				if a == nil {
+					continue
+				}
 				c.Reported = append(c.Reported, protoV4AddrPortToNetAddrPort(a))
 			}
 		}
@@ -329,6 +357,9 @@ func (r *RemoteList) CopyCache() *CacheMap {
 			}
 
 			for _, a := range mc.v6.reported {
+				if a == nil {
+					continue
+				}
 				c.Reported = append(c.Reported, protoV6AddrPortToNetAddrPort(a))
 			}
 		}
@@ -557,6 +588,9 @@ func (r *RemoteList) unlockedCollect() {
 			}
 
 			for _, v := range c.v4.reported {
+				if v == nil {
+					continue
+				}
 				u := protoV4AddrPortToNetAddrPort(v)
 				if !r.unlockedIsBad(u) {
 					addrs = append(addrs, u)
@@ -573,6 +607,9 @@ func (r *RemoteList) unlockedCollect() {
 			}
 
 			for _, v := range c.v6.reported {
+				if v == nil {
+					continue
+				}
 				u := protoV6AddrPortToNetAddrPort(v)
 				if !r.unlockedIsBad(u) {
 					addrs = append(addrs, u)

remote_list_test.go

@@ -6,8 +6,22 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
+// trackedHostnameResults builds a *hostnamesResults with a known cancel function and a
+// pre-populated ips map so tests can assert cancellation and verify previously-resolved
+// IPs survive a cancel without spinning up a real DNS resolver.
+func trackedHostnameResults(cancelFn func(), addrs ...string) *hostnamesResults {
+	hr := &hostnamesResults{cancelFn: cancelFn}
+	ips := map[netip.AddrPort]struct{}{}
+	for _, a := range addrs {
+		ips[netip.MustParseAddrPort(a)] = struct{}{}
+	}
+	hr.ips.Store(&ips)
+	return hr
+}
+
 func TestRemoteList_Rebuild(t *testing.T) {
 	rl := NewRemoteList([]netip.Addr{netip.MustParseAddr("0.0.0.0")}, nil)
 	rl.unlockedSetV4(
@@ -112,6 +126,81 @@ func TestRemoteList_Rebuild(t *testing.T) {
 	assert.Equal(t, "172.31.0.1:10101", rl.addrs[9].String())
 }
 
+func TestRemoteList_ResetForOwner(t *testing.T) {
+	ourselves := netip.MustParseAddr("10.0.0.1")
+	otherOwner := netip.MustParseAddr("10.0.0.2")
+	vpnAddr := netip.MustParseAddr("10.0.0.99")
+
+	rl := NewRemoteList([]netip.Addr{vpnAddr}, nil)
+	rl.unlockedSetV4(ourselves, vpnAddr,
+		[]*V4AddrPort{newIp4AndPortFromString("1.1.1.1:4242")},
+		func(netip.Addr, *V4AddrPort) bool { return true },
+	)
+	rl.unlockedSetV6(ourselves, vpnAddr,
+		[]*V6AddrPort{newIp6AndPortFromString("[1::1]:4242")},
+		func(netip.Addr, *V6AddrPort) bool { return true },
+	)
+	rl.unlockedSetV4(otherOwner, vpnAddr,
+		[]*V4AddrPort{newIp4AndPortFromString("2.2.2.2:4242")},
+		func(netip.Addr, *V4AddrPort) bool { return true },
+	)
+
+	canceled := 0
+	hr := trackedHostnameResults(func() { canceled++ }, "3.3.3.3:4242")
+	rl.Lock()
+	rl.unlockedSetHostnamesResults(hr)
+	rl.Unlock()
+
+	rl.ResetForOwner(ourselves)
+
+	rl.RLock()
+	defer rl.RUnlock()
+	assert.Empty(t, rl.cache[ourselves].v4.reported, "our v4 reported should be cleared")
+	assert.Empty(t, rl.cache[ourselves].v6.reported, "our v6 reported should be cleared")
+	assert.Len(t, rl.cache[otherOwner].v4.reported, 1, "other owner's contribution must be preserved")
+	assert.Equal(t, "2.2.2.2:4242", protoV4AddrPortToNetAddrPort(rl.cache[otherOwner].v4.reported[0]).String())
+	assert.Equal(t, 1, canceled, "DNS resolution goroutine should be canceled")
+	assert.Same(t, hr, rl.hr, "hostnamesResults must be preserved so DNS-resolved IPs keep feeding addrs until replaced")
+	assert.NotEmpty(t, rl.hr.GetAddrs(), "previously-resolved IPs should still be readable after cancel")
+	assert.True(t, rl.shouldRebuild, "shouldRebuild must be set so the next Rebuild recomputes addrs")
+}
+
+func TestRemoteList_ResetForOwner_NoEntry(t *testing.T) {
+	// An owner with no cache entry must not panic; shouldRebuild is still set and any
+	// existing hostnamesResults is canceled.
+	rl := NewRemoteList([]netip.Addr{netip.MustParseAddr("10.0.0.99")}, nil)
+	canceled := 0
+	rl.Lock()
+	rl.unlockedSetHostnamesResults(trackedHostnameResults(func() { canceled++ }, "3.3.3.3:4242"))
+	rl.Unlock()
+
+	rl.ResetForOwner(netip.MustParseAddr("10.0.0.1"))
+
+	rl.RLock()
+	defer rl.RUnlock()
+	assert.Equal(t, 1, canceled)
+	assert.True(t, rl.shouldRebuild)
+}
+
+func TestRemoteList_ClearHostnameResults(t *testing.T) {
+	rl := NewRemoteList([]netip.Addr{netip.MustParseAddr("10.0.0.99")}, nil)
+
+	canceled := 0
+	hr := trackedHostnameResults(func() { canceled++ }, "3.3.3.3:4242")
+	rl.Lock()
+	rl.unlockedSetHostnamesResults(hr)
+	rl.Unlock()
+	require.NotEmpty(t, hr.GetAddrs(), "hostnamesResults should have its fastrack IPs populated")
+
+	rl.ClearHostnameResults()
+
+	rl.RLock()
+	defer rl.RUnlock()
+	assert.Equal(t, 1, canceled, "DNS resolution goroutine should be canceled")
+	assert.Nil(t, rl.hr, "hostnamesResults should be dropped")
+	assert.True(t, rl.shouldRebuild)
+}
+
 func BenchmarkFullRebuild(b *testing.B) {
 	rl := NewRemoteList([]netip.Addr{netip.MustParseAddr("0.0.0.0")}, nil)
 	rl.unlockedSetV4(

service/service.go

@@ -43,12 +43,25 @@ type Service struct {
 	}
 }
 
-func New(control *nebula.Control) (*Service, error) {
+func New(control *nebula.Control) (_ *Service, reterr error) {
-	wait, err := control.Start()
+	// Check this before Start so a failure doesn't leave a running nebula
+	device, ok := control.Device().(*overlay.UserDevice)
+	if !ok {
+		return nil, errors.New("must be using user device")
+	}
+
+	err := control.Start()
 	if err != nil {
 		return nil, err
 	}
 
+	// Anything that fails after a successful Start must tear nebula back down
+	defer func() {
+		if reterr != nil {
+			control.Stop()
+		}
+	}()
+
 	ctx := control.Context()
 	eg, ctx := errgroup.WithContext(ctx)
 	s := Service{
@@ -57,11 +70,6 @@ func New(control *nebula.Control) (*Service, error) {
 	}
 	s.mu.listeners = map[uint16]*tcpListener{}
 
-	device, ok := control.Device().(*overlay.UserDevice)
-	if !ok {
-		return nil, errors.New("must be using user device")
-	}
-
 	s.ipstack = stack.New(stack.Options{
 		NetworkProtocols:   []stack.NetworkProtocolFactory{ipv4.NewProtocol, ipv6.NewProtocol},
 		TransportProtocols: []stack.TransportProtocolFactory{tcp.NewProtocol, udp.NewProtocol, icmp.NewProtocol4, icmp.NewProtocol6},
@@ -147,7 +155,7 @@ func New(control *nebula.Control) (*Service, error) {
 	// Add the nebula wait function to the group so a fatal reader error
 	// propagates out through errgroup.Wait().
 	eg.Go(func() error {
-		return wait()
+		return control.Wait()
 	})
 
 	return &s, nil