use P256 as the default curve for fips140 mode

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

.github/workflows/release.yml

@@ -12,9 +12,9 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
-          go-version: '1.25'
+          go-version: '1.26'
           check-latest: true
 
       - name: Build
@@ -38,9 +38,9 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
-          go-version: '1.25'
+          go-version: '1.26'
           check-latest: true
 
       - name: Build
@@ -78,9 +78,9 @@ jobs:
     steps:
       - uses: actions/checkout@v7
 
-      - uses: actions/setup-go@v6
+      - uses: actions/setup-go@v7
         with:
-          go-version: '1.25'
+          go-version: '1.26'
           check-latest: true
 
       - name: Import certificates

.github/workflows/smoke-extra.yml

@@ -32,9 +32,9 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: add hashicorp source
@@ -64,9 +64,9 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: add hashicorp source
@@ -90,9 +90,9 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     # WSL2 + Ubuntu so the smoke can run a real linux peer with its own

.github/workflows/smoke.yml

@@ -20,44 +20,24 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
-    - name: build
-      run: make bin-docker CGO_ENABLED=1 BUILD_ARGS=-race
+    - name: Smoke Docker
+      run: make smoke-docker
 
-    - name: setup docker image
-      working-directory: ./.github/workflows/smoke
-      run: ./build.sh
+    - name: Smoke Docker IPv6 overlay
+      run: make smoke-docker-ipv6
 
-    - name: run smoke
-      working-directory: ./.github/workflows/smoke
-      run: ./smoke.sh
+    - name: Smoke Relay Docker
+      run: make smoke-relay-docker
 
-    - name: setup docker image ipv6
-      working-directory: ./.github/workflows/smoke
-      run: SMOKE_OVERLAY_IPV6=1 ./build.sh
+    - name: Smoke Docker boringcrypto
+      run: make boringcrypto smoke-docker
 
-    - 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
-
-    - name: run smoke relay
-      working-directory: ./.github/workflows/smoke
-      run: ./smoke-relay.sh
-
-    - name: setup docker image for P256
-      working-directory: ./.github/workflows/smoke
-      run: NAME="smoke-p256" CURVE=P256 ./build.sh
-
-    - name: run smoke-p256
-      working-directory: ./.github/workflows/smoke
-      run: NAME="smoke-p256" ./smoke.sh
+    - name: Smoke Docker fips140
+      run: make fips140-all GOALS=smoke-docker
 
     timeout-minutes: 10

.github/workflows/test.yml

@@ -20,9 +20,9 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: Install goimports
@@ -42,7 +42,7 @@ jobs:
     - name: golangci-lint
       uses: golangci/golangci-lint-action@v9
       with:
-        version: v2.5
+        version: v2.12
 
   test:
     name: Test ${{ matrix.name }}
@@ -58,9 +58,14 @@ jobs:
           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"
+          build-cmd: make boringcrypto
+          test-cmd: make boringcrypto test
+          e2e-cmd: make boringcrypto e2evv
+        - name: linux-fips140
+          os: ubuntu-latest
+          build-cmd: make fips140-all
+          test-cmd: make fips140-all GOALS=test
+          e2e-cmd: make fips140-all GOALS=e2evv
         - name: linux-pkcs11
           os: ubuntu-latest
           build-cmd: make bin-pkcs11
@@ -80,9 +85,9 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: Build
@@ -125,9 +130,9 @@ jobs:
 
     - uses: actions/checkout@v7
 
-    - uses: actions/setup-go@v6
+    - uses: actions/setup-go@v7
       with:
-        go-version: '1.25'
+        go-version: '1.26'
         check-latest: true
 
     - name: Build ${{ matrix.name }}

Makefile

@@ -72,6 +72,17 @@ 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
 
+# Based on section 2.2 of the Go Cryptographic Module CVMP Security Policy #5247
+ALL_FIPS140 = linux-amd64-fips140 \
+	linux-arm64-fips140 \
+	windows-amd64-fips140 \
+	windows-arm64-fips140 \
+	darwin-arm64-fips140 \
+	freebsd-amd64-fips140 \
+	linux-arm-7-fips140 \
+	linux-mips64-fips140 \
+	linux-ppc64le-fips140
+
 e2e:
 	$(TEST_ENV) go test -tags=e2e_testing -count=1 $(TEST_FLAGS) ./e2e
 
@@ -137,6 +148,8 @@ release-netbsd: $(ALL_NETBSD:%=build/nebula-%.tar.gz)
 
 release-boringcrypto: build/nebula-linux-$(shell go env GOARCH)-boringcrypto.tar.gz
 
+release-fips140: $(ALL_FIPS140:%=build/nebula-%.tar.gz)
+
 BUILD_ARGS += -trimpath
 
 bin-windows: build/windows-amd64/nebula.exe build/windows-amd64/nebula-cert.exe
@@ -157,21 +170,20 @@ bin-freebsd-arm64: build/freebsd-arm64/nebula build/freebsd-arm64/nebula-cert
 bin-boringcrypto: build/linux-$(shell go env GOARCH)-boringcrypto/nebula build/linux-$(shell go env GOARCH)-boringcrypto/nebula-cert
 	mv $? .
 
+bin-fips140: build/linux-$(shell go env GOARCH)-fips140/nebula build/linux-$(shell go env GOARCH)-fips140/nebula-cert
+	mv $? .
+
 bin-pkcs11: BUILD_ARGS += -tags pkcs11
 bin-pkcs11: CGO_ENABLED = 1
 bin-pkcs11: bin
 
-# Build with the pprof debug server (serves on :6060). See startPprofServer.
-debug: BUILD_ARGS += -tags debug
-debug: bin
-
 bin:
-	go build $(BUILD_ARGS) -ldflags "$(LDFLAGS)" -o ./nebula${NEBULA_CMD_SUFFIX} ${NEBULA_CMD_PATH}
-	go build $(BUILD_ARGS) -ldflags "$(LDFLAGS)" -o ./nebula-cert${NEBULA_CMD_SUFFIX} ./cmd/nebula-cert
+	$(GOENV) go build $(BUILD_ARGS) -ldflags "$(LDFLAGS)" -o ./nebula${NEBULA_CMD_SUFFIX} ${NEBULA_CMD_PATH}
+	$(GOENV) go build $(BUILD_ARGS) -ldflags "$(LDFLAGS)" -o ./nebula-cert${NEBULA_CMD_SUFFIX} ./cmd/nebula-cert
 
 install:
-	go install $(BUILD_ARGS) -ldflags "$(LDFLAGS)" ${NEBULA_CMD_PATH}
-	go install $(BUILD_ARGS) -ldflags "$(LDFLAGS)" ./cmd/nebula-cert
+	$(GOENV) go install $(BUILD_ARGS) -ldflags "$(LDFLAGS)" ${NEBULA_CMD_PATH}
+	$(GOENV) go install $(BUILD_ARGS) -ldflags "$(LDFLAGS)" ./cmd/nebula-cert
 
 build/linux-arm-%: GOENV += GOARM=$(word 3, $(subst -, ,$*))
 build/linux-mips-%: GOENV += GOMIPS=$(word 3, $(subst -, ,$*))
@@ -182,8 +194,11 @@ build/linux-mips-softfloat/%: LDFLAGS += -s -w
 # boringcrypto
 build/linux-amd64-boringcrypto/%: GOENV += GOEXPERIMENT=boringcrypto CGO_ENABLED=1
 build/linux-arm64-boringcrypto/%: GOENV += GOEXPERIMENT=boringcrypto CGO_ENABLED=1
-build/linux-amd64-boringcrypto/%: LDFLAGS += -checklinkname=0
-build/linux-arm64-boringcrypto/%: LDFLAGS += -checklinkname=0
+
+# fips140
+FIPSVERSION = v1.0.0
+$(foreach _rule, $(ALL_FIPS140), build/$(_rule)/%): GOENV += GOFIPS140=$(FIPSVERSION)
+$(foreach _rule, $(ALL_FIPS140), build/$(_rule)/%): BUILD_ARGS += -tags fips140enforce
 
 build/%/nebula: .FORCE
 	GOOS=$(firstword $(subst -, , $*)) \
@@ -214,10 +229,7 @@ vet:
 	go vet $(VET_FLAGS) -v ./...
 
 test:
-	go test -v ./...
-
-test-boringcrypto:
-	GOEXPERIMENT=boringcrypto CGO_ENABLED=1 go test -ldflags "-checklinkname=0" -v ./...
+	$(TEST_ENV) go test $(TEST_FLAGS) -v ./...
 
 test-pkcs11:
 	CGO_ENABLED=1 go test -v -tags pkcs11 ./...
@@ -260,29 +272,72 @@ ifeq ($(words $(MAKECMDGOALS)),1)
 	@$(MAKE) service ${.DEFAULT_GOAL} --no-print-directory
 endif
 
+# Useful to chain together, like:
+# - make fips140 e2evv
+# - make fips140 smoke-docker
+# Use `release-fips140` to build release binaries
+fips140:
+	@echo > $(NULL_FILE)
+ifeq ($(strip $(GOFIPS140)),)
+	$(eval GOFIPS140 = $(FIPSVERSION))
+endif
+	$(eval GOENV += GOFIPS140=$(GOFIPS140))
+	$(eval BUILD_ARGS += -tags fips140enforce)
+	$(eval TEST_ENV += $(GOENV))
+	$(eval CURVE = P256)
+ifeq ($(words $(MAKECMDGOALS)),1)
+	@$(MAKE) fips140 GOFIPS140=$(GOFIPS140) ${.DEFAULT_GOAL} --no-print-directory
+endif
+
+# To test the future pending module, use like `make fips140-latest test`
+ALL_GOFIPS140 = v1.0.0 v1.26.0 latest
+define FIPS140_rule
+fips140-$(1): GOFIPS140 = $(1)
+fips140-$(1): fips140
+endef
+$(foreach _rule, $(ALL_GOFIPS140), $(eval $(call FIPS140_rule,$(_rule))))
+
+# Iterate and run the goals for all fips versions, like `make fips140-all GOALS=test`
+fips140-all:
+	@$(foreach _v,$(ALL_GOFIPS140),$(MAKE) fips140-$(_v) $(GOALS) &&) true
+
+# Useful to chain together, like:
+# - make boringcrypto e2evv
+# - make boringcrypto smoke-docker
+# Use `release-boringcrypto` or `bin-boringcrypto` to build release binaries
+boringcrypto:
+	@echo > $(NULL_FILE)
+	$(eval GOENV += GOEXPERIMENT=boringcrypto CGO_ENABLED=1)
+	$(eval TEST_ENV += $(GOENV))
+	$(eval CURVE = P256)
+ifeq ($(words $(MAKECMDGOALS)),1)
+	@$(MAKE) boringcrypto ${.DEFAULT_GOAL} --no-print-directory
+endif
+
 bin-docker: bin build/linux-amd64/nebula build/linux-amd64/nebula-cert
 
+smoke-docker: BUILD_ARGS += -race
+smoke-docker: GOENV += CGO_ENABLED=1
 smoke-docker: bin-docker
-	cd .github/workflows/smoke/ && ./build.sh
-	cd .github/workflows/smoke/ && ./smoke.sh
-	cd .github/workflows/smoke/ && NAME="smoke-p256" CURVE="P256" ./build.sh
-	cd .github/workflows/smoke/ && NAME="smoke-p256" ./smoke.sh
+	# This is so we can limit `fips140` smoke test to just P256 curve.
+	if [ "$(CURVE)" != "P256" ]; then cd .github/workflows/smoke/ && $(GOENV) ./build.sh; fi
+	if [ "$(CURVE)" != "P256" ]; then cd .github/workflows/smoke/ && $(GOENV) ./smoke.sh; fi
+	cd .github/workflows/smoke/ && $(GOENV) NAME="smoke-p256" CURVE="P256" ./build.sh
+	cd .github/workflows/smoke/ && $(GOENV) NAME="smoke-p256" ./smoke.sh
 
+smoke-relay-docker: BUILD_ARGS += -race
+smoke-relay-docker: GOENV += CGO_ENABLED=1
 smoke-relay-docker: bin-docker
-	cd .github/workflows/smoke/ && ./build-relay.sh
-	cd .github/workflows/smoke/ && ./smoke-relay.sh
+	cd .github/workflows/smoke/ && $(GOENV) ./build-relay.sh
+	cd .github/workflows/smoke/ && $(GOENV) ./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
-
 smoke-vagrant/%: bin-docker build/%/nebula
 	cd .github/workflows/smoke/ && ./build.sh $*
 	cd .github/workflows/smoke/ && ./smoke-vagrant.sh $*
 
 .FORCE:
-.PHONY: all all-linux all-freebsd all-openbsd all-netbsd all-darwin all-windows all-cross-linux all-cross-linux-arm all-cross-linux-mips all-cross-linux-other all-cross-darwin all-cross-windows bench bench-cpu bench-cpu-long bin debug 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 bin-windows bin-windows-arm64 bin-darwin bin-freebsd bin-freebsd-arm64 bin-boringcrypto bin-fips140 bin-pkcs11 bin-docker boringcrypto build-test-mobile docker e2e e2ev e2evv e2evvv e2evvvv e2e-bench fips140 fips140-all $(ALL_GOFIPS140:%=fips140-%) install proto release release-linux release-freebsd release-openbsd release-netbsd release-boringcrypto release-fips140 service smoke-docker smoke-relay-docker smoke-docker-ipv6 test test-pkcs11 test-cov-html vet smoke-vagrant/%
 .DEFAULT_GOAL := bin

README.md

@@ -145,17 +145,27 @@ To build nebula for a specific platform (ex, Windows):
 
 See the [Makefile](Makefile) for more details on build targets
 
-## Curve P256 and BoringCrypto
+## Curve P256 and FIPS 140-3 mode
 
 The default curve used for cryptographic handshakes and signatures is Curve25519. This is the recommended setting for most users. If your deployment has certain compliance requirements, you have the option of creating your CA using `nebula-cert ca -curve P256` to use NIST Curve P256. The CA will then sign certificates using ECDSA P256, and any hosts using these certificates will use P256 for ECDH handshakes.
 
-In addition, Nebula can be built using the [BoringCrypto GOEXPERIMENT](https://github.com/golang/go/blob/go1.20/src/crypto/internal/boring/README.md) by running either of the following make targets:
+Nebula can be built to support the [FIPS 140-3](https://go.dev/doc/security/fips140) mode of Go by running either of the following make targets. (This sets GOFIPS140=v1.0.0, which must be done at compile time so that the correct AES-GCM can be used for FIPS 140-3 enforcement mode).
+
+```sh
+make fips140
+make fips140 test
+make release-fips140
+```
+
+Nebula can also be built using the [BoringCrypto GOEXPERIMENT](https://github.com/golang/go/blob/go1.20/src/crypto/internal/boring/README.md) by running either of the following make targets.
 
 ```sh
 make bin-boringcrypto
 make release-boringcrypto
 ```
 
+NOTE: boringcrypto support is deprecated and will be removed in the next release. Users should migrate to the native FIPS 140-3 mode described above.
+
 This is not the recommended default deployment, but may be useful based on your compliance requirements.
 
 ## Credits

cmd/nebula-cert/ca.go

@@ -3,6 +3,7 @@ package main
 import (
 	"crypto/ecdsa"
 	"crypto/elliptic"
+	"crypto/fips140"
 	"crypto/rand"
 	"flag"
 	"fmt"
@@ -43,6 +44,13 @@ type caFlags struct {
 	subnets *string
 }
 
+func defaultCurve() string {
+	if fips140.Enforced() {
+		return "P256"
+	}
+	return "25519"
+}
+
 func newCaFlags() *caFlags {
 	cf := caFlags{set: flag.NewFlagSet("ca", flag.ContinueOnError)}
 	cf.set.Usage = func() {}
@@ -59,7 +67,7 @@ func newCaFlags() *caFlags {
 	cf.argonParallelism = cf.set.Uint("argon-parallelism", 4, "Optional: Argon2 parallelism parameter used for encrypted private key passphrase")
 	cf.argonIterations = cf.set.Uint("argon-iterations", 1, "Optional: Argon2 iterations parameter used for encrypted private key passphrase")
 	cf.encryption = cf.set.Bool("encrypt", false, "Optional: prompt for passphrase and write out-key in an encrypted format")
-	cf.curve = cf.set.String("curve", "25519", "EdDSA/ECDSA Curve (25519, P256)")
+	cf.curve = cf.set.String("curve", defaultCurve(), "EdDSA/ECDSA Curve (25519, P256)")
 	cf.p11url = p11Flag(cf.set)
 
 	cf.ips = cf.set.String("ips", "", "Deprecated, see -networks")

cmd/nebula-cert/fips140enforce.go

@@ -0,0 +1,5 @@
+//go:build fips140enforce
+
+//go:debug fips140=only
+
+package main

cmd/nebula-cert/keygen.go

@@ -24,7 +24,7 @@ func newKeygenFlags() *keygenFlags {
 	cf.set.Usage = func() {}
 	cf.outPubPath = cf.set.String("out-pub", "", "Required: path to write the public key to")
 	cf.outKeyPath = cf.set.String("out-key", "", "Required: path to write the private key to")
-	cf.curve = cf.set.String("curve", "25519", "ECDH Curve (25519, P256)")
+	cf.curve = cf.set.String("curve", defaultCurve(), "ECDH Curve (25519, P256)")
 	cf.p11url = p11Flag(cf.set)
 	return &cf
 }

cmd/nebula-service/fips140enforce.go

@@ -0,0 +1,5 @@
+//go:build fips140enforce
+
+//go:debug fips140=only
+
+package main

cmd/nebula/close_on_timer_test.go

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

cmd/nebula/fips140enforce.go

@@ -0,0 +1,5 @@
+//go:build fips140enforce
+
+//go:debug fips140=only
+
+package main

connection_manager_test.go

@@ -25,6 +25,7 @@ func newTestLighthouse() *LightHouse {
 	lighthouses := []netip.Addr{}
 	staticList := map[netip.Addr]struct{}{}
 
+	lh.localAddrsFn = func(*LocalAllowList) []netip.Addr { return nil }
 	lh.lighthouses.Store(&lighthouses)
 	lh.staticList.Store(&staticList)
 

connection_state.go

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

control.go

@@ -53,6 +53,7 @@ type Control struct {
 	statsStart             func()
 	dnsStart               func()
 	lighthouseStart        func()
+	networkChangeStart     func(rebind func())
 	connectionManagerStart func(context.Context)
 }
 
@@ -104,6 +105,9 @@ func (c *Control) Start() error {
 	if c.dnsStart != nil {
 		go c.dnsStart()
 	}
+	if c.networkChangeStart != nil {
+		go c.networkChangeStart(c.RebindUDPServer)
+	}
 	if c.connectionManagerStart != nil {
 		go c.connectionManagerStart(c.ctx)
 	}
@@ -198,7 +202,11 @@ func (c *Control) RebindUDPServer() {
 		return
 	}
 
-	_ = c.f.outside.Rebind()
+	// A failure here means we are likely still pinned to the interface we came up on, so the rest of this is
+	// unlikely to help. Say so instead of silently carrying on as if we rebound.
+	if err := c.f.outside.Rebind(); err != nil {
+		c.l.Error("Failed to rebind udp socket", "error", err)
+	}
 
 	// Trigger a lighthouse update, useful for mobile clients that should have an update interval of 0
 	c.f.lightHouse.SendUpdate()

control_lifecycle_test.go

@@ -11,8 +11,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/batch"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/test"
 	"github.com/slackhq/nebula/udp"
@@ -32,9 +30,9 @@ func newFakeDevice() *fakeDevice {
 
 // 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) {
+func (d *fakeDevice) Read(p []byte) (int, error) {
 	<-d.closedCh
-	return nil, io.EOF
+	return 0, io.EOF
 }
 
 func (d *fakeDevice) Write(p []byte) (int, error) { return len(p), nil }
@@ -51,8 +49,10 @@ 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 }
+func (d *fakeDevice) SupportsMultiqueue() bool              { return false }
+func (d *fakeDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, errors.New("unsupported")
+}
 
 // newReadyControl hand-builds the minimum Control that Main would have
 // produced right before Start, including the construction token NewInterface
@@ -78,7 +78,7 @@ func newReadyControl(t *testing.T) (*Control, *fakeDevice, *fakeConn) {
 		inside:     dev,
 		outside:    conn,
 		writers:    []udp.Conn{conn},
-		batchers:   make([]batch.RxBatcher, 1),
+		readers:    make([]io.ReadWriteCloser, 1),
 		routines:   1,
 		hostMap:    newHostMap(l),
 		lightHouse: lh,
@@ -109,8 +109,7 @@ func TestControl_StopBeforeStart(t *testing.T) {
 	require.NoError(t, c.Wait())
 
 	// A stopped control can never be started
-	err := c.Start()
-	require.ErrorIs(t, err, ErrAlreadyStopped)
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
 
 	// A second Stop is a harmless no-op
 	c.Stop()
@@ -144,29 +143,19 @@ type fakeConn struct {
 	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, _ func()) error { return nil }
-func (c *fakeConn) WriteTo(_ []byte, _ netip.AddrPort) error  { return nil }
-func (c *fakeConn) WriteBatch(_ [][]byte, _ []netip.AddrPort, _ []byte) 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 }
+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 (d *multiqueueDevice) SupportsMultiqueue() bool { return true }
 
 func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
 	dev := &multiqueueDevice{fakeDevice: newFakeDevice()}
@@ -177,7 +166,7 @@ func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
 		inside:   dev,
 		outside:  conn,
 		writers:  []udp.Conn{conn},
-		batchers: make([]batch.RxBatcher, 2),
+		readers:  make([]io.ReadWriteCloser, 2),
 		routines: 2,
 		l:        test.NewLogger(),
 	}
@@ -192,8 +181,7 @@ func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
 	}
 
 	// The second reader fails to open, everything must be released
-	err := c.Start()
-	require.Error(t, err)
+	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")
@@ -263,18 +251,15 @@ func TestControl_ConcurrentStopAndStart(t *testing.T) {
 	// panic and Wait must observe the final state
 	require.NoError(t, c.Wait())
 	assert.Equal(t, StateStopped, c.State())
-	err := c.Start()
-	require.ErrorIs(t, err, ErrAlreadyStopped)
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
 }
 
 func TestControl_StartStopLifecycle(t *testing.T) {
 	c, dev, conn := newReadyControl(t)
 
-	err := c.Start()
-	require.NoError(t, err)
+	require.NoError(t, c.Start())
 	assert.Equal(t, StateStarted, c.State())
-	err = c.Start()
-	require.ErrorIs(t, err, ErrAlreadyStarted)
+	require.ErrorIs(t, c.Start(), ErrAlreadyStarted)
 
 	// Stop must unpark the reader blocked in the device and release everything
 	c.Stop()
@@ -285,8 +270,7 @@ func TestControl_StartStopLifecycle(t *testing.T) {
 
 	// The reader drained off a closed device, that is not a fatal error
 	require.NoError(t, c.Wait())
-	err = c.Start()
-	require.ErrorIs(t, err, ErrAlreadyStopped)
+	require.ErrorIs(t, c.Start(), ErrAlreadyStopped)
 }
 
 func TestControl_RebindIsGatedByState(t *testing.T) {
@@ -296,8 +280,7 @@ func TestControl_RebindIsGatedByState(t *testing.T) {
 	c.RebindUDPServer()
 	assert.Equal(t, 0, conn.rebinds, "rebind before start must be a no-op")
 
-	err := c.Start()
-	require.NoError(t, err)
+	require.NoError(t, c.Start())
 	c.RebindUDPServer()
 	assert.Equal(t, 1, conn.rebinds, "rebind while started must reach the conn")
 

control_tester.go

@@ -108,7 +108,19 @@ func (c *Control) GetVpnAddrs() []netip.Addr {
 }
 
 func (c *Control) GetUDPAddr() netip.AddrPort {
-	return c.f.outside.(*udp.TesterConn).Addr
+	return c.f.outside.(*udp.TesterConn).GetAddr()
+}
+
+// SetUDPAddr moves this node to a new underlay address, standing in for a laptop waking up on a different
+// network. Register the new address with the router as well or nothing will route back.
+func (c *Control) SetUDPAddr(addr netip.AddrPort) {
+	c.f.outside.(*udp.TesterConn).SetAddr(addr)
+}
+
+// SetLocalAddrsFn replaces underlay address discovery so a test can advertise its simulated address instead of
+// whatever this machine's NICs happen to be. Call it before Start, SendUpdate reads it from the update worker.
+func (c *Control) SetLocalAddrsFn(fn func(*LocalAllowList) []netip.Addr) {
+	c.f.lightHouse.localAddrsFn = fn
 }
 
 func (c *Control) KillPendingTunnel(vpnIp netip.Addr) bool {

dns_server.go

@@ -97,8 +97,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	newAddr := getDnsServerAddr(c)
 
 	d.serverMu.Lock()
-	running := d.server
-	runningStarted := d.started
+	running := d.server != nil
 	sameAddr := d.addr == newAddr
 	d.addr = newAddr
 	d.enabled.Store(enabled)
@@ -112,7 +111,7 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 	}
 
 	if !enabled {
-		if running != nil {
+		if running {
 			d.Stop()
 		}
 		// Drop any records that accumulated while enabled; a later re-enable
@@ -121,12 +120,12 @@ func (d *dnsServer) reload(c *config.C, initial bool) error {
 		return nil
 	}
 
-	if running == nil {
+	if !running {
 		// Was disabled (or never started); bring it up now.
 		go d.Start()
 	} else if !sameAddr {
-		d.shutdownServer(running, runningStarted, "reload")
-		// Old Start goroutine has now exited; bring up a fresh listener on the new address.
+		// Stop clears the slot before shutting down, otherwise the Start below can find the dying server and refuse
+		d.Stop()
 		go d.Start()
 	}
 
@@ -162,7 +161,9 @@ func (d *dnsServer) Start() {
 
 	started := make(chan struct{})
 	d.serverMu.Lock()
-	if d.ctx.Err() != nil {
+	// Re-check enabled under the lock, a disable that raced our check above snapshots the slot under it too.
+	// Two reloads in quick succession can both spawn a Start, the loser would orphan the live listener past Stop
+	if d.ctx.Err() != nil || d.server != nil || !d.enabled.Load() {
 		d.serverMu.Unlock()
 		return
 	}
@@ -200,6 +201,14 @@ func (d *dnsServer) Start() {
 		close(started)
 	}
 
+	// Release our slot, unless a reload already replaced us, so a dead listener can't block a future Start
+	d.serverMu.Lock()
+	if d.server == server {
+		d.server = nil
+		d.started = nil
+	}
+	d.serverMu.Unlock()
+
 	if err != nil {
 		d.l.Warn("Failed to run the DNS responder", "error", err)
 	}

dns_server_test.go

@@ -194,14 +194,51 @@ func TestDnsServer_reload_initial_serveDnsWithoutLighthouse(t *testing.T) {
 }
 
 func TestDnsServer_reload_sameAddr_noOp(t *testing.T) {
+	port := freeUDPPort(t)
 	ds, c := newTestDnsServer(t)
-	setDnsConfig(c, "127.0.0.1", "0", true, true)
-
+	setDnsConfig(c, "127.0.0.1", port, true, true)
 	require.NoError(t, ds.reload(c, true))
-	// No server running yet, no addr change. Reload should not spawn anything.
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	before := ds.server
+	ds.serverMu.Unlock()
+	require.NotNil(t, before)
+
+	// Same address, so the running listener must be left alone rather than rebuilt under live queries
 	require.NoError(t, ds.reload(c, false))
 	assert.True(t, ds.enabled.Load())
-	assert.Nil(t, ds.server)
+
+	ds.serverMu.Lock()
+	after := ds.server
+	ds.serverMu.Unlock()
+	assert.Same(t, before, after, "a same-address reload must not restart the listener")
+
+	ds.Stop()
+}
+
+// The branch the old sameAddr test was accidentally hitting: enabled with nothing running means reload starts it.
+func TestDnsServer_reload_whenNotRunning_starts(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+
+	// initial only records config, it never starts anything
+	require.NoError(t, ds.reload(c, true))
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "the initial reload must not start a listener")
+	ds.serverMu.Unlock()
+
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	assert.NotNil(t, ds.server, "a reload with nothing running should bring DNS up")
+	ds.serverMu.Unlock()
+
+	ds.Stop()
 }
 
 func TestDnsServer_StartStop_lifecycle(t *testing.T) {
@@ -427,3 +464,168 @@ func waitFor(t *testing.T, cond func() bool) {
 	}
 	t.Fatal("timed out waiting for condition")
 }
+
+// Two reloads in quick succession, or a HUP before Control.Start, can race two Starts at the same listener.
+func TestDnsServer_Start_isIdempotent(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	first := ds.server
+	ds.serverMu.Unlock()
+	require.NotNil(t, first)
+
+	// If the second Start replaces the tracked server, Stop kills the wrong one and the port leaks
+	done := make(chan struct{})
+	go func() {
+		ds.Start()
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(time.Second * 5):
+		t.Fatal("second Start never returned")
+	}
+
+	ds.serverMu.Lock()
+	second := ds.server
+	ds.serverMu.Unlock()
+	assert.Same(t, first, second, "a second Start must not replace the running server")
+
+	// The real proof, after Stop the port must actually be free
+	ds.Stop()
+	waitFor(t, func() bool {
+		pc, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+		if err != nil {
+			return false
+		}
+		_ = pc.Close()
+		return true
+	})
+}
+
+// An address change must actually end up listening on the new port. Start's guard refuses when a server is already
+// installed, so reload has to clear the slot before shutting the old one down.
+func TestDnsServer_reload_addrChange_restarts(t *testing.T) {
+	first := freeUDPPort(t)
+	second := freeUDPPort(t)
+
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", first, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	go ds.Start()
+	waitForBind(t, ds)
+
+	// Cycle a few times, the failure this guards against depends on which goroutine wins serverMu
+	for i := range 8 {
+		want := second
+		if i%2 == 1 {
+			want = first
+		}
+		setDnsConfig(c, "127.0.0.1", want, true, true)
+		require.NoError(t, ds.reload(c, false))
+		waitForBind(t, ds)
+
+		ds.serverMu.Lock()
+		srv := ds.server
+		ds.serverMu.Unlock()
+		require.NotNil(t, srv, "reload left DNS down instead of restarting it")
+		require.Equal(t, "127.0.0.1:"+want, srv.Addr, "reload should be serving the new address")
+	}
+
+	// Land back on second so the port assertions below are meaningful
+	setDnsConfig(c, "127.0.0.1", second, true, true)
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	// The old port must be released and the new one actually held
+	waitFor(t, func() bool {
+		pc, err := net.ListenPacket("udp", "127.0.0.1:"+first)
+		if err != nil {
+			return false
+		}
+		_ = pc.Close()
+		return true
+	})
+	_, err := net.ListenPacket("udp", "127.0.0.1:"+second)
+	require.Error(t, err, "the new address should be bound by the DNS responder")
+
+	ds.Stop()
+}
+
+// A listener that dies on its own must release the slot, or a later same-addr reload sees it as running and no-ops.
+func TestDnsServer_Start_bindFailure_releasesSlot(t *testing.T) {
+	port := freeUDPPort(t)
+	blocker, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+	require.NoError(t, err)
+
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+
+	ds.Start() // returns once the bind fails
+
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "a listener that failed to bind must not stay parked in the slot")
+	ds.serverMu.Unlock()
+
+	// With the slot released, a reload can retry once the port frees up
+	require.NoError(t, blocker.Close())
+	require.NoError(t, ds.reload(c, false))
+	waitForBind(t, ds)
+
+	ds.serverMu.Lock()
+	assert.NotNil(t, ds.server, "a same-addr reload should retry after a failed bind")
+	ds.serverMu.Unlock()
+
+	ds.Stop()
+}
+
+// A disable that lands while Start is between its unlocked check and the guard must not leave a listener behind.
+func TestDnsServer_Start_refusesWhenDisabledUnderLock(t *testing.T) {
+	port := freeUDPPort(t)
+	ds, c := newTestDnsServer(t)
+	setDnsConfig(c, "127.0.0.1", port, true, true)
+	require.NoError(t, ds.reload(c, true))
+	require.True(t, ds.enabled.Load())
+
+	// Holding serverMu parks Start on the lock, the only way to land the disable in that window on purpose
+	ds.serverMu.Lock()
+
+	done := make(chan struct{})
+	go func() {
+		ds.Start()
+		close(done)
+	}()
+
+	select {
+	case <-done:
+		ds.serverMu.Unlock()
+		t.Fatal("Start returned early, the test never exercised the window")
+	case <-time.After(time.Millisecond * 100):
+	}
+
+	// The disable reload's critical section. It sees nothing running, so it never calls Stop.
+	ds.enabled.Store(false)
+	ds.serverMu.Unlock()
+
+	select {
+	case <-done:
+	case <-time.After(time.Second * 5):
+		t.Fatal("Start never returned")
+	}
+
+	ds.serverMu.Lock()
+	assert.Nil(t, ds.server, "Start must not install a listener a disable already cancelled")
+	ds.serverMu.Unlock()
+
+	pc, err := net.ListenPacket("udp", "127.0.0.1:"+port)
+	require.NoError(t, err, "an orphaned listener is still holding the port")
+	_ = pc.Close()
+}

e2e/handshakes_test.go

@@ -725,6 +725,70 @@ func TestReestablishRelays(t *testing.T) {
 
 }
 
+func TestRelayHandshakeOverDisestablishedEntry(t *testing.T) {
+	t.Parallel()
+	// If them tears down the tunnel while me keeps Established relay state, me's next
+	// handshake flows through the relay with no fresh CreateRelayRequest and lands on
+	// them's Disestablished terminal relay entry. them must re-establish that entry, or
+	// its first transmit deletes its only relay and the tunnel is born transmit-dead:
+	// them can receive but every send is silently dropped.
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version1, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+	myControl, myVpnIpNet, _, _ := newSimpleServer(cert.Version1, ca, caKey, "me     ", "10.128.0.1/24", m{"relay": m{"use_relays": true}})
+	relayControl, relayVpnIpNet, relayUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "relay  ", "10.128.0.128/24", m{"relay": m{"am_relay": true}})
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version1, ca, caKey, "them   ", "10.128.0.2/24", m{"relay": m{"use_relays": true}})
+
+	// Teach my how to get to the relay and that their can be reached via the relay
+	myControl.InjectLightHouseAddr(relayVpnIpNet[0].Addr(), relayUdpAddr)
+	myControl.InjectRelays(theirVpnIpNet[0].Addr(), []netip.Addr{relayVpnIpNet[0].Addr()})
+	relayControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+
+	// Build a router so we don't have to reason who gets which packet
+	r := router.NewR(t, myControl, relayControl, theirControl)
+	defer r.RenderFlow()
+
+	// Start the servers
+	myControl.Start()
+	relayControl.Start()
+	theirControl.Start()
+
+	t.Log("Trigger a handshake from me to them via the relay")
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("Hi from me")))
+
+	p := r.RouteForAllUntilTxTun(theirControl)
+	assertUdpPacket(t, []byte("Hi from me"), p, myVpnIpNet[0].Addr(), theirVpnIpNet[0].Addr(), 80, 80)
+	oldIdx := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false).LocalIndex
+
+	t.Log("Close the tunnel on them only, marking their relay entry Disestablished")
+	theirControl.CloseTunnel(myVpnIpNet[0].Addr(), true)
+
+	t.Log("Re-handshake from me, riding the still-Established relay state")
+	myControl.ReHandshake(theirVpnIpNet[0].Addr())
+	for {
+		h := myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false)
+		if h != nil && h.LocalIndex != oldIdx && h.RemoteIndex != 0 {
+			break
+		}
+		r.RouteForAllExitFunc(func(*udp.Packet, *nebula.Control) router.ExitType {
+			return router.RouteAndExit
+		})
+	}
+
+	hAtThem := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+	require.NotNil(t, hAtThem, "them should have completed the relayed handshake")
+	require.Equal(t, []netip.Addr{relayVpnIpNet[0].Addr()}, hAtThem.CurrentRelaysToMe, "them should know a relay for the new tunnel")
+
+	t.Log("Send from them to me; their only relay entry must survive the transmit")
+	theirControl.InjectTunPacket(BuildTunUDPPacket(myVpnIpNet[0].Addr(), 80, theirVpnIpNet[0].Addr(), 80, []byte("Hi from them")))
+	require.Never(t, func() bool {
+		h := theirControl.GetHostInfoByVpnAddr(myVpnIpNet[0].Addr(), false)
+		return h == nil || len(h.CurrentRelaysToMe) == 0
+	}, time.Second, 10*time.Millisecond, "them deleted its only relay entry; the tunnel is permanently transmit-dead")
+
+	p = r.RouteForAllUntilTxTun(myControl)
+	assertUdpPacket(t, []byte("Hi from them"), p, theirVpnIpNet[0].Addr(), myVpnIpNet[0].Addr(), 80, 80)
+	r.RenderHostmaps("Final hostmaps", myControl, relayControl, theirControl)
+}
+
 func TestStage1RaceRelays(t *testing.T) {
 	t.Parallel()
 	//NOTE: this is a race between me and relay resulting in a full tunnel from me to them via relay

e2e/helpers_test.go

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

e2e/rebind_test.go

@@ -0,0 +1,225 @@
+//go:build e2e_testing
+// +build e2e_testing
+
+package e2e
+
+import (
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/slackhq/nebula"
+	"github.com/slackhq/nebula/cert"
+	"github.com/slackhq/nebula/cert_test"
+	"github.com/slackhq/nebula/e2e/router"
+	"github.com/slackhq/nebula/header"
+	"github.com/slackhq/nebula/udp"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// reportedAddrs is what the lighthouse would hand a peer asking where vpnAddr is.
+func reportedAddrs(t *testing.T, lh *nebula.Control, vpnAddr netip.Addr) []netip.AddrPort {
+	t.Helper()
+	cm := lh.QueryLighthouse(vpnAddr)
+	if cm == nil {
+		return nil
+	}
+	var out []netip.AddrPort
+	for _, c := range *cm {
+		out = append(out, c.Reported...)
+		out = append(out, c.Learned...)
+	}
+	return out
+}
+
+// waitForLighthouseMsg routes until a lighthouse message lands on lh, or gives up. Reports whether one arrived.
+func waitForLighthouseMsg(t *testing.T, r *router.R, lh *nebula.Control, wait time.Duration) bool {
+	t.Helper()
+	h := &header.H{}
+	return r.RouteForAllExitFuncOrTimeout(wait, func(p *udp.Packet, c *nebula.Control) router.ExitType {
+		if c != lh {
+			return router.KeepRouting
+		}
+		// Punches are a single byte and never parse, they are just not what we are after
+		if err := h.Parse(p.Data); err != nil {
+			return router.KeepRouting
+		}
+		if h.Type == header.LightHouse {
+			return router.RouteAndExit
+		}
+		return router.KeepRouting
+	})
+}
+
+// A laptop that changes networks has to tell the lighthouse promptly, otherwise the lighthouse keeps handing peers
+// the old address and their punches land nowhere. On a long lighthouse interval the only thing that closes that
+// window is the rebind, which on darwin the network change monitor drives. The e2e build compiles the monitor out,
+// so we call RebindUDPServer directly, which is the same thing the monitor does.
+func TestRebindSendsLighthouseUpdate(t *testing.T) {
+	t.Parallel()
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	// 600s interval, so nothing scheduled can send an update during this test. A rebind is the only thing that can.
+	myControl, _, _, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	})
+
+	r := router.NewR(t, lhControl, myControl)
+	defer r.RenderFlow()
+
+	lhControl.Start()
+	myControl.Start()
+
+	// Let the startup registration finish, then clear everything it left behind
+	require.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5), "expected an initial registration")
+	r.RouteFor(time.Millisecond * 400)
+
+	// Nothing should be talking to the lighthouse on its own now
+	require.False(t, waitForLighthouseMsg(t, r, lhControl, time.Millisecond*200),
+		"nothing should reach the lighthouse before the rebind")
+
+	myControl.RebindUDPServer()
+
+	assert.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5),
+		"a rebind should push an update to the lighthouse rather than waiting out the interval")
+
+	lhControl.Stop()
+	myControl.Stop()
+}
+
+// The other half of a rebind: every live tunnel requeries the lighthouse on its next send. That query is what makes
+// the lighthouse tell the peer to punch toward our new address, which is the part that actually revives a tunnel
+// whose remote NAT state died while we were on a different network.
+func TestRebindRequeriesPeersOnNextSend(t *testing.T) {
+	t.Parallel()
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	lhCfg := m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+			// Without this the peers advertise this machine's real addresses and then try to punch at them,
+			// which the router has no route for.
+			"local_allow_list": m{
+				"10.0.0.0/24": true,
+				"::/0":        false,
+			},
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	}
+
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", lhCfg)
+	theirControl, theirVpnIpNet, theirUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "them", "10.128.0.3/24", lhCfg)
+
+	r := router.NewR(t, lhControl, myControl, theirControl)
+	defer r.RenderFlow()
+
+	lhControl.Start()
+	myControl.Start()
+	theirControl.Start()
+	r.RouteFor(time.Millisecond * 500)
+
+	// Point the peers at each other directly, this test is about the rebind and not about lighthouse discovery
+	myControl.InjectLightHouseAddr(theirVpnIpNet[0].Addr(), theirUdpAddr)
+	theirControl.InjectLightHouseAddr(myVpnIpNet[0].Addr(), myUdpAddr)
+
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("initial")))
+	r.RouteFor(time.Second)
+	require.NotNil(t, myControl.GetHostInfoByVpnAddr(theirVpnIpNet[0].Addr(), false), "expected a tunnel to them")
+	r.RouteFor(time.Millisecond * 300)
+
+	// Assert on what the peer sees rather than on lighthouse traffic. A query for them makes the lighthouse send
+	// them a punch notification, which is the whole point. Our own update to the lighthouse sends them nothing,
+	// so this cannot be satisfied by the update the rebind itself pushes.
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("quiet")))
+	require.False(t, waitForLighthouseMsg(t, r, theirControl, time.Millisecond*300),
+		"an ordinary send should not requery the lighthouse")
+
+	myControl.RebindUDPServer()
+	r.RouteFor(time.Millisecond * 300) // let the update the rebind itself sends pass by
+
+	myControl.InjectTunPacket(BuildTunUDPPacket(theirVpnIpNet[0].Addr(), 80, myVpnIpNet[0].Addr(), 80, []byte("after rebind")))
+	assert.True(t, waitForLighthouseMsg(t, r, theirControl, time.Second*5),
+		"the first send after a rebind should requery the lighthouse, which then tells the peer to punch at us")
+
+	lhControl.Stop()
+	myControl.Stop()
+	theirControl.Stop()
+}
+
+// The scenario this whole thing exists for: a laptop sleeps at the office and wakes up at home on a new address.
+// Until it tells the lighthouse, the lighthouse keeps handing peers the office address, so their punches land
+// nowhere and the tunnel stays dead. On a long interval the rebind is the only thing that closes that window.
+func TestRebindAdvertisesNewAddressAfterMove(t *testing.T) {
+	t.Parallel()
+	ca, _, caKey, _ := cert_test.NewTestCaCert(cert.Version2, cert.Curve_CURVE25519, time.Now(), time.Now().Add(10*time.Minute), nil, nil, []string{})
+
+	lhControl, lhVpnIpNet, lhUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "lh", "10.128.0.1/24", m{
+		"lighthouse": m{"am_lighthouse": true},
+	})
+
+	myControl, myVpnIpNet, myUdpAddr, _ := newSimpleServer(cert.Version2, ca, caKey, "me", "10.128.0.2/24", m{
+		"lighthouse": m{
+			"hosts":    []any{lhVpnIpNet[0].Addr().String()},
+			"interval": 600,
+		},
+		"static_host_map": m{
+			lhVpnIpNet[0].Addr().String(): []any{lhUdpAddr.String()},
+		},
+	})
+
+	// Advertise wherever we currently are rather than this machine's real NICs, read fresh each time so a move
+	// is picked up.
+	myControl.SetLocalAddrsFn(func(*nebula.LocalAllowList) []netip.Addr {
+		return []netip.Addr{myControl.GetUDPAddr().Addr()}
+	})
+
+	r := router.NewR(t, lhControl, myControl)
+	defer r.RenderFlow()
+
+	lhControl.Start()
+	myControl.Start()
+
+	require.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5), "expected an initial registration")
+	r.RouteFor(time.Millisecond * 400)
+
+	require.Contains(t, reportedAddrs(t, lhControl, myVpnIpNet[0].Addr()), myUdpAddr,
+		"the lighthouse should know the address we started on")
+
+	// Wake up somewhere else
+	newAddr := netip.MustParseAddrPort("10.0.0.99:4242")
+	myControl.SetUDPAddr(newAddr)
+	r.AddRoute(newAddr.Addr(), newAddr.Port(), myControl)
+
+	// Nothing has told the lighthouse, and with interval 600 nothing scheduled will
+	r.RouteFor(time.Millisecond * 400)
+	require.NotContains(t, reportedAddrs(t, lhControl, myVpnIpNet[0].Addr()), newAddr,
+		"the lighthouse should still be handing out the old address before the rebind")
+
+	myControl.RebindUDPServer()
+	require.True(t, waitForLighthouseMsg(t, r, lhControl, time.Second*5), "expected an update after the rebind")
+	r.RouteFor(time.Millisecond * 400)
+
+	assert.Contains(t, reportedAddrs(t, lhControl, myVpnIpNet[0].Addr()), newAddr,
+		"after the rebind the lighthouse should hand peers our new address")
+
+	lhControl.Stop()
+	myControl.Stop()
+}

e2e/router/router.go

@@ -114,6 +114,28 @@ type packet struct {
 	packet *udp.Packet
 	tun    bool // a packet pulled off a tun device
 	rx     bool // the packet was received by a udp device
+
+	// h is the nebula header, parsed once when the packet is recorded. parseErr says why there isn't one, which
+	// the flow log reports rather than hiding. Punchy sends a single byte, so an unparseable packet is normal.
+	h        header.H
+	parseErr error
+}
+
+// fromAddr and toAddr are the addresses this packet actually travelled between. Reading them off the control
+// instead would misreport the whole history once a test moves a node. Tun packets are synthesized without
+// addresses, so they fall back to the control.
+func (p *packet) fromAddr() netip.AddrPort {
+	if p.tun || !p.packet.From.IsValid() {
+		return p.from.GetUDPAddr()
+	}
+	return p.packet.From
+}
+
+func (p *packet) toAddr() netip.AddrPort {
+	if p.tun || !p.packet.To.IsValid() {
+		return p.to.GetUDPAddr()
+	}
+	return p.packet.To
 }
 
 func (p *packet) WasReceived() {
@@ -249,7 +271,7 @@ func (r *R) renderFlow() {
 			continue
 		}
 
-		addr := e.packet.from.GetUDPAddr()
+		addr := e.packet.fromAddr()
 		if _, ok := participants[addr]; ok {
 			continue
 		}
@@ -268,7 +290,6 @@ func (r *R) renderFlow() {
 	}
 
 	// Print packets
-	h := &header.H{}
 	for _, e := range r.flow {
 		if e.packet == nil {
 			//fmt.Fprintf(f, "    note over %s: %s\n", strings.Join(participantsVals, ", "), e.note)
@@ -280,21 +301,22 @@ func (r *R) renderFlow() {
 			fmt.Fprintln(f, r.formatUdpPacket(p))
 
 		} else {
-			if err := h.Parse(p.packet.Data); err != nil {
-				panic(err)
-			}
-
 			line := "--x"
 			if p.rx {
 				line = "->>"
 			}
 
-			fmt.Fprintf(f,
-				"    %s%s%s: %s(%s), index %v, counter: %v\n",
-				normalizeName(p.from.GetUDPAddr().String()),
+			detail := fmt.Sprintf("%s(%s), index %v, counter: %v",
+				p.h.TypeName(), p.h.SubTypeName(), p.h.RemoteIndex, p.h.MessageCounter)
+			if p.parseErr != nil {
+				detail = fmt.Sprintf("unparsed, %v (%d bytes)", p.parseErr, len(p.packet.Data))
+			}
+
+			fmt.Fprintf(f, "    %s%s%s: %s\n",
+				normalizeName(p.fromAddr().String()),
 				line,
-				normalizeName(p.to.GetUDPAddr().String()),
-				h.TypeName(), h.SubTypeName(), h.RemoteIndex, h.MessageCounter,
+				normalizeName(p.toAddr().String()),
+				detail,
 			)
 		}
 	}
@@ -408,29 +430,34 @@ func (r *R) unlockedInjectFlow(from, to *nebula.Control, p *udp.Packet, tun bool
 
 	r.renderHostmaps(fmt.Sprintf("Packet %v", len(r.flow)))
 
-	if len(r.ignoreFlows) > 0 {
-		var h header.H
-		err := h.Parse(p.Data)
-		if err != nil {
-			panic(err)
-		}
+	var h header.H
+	var parseErr error
+	if !tun {
+		parseErr = h.Parse(p.Data)
+	}
 
-		for _, i := range r.ignoreFlows {
-			if !tun {
-				if i.messageType == h.Type && i.subType == h.Subtype {
-					return nil
-				}
-			} else if i.tun.HasValue && i.tun.IsTrue {
+	// Decide before copying, the copy comes from a freelist and an ignored packet would never be released
+	for _, i := range r.ignoreFlows {
+		if tun {
+			if i.tun.HasValue && i.tun.IsTrue {
 				return nil
 			}
+			continue
+		}
+
+		// A packet we could not parse has no type to match against, so no rule can ignore it
+		if parseErr == nil && i.messageType == h.Type && i.subType == h.Subtype {
+			return nil
 		}
 	}
 
 	fp := &packet{
-		from:   from,
-		to:     to,
-		packet: p.Copy(),
-		tun:    tun,
+		from:     from,
+		to:       to,
+		packet:   p.Copy(),
+		tun:      tun,
+		h:        h,
+		parseErr: parseErr,
 	}
 
 	r.flow = append(r.flow, flowEntry{packet: fp})
@@ -690,6 +717,81 @@ func (r *R) RouteUntilAfterMsgType(sender *nebula.Control, msgType header.Messag
 	})
 }
 
+// RouteFor routes everything that shows up for the given duration and then returns. Use it to let a test settle
+// deterministically rather than sleeping and hoping: a single FlushAll races a completing handshake, which queues
+// more packets right behind it.
+func (r *R) RouteFor(d time.Duration) {
+	r.RouteForAllExitFuncOrTimeout(d, func(*udp.Packet, *nebula.Control) ExitType {
+		return KeepRouting
+	})
+}
+
+// RouteForAllExitFuncOrTimeout is RouteForAllExitFunc with a deadline, reporting whether whatDo asked to exit
+// before time ran out. The unbounded version blocks forever on a quiet network, so this is what a test needs to
+// assert that something does NOT happen, or to route for a fixed settling period.
+func (r *R) RouteForAllExitFuncOrTimeout(timeout time.Duration, whatDo ExitFunc) bool {
+	sc := make([]reflect.SelectCase, 0, len(r.controls)+1)
+	cm := make([]*nebula.Control, 0, len(r.controls))
+
+	for _, c := range r.controls {
+		sc = append(sc, reflect.SelectCase{
+			Dir:  reflect.SelectRecv,
+			Chan: reflect.ValueOf(c.GetUDPTxChan()),
+			Send: reflect.Value{},
+		})
+		cm = append(cm, c)
+	}
+
+	timer := time.NewTimer(timeout)
+	defer timer.Stop()
+	sc = append(sc, reflect.SelectCase{
+		Dir:  reflect.SelectRecv,
+		Chan: reflect.ValueOf(timer.C),
+		Send: reflect.Value{},
+	})
+
+	for {
+		x, rx, _ := reflect.Select(sc)
+		if x == len(cm) {
+			return false
+		}
+
+		r.Lock()
+		p := rx.Interface().(*udp.Packet)
+		receiver := r.getControl(cm[x].GetUDPAddr(), p.To, p)
+		if receiver == nil {
+			r.Unlock()
+			panic("Can't RouteForAllExitFuncOrTimeout for host: " + p.To.String())
+		}
+
+		e := whatDo(p, receiver)
+		switch e {
+		case ExitNow:
+			r.Unlock()
+			p.Release()
+			return true
+
+		case RouteAndExit:
+			fp := r.unlockedInjectFlow(cm[x], receiver, p, false)
+			receiver.InjectUDPPacket(p)
+			fp.WasReceived()
+			r.Unlock()
+			p.Release()
+			return true
+
+		case KeepRouting:
+			fp := r.unlockedInjectFlow(cm[x], receiver, p, false)
+			receiver.InjectUDPPacket(p)
+			fp.WasReceived()
+
+		default:
+			panic(fmt.Sprintf("Unknown exitFunc return: %v", e))
+		}
+		r.Unlock()
+		p.Release()
+	}
+}
+
 func (r *R) RouteForAllUntilAfterMsgTypeTo(receiver *nebula.Control, msgType header.MessageType, subType header.MessageSubType) {
 	h := &header.H{}
 	r.RouteForAllExitFunc(func(p *udp.Packet, r *nebula.Control) ExitType {

ecn_inner_test.go

@@ -1,188 +0,0 @@
-package nebula
-
-import (
-	"encoding/binary"
-	"log/slog"
-	"testing"
-
-	"golang.org/x/net/ipv4"
-)
-
-func TestInnerECN(t *testing.T) {
-	cases := []struct {
-		name string
-		pkt  []byte
-		want byte
-	}{
-		{"empty", nil, 0},
-		{"v4_NotECT", v4WithToS(0x00), 0x00},
-		{"v4_ECT0", v4WithToS(0x02), 0x02},
-		{"v4_ECT1", v4WithToS(0x01), 0x01},
-		{"v4_CE", v4WithToS(0x03), 0x03},
-		{"v4_DSCP_then_NotECT", v4WithToS(0x88 | 0x00), 0x00},
-		{"v4_DSCP_then_CE", v4WithToS(0x88 | 0x03), 0x03},
-		{"v6_NotECT", v6WithTC(0x00), 0x00},
-		{"v6_ECT0", v6WithTC(0x02), 0x02},
-		{"v6_CE", v6WithTC(0x03), 0x03},
-		{"v6_DSCP_then_CE", v6WithTC(0x88 | 0x03), 0x03},
-		{"unknown_version", []byte{0xa5, 0xff}, 0},
-	}
-	for _, c := range cases {
-		t.Run(c.name, func(t *testing.T) {
-			got := innerECN(c.pkt)
-			if got != c.want {
-				t.Errorf("innerECN=0x%02x want 0x%02x", got, c.want)
-			}
-		})
-	}
-}
-
-// v4WithToS returns a 2-byte slice tall enough for innerECN: byte 0 carries
-// version=4 in the high nibble, byte 1 is the full ToS so we exercise both
-// the DSCP and ECN portions through the byte 1 mask.
-func v4WithToS(tos byte) []byte {
-	return []byte{0x45, tos}
-}
-
-// v6WithTC builds a 2-byte slice that places a known traffic class value
-// across bytes 0 (high nibble of TC) and 1 (low nibble of TC). innerECN
-// extracts ECN as (b[1]>>4)&0x03, which corresponds to TC[1:0].
-func v6WithTC(tc byte) []byte {
-	return []byte{0x60 | (tc>>4)&0x0f, (tc & 0x0f) << 4}
-}
-
-func TestApplyOuterECN(t *testing.T) {
-	silent := slog.New(slog.DiscardHandler)
-	hi := &HostInfo{}
-
-	// Build a v4 packet helper with a given inner ECN field.
-	v4 := func(innerECN byte) []byte {
-		// 20-byte minimal IPv4 header with ToS = innerECN (DSCP zeroed).
-		return []byte{
-			0x45, innerECN, 0, 28,
-			0, 0, 0x40, 0,
-			64, 6, 0, 0,
-			10, 0, 0, 1,
-			10, 0, 0, 2,
-		}
-	}
-	// Build a v6 packet helper with a given inner ECN field. ECN occupies
-	// TC[1:0] which sit at byte 1 mask 0x30.
-	v6 := func(innerECN byte) []byte {
-		// 40-byte minimal IPv6 header with TC[1:0] = innerECN.
-		pkt := make([]byte, 40)
-		pkt[0] = 0x60                   // version=6, TC[7:4]=0
-		pkt[1] = (innerECN & 0x03) << 4 // TC[3:0]: low 2 bits = ECN, top 2 = DSCP-low (0)
-		return pkt
-	}
-
-	type cell struct {
-		outer    byte
-		inner    byte
-		wantECN  byte
-		wantSame bool // expect inner unchanged (true => verify the byte didn't move)
-	}
-
-	// RFC 6040 normal-mode combine table. Only outer==CE causes mutation.
-	table := []cell{
-		{ecnNotECT, ecnNotECT, ecnNotECT, true},
-		{ecnNotECT, ecnECT0, ecnECT0, true},
-		{ecnNotECT, ecnECT1, ecnECT1, true},
-		{ecnNotECT, ecnCE, ecnCE, true},
-
-		{ecnECT0, ecnNotECT, ecnNotECT, true},
-		{ecnECT0, ecnECT0, ecnECT0, true},
-		{ecnECT0, ecnECT1, ecnECT1, true},
-		{ecnECT0, ecnCE, ecnCE, true},
-
-		{ecnECT1, ecnNotECT, ecnNotECT, true},
-		{ecnECT1, ecnECT0, ecnECT0, true},
-		{ecnECT1, ecnECT1, ecnECT1, true},
-		{ecnECT1, ecnCE, ecnCE, true},
-
-		{ecnCE, ecnNotECT, ecnNotECT, true}, // legacy: log, leave alone
-		{ecnCE, ecnECT0, ecnCE, false},      // CE folded in
-		{ecnCE, ecnECT1, ecnCE, false},
-		{ecnCE, ecnCE, ecnCE, true},
-	}
-
-	for _, c := range table {
-		t.Run("v4", func(t *testing.T) {
-			pkt := v4(c.inner)
-			applyOuterECN(pkt, c.outer, hi, silent)
-			got := pkt[1] & 0x03
-			if got != c.wantECN {
-				t.Errorf("v4 outer=0x%02x inner=0x%02x: got 0x%02x want 0x%02x", c.outer, c.inner, got, c.wantECN)
-			}
-		})
-		t.Run("v6", func(t *testing.T) {
-			pkt := v6(c.inner)
-			applyOuterECN(pkt, c.outer, hi, silent)
-			got := (pkt[1] >> 4) & 0x03
-			if got != c.wantECN {
-				t.Errorf("v6 outer=0x%02x inner=0x%02x: got 0x%02x want 0x%02x", c.outer, c.inner, got, c.wantECN)
-			}
-		})
-	}
-}
-
-// TestApplyOuterECN_IPv4ChecksumStaysValid guards against H1: folding an outer
-// CE mark into the inner IPv4 ToS byte must keep the IPv4 header checksum valid.
-// The passthrough emit paths write the packet verbatim, so a stale checksum
-// turns an underlay congestion mark into packet loss at the receiver.
-func TestApplyOuterECN_IPv4ChecksumStaysValid(t *testing.T) {
-	silent := slog.New(slog.DiscardHandler)
-	hi := &HostInfo{}
-
-	// 20-byte IPv4 header with DSCP=0x88 and inner ECN = ECT(0). Folding CE
-	// flips only the low two bits of the ToS byte while leaving DSCP intact.
-	pkt := []byte{
-		0x45, 0x88 | ecnECT0, 0, 40,
-		0x1c, 0x46, 0x40, 0x00,
-		64, 6, 0, 0,
-		10, 0, 0, 1,
-		10, 0, 0, 2,
-	}
-	// Stamp a correct header checksum before the fold.
-	binary.BigEndian.PutUint16(pkt[10:12], ipv4HeaderChecksum(pkt[:ipv4.HeaderLen]))
-	if !ipv4HeaderChecksumValid(pkt[:ipv4.HeaderLen]) {
-		t.Fatal("test setup: initial header checksum invalid")
-	}
-
-	applyOuterECN(pkt, ecnCE, hi, silent)
-
-	// CE folded in, DSCP preserved.
-	if got, want := pkt[1], byte(0x88|ecnCE); got != want {
-		t.Fatalf("ToS after fold = 0x%02x, want 0x%02x", got, want)
-	}
-	// The incremental RFC 1624 update must leave the checksum valid and equal
-	// to a full recompute over the mutated header.
-	if !ipv4HeaderChecksumValid(pkt[:ipv4.HeaderLen]) {
-		t.Fatalf("IPv4 header checksum invalid after CE fold: 0x%04x", binary.BigEndian.Uint16(pkt[10:12]))
-	}
-	if got, want := binary.BigEndian.Uint16(pkt[10:12]), ipv4HeaderChecksum(pkt[:ipv4.HeaderLen]); got != want {
-		t.Fatalf("checksum = 0x%04x, full recompute = 0x%04x", got, want)
-	}
-}
-
-// ipv4HeaderChecksum computes the RFC 1071 IPv4 header checksum over hdr,
-// treating the checksum field (bytes 10:12) as zero.
-func ipv4HeaderChecksum(hdr []byte) uint16 {
-	var sum uint32
-	for i := 0; i+1 < len(hdr); i += 2 {
-		if i == 10 {
-			continue // checksum field
-		}
-		sum += uint32(hdr[i])<<8 | uint32(hdr[i+1])
-	}
-	for sum > 0xffff {
-		sum = (sum >> 16) + (sum & 0xffff)
-	}
-	return ^uint16(sum)
-}
-
-// ipv4HeaderChecksumValid reports whether the stored checksum matches a fresh
-// computation over the header.
-func ipv4HeaderChecksumValid(hdr []byte) bool {
-	return binary.BigEndian.Uint16(hdr[10:12]) == ipv4HeaderChecksum(hdr)
-}

examples/config.yml

@@ -146,6 +146,14 @@ listen:
   # Default true; set to false to leave WDF in charge of inbound decisions on the listener port. Not reloadable.
   #windows_bypass_wdf: true
 
+  # On macOS only
+  # macOS scopes the udp socket to the interface it was created on, so moving between networks (wifi to wired,
+  # office to home) leaves Nebula sending out an interface that no longer has a route. When true, Nebula watches
+  # the routing socket and rebinds the listener once the change settles.
+  # iOS does not use this, the host app drives the same rebind itself.
+  # Default true. Not reloadable.
+  #rebind_on_network_change: true
+
   # By default, Nebula replies to packets it has no tunnel for with a "recv_error" packet. This packet helps speed up reconnection
   # in the case that Nebula on either side did not shut down cleanly. This response can be abused as a way to discover if Nebula is running
   # on a host though. This option lets you configure if you want to send "recv_error" packets always, never, or only to private network remotes.
@@ -254,28 +262,6 @@ 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
-  # batched 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.
-  #
-  # When cpu_affinity is unset, nebula picks CPUs that do NOT service any physical NIC's interrupts (read from
-  # /sys/class/net/*/device/msi_irqs and /proc/irq/*/effective_affinity_list): an encrypt thread pinned onto a core
-  # that also runs NAPI for a NIC RX queue fights the softirq for the core and collapses throughput for flows hashed
-  # to that queue. If the NIC's vectors blanket every allowed CPU (many drivers default to one queue per core) the
-  # avoidance logs and falls back to the old spread; narrow the NIC's queue/IRQ spread (e.g. `ethtool -X <dev>
-  # equal N`) or set cpu_affinity explicitly to benefit.
-  #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. Setting this disables the automatic NIC-IRQ avoidance described under pin_threads — prefer CPUs that don't
-  # service your underlay NIC's RX queue IRQs. 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
@@ -412,16 +398,6 @@ logging:
   # This setting is reloadable
   #inactivity_timeout: 10m
 
-  # ecn (default true) propagates ECN (Explicit Congestion Notification) across the tunnel per RFC 6040: the inner
-  # packet's ECN codepoint is copied onto the outer carrier header on encapsulation, and an outer CE ("congestion
-  # experienced") mark is folded back into the inner header on decapsulation. On linux it additionally stamps
-  # RTAX_FEATURE_ECN on the routes nebula installs, so the kernel actively negotiates ECN for connections to mesh
-  # prefixes. Disable this only when an underlay middlebox mangles or clears ECN bits unpredictably.
-  # This setting is reloadable, BUT flipping it at runtime only updates the datapath (the inner<->outer copy/combine).
-  # The RTAX_FEATURE_ECN flag on already-installed routes is NOT revisited on reload, so nebula must be restarted for
-  # the route half of this setting to take effect.
-  #ecn: true
-
 # Nebula security group configuration
 firewall:
   # Action to take when a packet is not allowed by the firewall rules.

examples/service_scripts/nebula.service

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

firewall/cache.go

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

firewall/cache_test.go

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

go.mod

@@ -1,6 +1,6 @@
 module github.com/slackhq/nebula
 
-go 1.25.0
+go 1.26.0
 
 require (
 	dario.cat/mergo v1.0.2
@@ -24,12 +24,12 @@ require (
 	github.com/vishvananda/netlink v1.3.1
 	go.uber.org/goleak v1.3.0
 	go.yaml.in/yaml/v3 v3.0.4
-	golang.org/x/crypto v0.53.0
+	golang.org/x/crypto v0.54.0
 	golang.org/x/exp v0.0.0-20230725093048-515e97ebf090
-	golang.org/x/net v0.56.0
-	golang.org/x/sync v0.21.0
-	golang.org/x/sys v0.46.0
-	golang.org/x/term v0.44.0
+	golang.org/x/net v0.57.0
+	golang.org/x/sync v0.22.0
+	golang.org/x/sys v0.47.0
+	golang.org/x/term v0.45.0
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2
 	golang.zx2c4.com/wireguard v0.0.0-20230325221338-052af4a8072b
 	golang.zx2c4.com/wireguard/windows v1.0.1

go.sum

@@ -162,8 +162,8 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
 golang.org/x/crypto v0.0.0-20210322153248-0c34fe9e7dc2/go.mod h1:T9bdIzuCu7OtxOm1hfPfRQxPLYneinmdGuTeoZ9dtd4=
-golang.org/x/crypto v0.53.0 h1:QZ4Muo8THX6CizN2vPPd5fBGHyogrdK9fG4wLPFUsto=
-golang.org/x/crypto v0.53.0/go.mod h1:DNLU434OwVakk9PzuwV8w62mAJpRJL3vsgcfp4Qnsio=
+golang.org/x/crypto v0.54.0 h1:YLIA59K4fiNzHzjnZt2tUJQjQtUWfWbeHBqKtk3eScw=
+golang.org/x/crypto v0.54.0/go.mod h1:KWL8ny2AZdGR2cWmzeHrp2azQPGogOv+HeQaVEXC2dk=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090 h1:Di6/M8l0O2lCLc6VVRWhgCiApHV8MnQurBnFSHsQtNY=
 golang.org/x/exp v0.0.0-20230725093048-515e97ebf090/go.mod h1:FXUEEKJgO7OQYeo8N01OfiKP8RXMtf6e8aTskBGqWdc=
 golang.org/x/lint v0.0.0-20200302205851-738671d3881b/go.mod h1:3xt1FjdF8hUf6vQPIChWIBhFzV8gjjsPE/fR3IyQdNY=
@@ -182,8 +182,8 @@ golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLL
 golang.org/x/net v0.0.0-20200625001655-4c5254603344/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA=
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
-golang.org/x/net v0.56.0 h1:Rw8j/hFzGvJUZwNBXnAtf5sVDVt+65SK2C7IxCxZt5o=
-golang.org/x/net v0.56.0/go.mod h1:D3Ku6r+V6JROoZK144D2XfMHFcMq/0zSfLelVTCFKec=
+golang.org/x/net v0.57.0 h1:K5+3DljvIuDG9/Jv9rvyMywYNFCQ9RSUY6OOTTkT+tE=
+golang.org/x/net v0.57.0/go.mod h1:KpXc8iv+r3XplLAG/f7Jsf9RPszJzdR0f58q9vGOuEU=
 golang.org/x/oauth2 v0.0.0-20190226205417-e64efc72b421/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20181221193216-37e7f081c4d4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -191,8 +191,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201207232520-09787c993a3a/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.0.0-20180905080454-ebe1bf3edb33/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20181116152217-5ac8a444bdc5/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
@@ -208,11 +208,11 @@ golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7w
 golang.org/x/sys v0.0.0-20210603081109-ebe580a85c40/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.2.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.10.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=
-golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=
+golang.org/x/term v0.45.0 h1:NwWyBmoJCbfTHpxrWoZ9C6/VxOf7ic219I8xZZFdrf0=
+golang.org/x/term v0.45.0/go.mod h1:9aqxs0blBcrm/n0L9QW0aRVD+ktan8ssZromtqJC43w=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.2/go.mod h1:bEr9sfX3Q8Zfm5fL9x+3itogRgK3+ptLWKqgva+5dAk=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=

handshake_manager.go

@@ -295,7 +295,13 @@ func (hm *HandshakeManager) handleOutbound(vpnIp netip.Addr, lighthouseTriggered
 		hm.messageMetrics.Tx(header.Handshake, hh.machine.Subtype(), 1)
 		err := hm.outside.WriteTo(stage0, addr)
 		if err != nil {
-			hostinfo.logger(hm.l).Error("Failed to send handshake message",
+			// These repeat every attempt, so match the success log below and only shout when the remotes changed
+			level := slog.LevelDebug
+			if remotesHaveChanged {
+				level = slog.LevelError
+			}
+
+			hostinfo.logger(hm.l).Log(context.Background(), level, "Failed to send handshake message",
 				"udpAddr", addr,
 				"initiatorIndex", hostinfo.localIndexId,
 				"handshake", hsFields,
@@ -529,7 +535,9 @@ func (hm *HandshakeManager) DeleteHostInfo(hostinfo *HostInfo) {
 
 func (hm *HandshakeManager) unlockedDeleteHostInfo(hostinfo *HostInfo) {
 	for _, addr := range hostinfo.vpnAddrs {
-		delete(hm.vpnIps, addr)
+		if cur, ok := hm.vpnIps[addr]; ok && cur.hostinfo == hostinfo {
+			delete(hm.vpnIps, addr)
+		}
 	}
 
 	if len(hm.vpnIps) == 0 {
@@ -967,7 +975,6 @@ func (hm *HandshakeManager) continueHandshake(via ViaSender, hh *HandshakeHostIn
 		nb := make([]byte, 12, 12)
 		out := make([]byte, mtu)
 		for _, cp := range hh.packetStore {
-			//todo use a sendbatcher
 			cp.callback(cp.messageType, cp.messageSubType, hostinfo, cp.packet, nb, out)
 		}
 		f.cachedPacketMetrics.sent.Inc(int64(len(hh.packetStore)))
@@ -1078,7 +1085,7 @@ func (hm *HandshakeManager) sendHandshakeResponse(via ViaSender, msg []byte, hos
 		hostinfo.relayState.InsertRelayTo(via.relayHI.vpnAddrs[0])
 		// We received a valid handshake on this relay, so make sure the relay
 		// state reflects that, in case it had been marked Disestablished.
-		via.relayHI.relayState.UpdateRelayForByIdxState(via.remoteIdx, Established)
+		via.relayHI.relayState.UpdateRelayForByIdxState(via.relay.LocalIndex, Established)
 		f.SendVia(via.relayHI, via.relay, msg, make([]byte, 12), make([]byte, mtu), false)
 		f.l.Info("Handshake message sent", append(logFields, "relay", via.relayHI.vpnAddrs[0])...)
 	}

hostmap.go

@@ -287,7 +287,6 @@ type HostInfo struct {
 type ViaSender struct {
 	UdpAddr   netip.AddrPort
 	relayHI   *HostInfo // relayHI is the host info object of the relay
-	remoteIdx uint32    // remoteIdx is the index included in the header of the received packet
 	relay     *Relay    // relay contains the rest of the relay information, including the PeerIP of the host trying to communicate with us.
 	IsRelayed bool      // IsRelayed is true if the packet was sent through a relay
 }

inside.go

@@ -2,7 +2,6 @@ package nebula
 
 import (
 	"context"
-	"io"
 	"log/slog"
 	"net/netip"
 
@@ -10,24 +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/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
-func (f *Interface) consumeInsidePacket(pkt tio.Packet, fwPacket *firewall.Packet, nb []byte, sendBatch batch.TxBatcher, rejectBuf []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. Every consumer below
-	// (parse, self-forward, handshake cache, sendInsideMessage) reads it
-	// synchronously; do not retain pkt outside this call. If a future
-	// caller needs to keep the packet, use pkt.Clone() to detach it from
-	// the borrow.
-	//
-	// pkt.Bytes is either one IP datagram (GSO zero) or a TSO/USO
-	// superpacket. In both cases the L3+L4 headers at the start describe
-	// the same 5-tuple every segment will share, so a single newPacket /
-	// firewall check covers the whole superpacket.
-	packet := pkt.Bytes
+func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet, nb, out []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(packet, false, fwPacket)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
@@ -52,14 +37,7 @@ func (f *Interface) consumeInsidePacket(pkt tio.Packet, fwPacket *firewall.Packe
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
 		// TUN device.
 		if immediatelyForwardToSelf {
-			// Write copies into the kernel queue synchronously, so seg's lifetime ends at return.
-			// A self-forwarded superpacket would be re-handed to the
-			// kernel as one giant blob; segment first so the loopback
-			// path sees one IP datagram per Write.
-			err := tio.SegmentSuperpacket(pkt, func(seg []byte) error {
-				_, werr := f.queues[q].Write(seg)
-				return werr
-			})
+			_, err := f.readers[q].Write(packet)
 			if err != nil {
 				f.l.Error("Failed to forward to tun", "error", err)
 			}
@@ -75,23 +53,11 @@ func (f *Interface) consumeInsidePacket(pkt tio.Packet, fwPacket *firewall.Packe
 	}
 
 	hostinfo, ready := f.getOrHandshakeConsiderRouting(fwPacket, func(hh *HandshakeHostInfo) {
-		// borrowed: SegmentSuperpacket builds each segment in the kernel-supplied pkt
-		// bytes underneath. cachePacket explicitly copies its argument (handshake_manager.go cachePacket),
-		// so retaining segments past the loop is safe.
-		err := tio.SegmentSuperpacket(pkt, 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,
-			)
-		}
+		hh.cachePacket(f.l, header.Message, 0, packet, f.sendMessageNow, f.cachedPacketMetrics)
 	})
 
 	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,
@@ -107,9 +73,10 @@ func (f *Interface) consumeInsidePacket(pkt tio.Packet, fwPacket *firewall.Packe
 
 	dropReason := f.firewall.Drop(*fwPacket, false, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason == nil {
-		f.sendInsideMessage(hostinfo, pkt, nb, sendBatch, rejectBuf, q)
+		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,
@@ -119,150 +86,6 @@ func (f *Interface) consumeInsidePacket(pkt tio.Packet, fwPacket *firewall.Packe
 	}
 }
 
-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.GetRemote(),
-			"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: SegmentSuperpacket 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 tio.Packet, nb []byte, sendBatch batch.TxBatcher, rejectBuf []byte, q int) {
-	ci := hostinfo.ConnectionState
-	if ci.eKey == nil {
-		return
-	}
-
-	remote := hostinfo.GetRemote()
-	ecnEnabled := f.ecnEnabled.Load()
-	if hostinfo.lastRebindCount != f.rebindCount {
-		//NOTE: there is an update hole if a tunnel isn't used and exactly 256 rebinds occur before the tunnel is
-		// finally used again. This tunnel would eventually be torn down and recreated if this action didn't help.
-		f.lightHouse.QueryServer(hostinfo.vpnAddrs[0])
-		hostinfo.lastRebindCount = f.rebindCount
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("Lighthouse update triggered for punch due to rebind counter",
-				"vpnAddrs", hostinfo.vpnAddrs,
-			)
-		}
-	}
-
-	if !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 = tio.SegmentSuperpacket(pkt, func(seg []byte) error {
-			//relay header + header + plaintext + AEAD tag (16 bytes for both AES-GCM and ChaCha20-Poly1305) + relay tag
-			scratch := sendBatch.Reserve(header.Len + header.Len + len(seg) + 16 + 16)
-
-			innerPacket := f.sendInsideEncrypt(hostinfo, ci, seg, scratch[header.Len:], nb)
-			if innerPacket == nil {
-				return nil
-			}
-
-			//now we need to do a relay-encrypt:
-			toSend, err := f.prepareSendVia(relayHostInfo, relay, innerPacket, nb, scratch, true)
-			if err != nil {
-				//already logged
-				return nil
-			}
-
-			var ecn byte
-			if ecnEnabled {
-				ecn = innerECN(seg)
-			}
-			sendBatch.Commit(toSend, relayHostInfo.GetRemote(), ecn)
-			return nil
-		})
-		if err != nil {
-			hostinfo.logger(f.l).Error("Failed to segment superpacket for relay send", "error", err)
-		}
-		return
-	}
-
-	err := tio.SegmentSuperpacket(pkt, func(seg []byte) error {
-		// header + plaintext + AEAD tag (16 bytes for both AES-GCM and ChaCha20-Poly1305)
-		scratch := sendBatch.Reserve(header.Len + len(seg) + 16)
-
-		out := f.sendInsideEncrypt(hostinfo, ci, seg, scratch, nb)
-		if out == nil {
-			return nil
-		}
-
-		var ecn byte
-		if ecnEnabled {
-			ecn = innerECN(seg)
-		}
-		sendBatch.Commit(out, remote, ecn)
-		return nil
-	})
-	if err != nil {
-		hostinfo.logger(f.l).Error("Failed to segment superpacket for send",
-			"error", err,
-		)
-	}
-}
-
-// innerECN returns the 2-bit IP-level ECN codepoint of an inner IPv4 or IPv6
-// packet, or 0 if pkt is too short or its IP version is unrecognized. Used at
-// encap to copy the inner codepoint onto the outer carrier per RFC 6040.
-func innerECN(pkt []byte) byte {
-	if len(pkt) < 2 {
-		return 0
-	}
-	switch pkt[0] >> 4 {
-	case 4:
-		return pkt[1] & 0x03
-	case 6:
-		return (pkt[1] >> 4) & 0x03
-	}
-	return 0
-}
-
 func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 	if !f.firewall.OutboundSendReject {
 		return
@@ -273,7 +96,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 		return
 	}
 
-	_, err := f.queues[q].Write(out)
+	_, err := f.readers[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}
@@ -452,13 +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()
@@ -480,7 +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.
@@ -500,37 +331,13 @@ 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 {
-		// already logged by prepareSendVia
 		return
 	}
-
-	err = f.writers[0].WriteTo(toSend, via.GetRemote())
+	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) {
@@ -601,7 +408,7 @@ func (f *Interface) sendNoMetrics(t header.MessageType, st header.MessageSubType
 		if err != nil {
 			hostinfo.logger(f.l).Error("Failed to write outgoing packet",
 				"error", err,
-				"udpAddr", remote,
+				"udpAddr", hr,
 			)
 		}
 	} else {

interface.go

@@ -2,11 +2,12 @@ package nebula
 
 import (
 	"context"
+	"crypto/fips140"
 	"errors"
 	"fmt"
+	"io"
 	"log/slog"
 	"net/netip"
-	"runtime"
 	"slices"
 	"sync"
 	"sync/atomic"
@@ -14,15 +15,12 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/rcrowley/go-metrics"
-	"github.com/slackhq/nebula/util"
 
 	"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"
 )
 
@@ -52,19 +50,7 @@ type InterfaceConfig struct {
 	reQueryWait     time.Duration
 
 	ConntrackCacheTimeout time.Duration
-
-	// CpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
-	// should pin to. Queue i pins to CpuAffinity[i % len(CpuAffinity)] —
-	// shorter lists than `routines` cycle. Empty list keeps the default
-	// pin-to-(i % NumCPU) behavior. 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 sendmmsg on one XPS-selected NIC TX ring so per-flow
-	// packets stay ordered on the wire.
-	PinThreads bool
-
-	l *slog.Logger
+	l                     *slog.Logger
 }
 
 type Interface struct {
@@ -88,21 +74,7 @@ type Interface struct {
 	routines              int
 	disconnectInvalid     atomic.Bool
 	closed                atomic.Bool
-	// cpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
-	// should pin to. Queue i pins to cpuAffinity[i % len(cpuAffinity)].
-	// Empty falls back to the default pin-to-(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
-	// ecnEnabled gates RFC 6040 underlay ECN propagation. When true,
-	// inside.go copies the inner ECN onto the outer carrier on encap and
-	// decryptToTun folds outer CE into the inner header on decap. Toggle
-	// via tunnels.ecn (default true).
-	ecnEnabled   atomic.Bool
-	relayManager *relayManager
+	relayManager          *relayManager
 
 	tryPromoteEvery atomic.Uint32
 	reQueryEvery    atomic.Uint32
@@ -119,12 +91,8 @@ type Interface struct {
 
 	ctx     context.Context
 	writers []udp.Conn
-	queues  []tio.Queue
-	// batchers is one per tun queue, wrapping queues[i].
-	// decryptToTun sends plaintext into the batch.RxBatcher;
-	// listenOut calls its Flush at the end of each UDP recvmmsg batch.
-	batchers []batch.RxBatcher
-	wg       sync.WaitGroup
+	readers []io.ReadWriteCloser
+	wg      sync.WaitGroup
 
 	// fatalErr holds the first unexpected reader error that caused shutdown.
 	// nil means "no fatal error" (yet)
@@ -222,7 +190,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		routines:              c.routines,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		batchers:              make([]batch.RxBatcher, c.routines),
+		readers:               make([]io.ReadWriteCloser, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -231,8 +199,6 @@ 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,
@@ -273,42 +239,30 @@ func (f *Interface) activate() error {
 		"build", f.version,
 		"udpAddr", addr,
 		"boringcrypto", boringEnabled(),
+		"fips140Version", fips140.Version(),
+		"fips140Enabled", fips140.Enabled(),
+		"fips140Enforced", fips140.Enforced(),
 	)
 
-	if f.routines > 1 && !f.outside.SupportsMultipleReaders() {
-		f.routines = 1
-		f.l.Warn("multiple udp readers are not supported on this platform, falling back to a single routine")
+	if f.routines > 1 {
+		if !f.inside.SupportsMultiqueue() || !f.outside.SupportsMultipleReaders() {
+			f.routines = 1
+			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))
 
-	for i := range f.queues {
-		caps := tio.QueueCapabilities(f.queues[i])
-		if caps.TSO || caps.USO {
-			// Multi-lane: TCP gets coalesced when TSO is on, UDP when USO
-			// is on, everything else (and either lane disabled) falls
-			// through to passthrough so non-IP / non-TCP-UDP traffic still
-			// reaches the TUN.
-			arena := batch.NewArena(batch.DefaultMultiArenaCap)
-			f.batchers[i] = batch.NewMultiCoalescer(f.queues[i], f.l, arena, caps.TSO, caps.USO)
-		} else {
-			arena := batch.NewArena(batch.DefaultPassthroughArenaCap)
-			f.batchers[i] = batch.NewPassthrough(f.queues[i], arena.Reserve, arena.Reset)
+	// Prepare n tun queues
+	var reader io.ReadWriteCloser = f.inside
+	for i := 0; i < f.routines; i++ {
+		if i > 0 {
+			reader, err = f.inside.NewMultiQueueReader()
+			if err != nil {
+				return err
+			}
 		}
+		f.readers[i] = reader
 	}
 
 	// On error the caller owns the cleanup, Control.Start cancels the service context
@@ -331,7 +285,7 @@ func (f *Interface) run() {
 	// Launch n queues to read packets from tun dev
 	for i := 0; i < f.routines; i++ {
 		f.wg.Go(func() {
-			f.listenIn(f.queues[i], i)
+			f.listenIn(f.readers[i], i)
 		})
 	}
 
@@ -366,22 +320,14 @@ 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) {
-		plaintext := f.batchers[i].Reserve(len(payload))
-		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get(), meta)
-	}
-
-	flusher := func() {
-		if err := f.batchers[i].Flush(); err != nil {
-			f.l.Error("Failed to flush tun coalescer", "error", err)
-		}
-	}
-
-	err := li.ListenOut(listener, flusher)
+	err := li.ListenOut(func(fromUdpAddr netip.AddrPort, payload []byte) {
+		f.readOutsidePackets(ViaSender{UdpAddr: fromUdpAddr}, plaintext[:0], payload, h, fwPacket, lhh, nb, i, ctCache.Get())
+	})
 
 	// An error after teardown began is shutdown noise, the closed flag covers resources
 	// Close releases itself and the cancelled ctx covers ones torn down by their owners
@@ -394,38 +340,16 @@ func (f *Interface) listenOut(i int) {
 	f.l.Debug("underlay reader is done", "reader", i)
 }
 
-func (f *Interface) listenIn(queue tio.Queue, i int) {
-	// Pinning this thread (and goroutine) to a single CPU keeps every sendmmsg from this goroutine going through the
-	// same TX ring on the nic, so the wire sees per-flow order. Skip entirely when tun.pin_threads is false.
-	if f.pinThreads {
-		var cpu int
-		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)
-		}
-	}
-
-	rejectBuf := make([]byte, mtu)
-	arenaSize := batch.SendBatchCap * (udp.MTU + 32)
-	sb := batch.NewSendBatch(f.writers[i], batch.SendBatchCap, arenaSize)
+func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
+	packet := make([]byte, mtu)
+	out := make([]byte, mtu)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
 
 	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 
 	for {
-		pkts, err := queue.Read()
+		n, err := reader.Read(packet)
 		if err != nil {
 			// Same shutdown noise handling as listenOut
 			if !f.closed.Load() && f.ctx.Err() == nil {
@@ -435,20 +359,7 @@ func (f *Interface) listenIn(queue tio.Queue, i int) {
 			break
 		}
 
-		for _, pkt := range pkts {
-			f.consumeInsidePacket(pkt, fwPacket, nb, sb, rejectBuf, i, conntrackCache.Get())
-			// Flush incrementally once a full sendmmsg batch has
-			// accumulated so the first packets of a deep read drain
-			// hit the wire while the rest are still being encrypted.
-			if sb.Len() >= batch.SendBatchCap {
-				if err := sb.Flush(); err != nil {
-					f.l.Error("Failed to write outgoing batch", "error", err, "writer", i)
-				}
-			}
-		}
-		if err := sb.Flush(); err != nil {
-			f.l.Error("Failed to write outgoing batch", "error", err, "writer", i)
-		}
+		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get())
 	}
 
 	f.l.Debug("overlay reader is done", "reader", i)
@@ -460,7 +371,6 @@ func (f *Interface) RegisterConfigChangeCallbacks(c *config.C) {
 	c.RegisterReloadCallback(f.reloadAcceptRecvError)
 	c.RegisterReloadCallback(f.reloadDisconnectInvalid)
 	c.RegisterReloadCallback(f.reloadMisc)
-	c.RegisterReloadCallback(f.reloadEcn)
 
 	for _, udpConn := range f.writers {
 		c.RegisterReloadCallback(udpConn.ReloadConfig)
@@ -593,23 +503,6 @@ func (f *Interface) reloadMisc(c *config.C) {
 	}
 }
 
-// reloadEcn syncs Interface.ecnEnabled with the tunnels.ecn config knob.
-// Default is enabled (RFC 6040 normal mode); set false on the rare path
-// where an underlay middlebox rewrites or drops ECN bits unpredictably.
-func (f *Interface) reloadEcn(c *config.C) {
-	initial := c.InitialLoad()
-	if initial || c.HasChanged("tunnels.ecn") {
-		v := c.GetBool("tunnels.ecn", true)
-		changed := f.ecnEnabled.Swap(v) != v
-		if !initial {
-			f.l.Info("tunnels.ecn changed", "enabled", v)
-			if changed {
-				f.l.Warn("tunnels.ecn datapath toggled, but route-level ECN negotiation (RTAX_FEATURE_ECN) retains its previous state until nebula is restarted", "enabled", v)
-			}
-		}
-	}
-}
-
 func (f *Interface) emitStats(ctx context.Context, i time.Duration) {
 	ticker := time.NewTicker(i)
 	defer ticker.Stop()

iputil/packet_test.go

@@ -1,7 +1,6 @@
 package iputil
 
 import (
-	"bytes"
 	"encoding/binary"
 	"net"
 	"testing"
@@ -180,46 +179,6 @@ func Test_CreateRejectPacket_NoICMPError(t *testing.T) {
 	}
 }
 
-// Test_CreateRejectPacket_RespectsCap ensures it is impossible for
-// an oversized ICMPv6 reject to overwrite the neighbor segment's bytes.
-func Test_CreateRejectPacket_RespectsCap(t *testing.T) {
-	src := net.ParseIP("fd00::1")
-	dst := net.ParseIP("fd00::2")
-
-	// Inner IPv6 UDP packet. An ICMPv6 reject copies the whole inner packet
-	// plus a 48-byte header (40 IPv6 + 8 ICMPv6), so it needs 48 more bytes
-	// than the inner packet length.
-	inner := makeIPv6Packet(src, dst, 17, make([]byte, 20))
-
-	// The ciphertext scratch reused as the reject buffer is the received
-	// datagram: 16-byte Nebula header + inner + 16-byte AEAD tag. That is only
-	// 32 bytes of slack, so a full ICMPv6 reject overruns it by 16 bytes.
-	const nebulaOverhead = 32
-	segLen := len(inner) + nebulaOverhead
-
-	// Shared backing row laid out as [segment][neighbor's 16-byte Nebula header].
-	const neighborHdr = 16
-	sentinel := bytes.Repeat([]byte{0xAB}, neighborHdr)
-
-	// Uncapped: the slice's capacity reaches into the neighbor, reproducing
-	// the overrun that silently drops the neighbor packet.
-	backing := make([]byte, segLen+neighborHdr)
-	copy(backing[segLen:], sentinel)
-	reject := CreateRejectPacket(inner, backing[:segLen])
-	assert.NotNil(t, reject, "uncapped buffer reaches into the neighbor, so the reject is built")
-	assert.NotEqual(t, sentinel, backing[segLen:segLen+neighborHdr],
-		"without the cap the oversized reject overruns into the neighbor segment")
-
-	// Capped (the fix): cap==len, so the builder cannot exceed the segment. The
-	// reject does not fit, so it is refused rather than corrupting the neighbor.
-	backing = make([]byte, segLen+neighborHdr)
-	copy(backing[segLen:], sentinel)
-	reject = CreateRejectPacket(inner, backing[:segLen:segLen])
-	assert.Nil(t, reject, "capped segment is 16 bytes too small for a full ICMPv6 reject, so it is refused")
-	assert.Equal(t, sentinel, backing[segLen:segLen+neighborHdr],
-		"capped segment must leave the neighbor untouched")
-}
-
 func makeIPv6Packet(src, dst net.IP, nextHeader uint8, payload []byte) []byte {
 	b := make([]byte, ipv6.HeaderLen+len(payload))
 	b[0] = ipv6.Version << 4

lighthouse.go

@@ -36,6 +36,10 @@ type LightHouse struct {
 	myVpnNetworksTable *bart.Lite
 	punchy             *Punchy
 
+	// localAddrsFn enumerates the underlay addresses we advertise. It is a field so tests can supply simulated
+	// addresses rather than whatever this machine's NICs happen to be. Set it before Start.
+	localAddrsFn func(*LocalAllowList) []netip.Addr
+
 	// Local cache of answers from light houses
 	// map of vpn addr to answers
 	addrMap map[netip.Addr]*RemoteList
@@ -107,6 +111,10 @@ func NewLightHouseFromConfig(ctx context.Context, l *slog.Logger, c *config.C, c
 		queryChan:          make(chan netip.Addr, c.GetUint32("handshakes.query_buffer", 64)),
 		l:                  l,
 	}
+	h.localAddrsFn = func(al *LocalAllowList) []netip.Addr {
+		return localAddrs(h.l, al)
+	}
+
 	lighthouses := make([]netip.Addr, 0)
 	h.lighthouses.Store(&lighthouses)
 	staticList := make(map[netip.Addr]struct{})
@@ -918,7 +926,7 @@ func (lh *LightHouse) SendUpdate() {
 	}
 
 	lal := lh.GetLocalAllowList()
-	for _, e := range localAddrs(lh.l, lal) {
+	for _, e := range lh.localAddrsFn(lal) {
 		if lh.myVpnNetworksTable.Contains(e) {
 			continue
 		}

main.go

@@ -7,7 +7,6 @@ import (
 	"net"
 	"net/netip"
 	"runtime/debug"
-	"slices"
 	"strings"
 	"time"
 
@@ -34,9 +33,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		buildVersion = moduleVersion()
 	}
 
-	// Debug builds (-tags debug) serve pprof on :6060; a no-op otherwise.
-	startPprofServer(ctx, l)
-
 	// Print the config if in test, the exit comes later
 	if configTest {
 		b, err := yaml.Marshal(c.Settings)
@@ -214,12 +210,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		l.Warn("Failed to start DNS responder", "error", err)
 	}
 
-	pinThreads := c.GetBool("tun.pin_threads", true)
-	cpuAffinity := parseCpuAffinity(c, l, routines)
-	if pinThreads && len(cpuAffinity) == 0 && !configTest {
-		cpuAffinity = defaultCPUAffinityAvoidingIRQs(l, routines)
-	}
-
 	ifConfig := &InterfaceConfig{
 		HostMap:               hostMap,
 		Inside:                tun,
@@ -241,8 +231,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		relayManager:          NewRelayManager(ctx, l, hostMap, c),
 		punchy:                punchy,
 		ConntrackCacheTimeout: conntrackCacheTimeout,
-		CpuAffinity:           cpuAffinity,
-		PinThreads:            pinThreads,
 		l:                     l,
 	}
 
@@ -260,7 +248,6 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		ifce.reloadDisconnectInvalid(c)
 		ifce.reloadSendRecvError(c)
 		ifce.reloadAcceptRecvError(c)
-		ifce.reloadEcn(c)
 
 		handshakeManager.f = ifce
 		go handshakeManager.Run(ctx)
@@ -281,6 +268,8 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 
 	attachCommands(l, c, ssh, ifce)
 
+	networkChanges := udp.NewNetworkChangeMonitor(ctx, l, c)
+
 	return &Control{
 		state:                  StateReady,
 		f:                      ifce,
@@ -291,125 +280,11 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		statsStart:             stats.Start,
 		dnsStart:               ds.Start,
 		lighthouseStart:        lightHouse.StartUpdateWorker,
+		networkChangeStart:     networkChanges.Start,
 		connectionManagerStart: connManager.Start,
 	}, nil
 }
 
-// parseCpuAffinity reads `tun.cpu_affinity` from the config — a list of
-// integer CPU IDs, one per TUN reader goroutine. Empty / unset returns nil
-// (listenIn falls back to 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
-}
-
-// defaultCPUAffinityAvoidingIRQs picks the default pin set for the tun
-// readers when tun.cpu_affinity is unset: allowed CPUs that do NOT service
-// any physical NIC's interrupts. The stock allowed[i] spread pins the
-// encrypt threads onto exactly the cores most drivers affine their first RX
-// queue IRQs to, so whenever a flow's RSS queue fires on a core hosting a
-// tun reader, NAPI and encrypt fight for the core and per-flow throughput
-// drops (measured: REV 8.4 vs 10.2 Gbps on the same hardware, 2026-07-14).
-//
-// Returns nil — keeping the old allowed[i] fallback in listenIn — when IRQ
-// info is unavailable or when there aren't enough IRQ-free CPUs to give
-// every routine its own core: silently doubling readers up on fewer cores
-// is worse than the occasional IRQ collision. NICs whose vectors blanket
-// every CPU (e.g. mlx5 defaults to one queue per core) make avoidance
-// impossible; narrowing the NIC's spread (ethtool -X <dev> equal N, or
-// /proc/irq/*/smp_affinity) or setting tun.cpu_affinity explicitly makes it
-// effective.
-func defaultCPUAffinityAvoidingIRQs(l *slog.Logger, routines int) []int {
-	irq, err := util.NICIRQCPUs()
-	if err != nil || len(irq) == 0 {
-		return nil
-	}
-	allowed, err := util.AllowedCPUs()
-	if err != nil {
-		return nil
-	}
-	cpus := chooseIRQFreeCPUs(allowed, irq, routines)
-	if cpus == nil {
-		l.Info("not enough CPUs are free of NIC IRQs to give every tun reader its own; using the default spread",
-			"routines", routines, "allowed", len(allowed), "irqCPUs", len(irq))
-		return nil
-	}
-	l.Info("pinning tun readers to CPUs clear of NIC IRQs", "cpus", cpus)
-	return cpus
-}
-
-// chooseIRQFreeCPUs returns the first `routines` allowed CPUs not present in
-// irq, or nil if fewer than `routines` qualify.
-func chooseIRQFreeCPUs(allowed []int, irq map[int]bool, routines int) []int {
-	free := make([]int, 0, routines)
-	for _, cpu := range allowed {
-		if irq[cpu] {
-			continue
-		}
-		free = append(free, cpu)
-		if len(free) == routines {
-			return free
-		}
-	}
-	return nil
-}
-
 func moduleVersion() string {
 	info, ok := debug.ReadBuildInfo()
 	if !ok {

main_test.go

@@ -1,71 +0,0 @@
-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 TestChooseIRQFreeCPUs(t *testing.T) {
-	irq := map[int]bool{0: true, 1: true, 2: true, 3: true}
-
-	// Plenty of IRQ-free CPUs: take the first `routines` of them in order.
-	assert.Equal(t, []int{4, 5}, chooseIRQFreeCPUs([]int{0, 1, 2, 3, 4, 5, 6}, irq, 2))
-
-	// Exactly enough.
-	assert.Equal(t, []int{4, 5, 6}, chooseIRQFreeCPUs([]int{0, 1, 2, 3, 4, 5, 6}, irq, 3))
-
-	// Not enough IRQ-free CPUs: nil, caller keeps the old default rather
-	// than doubling readers up on shared cores.
-	assert.Nil(t, chooseIRQFreeCPUs([]int{0, 1, 2, 3, 4}, irq, 2))
-
-	// No IRQ info at all behaves like a plain prefix of allowed.
-	assert.Equal(t, []int{0, 1}, chooseIRQFreeCPUs([]int{0, 1, 2}, map[int]bool{}, 2))
-
-	// Non-contiguous allowed set (cgroup cpuset) with holes.
-	assert.Equal(t, []int{9, 12}, chooseIRQFreeCPUs([]int{1, 3, 9, 12}, map[int]bool{1: true, 3: true}, 2))
-}
-
-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/boring.go

@@ -4,77 +4,16 @@
 package noiseutil
 
 import (
-	"crypto/aes"
-	"crypto/cipher"
-	"encoding/binary"
-
-	// unsafe needed for go:linkname
-	_ "unsafe"
+	"crypto/boring"
 
 	"github.com/flynn/noise"
 )
 
+var CipherAESGCM noise.CipherFunc = CipherAESGCMFIPS140
+
 // EncryptLockNeeded indicates if calls to Encrypt need a lock
 // This is true for boringcrypto because the Seal function verifies that the
 // nonce is strictly increasing.
 const EncryptLockNeeded = true
 
-// NewGCMTLS is no longer exposed in go1.19+, so we need to link it in
-// See: https://github.com/golang/go/issues/56326
-//
-// NewGCMTLS is the internal method used with boringcrypto that provides a
-// validated mode of AES-GCM which enforces the nonce is strictly
-// monotonically increasing.  This is the TLS 1.2 specification for nonce
-// generation (which also matches the method used by the Noise Protocol)
-//
-// - https://github.com/golang/go/blob/go1.19/src/crypto/tls/cipher_suites.go#L520-L522
-// - https://github.com/golang/go/blob/go1.19/src/crypto/internal/boring/aes.go#L235-L237
-// - https://github.com/golang/go/blob/go1.19/src/crypto/internal/boring/aes.go#L250
-// - https://github.com/google/boringssl/blob/ae223d6138807a13006342edfeef32e813246b39/include/openssl/aead.h#L379-L381
-// - https://github.com/google/boringssl/blob/ae223d6138807a13006342edfeef32e813246b39/crypto/fipsmodule/cipher/e_aes.c#L1082-L1093
-//
-//go:linkname newGCMTLS crypto/internal/boring.NewGCMTLS
-func newGCMTLS(c cipher.Block) (cipher.AEAD, error)
-
-type cipherFn struct {
-	fn   func([32]byte) noise.Cipher
-	name string
-}
-
-func (c cipherFn) Cipher(k [32]byte) noise.Cipher { return c.fn(k) }
-func (c cipherFn) CipherName() string             { return c.name }
-
-// CipherAESGCM is the AES256-GCM AEAD cipher (using NewGCMTLS when GoBoring is present)
-var CipherAESGCM noise.CipherFunc = cipherFn{cipherAESGCMBoring, "AESGCM"}
-
-func cipherAESGCMBoring(k [32]byte) noise.Cipher {
-	c, err := aes.NewCipher(k[:])
-	if err != nil {
-		panic(err)
-	}
-	gcm, err := newGCMTLS(c)
-	if err != nil {
-		panic(err)
-	}
-	return aeadCipher{
-		gcm,
-		func(n uint64) []byte {
-			var nonce [12]byte
-			binary.BigEndian.PutUint64(nonce[4:], n)
-			return nonce[:]
-		},
-	}
-}
-
-type aeadCipher struct {
-	cipher.AEAD
-	nonce func(uint64) []byte
-}
-
-func (c aeadCipher) Encrypt(out []byte, n uint64, ad, plaintext []byte) []byte {
-	return c.Seal(out, c.nonce(n), plaintext, ad)
-}
-
-func (c aeadCipher) Decrypt(out []byte, n uint64, ad, ciphertext []byte) ([]byte, error) {
-	return c.Open(out, c.nonce(n), ciphertext, ad)
-}
+var boringEnabled = boring.Enabled()

noiseutil/boring_test.go

@@ -4,8 +4,6 @@
 package noiseutil
 
 import (
-	"crypto/boring"
-	"encoding/hex"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -14,33 +12,3 @@ import (
 func TestEncryptLockNeeded(t *testing.T) {
 	assert.True(t, EncryptLockNeeded)
 }
-
-// Ensure NewGCMTLS validates the nonce is non-repeating
-func TestNewGCMTLS(t *testing.T) {
-	assert.True(t, boring.Enabled())
-
-	// Test Case 16 from GCM Spec:
-	//  - (now dead link): http://csrc.nist.gov/groups/ST/toolkit/BCM/documents/proposedmodes/gcm/gcm-spec.pdf
-	//  - as listed in boringssl tests: https://github.com/google/boringssl/blob/fips-20220613/crypto/cipher_extra/test/cipher_tests.txt#L412-L418
-	key, _ := hex.DecodeString("feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308")
-	iv, _ := hex.DecodeString("cafebabefacedbaddecaf888")
-	plaintext, _ := hex.DecodeString("d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39")
-	aad, _ := hex.DecodeString("feedfacedeadbeeffeedfacedeadbeefabaddad2")
-	expected, _ := hex.DecodeString("522dc1f099567d07f47f37a32a84427d643a8cdcbfe5c0c97598a2bd2555d1aa8cb08e48590dbb3da7b08b1056828838c5f61e6393ba7a0abcc9f662")
-	expectedTag, _ := hex.DecodeString("76fc6ece0f4e1768cddf8853bb2d551b")
-
-	expected = append(expected, expectedTag...)
-
-	var keyArray [32]byte
-	copy(keyArray[:], key)
-	c := CipherAESGCM.Cipher(keyArray)
-	aead := c.(aeadCipher).AEAD
-
-	dst := aead.Seal([]byte{}, iv, plaintext, aad)
-	assert.Equal(t, expected, dst)
-
-	// We expect this to fail since we are re-encrypting with a repeat IV
-	assert.PanicsWithError(t, "boringcrypto: EVP_AEAD_CTX_seal failed", func() {
-		dst = aead.Seal([]byte{}, iv, plaintext, aad)
-	})
-}

noiseutil/cipher_state.go

@@ -29,8 +29,11 @@ type CipherState interface {
 // NewCipherState wraps the post-handshake noise.CipherState in the per-cipher type that matches cipherFunc.
 // cipherFunc must be the same cipher used to build the noise CipherSuite that produced s.
 func NewCipherState(s *noise.CipherState, cipherFunc noise.CipherFunc) CipherState {
+	if cs, ok := s.Cipher().(CipherState); ok {
+		return cs
+	}
 	switch cipherFunc.CipherName() {
-	case CipherAESGCM.CipherName():
+	case noise.CipherAESGCM.CipherName():
 		return NewCipherStateAESGCM(s)
 	case noise.CipherChaChaPoly.CipherName():
 		return NewCipherStateChaChaPoly(s)

noiseutil/cipher_state_test.go

@@ -1,6 +1,7 @@
 package noiseutil
 
 import (
+	"crypto/fips140"
 	"testing"
 
 	"github.com/flynn/noise"
@@ -10,24 +11,30 @@ import (
 
 func TestCipherStateAESGCMRoundtrip(t *testing.T) {
 	enc, dec := buildCipherStates(t, CipherAESGCM)
-	roundtrip(t, NewCipherStateAESGCM(enc), NewCipherStateAESGCM(dec))
+	roundtrip(t, NewCipherState(enc, CipherAESGCM), NewCipherState(dec, CipherAESGCM))
 }
 
 func TestCipherStateChaChaPolyRoundtrip(t *testing.T) {
 	enc, dec := buildCipherStates(t, noise.CipherChaChaPoly)
-	roundtrip(t, NewCipherStateChaChaPoly(enc), NewCipherStateChaChaPoly(dec))
+	roundtrip(t, NewCipherState(enc, noise.CipherChaChaPoly), NewCipherState(dec, noise.CipherChaChaPoly))
 }
 
 func TestNewCipherStateDispatch(t *testing.T) {
 	encA, _ := buildCipherStates(t, CipherAESGCM)
 	encC, _ := buildCipherStates(t, noise.CipherChaChaPoly)
 
-	assert.IsType(t, &CipherStateAESGCM{}, NewCipherState(encA, CipherAESGCM))
+	if !boringEnabled && !fips140.Enabled() {
+		assert.IsType(t, &CipherStateAESGCM{}, NewCipherState(encA, CipherAESGCM))
+	} else {
+		// fips140
+		assert.IsType(t, encA.Cipher(), NewCipherState(encA, CipherAESGCM))
+	}
+
 	assert.IsType(t, &CipherStateChaChaPoly{}, NewCipherState(encC, noise.CipherChaChaPoly))
 }
 
 func TestNewCipherStateUnsupportedPanics(t *testing.T) {
-	enc, _ := buildCipherStates(t, CipherAESGCM)
+	enc, _ := buildCipherStates(t, noise.CipherChaChaPoly)
 	assert.Panics(t, func() {
 		NewCipherState(enc, fakeCipher{})
 	})

noiseutil/fips140.go

@@ -0,0 +1,179 @@
+package noiseutil
+
+import (
+	"bytes"
+	"crypto/cipher"
+	"crypto/fips140"
+	"encoding/binary"
+	"fmt"
+	"reflect"
+	"runtime"
+	"unsafe"
+
+	// unsafe needed for go:linkname
+	_ "crypto/tls"
+	_ "unsafe"
+
+	"github.com/flynn/noise"
+)
+
+// TODO: Use NewGCMWithCounterNonce or NewGCMForQUIC once available:
+// - https://github.com/golang/go/issues/73110
+// - https://github.com/golang/go/issues/79219
+// Using tls.aeadAESGCMTLS13 gives us the TLS 1.3 GCM, which also verifies
+// that the nonce is strictly increasing. This works for both boringcrypto
+// and fips140.
+//
+//go:linkname aeadAESGCMTLS13 crypto/tls.aeadAESGCMTLS13
+func aeadAESGCMTLS13(key, noncePrefix []byte) cipher.AEAD
+
+type cipherFn struct {
+	fn   func([32]byte) noise.Cipher
+	name string
+}
+
+func (c cipherFn) Cipher(k [32]byte) noise.Cipher { return c.fn(k) }
+func (c cipherFn) CipherName() string             { return c.name }
+
+// CipherAESGCMFIPS140 is the AES256-GCM AEAD cipher (using tls.aeadAESGCMTLS13, for both boringcrypto and fips140)
+var CipherAESGCMFIPS140 noise.CipherFunc = cipherFn{cipherAESGCMFIPS140, "AESGCM"}
+
+// tls.aeadAESGCMTLS13 uses a 4 byte static prefix and an 8 byte XOR mask
+var emptyNonce = []byte{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
+
+func cipherAESGCMFIPS140(k [32]byte) noise.Cipher {
+	gcm := aeadAESGCMTLS13(k[:], emptyNonce)
+	gcm = extractFIPSAEAD(gcm)
+	return &aeadGCMFIPS140Cipher{
+		AEAD: gcm,
+	}
+}
+
+type aeadGCMFIPS140Cipher struct {
+	cipher.AEAD
+	ready bool
+}
+
+// Extract the internal FIPS GCM implementation from the tls wrapper. The TLS
+// wrapper is not thread safe around Open, so instead of locking around it we
+// can grab the internal implementation that is thread safe. This is the FIPS
+// module implementation: `crypto/internal/fips140/aes/gcm.GCMWithXORCounterNonce`
+//
+// - https://github.com/golang/go/blob/go1.26.4/src/crypto/internal/fips140/aes/gcm/gcm_nonces.go#L212-L287
+//
+// The wrapper is struct `crypto/tls.xorNonceAEAD` , with field `aead`:
+//
+// - https://github.com/golang/go/blob/go1.26.4/src/crypto/tls/cipher_suites.go#L482-L487
+//
+// This can be cleaned up once these FIPS implementations are exposed directly:
+//
+// - https://github.com/golang/go/issues/73110
+func extractFIPSAEAD(xorNonceAEAD cipher.AEAD) cipher.AEAD {
+	r := reflect.ValueOf(xorNonceAEAD)
+	v := r.Elem().FieldByName("aead")
+	if !v.IsValid() {
+		// The internal crypto/tls.xorNonceAEAD struct no longer has an `aead`
+		// field. This can only happen on a Go version this code was not built
+		// against; the package init() self-test guards against ever reaching
+		// this at runtime, so this is a defensive fail-fast.
+		panic(fmt.Sprintf("noiseutil: could not extract FIPS AEAD from %T on %s: no `aead` field (incompatible Go version)", xorNonceAEAD, runtime.Version()))
+	}
+	v2 := reflect.NewAt(v.Type(), unsafe.Pointer(v.UnsafeAddr())).Elem()
+	aead, ok := v2.Interface().(cipher.AEAD)
+	if !ok {
+		panic(fmt.Sprintf("noiseutil: extracted FIPS `aead` field is %s, not a cipher.AEAD, on %s (incompatible Go version)", v2.Type(), runtime.Version()))
+	}
+	return aead
+}
+
+func (c *aeadGCMFIPS140Cipher) init(nonce []byte) {
+	// GCMWithXORCounterNonce expects that the first call to Seal
+	// is with a counter of `0`, this is how it extracts the nonce mask.
+	// We can clean this up in the future when NewGCMWithCounterNonce or
+	// NewGCMForQUIC are available:
+	if !bytes.Equal(emptyNonce, nonce) {
+		c.AEAD.Seal([]byte{}, emptyNonce, []byte{}, []byte{})
+	}
+	c.ready = true
+}
+
+func (c *aeadGCMFIPS140Cipher) Seal(dst, nonce, plaintext, additionalData []byte) []byte {
+	if !c.ready {
+		c.init(nonce)
+	}
+	return c.AEAD.Seal(dst, nonce, plaintext, additionalData)
+}
+
+func (c *aeadGCMFIPS140Cipher) Encrypt(out []byte, n uint64, ad, plaintext []byte) []byte {
+	return c.Seal(out, aeadGCMFIPS140CipherNonce(n), plaintext, ad)
+}
+
+func (c *aeadGCMFIPS140Cipher) Decrypt(out []byte, n uint64, ad, ciphertext []byte) ([]byte, error) {
+	return c.Open(out, aeadGCMFIPS140CipherNonce(n), ciphertext, ad)
+}
+
+func (c *aeadGCMFIPS140Cipher) EncryptDanger(out, ad, plaintext []byte, n uint64, nb []byte) ([]byte, error) {
+	binary.BigEndian.PutUint64(nb[4:], n)
+	out = c.Seal(out, nb, plaintext, ad)
+	return out, nil
+}
+
+func (c *aeadGCMFIPS140Cipher) DecryptDanger(out, ad, ciphertext []byte, n uint64, nb []byte) ([]byte, error) {
+	binary.BigEndian.PutUint64(nb[4:], n)
+	return c.Open(out, nb, ciphertext, ad)
+}
+
+func aeadGCMFIPS140CipherNonce(n uint64) []byte {
+	// GCMWithXORCounterNonce uses a 4 byte static prefix and an 8 byte nonce
+	var nonce [12]byte
+	binary.BigEndian.PutUint64(nonce[4:], n)
+	return nonce[:]
+}
+
+// init validates the go:linkname + reflection extraction and the nonce-reuse
+// protection at startup, in every build. cipherAESGCMFIPS140 relies on unexported
+// crypto/tls and crypto/internal/fips140 internals; if a future Go version changes
+// those, this fails fast with a clear message instead of panicking per-handshake
+// (or, worse, silently losing the strictly-increasing nonce check that is the whole
+// point of using this cipher). Because this file has no build tag, this self-test
+// runs even in non-FIPS builds, so the default CI lane catches an incompatible Go.
+func init() {
+	var key [32]byte
+	c := cipherAESGCMFIPS140(key)
+
+	// Verify the extracted AEAD produces a working encrypt/decrypt roundtrip.
+	plaintext := []byte("nebula fips140 self-test")
+	ad := []byte("ad")
+	ct := c.Encrypt(nil, 1, ad, plaintext)
+	pt, err := c.Decrypt(nil, 1, ad, ct)
+	if err != nil {
+		panic(fmt.Sprintf("noiseutil: FIPS AES-GCM self-test roundtrip failed on %s: %v", runtime.Version(), err))
+	}
+	if !bytes.Equal(pt, plaintext) {
+		panic(fmt.Sprintf("noiseutil: FIPS AES-GCM self-test roundtrip returned wrong plaintext on %s", runtime.Version()))
+	}
+
+	// Verify the nonce-reuse protection still fires: re-encrypting with the same
+	// counter must panic. This is the guarantee we depend on for nonce safety, so
+	// if the extraction ever silently yields an AEAD without it, refuse to start.
+	// The strictly-increasing nonce check only exists under boringcrypto/fips140;
+	// in a plain build aeadAESGCMTLS13 wraps a standard GCM that does not enforce
+	// it (and CipherAESGCMFIPS140 is unused there anyway), so only assert it when
+	// one of those modes is active.
+	if (boringEnabled || fips140.Enabled()) && !reusePanics(c) {
+		panic(fmt.Sprintf("noiseutil: FIPS AES-GCM self-test did not reject a reused nonce on %s; nonce-reuse protection is missing (incompatible Go version)", runtime.Version()))
+	}
+}
+
+// reusePanics reports whether re-encrypting with an already-used counter panics,
+// as GCMWithXORCounterNonce is expected to.
+func reusePanics(c noise.Cipher) (panicked bool) {
+	c.Encrypt(nil, 2, nil, nil)
+	defer func() {
+		if recover() != nil {
+			panicked = true
+		}
+	}()
+	c.Encrypt(nil, 2, nil, nil)
+	return false
+}

noiseutil/fips140_test.go

@@ -0,0 +1,48 @@
+package noiseutil
+
+import (
+	"crypto/cipher"
+	"crypto/fips140"
+	"encoding/hex"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+// Ensure NewAESGCM validates the nonce is non-repeating
+func TestNewAESGCM(t *testing.T) {
+	if !boringEnabled && !fips140.Enabled() {
+		t.Skip("TestNewAESGCM is only for fips140/boringcrypto")
+	}
+
+	key, _ := hex.DecodeString("feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308")
+	iv, _ := hex.DecodeString("00000000facedbaddecaf888")
+	plaintext, _ := hex.DecodeString("d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39")
+	aad, _ := hex.DecodeString("feedfacedeadbeeffeedfacedeadbeefabaddad2")
+	expected, _ := hex.DecodeString("6a65c2edd45bd63c7e29f40e3d2ed8ba2b99f4c83135383d5676652f255059ceb24863ff10afb1089db701245da87fb88d3acd5f9dd0770cac220c3c04145caf25e190aeb775e7080401c628")
+
+	var keyArray [32]byte
+	copy(keyArray[:], key)
+	c := CipherAESGCM.Cipher(keyArray)
+	aead := c.(cipher.AEAD)
+
+	dst := aead.Seal([]byte{}, iv, plaintext, aad)
+	t.Logf("%x", dst)
+	assert.Equal(t, expected, dst)
+
+	// We expect this to fail since we are re-encrypting with a repeat IV
+	switch {
+	case boringEnabled:
+		assert.PanicsWithError(t, "boringcrypto: EVP_AEAD_CTX_seal failed", func() {
+			dst = aead.Seal([]byte{}, iv, plaintext, aad)
+		})
+	case fips140.Version() == "v1.0.0":
+		assert.PanicsWithValue(t, "crypto/cipher: counter decreased", func() {
+			dst = aead.Seal([]byte{}, iv, plaintext, aad)
+		})
+	default:
+		assert.PanicsWithValue(t, "crypto/cipher: counter decreased or remained the same", func() {
+			dst = aead.Seal([]byte{}, iv, plaintext, aad)
+		})
+	}
+}

noiseutil/fips140enforce.go

@@ -0,0 +1,13 @@
+//go:build fips140enforce
+
+package noiseutil
+
+import (
+	"crypto/fips140"
+)
+
+func init() {
+	if !fips140.Enforced() {
+		panic("Nebula compiled with fips140 expects FIPS140 to be enforced. Do not set GODEBUG=fips140, or if you do it must be set as GODEBUG=fips140=only")
+	}
+}

noiseutil/notboring.go

@@ -1,14 +1,25 @@
 //go:build !boringcrypto
-// +build !boringcrypto
 
 package noiseutil
 
 import (
+	"crypto/fips140"
+
 	"github.com/flynn/noise"
 )
 
 // EncryptLockNeeded indicates if calls to Encrypt need a lock
-const EncryptLockNeeded = false
+var EncryptLockNeeded = fips140.Enabled()
 
-// CipherAESGCM is the standard noise.CipherAESGCM when boringcrypto is not enabled
-var CipherAESGCM noise.CipherFunc = noise.CipherAESGCM
+var CipherAESGCM noise.CipherFunc = initAESGCM()
+
+func initAESGCM() noise.CipherFunc {
+	if fips140.Enabled() {
+		return CipherAESGCMFIPS140
+	} else {
+		return noise.CipherAESGCM
+	}
+
+}
+
+var boringEnabled = false

noiseutil/notboring_test.go

@@ -1,14 +0,0 @@
-//go:build !boringcrypto
-// +build !boringcrypto
-
-package noiseutil
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-)
-
-func TestEncryptLockNeeded(t *testing.T) {
-	assert.False(t, EncryptLockNeeded)
-}

outside.go

@@ -13,7 +13,6 @@ import (
 
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
-	"github.com/slackhq/nebula/udp"
 	"golang.org/x/net/ipv4"
 )
 
@@ -23,7 +22,7 @@ const (
 
 var ErrOutOfWindow = errors.New("out of window packet")
 
-func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache, meta udp.RxMeta) {
+func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache) {
 	err := h.Parse(packet)
 	if err != nil {
 		// Hole punch packets are 0 or 1 byte big, so lets ignore printing those errors
@@ -103,26 +102,31 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 		return
 	}
 
+	if len(packet) < header.Len+hostinfo.ConnectionState.dKey.Overhead() {
+		f.messageMetrics.RxInvalid(1)
+		if f.l.Enabled(context.Background(), slog.LevelDebug) {
+			f.l.Debug("packet too small", "from", via, "length", len(packet))
+		}
+		return
+	}
+
 	// All remaining packets are encrypted
-	ci := hostinfo.ConnectionState
-	if !ci.window.Check(f.l, h.MessageCounter) {
-		return
-	}
-
-	// Relay packets are special
 	if isMessageRelay {
-		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache, meta)
+		// Relay packets are special, this branch should always early-return
+		if err = hostinfo.ConnectionState.VerifyRelay(f.l, h.MessageCounter, packet, nb); err != nil {
+			if f.l.Enabled(context.Background(), slog.LevelDebug) {
+				hostinfo.logger(f.l).Debug("Failed to verify relay packet", "error", err, "from", via, "header", h)
+			}
+			return
+		}
+		f.handleOutsideRelayPacket(hostinfo, via, out, packet, h, fwPacket, lhf, nb, q, localCache)
 		return
 	}
 
-	out, err = f.decrypt(hostinfo, h.MessageCounter, out, packet, h, nb)
+	out, err = hostinfo.ConnectionState.Decrypt(f.l, h.MessageCounter, out, packet, nb)
 	if err != nil {
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("Failed to decrypt packet",
-				"error", err,
-				"from", via,
-				"header", h,
-			)
+			hostinfo.logger(f.l).Debug("Failed to decrypt packet", "error", err, "from", via, "header", h)
 		}
 		return
 	}
@@ -135,7 +139,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 	case header.Message:
 		switch h.Subtype {
 		case header.MessageNone:
-			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache, meta)
+			f.handleOutsideMessagePacket(hostinfo, out, packet, fwPacket, nb, q, localCache)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected message subtype seen", "from", via, "header", h)
 			return
@@ -151,7 +155,7 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 			// No-op, useful for the Roaming and connectionManager side-effects above
 		case header.TestRequest:
 			//recycle the input packet ciphertext as our output buffer
-			f.send(header.Test, header.TestReply, ci, hostinfo, out, nb, packet)
+			f.send(header.Test, header.TestReply, hostinfo.ConnectionState, hostinfo, out, nb, packet)
 		default:
 			hostinfo.logger(f.l).Error("IsValidSubType was true, but unexpected test subtype seen", "from", via, "header", h)
 			return
@@ -169,28 +173,9 @@ func (f *Interface) readOutsidePackets(via ViaSender, out []byte, packet []byte,
 	}
 }
 
-func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender, out []byte, packet []byte, h *header.H, fwPacket *firewall.Packet, lhf *LightHouseHandler, nb []byte, q int, localCache firewall.ConntrackCache, meta udp.RxMeta) {
-	// The entire body is sent as AD, not encrypted.
-	// The packet consists of a 16-byte parsed Nebula header, Associated Data-protected payload, and a trailing 16-byte AEAD signature value.
-	// The packet is guaranteed to be at least 16 bytes at this point, b/c it got past the h.Parse() call above. If it's
-	// otherwise malformed (meaning, there is no trailing 16 byte AEAD value), then this will result in at worst a 0-length slice
-	// which will gracefully fail in the DecryptDanger call.
-	signedPayload := packet[:len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
-	signatureValue := packet[len(packet)-hostinfo.ConnectionState.dKey.Overhead():]
-	var err error
-	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, signedPayload, signatureValue, h.MessageCounter, nb)
-	if err != nil {
-		return
-	}
-	// Advance the replay window now that the frame is authenticated
-	if !hostinfo.ConnectionState.window.Update(f.l, h.MessageCounter) {
-		if f.l.Enabled(context.Background(), slog.LevelDebug) {
-			hostinfo.logger(f.l).Debug("dropping out of window relay packet", "header", h)
-		}
-		return
-	}
-	// Successfully validated the thing. Get rid of the Relay header.
-	signedPayload = signedPayload[header.Len:]
+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) {
+	// Successfully validated the thing. Get rid of the Relay header and the AEAD tag
+	signedPayload := packet[header.Len : len(packet)-hostinfo.ConnectionState.dKey.Overhead()]
 	// Pull the Roaming parts up here, and return in all call paths.
 	f.handleHostRoaming(hostinfo, via)
 	// Track usage of both the HostInfo and the Relay for the received & authenticated packet
@@ -214,11 +199,10 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		via = ViaSender{
 			UdpAddr:   via.UdpAddr,
 			relayHI:   hostinfo,
-			remoteIdx: relay.RemoteIndex,
 			relay:     relay,
 			IsRelayed: true,
 		}
-		f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache, meta)
+		f.readOutsidePackets(via, out[:0], signedPayload, h, fwPacket, lhf, nb, q, localCache)
 	case ForwardingType:
 		// Find the target HostInfo relay object
 		targetHI, targetRelay, err := f.hostMap.QueryVpnAddrsRelayFor(hostinfo.vpnAddrs, relay.PeerAddr)
@@ -235,9 +219,10 @@ func (f *Interface) handleOutsideRelayPacket(hostinfo *HostInfo, via ViaSender,
 		if targetRelay.State == Established {
 			switch targetRelay.Type {
 			case ForwardingType:
-				// Forward this packet through the relay tunnel
-				// Find the target HostInfo //todo it would potentially be nice to batch these
-				f.SendVia(targetHI, targetRelay, signedPayload, nb, out, false)
+				// Forward this packet through the relay tunnel, rebuilding it in place.
+				// Encode overwrites the old outer header, and the new AEAD tag lands where the old one was
+				fwdBuf := packet[:0:len(packet)] // Cap to len(packet) to protect memory from a larger parent buffer
+				f.SendVia(targetHI, targetRelay, signedPayload, nb, fwdBuf, true)
 			case TerminalType:
 				hostinfo.logger(f.l).Error("Unexpected Relay Type of Terminal")
 				return
@@ -504,91 +489,7 @@ func parseV4(data []byte, incoming bool, fp *firewall.Packet) error {
 	return nil
 }
 
-func (f *Interface) decrypt(hostinfo *HostInfo, mc uint64, out []byte, packet []byte, h *header.H, nb []byte) ([]byte, error) {
-	var err error
-	out, err = hostinfo.ConnectionState.dKey.DecryptDanger(out, packet[:header.Len], packet[header.Len:], mc, nb)
-	if err != nil {
-		return nil, err
-	}
-
-	if !hostinfo.ConnectionState.window.Update(f.l, mc) {
-		return nil, ErrOutOfWindow
-	}
-
-	return out, nil
-}
-
-// 2-bit IP-level ECN codepoints (lower bits of IPv4 ToS / IPv6 TC).
-const (
-	ecnNotECT = 0x00
-	ecnECT1   = 0x01
-	ecnECT0   = 0x02
-	ecnCE     = 0x03
-)
-
-// applyOuterECN folds an outer CE mark from the underlay into the inner
-// IP header per RFC 6040 normal mode. It mutates pkt[1] in place. Other
-// codepoints are advisory only and leave the inner unchanged.
-//
-// Merge cases (outer × inner → action):
-//
-//	outer != CE                : no-op (inner is authoritative)
-//	outer == CE, inner Not-ECT : log; cannot propagate to a non-ECN host
-//	outer == CE, inner ECT/CE  : rewrite inner ECN to CE
-func applyOuterECN(pkt []byte, outerECN byte, hostinfo *HostInfo, l *slog.Logger) {
-	if outerECN&ecnCE != ecnCE || len(pkt) < 2 {
-		return
-	}
-	switch pkt[0] >> 4 {
-	case 4:
-		switch pkt[1] & 0x03 {
-		case ecnNotECT:
-			if l.Enabled(context.Background(), slog.LevelDebug) {
-				hostinfo.logger(l).Debug("RFC 6040: outer CE on inner Not-ECT, leaving inner unchanged")
-			}
-		case ecnCE:
-			// Already CE.
-		default:
-			// Rewriting the ToS byte invalidates the IPv4 header checksum, so
-			// patch it incrementally per RFC 1624 (HC' = ~(~HC + ~m + m')). The
-			// ToS is the low byte of the 16-bit word at pkt[0:2]; the header
-			// checksum lives at pkt[10:12]. A header too short to carry a
-			// checksum can't be fixed up here, so leave it for newPacket to
-			// reject rather than emit a mangled packet.
-			if len(pkt) < ipv4.HeaderLen {
-				return
-			}
-			m := binary.BigEndian.Uint16(pkt[0:2])
-			pkt[1] = (pkt[1] &^ 0x03) | ecnCE
-			mNew := binary.BigEndian.Uint16(pkt[0:2])
-			sum := uint32(^binary.BigEndian.Uint16(pkt[10:12])) + uint32(^m) + uint32(mNew)
-			for sum > 0xffff {
-				sum = (sum >> 16) + (sum & 0xffff)
-			}
-			binary.BigEndian.PutUint16(pkt[10:12], ^uint16(sum))
-		}
-	case 6:
-		switch (pkt[1] >> 4) & 0x03 {
-		case ecnNotECT:
-			if l.Enabled(context.Background(), slog.LevelDebug) {
-				hostinfo.logger(l).Debug("RFC 6040: outer CE on inner Not-ECT, leaving inner unchanged")
-			}
-		case ecnCE:
-			// Already CE.
-		default:
-			pkt[1] = (pkt[1] &^ 0x30) | (ecnCE << 4)
-		}
-	}
-}
-
-func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache, meta udp.RxMeta) {
-	// RFC 6040 normal-mode combine: fold any outer CE mark stamped by the
-	// underlay into the inner header before firewall + TUN write. Other
-	// outer codepoints are advisory only — we keep the inner unchanged.
-	if f.ecnEnabled.Load() {
-		applyOuterECN(out, meta.OuterECN, hostinfo, f.l)
-	}
-
+func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, packet []byte, fwPacket *firewall.Packet, nb []byte, q int, localCache firewall.ConntrackCache) {
 	err := newPacket(out, true, fwPacket)
 	if err != nil {
 		hostinfo.logger(f.l).Warn("Error while validating inbound packet",
@@ -601,10 +502,8 @@ func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, p
 	dropReason := f.firewall.Drop(*fwPacket, true, hostinfo, f.pki.GetCAPool(), localCache)
 	if dropReason != nil {
 		// NOTE: We give `packet` as the `out` here since we already decrypted from it and we don't need it anymore
-		// This gives us a buffer to build the reject packet in. With UDP GRO this is a single segment of a shared
-		// recvmmsg row whose capacity runs to the end of the whole row, so cap it to its own length (cap==len) to
-		// keep the reject builder from writing past this segment into the next, not-yet-processed coalesced segment.
-		f.rejectOutside(out, hostinfo.ConnectionState, hostinfo, nb, packet[:len(packet):len(packet)], q)
+		// This gives us a buffer to build the reject packet in
+		f.rejectOutside(out, hostinfo.ConnectionState, hostinfo, nb, packet, q)
 		if f.l.Enabled(context.Background(), slog.LevelDebug) {
 			hostinfo.logger(f.l).Debug("dropping inbound packet",
 				"fwPacket", fwPacket,
@@ -614,7 +513,7 @@ func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, p
 		return
 	}
 
-	err = f.batchers[q].Commit(out)
+	_, err = f.readers[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}

overlay/batch/batch.go

@@ -1,28 +0,0 @@
-package batch
-
-import "net/netip"
-
-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
-}
-
-type TxBatcher interface {
-	// Reserve creates a pkt to borrow
-	Reserve(sz int) []byte
-	// Commit borrows pkt and records its destination plus the 2-bit
-	// IP-level ECN codepoint to set on the outer (carrier) header. The
-	// caller must keep pkt valid until the next Flush. Pass 0 (Not-ECT)
-	// to leave the outer ECN field unset.
-	Commit(pkt []byte, dst netip.AddrPort, outerECN byte)
-	// Flush emits every queued packet via the underlying batch writer in arrival order.
-	// Returns an errors.Join of one or more errors.
-	// After Flush returns, borrowed payload slices may be recycled.
-	Flush() error
-}

overlay/batch/coalesce_core.go

@@ -1,181 +0,0 @@
-package batch
-
-import (
-	"bytes"
-	"encoding/binary"
-)
-
-// flowKey identifies a transport flow by {src, dst, sport, dport, family}.
-// Comparable, so map lookups and linear scans over the slot list stay tight.
-// Shared by the TCP and UDP coalescers; each coalescer keeps its own
-// openSlots map, so a TCP and UDP flow on the same 5-tuple-without-proto
-// never alias.
-type flowKey struct {
-	src, dst     [16]byte
-	sport, dport uint16
-	isV6         bool
-}
-
-// initialSlots is the starting capacity of the slot pool. One flow per
-// packet is the worst case so this matches a typical carrier-side
-// recvmmsg batch on the encrypted UDP socket.
-const initialSlots = 64
-
-// parsedIP is the IP-level result of parseIPPrologue. The caller layers
-// L4-specific parsing (TCP / UDP) on top.
-type parsedIP struct {
-	fk       flowKey
-	ipHdrLen int
-	// pkt is the original buffer trimmed to the IP-declared total length.
-	// Anything below the IP layer (transport parsers) should slice into
-	// pkt rather than the unbounded original.
-	pkt []byte
-}
-
-// parseIPPrologue extracts the IP-level fields the coalescers care about:
-// IHL/payload length, version, src/dst addresses, and the L4 protocol byte.
-// Returns ok=false for malformed input, IPv4 with options or fragmentation,
-// or IPv6 with extension headers (all rejected by both coalescers in
-// identical ways before this refactor).
-//
-// On success, p.pkt is len-trimmed to the IP-declared length so callers
-// don't have to repeat the trim. wantProto is the IANA protocol number to
-// require (6 for TCP, 17 for UDP); ok=false for any other value.
-func parseIPPrologue(pkt []byte, wantProto byte) (parsedIP, bool) {
-	var p parsedIP
-	if len(pkt) < 20 {
-		return p, false
-	}
-	v := pkt[0] >> 4
-	switch v {
-	case 4:
-		ihl := int(pkt[0]&0x0f) * 4
-		if ihl != 20 {
-			return p, false
-		}
-		if pkt[9] != wantProto {
-			return p, false
-		}
-		// Reject actual fragmentation (MF or non-zero frag offset).
-		if binary.BigEndian.Uint16(pkt[6:8])&0x3fff != 0 {
-			return p, false
-		}
-		totalLen := int(binary.BigEndian.Uint16(pkt[2:4]))
-		if totalLen > len(pkt) || totalLen < ihl {
-			return p, false
-		}
-		p.ipHdrLen = 20
-		p.fk.isV6 = false
-		copy(p.fk.src[:4], pkt[12:16])
-		copy(p.fk.dst[:4], pkt[16:20])
-		p.pkt = pkt[:totalLen]
-	case 6:
-		if len(pkt) < 40 {
-			return p, false
-		}
-		if pkt[6] != wantProto {
-			return p, false
-		}
-		payloadLen := int(binary.BigEndian.Uint16(pkt[4:6]))
-		if 40+payloadLen > len(pkt) {
-			return p, false
-		}
-		p.ipHdrLen = 40
-		p.fk.isV6 = true
-		copy(p.fk.src[:], pkt[8:24])
-		copy(p.fk.dst[:], pkt[24:40])
-		p.pkt = pkt[:40+payloadLen]
-	default:
-		return p, false
-	}
-	return p, true
-}
-
-// ipHeadersMatch compares the IP portion of two packet header prefixes for
-// byte-for-byte equality on every field that must be identical across
-// coalesced segments. Size/IPID/IPCsum are masked out. The full DSCP/ECN
-// byte (IPv4 ToS / IPv6 traffic class) is compared, matching Linux kernel
-// GRO: segments with differing ECN codepoints must not coalesce, otherwise
-// ORing e.g. ECT(0) with ECT(1) would fabricate a false CE (congestion)
-// mark or mark a Not-ECT flow as ECN-capable.
-//
-// The transport (L4) portion of the header is checked separately by the
-// per-protocol matcher.
-func ipHeadersMatch(a, b []byte, isV6 bool) bool {
-	if isV6 {
-		// IPv6: byte 0 = version/TC[7:4], byte 1 = TC[3:0]/flow[19:16],
-		// bytes [2:4] = flow[15:0], [6:8] = next_hdr/hop, [8:40] = src+dst.
-		// Compare byte 1 fully so ECN (TC[1:0]) must match. Skip [4:6] payload_len.
-		if a[0] != b[0] {
-			return false
-		}
-		if a[1] != b[1] {
-			return false
-		}
-		if !bytes.Equal(a[2:4], b[2:4]) {
-			return false
-		}
-		if !bytes.Equal(a[6:40], b[6:40]) {
-			return false
-		}
-		return true
-	}
-	// IPv4: byte 0 = version/IHL, byte 1 = DSCP(6)|ECN(2),
-	// [6:10] flags/fragoff/TTL/proto, [12:20] src+dst.
-	// Compare byte 1 fully so ECN must match.
-	// Skip [2:4] total len, [4:6] id, [10:12] csum.
-	if a[0] != b[0] {
-		return false
-	}
-	if a[1] != b[1] {
-		return false
-	}
-	if !bytes.Equal(a[6:10], b[6:10]) {
-		return false
-	}
-	if !bytes.Equal(a[12:20], b[12:20]) {
-		return false
-	}
-	return true
-}
-
-// Arena is an injectable byte-slab that hands out non-overlapping borrowed
-// slices via Reserve and releases them in bulk via Reset.
-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]
-}
-
-// Reserver hands out an sz-byte slice valid until its Resetter runs.
-type Reserver func(sz int) []byte
-
-// Resetter clears all reservations held by a Reserver. Only the arena's
-// owner holds one; lanes inside a MultiCoalescer get nil.
-type Resetter func()

overlay/batch/multi_coalesce.go

@@ -1,132 +0,0 @@
-package batch
-
-import (
-	"errors"
-	"io"
-	"log/slog"
-)
-
-// MultiCoalescer fans plaintext packets out to lane-specific batchers based
-// on the IP/L4 protocol of the packet, sharing a single Reserve arena
-// across lanes so the caller's allocation pattern is unchanged.
-//
-// Lanes are processed independently: the TCP coalescer only sees TCP, the
-// UDP coalescer only sees UDP, and the passthrough lane handles everything
-// else. Per-flow arrival order is preserved because a single 5-tuple only
-// ever lands in one lane and each lane preserves its own slot order.
-//
-// Cross-lane order is NOT preserved across the TCP/UDP/passthrough split.
-// This is acceptable because the carrier-side recvmmsg path already
-// stable-sorts by (peer, message counter) before delivering plaintext
-// here, so replay-window invariants are unaffected, and apps observe
-// correct per-flow ordering — which is all the IP layer guarantees anyway.
-// Do not "fix" this by interleaving lane outputs at flush time; that
-// negates the entire point of coalescing (each lane needs to see runs of
-// adjacent same-flow packets to coalesce them).
-type MultiCoalescer struct {
-	tcp *TCPCoalescer
-	udp *UDPCoalescer
-	pt  *Passthrough
-	// arena is owned by the Multi: lanes get only its Reserve (nil Resetter)
-	// and Flush resets it exactly once after every lane has drained.
-	arena *Arena
-}
-
-// DefaultMultiArenaCap is the recommended arena capacity for a Multi-lane
-// batcher: 64 slots × 65535 bytes ≈ 4 MiB, enough to hold one recvmmsg
-// burst worth of MTU-sized packets without the arena growing.
-const DefaultMultiArenaCap = initialSlots * 65535
-
-// NewMultiCoalescer builds a multi-lane batcher. tcpEnabled lets the caller
-// opt out of TCP coalescing (e.g. when the queue can't do TSO); udpEnabled
-// likewise gates UDP coalescing (only enable when USO was negotiated).
-// Either lane disabled redirects its traffic into the passthrough lane.
-// arena is the single backing slab shared across every lane; the caller
-// pre-sizes it via NewArena so the hot path never allocates.
-func NewMultiCoalescer(w io.Writer, l *slog.Logger, arena *Arena, tcpEnabled, udpEnabled bool) *MultiCoalescer {
-	m := &MultiCoalescer{
-		pt:    NewPassthrough(w, arena.Reserve, nil),
-		arena: arena,
-	}
-	if tcpEnabled {
-		m.tcp = NewTCPCoalescer(w, l, arena.Reserve, nil)
-	}
-	if udpEnabled {
-		m.udp = NewUDPCoalescer(w, arena.Reserve, nil)
-	}
-	return m
-}
-
-func (m *MultiCoalescer) Reserve(sz int) []byte {
-	return m.arena.Reserve(sz)
-}
-
-// Commit dispatches pkt to the appropriate lane based on IP version + L4
-// proto. Borrowed slice contract is identical to the single-lane batchers,
-// pkt must remain valid until the next Flush.
-//
-// On the success path the IP/TCP-or-UDP parse happens here once and the
-// parsed struct is handed to the lane via commitParsed so the lane doesn't
-// re-walk the header.
-func (m *MultiCoalescer) Commit(pkt []byte) error {
-	if len(pkt) < 20 {
-		return m.pt.Commit(pkt)
-	}
-	v := pkt[0] >> 4
-	var proto byte
-	switch v {
-	case 4:
-		proto = pkt[9]
-	case 6:
-		if len(pkt) < 40 {
-			return m.pt.Commit(pkt)
-		}
-		proto = pkt[6]
-	default:
-		return m.pt.Commit(pkt)
-	}
-	switch proto {
-	case ipProtoTCP:
-		if m.tcp != nil {
-			info, ok := parseTCPBase(pkt)
-			if !ok {
-				// Malformed/unsupported TCP shape (IP options, fragments, ...).
-				// Handle this via passthrough support in the TCP coalescer, to attempt to preserve flow order.
-				m.tcp.addPassthrough(pkt)
-				return nil
-			}
-			return m.tcp.commitParsed(pkt, info)
-		}
-	case ipProtoUDP:
-		if m.udp != nil {
-			info, ok := parseUDP(pkt)
-			if !ok {
-				m.udp.addPassthrough(pkt) //we could also m.pt.Commit() here I guess?
-				return nil
-			}
-			return m.udp.commitParsed(pkt, info)
-		}
-	}
-	return m.pt.Commit(pkt)
-}
-
-// Flush drains every lane in a fixed order, then resets the shared arena once.
-// A lane error doesn't stop the remaining lanes; the joined errors are returned.
-func (m *MultiCoalescer) Flush() error {
-	var errs []error
-	if m.tcp != nil {
-		if err := m.tcp.Flush(); err != nil {
-			errs = append(errs, err)
-		}
-	}
-	if m.udp != nil {
-		if err := m.udp.Flush(); err != nil {
-			errs = append(errs, err)
-		}
-	}
-	if err := m.pt.Flush(); err != nil {
-		errs = append(errs, err)
-	}
-	m.arena.Reset()
-	return errors.Join(errs...)
-}

overlay/batch/multi_coalesce_test.go

@@ -1,96 +0,0 @@
-package batch
-
-import (
-	"testing"
-
-	"github.com/slackhq/nebula/test"
-)
-
-// TestMultiCoalescerRoutesByProto confirms TCP/UDP/other land in the right
-// lane: TCP and UDP get coalesced when their lanes are enabled, anything
-// else (ICMP here) falls through to plain Write.
-func TestMultiCoalescerRoutesByProto(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	m := NewMultiCoalescer(w, test.NewLogger(), NewArena(0), true, true)
-
-	tcpPay := make([]byte, 1200)
-	udpPay := make([]byte, 1200)
-	icmp := make([]byte, 28)
-	icmp[0] = 0x45
-	icmp[2] = 0
-	icmp[3] = 28
-	icmp[9] = 1
-
-	if err := m.Commit(buildTCPv4(1000, tcpAck, tcpPay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Commit(buildTCPv4(2200, tcpAck, tcpPay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Commit(buildUDPv4(2000, 53, udpPay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Commit(buildUDPv4(2000, 53, udpPay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Commit(icmp); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	// 1 TCP super (2 segments) + 1 UDP super (2 segments) = 2 gso writes.
-	if len(w.gsoWrites) != 2 {
-		t.Fatalf("want 2 gso writes (one TCP + one UDP), got %d", len(w.gsoWrites))
-	}
-	if len(w.writes) != 1 {
-		t.Fatalf("want 1 plain write (ICMP), got %d", len(w.writes))
-	}
-}
-
-// TestMultiCoalescerDisabledUDPFallsThrough verifies that when the UDP lane
-// is disabled (e.g. kernel doesn't support USO), UDP packets still reach
-// the kernel via the passthrough lane rather than being lost.
-func TestMultiCoalescerDisabledUDPFallsThrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	m := NewMultiCoalescer(w, test.NewLogger(), NewArena(0), true, false) // TSO on, USO off
-
-	if err := m.Commit(buildUDPv4(1000, 53, make([]byte, 800))); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Commit(buildUDPv4(1000, 53, make([]byte, 800))); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 0 {
-		t.Errorf("UDP must NOT be coalesced when USO disabled, got %d gso writes", len(w.gsoWrites))
-	}
-	if len(w.writes) != 2 {
-		t.Errorf("UDP must pass through as 2 plain writes, got %d", len(w.writes))
-	}
-}
-
-// TestMultiCoalescerDisabledTCPFallsThrough mirrors the TSO=off case.
-func TestMultiCoalescerDisabledTCPFallsThrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	m := NewMultiCoalescer(w, test.NewLogger(), NewArena(0), false, true) // TSO off, USO on
-
-	pay := make([]byte, 1200)
-	if err := m.Commit(buildTCPv4(1000, tcpAck, pay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Commit(buildTCPv4(2200, tcpAck, pay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := m.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 0 {
-		t.Errorf("TCP must NOT be coalesced when TSO disabled, got %d gso writes", len(w.gsoWrites))
-	}
-	if len(w.writes) != 2 {
-		t.Errorf("TCP must pass through as 2 plain writes, got %d", len(w.writes))
-	}
-}

overlay/batch/passthrough.go

@@ -1,63 +0,0 @@
-package batch
-
-import (
-	"io"
-
-	"github.com/slackhq/nebula/udp"
-)
-
-// Passthrough is a RxBatcher that doesn't batch anything, it just accumulates and then sends packets.
-type Passthrough struct {
-	out      io.Writer
-	slots    [][]byte
-	reserver Reserver
-	resetter Resetter
-	cursor   int
-}
-
-const passthroughBaseNumSlots = 128
-
-// DefaultPassthroughArenaCap is the recommended arena capacity for a
-// standalone Passthrough batcher: 128 slots × udp.MTU ≈ 1.1 MiB.
-const DefaultPassthroughArenaCap = passthroughBaseNumSlots * udp.MTU
-
-func NewPassthrough(w io.Writer, reserver Reserver, resetter Resetter) *Passthrough {
-	return &Passthrough{
-		out:      w,
-		slots:    make([][]byte, 0, passthroughBaseNumSlots),
-		reserver: reserver,
-		resetter: resetter,
-	}
-}
-
-func (p *Passthrough) Reserve(sz int) []byte {
-	return p.reserver(sz)
-}
-
-func (p *Passthrough) Commit(pkt []byte) error {
-	p.slots = append(p.slots, pkt)
-	return nil
-}
-
-// Flush drains every queued packet and calls the configured Resetter
-func (p *Passthrough) Flush() error {
-	firstErr := p.drain()
-	if p.resetter != nil {
-		p.resetter()
-	}
-	return firstErr
-}
-
-// drain writes out every queued packet and clears the slot list.
-func (p *Passthrough) drain() 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]
-	return firstErr
-}

overlay/batch/tcp_coalesce.go

@@ -1,728 +0,0 @@
-package batch
-
-import (
-	"bytes"
-	"context"
-	"encoding/binary"
-	"io"
-	"log/slog"
-	"net/netip"
-	"slices"
-
-	"github.com/slackhq/nebula/overlay/tio"
-)
-
-// ipProtoTCP is the IANA protocol number for TCP. Defined here to help Windows out.
-const ipProtoTCP = 6
-
-// tcpCoalesceBufSize caps total bytes per superpacket. Mirrors the kernel's
-// sk_gso_max_size of ~64KiB; anything beyond this would be rejected anyway.
-const tcpCoalesceBufSize = 65535
-
-// tcpCoalesceMaxSegs caps how many segments we'll coalesce into a single
-// superpacket. Keeping this well below the kernel's TSO ceiling bounds latency.
-const tcpCoalesceMaxSegs = 64
-
-// tcpCoalesceHdrCap is the scratch space we copy a seed's IP+TCP header
-// into. IPv6 (40) + TCP with full options (60) = 100 bytes.
-const tcpCoalesceHdrCap = 100
-
-// coalesceSlot is one entry in the coalescer's ordered event queue. When
-// passthrough is true the slot holds a single borrowed packet that must be
-// emitted verbatim (non-TCP, non-admissible TCP, or oversize seed). When
-// passthrough is false the slot is an in-progress coalesced superpacket:
-// hdrBuf is a mutable copy of the seed's IP+TCP header (we patch total
-// length and pseudo-header partial at flush), and payIovs are *borrowed*
-// slices from the caller's plaintext buffers — no payload is ever copied.
-// The caller (listenOut) must keep those buffers alive until Flush.
-type coalesceSlot struct {
-	passthrough bool
-	rawPkt      []byte // borrowed when passthrough
-
-	fk       flowKey
-	hdrBuf   [tcpCoalesceHdrCap]byte
-	hdrLen   int
-	ipHdrLen int
-	isV6     bool
-	gsoSize  int
-	numSeg   int
-	totalPay int
-	nextSeq  uint32
-	// psh closes the chain: set when the last-accepted segment had PSH or
-	// was sub-gsoSize. No further appends after that.
-	psh     bool
-	payIovs [][]byte
-}
-
-// TCPCoalescer accumulates adjacent in-flow TCP data segments across
-// multiple concurrent flows and emits each flow's run as a single TSO
-// superpacket via tio.GSOWriter. All output — coalesced or not — is
-// deferred until Flush so arrival order is preserved on the wire. Owns
-// no locks; one coalescer per TUN write queue.
-type TCPCoalescer struct {
-	plainW io.Writer
-	gsoW   tio.GSOWriter // nil when the queue doesn't support TSO
-
-	// slots is the ordered event queue. Flush walks it once and emits each
-	// entry as either a WriteGSO (coalesced) or a plainW.Write (passthrough).
-	slots []*coalesceSlot
-	// openSlots maps a flow key to its most recent non-sealed slot, so new
-	// segments can extend an in-progress superpacket in O(1). Slots are
-	// removed from this map when they close (PSH or short-last-segment),
-	// when a non-admissible packet for that flow arrives, or in Flush.
-	openSlots map[flowKey]*coalesceSlot
-	// lastSlot caches the most recently touched open slot. Steady-state
-	// bulk traffic is dominated by a single flow, so comparing the
-	// incoming key against the cached slot's own fk lets the hot path
-	// skip the map lookup (and the aeshash of a 38-byte key) entirely.
-	// Kept in lockstep with openSlots: nil whenever the slot it pointed
-	// at is removed/sealed.
-	lastSlot *coalesceSlot
-	pool     []*coalesceSlot // free list for reuse
-	reserver Reserver
-	resetter Resetter
-	l        *slog.Logger
-}
-
-func NewTCPCoalescer(w io.Writer, l *slog.Logger, reserver Reserver, resetter Resetter) *TCPCoalescer {
-	c := &TCPCoalescer{
-		plainW:    w,
-		slots:     make([]*coalesceSlot, 0, initialSlots),
-		openSlots: make(map[flowKey]*coalesceSlot, initialSlots),
-		pool:      make([]*coalesceSlot, 0, initialSlots),
-		reserver:  reserver,
-		resetter:  resetter,
-		l:         l,
-	}
-	if gw, ok := tio.SupportsGSO(w, tio.GSOProtoTCP); ok {
-		c.gsoW = gw
-	}
-	return c
-}
-
-// parsedTCP holds the fields extracted from a single parse so later steps
-// (admission, slot lookup, canAppend) don't re-walk the header.
-type parsedTCP struct {
-	fk        flowKey
-	ipHdrLen  int
-	tcpHdrLen int
-	hdrLen    int
-	payLen    int
-	seq       uint32
-	flags     byte
-}
-
-// parseTCPBase extracts the flow key and IP/TCP offsets for any TCP packet,
-// regardless of whether it's admissible for coalescing. Returns ok=false
-// for non-TCP or malformed input.
-// Accepts IPv4 (no options or fragmentation) and IPv6 (no extension headers).
-func parseTCPBase(pkt []byte) (parsedTCP, bool) {
-	var p parsedTCP
-	ip, ok := parseIPPrologue(pkt, ipProtoTCP)
-	if !ok {
-		return p, false
-	}
-	pkt = ip.pkt
-	p.fk = ip.fk
-	p.ipHdrLen = ip.ipHdrLen
-
-	if len(pkt) < p.ipHdrLen+20 {
-		return p, false
-	}
-	tcpOff := int(pkt[p.ipHdrLen+12]>>4) * 4
-	if tcpOff < 20 || tcpOff > 60 {
-		return p, false
-	}
-	if len(pkt) < p.ipHdrLen+tcpOff {
-		return p, false
-	}
-	p.tcpHdrLen = tcpOff
-	p.hdrLen = p.ipHdrLen + tcpOff
-	p.payLen = len(pkt) - p.hdrLen
-	p.seq = binary.BigEndian.Uint32(pkt[p.ipHdrLen+4 : p.ipHdrLen+8])
-	p.flags = pkt[p.ipHdrLen+13]
-	p.fk.sport = binary.BigEndian.Uint16(pkt[p.ipHdrLen : p.ipHdrLen+2])
-	p.fk.dport = binary.BigEndian.Uint16(pkt[p.ipHdrLen+2 : p.ipHdrLen+4])
-	return p, true
-}
-
-// TCP flag bits (byte 13 of the TCP header). Only the bits actually consulted
-// by the coalescer are named; FIN/SYN/RST/URG/CWR are rejected via the
-// negative mask in coalesceable, not by name.
-const (
-	tcpFlagPsh = 0x08
-	tcpFlagAck = 0x10
-	tcpFlagEce = 0x40
-)
-
-// coalesceable reports whether a parsed TCP segment is eligible for
-// coalescing. Accepts ACK, ACK|PSH, ACK|ECE, ACK|PSH|ECE with a
-// non-empty payload. CWR is excluded because it marks a one-shot
-// congestion-window-reduced transition the receiver must observe at a
-// segment boundary.
-func (p parsedTCP) coalesceable() bool {
-	if p.flags&tcpFlagAck == 0 {
-		return false
-	}
-	if p.flags&^(tcpFlagAck|tcpFlagPsh|tcpFlagEce) != 0 {
-		return false
-	}
-	return p.payLen > 0
-}
-
-func (c *TCPCoalescer) Reserve(sz int) []byte {
-	return c.reserver(sz)
-}
-
-// Commit borrows pkt. The caller must keep pkt valid until the next Flush.
-func (c *TCPCoalescer) Commit(pkt []byte) error {
-	if c.gsoW == nil {
-		c.addPassthrough(pkt)
-		return nil
-	}
-	info, ok := parseTCPBase(pkt)
-	if !ok {
-		c.addPassthrough(pkt)
-		return nil
-	}
-	return c.commitParsed(pkt, info)
-}
-
-// commitParsed is the post-parse half of Commit. The caller must have
-// already verified parseTCPBase succeeded (info is a valid TCP parse).
-// Used by MultiCoalescer.Commit to avoid re-walking the IP/TCP header
-// after the dispatcher has already done so.
-func (c *TCPCoalescer) commitParsed(pkt []byte, info parsedTCP) error {
-	if c.gsoW == nil {
-		c.addPassthrough(pkt)
-		return nil
-	}
-	if !info.coalesceable() {
-		// TCP but not admissible (SYN/FIN/RST/URG/CWR or zero-payload).
-		// Seal this flow's open slot so later in-flow packets don't extend
-		// it and accidentally reorder past this passthrough.
-		if last := c.lastSlot; last != nil && last.fk == info.fk {
-			c.lastSlot = nil
-		}
-		delete(c.openSlots, info.fk)
-		c.addPassthrough(pkt)
-		return nil
-	}
-
-	// Single-flow fast path: with only one open flow the cache hits every
-	// packet, and len(openSlots)==1 lets us skip the 38-byte fk compare
-	// when there are multiple flows in flight (where the hit rate would
-	// be ~0 and the compare is pure overhead).
-	var open *coalesceSlot
-	if last := c.lastSlot; last != nil && len(c.openSlots) == 1 && last.fk == info.fk {
-		open = last
-	} else {
-		open = c.openSlots[info.fk]
-	}
-	if open != nil {
-		if c.canAppend(open, pkt, info) {
-			c.appendPayload(open, pkt, info)
-			if open.psh {
-				delete(c.openSlots, info.fk)
-				c.lastSlot = nil
-			} else {
-				c.lastSlot = open
-			}
-			return nil
-		}
-		// Can't extend — seal it and fall through to seed a fresh slot.
-		delete(c.openSlots, info.fk)
-		if c.lastSlot == open {
-			c.lastSlot = nil
-		}
-	}
-	c.seed(pkt, info)
-	return nil
-}
-
-// Flush emits every queued event in (per-flow) seq order.
-func (c *TCPCoalescer) Flush() error {
-	first := c.drain()
-	if c.resetter != nil {
-		c.resetter()
-	}
-	return first
-}
-
-// drain emits every queued slot (reordering/merging coalesced runs first)
-// and clears the slot state.
-func (c *TCPCoalescer) drain() error {
-	c.reorderForFlush()
-	var first error
-	for _, s := range c.slots {
-		var err error
-		if s.passthrough {
-			_, err = c.plainW.Write(s.rawPkt)
-		} else {
-			err = c.flushSlot(s)
-		}
-		if err != nil && first == nil {
-			first = err
-		}
-		c.release(s)
-	}
-	clear(c.slots)
-	c.slots = c.slots[:0]
-	clear(c.openSlots)
-	c.lastSlot = nil
-
-	return first
-}
-
-func (c *TCPCoalescer) addPassthrough(pkt []byte) {
-	s := c.take()
-	s.passthrough = true
-	s.rawPkt = pkt
-	c.slots = append(c.slots, s)
-}
-
-func (c *TCPCoalescer) seed(pkt []byte, info parsedTCP) {
-	if info.hdrLen > tcpCoalesceHdrCap || info.hdrLen+info.payLen > tcpCoalesceBufSize {
-		// Pathological shape — can't fit our scratch, emit as-is.
-		c.addPassthrough(pkt)
-		return
-	}
-	s := c.take()
-	s.passthrough = false
-	s.rawPkt = nil
-	copy(s.hdrBuf[:], pkt[:info.hdrLen])
-	s.hdrLen = info.hdrLen
-	s.ipHdrLen = info.ipHdrLen
-	s.isV6 = info.fk.isV6
-	s.fk = info.fk
-	s.gsoSize = info.payLen
-	s.numSeg = 1
-	s.totalPay = info.payLen
-	s.nextSeq = info.seq + uint32(info.payLen)
-	s.psh = info.flags&tcpFlagPsh != 0
-	s.payIovs = append(s.payIovs[:0], pkt[info.hdrLen:info.hdrLen+info.payLen])
-	c.slots = append(c.slots, s)
-	if !s.psh {
-		c.openSlots[info.fk] = s
-		c.lastSlot = s
-	} else if last := c.lastSlot; last != nil && last.fk == info.fk {
-		// PSH-on-seed seals the slot immediately. Any prior cached open
-		// slot for this flow has just been sealed-and-replaced by this
-		// passthrough-shaped seed, so drop the cache too.
-		c.lastSlot = nil
-	}
-}
-
-// canAppend reports whether info's packet extends the slot's seed: same
-// header shape and stable contents, adjacent seq, not oversized, chain not closed.
-func (c *TCPCoalescer) canAppend(s *coalesceSlot, pkt []byte, info parsedTCP) bool {
-	if s.psh {
-		return false
-	}
-	if info.hdrLen != s.hdrLen {
-		return false
-	}
-	if info.seq != s.nextSeq {
-		return false
-	}
-	if s.numSeg >= tcpCoalesceMaxSegs {
-		return false
-	}
-	if info.payLen > s.gsoSize {
-		return false
-	}
-	if s.hdrLen+s.totalPay+info.payLen > tcpCoalesceBufSize {
-		return false
-	}
-	// ECE state must be stable across a burst — receivers expect the
-	// flag set on every segment of a CE-echoing window or none.
-	seedFlags := s.hdrBuf[s.ipHdrLen+13]
-	if (seedFlags^info.flags)&tcpFlagEce != 0 {
-		return false
-	}
-	if !headersMatch(s.hdrBuf[:s.hdrLen], pkt[:info.hdrLen], s.isV6, s.ipHdrLen) {
-		return false
-	}
-	return true
-}
-
-func (c *TCPCoalescer) appendPayload(s *coalesceSlot, pkt []byte, info parsedTCP) {
-	s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
-	s.numSeg++
-	s.totalPay += info.payLen
-	s.nextSeq = info.seq + uint32(info.payLen)
-	if info.flags&tcpFlagPsh != 0 {
-		// Propagate PSH into the seed header so kernel TSO sets it on the
-		// last segment. Without this the sender's push signal is dropped.
-		s.hdrBuf[s.ipHdrLen+13] |= tcpFlagPsh
-	}
-	if info.payLen < s.gsoSize || info.flags&tcpFlagPsh != 0 {
-		s.psh = true
-	}
-}
-
-func (c *TCPCoalescer) take() *coalesceSlot {
-	if n := len(c.pool); n > 0 {
-		s := c.pool[n-1]
-		c.pool[n-1] = nil
-		c.pool = c.pool[:n-1]
-		return s
-	}
-	return &coalesceSlot{}
-}
-
-func (c *TCPCoalescer) release(s *coalesceSlot) {
-	s.passthrough = false
-	s.rawPkt = nil
-	clear(s.payIovs)
-	s.payIovs = s.payIovs[:0]
-	s.numSeg = 0
-	s.totalPay = 0
-	s.psh = false
-	c.pool = append(c.pool, s)
-}
-
-// flushSlot patches the header and calls WriteGSO. Does not remove the slot from c.slots.
-func (c *TCPCoalescer) flushSlot(s *coalesceSlot) error {
-	total := s.hdrLen + s.totalPay
-	l4Len := total - s.ipHdrLen
-	hdr := s.hdrBuf[:s.hdrLen]
-
-	if s.isV6 {
-		binary.BigEndian.PutUint16(hdr[4:6], uint16(l4Len))
-	} else {
-		binary.BigEndian.PutUint16(hdr[2:4], uint16(total))
-		hdr[10] = 0
-		hdr[11] = 0
-		binary.BigEndian.PutUint16(hdr[10:12], ipv4HdrChecksum(hdr[:s.ipHdrLen]))
-	}
-
-	var psum uint32
-	if s.isV6 {
-		psum = pseudoSumIPv6(hdr[8:24], hdr[24:40], ipProtoTCP, l4Len)
-	} else {
-		psum = pseudoSumIPv4(hdr[12:16], hdr[16:20], ipProtoTCP, l4Len)
-	}
-	tcsum := s.ipHdrLen + 16
-	binary.BigEndian.PutUint16(hdr[tcsum:tcsum+2], foldOnceNoInvert(psum))
-
-	return c.gsoW.WriteGSO(hdr[:s.ipHdrLen], hdr[s.ipHdrLen:], s.payIovs, tio.GSOProtoTCP)
-}
-
-// headersMatch compares two IP+TCP header prefixes for byte-for-byte
-// equality on every field that must be identical across coalesced
-// segments. Size/IPID/IPCsum/seq/flags/tcpCsum are masked out.
-func headersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
-	if len(a) != len(b) {
-		return false
-	}
-	if !ipHeadersMatch(a, b, isV6) {
-		return false
-	}
-	// TCP: compare [0:4] ports, [8:13] ack+dataoff, [14:16] window,
-	// [18:tcpHdrLen] options (incl. urgent).
-	tcp := ipHdrLen
-	if !bytes.Equal(a[tcp:tcp+4], b[tcp:tcp+4]) {
-		return false
-	}
-	if !bytes.Equal(a[tcp+8:tcp+13], b[tcp+8:tcp+13]) {
-		return false
-	}
-	if !bytes.Equal(a[tcp+14:tcp+16], b[tcp+14:tcp+16]) {
-		return false
-	}
-	if !bytes.Equal(a[tcp+18:], b[tcp+18:]) {
-		return false
-	}
-	return true
-}
-
-// reorderForFlush neutralizes wire-side reorder that the rxOrder buffer
-// couldn't catch (anything crossing a recvmmsg batch boundary). Without
-// this pass a small wire reorder — counter 250 arriving in batch K when
-// 200..249 are coming in batch K+1 — would seed an out-of-seq slot first
-// and emit it ahead of the lower-seq slot, manifesting at the inner TCP
-// receiver as a much larger reorder than the wire actually had.
-//
-// Two phases:
-//  1. Sort each passthrough-bounded segment of c.slots by (flow, seq).
-//     Cross-flow ordering inside a segment isn't preserved (it never was
-//     and doesn't matter for any single flow's TCP correctness).
-//  2. Sweep once and merge adjacent same-flow slots whose ranges are now
-//     contiguous AND whose tail is gsoSize-aligned. The tail constraint
-//     matters because the kernel TSO splitter chops at gsoSize from the
-//     start of the merged payload — a short segment in the middle would
-//     desynchronize every later segment.
-//
-// Passthrough slots act as barriers: the merge check skips them on either
-// side, so a SYN/FIN/RST/CWR is never reordered relative to its flow's
-// data.
-func (c *TCPCoalescer) reorderForFlush() {
-	if len(c.slots) <= 1 {
-		return
-	}
-	runStart := 0
-	for i := 0; i <= len(c.slots); i++ {
-		if i < len(c.slots) && !c.slots[i].passthrough {
-			continue
-		}
-		c.sortRun(c.slots[runStart:i])
-		runStart = i + 1
-	}
-	out := c.slots[:0]
-	logged := false
-	for _, s := range c.slots {
-		if n := len(out); n > 0 {
-			prev := out[n-1]
-			if !prev.passthrough && !s.passthrough && prev.fk == s.fk {
-				// Same-flow neighbors after sort. If they aren't seq-
-				// contiguous it's a real gap — packets the wire reordered
-				// across batches, or actual loss before nebula. Log it so
-				// the operator can quantify how often it happens; the data
-				// itself still emits in seq order, kernel TCP handles the
-				// gap via its OOO queue.
-				if c.l.Enabled(context.Background(), slog.LevelDebug) {
-					if prev.nextSeq != slotSeedSeq(s) {
-						logged = true
-						gap := int64(slotSeedSeq(s)) - int64(prev.nextSeq)
-						c.l.Debug("tcp coalesce: cross-slot seq gap",
-							"src", flowKeyAddr(s.fk, false),
-							"dst", flowKeyAddr(s.fk, true),
-							"sport", s.fk.sport,
-							"dport", s.fk.dport,
-							"prev_seed_seq", slotSeedSeq(prev),
-							"prev_next_seq", prev.nextSeq,
-							"this_seed_seq", slotSeedSeq(s),
-							"gap_bytes", gap,
-							"prev_seg_count", prev.numSeg,
-							"prev_total_pay", prev.totalPay,
-						)
-					}
-				}
-
-				if canMergeSlots(prev, s) {
-					mergeSlots(prev, s)
-					c.release(s)
-					continue
-				}
-			}
-		}
-		out = append(out, s)
-	}
-	if logged {
-		c.l.Warn("==== end of batch ====")
-	}
-	c.slots = out
-}
-
-// flowKeyAddr returns the src or dst address from fk as a netip.Addr for
-// logging. Only used on the cold gap-log path so the netip allocation
-// doesn't matter.
-func flowKeyAddr(fk flowKey, dst bool) netip.Addr {
-	src := fk.src
-	if dst {
-		src = fk.dst
-	}
-	if fk.isV6 {
-		return netip.AddrFrom16(src)
-	}
-	var v4 [4]byte
-	copy(v4[:], src[:4])
-	return netip.AddrFrom4(v4)
-}
-
-// sortRun stable-sorts run by (flowKey, seedSeq) so each flow's slots
-// cluster together in seq order, ready for the merge sweep. Stable so
-// equal-key slots keep their original relative position (defensive — a
-// duplicate seedSeq would already mean something's wrong upstream).
-func (c *TCPCoalescer) sortRun(run []*coalesceSlot) {
-	if len(run) <= 1 {
-		return
-	}
-	// slices.SortStableFunc with a free, non-capturing comparator avoids the
-	// reflection + closure-escape allocations that sort.SliceStable forces.
-	slices.SortStableFunc(run, compareCoalesceSlots)
-}
-
-func compareCoalesceSlots(a, b *coalesceSlot) int {
-	if cmp := flowKeyCompare(a.fk, b.fk); cmp != 0 {
-		return cmp
-	}
-	aSeq, bSeq := slotSeedSeq(a), slotSeedSeq(b)
-	if aSeq == bSeq {
-		return 0
-	}
-	if tcpSeqLess(aSeq, bSeq) {
-		return -1
-	}
-	return 1
-}
-
-// slotSeedSeq returns the TCP seq of the slot's seed (first segment).
-// nextSeq tracks the seq just past the last appended byte; subtracting
-// totalPay walks back to the seed. uint32 wraparound is the right TCP
-// arithmetic so no special-casing is needed.
-func slotSeedSeq(s *coalesceSlot) uint32 {
-	return s.nextSeq - uint32(s.totalPay)
-}
-
-// tcpSeqLess reports whether a precedes b in TCP serial-number arithmetic
-// (RFC 1323 §2.3). The signed int32 cast turns the modular subtraction
-// into the right comparison even across the 2^32 wrap.
-func tcpSeqLess(a, b uint32) bool {
-	return int32(a-b) < 0
-}
-
-// flowKeyCompare orders flowKeys deterministically. The exact ordering
-// is irrelevant — only that same-flow slots cluster together so the
-// post-sort sweep can merge contiguous pairs.
-func flowKeyCompare(a, b flowKey) int {
-	// Cheap scalar fields first so most non-matching keys short-circuit
-	// without ever calling bytes.Compare. sport is the ephemeral port on
-	// egress flows and discriminates fastest. For matching keys (same
-	// flow), array equality on src/dst inlines to word-sized compares,
-	// so we only pay bytes.Compare when the arrays actually differ.
-	if a.sport != b.sport {
-		if a.sport < b.sport {
-			return -1
-		}
-		return 1
-	}
-	if a.dport != b.dport {
-		if a.dport < b.dport {
-			return -1
-		}
-		return 1
-	}
-	if a.dst != b.dst {
-		return bytes.Compare(a.dst[:], b.dst[:])
-	}
-	if a.src != b.src {
-		return bytes.Compare(a.src[:], b.src[:])
-	}
-	if a.isV6 != b.isV6 {
-		if !a.isV6 {
-			return -1
-		}
-		return 1
-	}
-	return 0
-}
-
-// canMergeSlots reports whether s can fold into prev as one merged TSO
-// superpacket. Same flow, contiguous TCP byte range, equal gsoSize, and
-// fits within the kernel TSO limits. The tail-of-prev check rejects any
-// merge whose first slot ended on a sub-gsoSize segment — kernel TSO
-// would split the merged skb at gsoSize boundaries from the start, so a
-// short segment in the middle would corrupt every later segment. PSH and
-// ECE state must agree across both slots: PSH is a semantic delimiter
-// (preserving the sender's push boundary) and ECE state must be uniform
-// across a window (the same rule canAppend enforces for in-flow appends).
-// The IP-level ECN codepoint must also match: this check calls headersMatch
-// → ipHeadersMatch, which compares the full DSCP/ECN byte, so two slots with
-// differing ECN marks stay separate superpackets, each keeping its own mark.
-//
-// Note: a slot sealed by reorder (canAppend returned false on seq
-// mismatch) keeps psh=false, so this restriction does not block the
-// reorder-fix merge — only legitimate PSH-set seals.
-func canMergeSlots(prev, s *coalesceSlot) bool {
-	if prev.psh {
-		return false
-	}
-	if prev.fk != s.fk {
-		return false
-	}
-	if prev.gsoSize != s.gsoSize {
-		return false
-	}
-	if prev.nextSeq != slotSeedSeq(s) {
-		return false
-	}
-	if prev.numSeg+s.numSeg > tcpCoalesceMaxSegs {
-		return false
-	}
-	if prev.hdrLen+prev.totalPay+s.totalPay > tcpCoalesceBufSize {
-		return false
-	}
-	if len(prev.payIovs[len(prev.payIovs)-1]) != prev.gsoSize {
-		return false
-	}
-	prevFlags := prev.hdrBuf[prev.ipHdrLen+13]
-	sFlags := s.hdrBuf[s.ipHdrLen+13]
-	if (prevFlags^sFlags)&tcpFlagEce != 0 {
-		return false
-	}
-	if !headersMatch(prev.hdrBuf[:prev.hdrLen], s.hdrBuf[:s.hdrLen], prev.isV6, prev.ipHdrLen) {
-		return false
-	}
-	return true
-}
-
-// mergeSlots folds src into dst in place: payIovs concatenated, counters
-// and totals updated, PSH OR'd into the seed header so the push signal is
-// not lost. The seed header's seq, gsoSize, and fk are unchanged. Caller
-// is responsible for releasing src (it's no longer in c.slots after this call).
-func mergeSlots(dst, src *coalesceSlot) {
-	dst.payIovs = append(dst.payIovs, src.payIovs...)
-	dst.numSeg += src.numSeg
-	dst.totalPay += src.totalPay
-	dst.nextSeq = src.nextSeq
-	if src.psh {
-		dst.psh = true
-		dst.hdrBuf[dst.ipHdrLen+13] |= tcpFlagPsh
-	}
-}
-
-// ipv4HdrChecksum computes the IPv4 header checksum over hdr (which must
-// already have its checksum field zeroed) and returns the folded/inverted
-// 16-bit value to store.
-func ipv4HdrChecksum(hdr []byte) uint16 {
-	var sum uint32
-	for i := 0; i+1 < len(hdr); i += 2 {
-		sum += uint32(binary.BigEndian.Uint16(hdr[i : i+2]))
-	}
-	if len(hdr)%2 == 1 {
-		sum += uint32(hdr[len(hdr)-1]) << 8
-	}
-	for sum>>16 != 0 {
-		sum = (sum & 0xffff) + (sum >> 16)
-	}
-	return ^uint16(sum)
-}
-
-// pseudoSumIPv4 / pseudoSumIPv6 build the L4 pseudo-header partial sum
-// expected by the virtio NEEDS_CSUM kernel path: the 32-bit accumulator
-// before folding. proto selects the L4 (TCP or UDP); the UDP coalescer
-// reuses these helpers.
-func pseudoSumIPv4(src, dst []byte, proto byte, l4Len int) uint32 {
-	var sum uint32
-	sum += uint32(binary.BigEndian.Uint16(src[0:2]))
-	sum += uint32(binary.BigEndian.Uint16(src[2:4]))
-	sum += uint32(binary.BigEndian.Uint16(dst[0:2]))
-	sum += uint32(binary.BigEndian.Uint16(dst[2:4]))
-	sum += uint32(proto)
-	sum += uint32(l4Len)
-	return sum
-}
-
-func pseudoSumIPv6(src, dst []byte, proto byte, l4Len int) uint32 {
-	var sum uint32
-	for i := 0; i < 16; i += 2 {
-		sum += uint32(binary.BigEndian.Uint16(src[i : i+2]))
-		sum += uint32(binary.BigEndian.Uint16(dst[i : i+2]))
-	}
-	sum += uint32(l4Len >> 16)
-	sum += uint32(l4Len & 0xffff)
-	sum += uint32(proto)
-	return sum
-}
-
-// foldOnceNoInvert folds the 32-bit accumulator to 16 bits and returns it
-// unchanged (no one's complement). This is what virtio NEEDS_CSUM wants in
-// the L4 checksum field — the kernel will add the payload sum and invert.
-func foldOnceNoInvert(sum uint32) uint16 {
-	for sum>>16 != 0 {
-		sum = (sum & 0xffff) + (sum >> 16)
-	}
-	return uint16(sum)
-}

overlay/batch/tcp_coalesce_bench_test.go

@@ -1,241 +0,0 @@
-package batch
-
-import (
-	"encoding/binary"
-	"runtime"
-	"testing"
-
-	"github.com/slackhq/nebula/overlay/tio"
-	"github.com/slackhq/nebula/test"
-)
-
-// nopTunWriter is a zero-alloc tio.GSOWriter for benchmarks. Discards
-// everything but satisfies the interface the coalescer detects.
-type nopTunWriter struct{}
-
-func (nopTunWriter) Write(p []byte) (int, error) { return len(p), nil }
-func (nopTunWriter) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte, _ tio.GSOProto) error {
-	return nil
-}
-func (nopTunWriter) Capabilities() tio.Capabilities {
-	return tio.Capabilities{TSO: true, USO: true}
-}
-
-// buildTCPv4BulkFlow returns a slice of N adjacent ACK-only TCP segments
-// on a single 5-tuple, each carrying payloadLen bytes. Seq numbers are
-// contiguous so every packet is coalesceable onto the previous one.
-func buildTCPv4BulkFlow(n, payloadLen int) [][]byte {
-	pkts := make([][]byte, n)
-	pay := make([]byte, payloadLen)
-	seq := uint32(1000)
-	for i := range n {
-		pkts[i] = buildTCPv4(seq, tcpAck, pay)
-		seq += uint32(payloadLen)
-	}
-	return pkts
-}
-
-// buildTCPv4Interleaved returns nFlows * perFlow packets with per-flow
-// seq continuity but round-robin across flows — worst case for any
-// "last-slot" cache.
-func buildTCPv4Interleaved(nFlows, perFlow, payloadLen int) [][]byte {
-	pay := make([]byte, payloadLen)
-	seqs := make([]uint32, nFlows)
-	for i := range seqs {
-		seqs[i] = uint32(1000 + i*1000000)
-	}
-	pkts := make([][]byte, 0, nFlows*perFlow)
-	for range perFlow {
-		for f := range nFlows {
-			sport := uint16(10000 + f)
-			pkts = append(pkts, buildTCPv4Ports(sport, 2000, seqs[f], tcpAck, pay))
-			seqs[f] += uint32(payloadLen)
-		}
-	}
-	return pkts
-}
-
-// buildICMPv4 returns a minimal non-TCP packet that takes the passthrough
-// branch in Commit.
-func buildICMPv4() []byte {
-	pkt := make([]byte, 28)
-	pkt[0] = 0x45
-	binary.BigEndian.PutUint16(pkt[2:4], 28)
-	pkt[9] = 1 // ICMP
-	copy(pkt[12:16], []byte{10, 0, 0, 1})
-	copy(pkt[16:20], []byte{10, 0, 0, 2})
-	return pkt
-}
-
-// runCommitBench drives Commit over pkts batchSize at a time, flushing
-// between batches, and reports per-packet cost.
-func runCommitBench(b *testing.B, pkts [][]byte, batchSize int) {
-	b.Helper()
-	arena := NewArena(0)
-	c := NewTCPCoalescer(nopTunWriter{}, test.NewLogger(), arena.Reserve, arena.Reset)
-	b.ReportAllocs()
-	b.SetBytes(int64(len(pkts[0])))
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		pkt := pkts[i%len(pkts)]
-		if err := c.Commit(pkt); err != nil {
-			b.Fatal(err)
-		}
-		if (i+1)%batchSize == 0 {
-			if err := c.Flush(); err != nil {
-				b.Fatal(err)
-			}
-		}
-	}
-	// Drain any trailing partial batch so slot state doesn't leak across runs.
-	_ = c.Flush()
-}
-
-// BenchmarkCommitSingleFlow is the bulk-TCP steady state: one flow,
-// contiguous seq, 1200-byte payloads. Every packet past the seed should
-// append onto the open slot. This is the case we most care about.
-func BenchmarkCommitSingleFlow(b *testing.B) {
-	pkts := buildTCPv4BulkFlow(tcpCoalesceMaxSegs, 1200)
-	runCommitBench(b, pkts, tcpCoalesceMaxSegs)
-}
-
-// BenchmarkCommitInterleaved4 has 4 concurrent bulk flows round-robined.
-// A single-entry fast-path cache will miss on every packet; an N-way
-// cache or map lookup carries the weight.
-func BenchmarkCommitInterleaved4(b *testing.B) {
-	pkts := buildTCPv4Interleaved(4, tcpCoalesceMaxSegs, 1200)
-	runCommitBench(b, pkts, len(pkts))
-}
-
-// BenchmarkCommitInterleaved16 stresses the map at higher flow counts.
-func BenchmarkCommitInterleaved16(b *testing.B) {
-	pkts := buildTCPv4Interleaved(16, tcpCoalesceMaxSegs, 1200)
-	runCommitBench(b, pkts, len(pkts))
-}
-
-// BenchmarkCommitPassthrough exercises the non-TCP branch: parseTCPBase
-// bails early and addPassthrough is the only work.
-func BenchmarkCommitPassthrough(b *testing.B) {
-	pkt := buildICMPv4()
-	pkts := make([][]byte, 64)
-	for i := range pkts {
-		pkts[i] = pkt
-	}
-	runCommitBench(b, pkts, 64)
-}
-
-// BenchmarkCommitNonCoalesceableTCP sends SYN|ACK packets on one flow.
-// Each packet takes the "TCP but not admissible" branch which does a
-// map delete + passthrough. Measures the seal-without-slot cost.
-func BenchmarkCommitNonCoalesceableTCP(b *testing.B) {
-	pay := make([]byte, 0)
-	pkts := make([][]byte, 64)
-	for i := range pkts {
-		pkts[i] = buildTCPv4(uint32(1000+i), tcpSyn|tcpAck, pay)
-	}
-	runCommitBench(b, pkts, 64)
-}
-
-// runMultiCommitBench drives MultiCoalescer.Commit. The dispatcher does
-// the IP/L4 parse once and passes the parsed struct to the lane, so this
-// is the bench that shows the savings of skipping the lane's re-parse.
-func runMultiCommitBench(b *testing.B, pkts [][]byte, batchSize int) {
-	b.Helper()
-	m := NewMultiCoalescer(nopTunWriter{}, test.NewLogger(), NewArena(0), true, true)
-	b.ReportAllocs()
-	b.SetBytes(int64(len(pkts[0])))
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		pkt := pkts[i%len(pkts)]
-		if err := m.Commit(pkt); err != nil {
-			b.Fatal(err)
-		}
-		if (i+1)%batchSize == 0 {
-			if err := m.Flush(); err != nil {
-				b.Fatal(err)
-			}
-		}
-	}
-	_ = m.Flush()
-}
-
-// BenchmarkMultiCommitSingleFlow is the multi-lane analogue of
-// BenchmarkCommitSingleFlow — same workload but routed through the
-// dispatcher. The delta vs the single-lane bench measures dispatcher
-// overhead.
-func BenchmarkMultiCommitSingleFlow(b *testing.B) {
-	pkts := buildTCPv4BulkFlow(tcpCoalesceMaxSegs, 1200)
-	runMultiCommitBench(b, pkts, tcpCoalesceMaxSegs)
-}
-
-// BenchmarkMultiCommitInterleaved4 mirrors BenchmarkCommitInterleaved4
-// through the dispatcher.
-func BenchmarkMultiCommitInterleaved4(b *testing.B) {
-	pkts := buildTCPv4Interleaved(4, tcpCoalesceMaxSegs, 1200)
-	runMultiCommitBench(b, pkts, len(pkts))
-}
-
-// flowKeyPair is one comparison input for the flowKeyCompare bench.
-type flowKeyPair struct{ a, b flowKey }
-
-// makeFlowKey builds an IPv4 flowKey from compact inputs.
-func makeFlowKey(srcLow, dstLow uint32, sport, dport uint16) flowKey {
-	var fk flowKey
-	binary.BigEndian.PutUint32(fk.src[12:16], srcLow)
-	binary.BigEndian.PutUint32(fk.dst[12:16], dstLow)
-	fk.sport = sport
-	fk.dport = dport
-	return fk
-}
-
-// flowKeyCases are the workload mixes flowKeyCompare sees in practice.
-//   - sameFlow: equal keys; tests the equal-path cost (sort runs hit this
-//     repeatedly when many segments share a flow).
-//   - sportDiffers: same src/dst/dport, different sport — the typical
-//     "sibling flows from one host to one server" pattern.
-//   - dstDiffers: same src/sport/dport, different dst — outbound to many
-//     servers from a fixed local port.
-//   - allDiffer: every field differs; worst case for short-circuiting.
-func flowKeyCases() map[string][]flowKeyPair {
-	const n = 64
-	cases := map[string][]flowKeyPair{
-		"sameFlow":     make([]flowKeyPair, n),
-		"sportDiffers": make([]flowKeyPair, n),
-		"dstDiffers":   make([]flowKeyPair, n),
-		"allDiffer":    make([]flowKeyPair, n),
-	}
-	for i := range n {
-		base := makeFlowKey(0x0a000001, 0x0a000002, 40000, 443)
-		cases["sameFlow"][i] = flowKeyPair{a: base, b: base}
-		cases["sportDiffers"][i] = flowKeyPair{
-			a: base,
-			b: makeFlowKey(0x0a000001, 0x0a000002, uint16(40001+i), 443),
-		}
-		cases["dstDiffers"][i] = flowKeyPair{
-			a: base,
-			b: makeFlowKey(0x0a000001, uint32(0x0a000002+i+1), 40000, 443),
-		}
-		cases["allDiffer"][i] = flowKeyPair{
-			a: makeFlowKey(uint32(0x0a000001+i), uint32(0x0a000002+i), uint16(40000+i), uint16(80+i)),
-			b: makeFlowKey(uint32(0x0b000001+i), uint32(0x0b000002+i), uint16(50000+i), uint16(443+i)),
-		}
-	}
-	return cases
-}
-
-// BenchmarkFlowKeyCompare measures flowKeyCompare across the workloads
-// the sort step actually sees. Use this to compare reorderings.
-func BenchmarkFlowKeyCompare(b *testing.B) {
-	for name, pairs := range flowKeyCases() {
-		b.Run(name, func(b *testing.B) {
-			b.ReportAllocs()
-			b.ResetTimer()
-			var sink int
-			for i := 0; i < b.N; i++ {
-				p := pairs[i&(len(pairs)-1)]
-				sink += flowKeyCompare(p.a, p.b)
-			}
-			runtime.KeepAlive(sink)
-		})
-	}
-}

overlay/batch/tx_batch.go

@@ -1,60 +0,0 @@
-package batch
-
-import "net/netip"
-
-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.
-// Slots are backed by an Arena (see its docs)
-type SendBatch struct {
-	out   batchWriter
-	bufs  [][]byte
-	dsts  []netip.AddrPort
-	ecns  []byte
-	arena *Arena
-}
-
-// NewSendBatch makes a SendBatch with batchCap slots and an arenaSize byte buffer for slices to back those slots
-func NewSendBatch(out batchWriter, batchCap, arenaSize int) *SendBatch {
-	return &SendBatch{
-		out:   out,
-		bufs:  make([][]byte, 0, batchCap),
-		dsts:  make([]netip.AddrPort, 0, batchCap),
-		ecns:  make([]byte, 0, batchCap),
-		arena: NewArena(arenaSize),
-	}
-}
-
-func (b *SendBatch) Reserve(sz int) []byte {
-	return b.arena.Reserve(sz)
-}
-
-// Len reports how many packets are queued for the next Flush. Callers use
-// it to flush incrementally once a full sendmmsg batch has accumulated,
-// bounding how long the first packet of a large read batch waits.
-func (b *SendBatch) Len() int { return len(b.bufs) }
-
-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

@@ -1,124 +0,0 @@
-package batch
-
-import (
-	"net/netip"
-	"testing"
-)
-
-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, 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, 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, 4)
-	ap := netip.MustParseAddrPort("10.0.0.1:80")
-
-	s1 := b.Reserve(4)
-	pkt1 := append(s1[:0], 0x11, 0x22, 0x33, 0x44)
-	b.Commit(pkt1, ap, 0)
-
-	s2 := b.Reserve(8) // exceeds remaining cap, triggers grow
-	pkt2 := append(s2[:0], 0xA, 0xB, 0xC, 0xD, 0xE)
-	b.Commit(pkt2, ap, 0)
-
-	// pkt1 must still be intact even though backing reallocated.
-	if pkt1[0] != 0x11 || pkt1[3] != 0x44 {
-		t.Fatalf("first packet corrupted by grow: %x", pkt1)
-	}
-
-	if err := b.Flush(); err != nil {
-		t.Fatalf("Flush: %v", err)
-	}
-	if len(fw.bufs) != 2 {
-		t.Fatalf("got %d bufs want 2", len(fw.bufs))
-	}
-	if fw.bufs[0][0] != 0x11 || fw.bufs[0][3] != 0x44 {
-		t.Errorf("first packet on the wire: %x", fw.bufs[0])
-	}
-	if fw.bufs[1][0] != 0xA || fw.bufs[1][4] != 0xE {
-		t.Errorf("second packet on the wire: %x", fw.bufs[1])
-	}
-}

overlay/batch/udp_coalesce.go

@@ -1,355 +0,0 @@
-package batch
-
-import (
-	"encoding/binary"
-	"io"
-
-	"github.com/slackhq/nebula/overlay/tio"
-)
-
-// ipProtoUDP is the IANA protocol number for UDP.
-const ipProtoUDP = 17
-
-// udpCoalesceBufSize caps total bytes per UDP superpacket. Mirrors the
-// kernel's gso_max_size; payloads beyond this are emitted as-is.
-const udpCoalesceBufSize = 65535
-
-// udpCoalesceMaxSegs caps how many segments we'll coalesce. Kernel UDP-GSO
-// accepts up to 64 segments per skb (UDP_MAX_SEGMENTS); stay under that.
-const udpCoalesceMaxSegs = 64
-
-// udpCoalesceHdrCap is the scratch space we copy a seed's IP+UDP header
-// into. IPv6 (40) + UDP (8) = 48; round up for safety.
-const udpCoalesceHdrCap = 64
-
-// udpSlot is one entry in the UDPCoalescer's ordered event queue.
-type udpSlot struct {
-	passthrough bool
-	rawPkt      []byte // borrowed when passthrough
-
-	fk       flowKey
-	hdrBuf   [udpCoalesceHdrCap]byte
-	hdrLen   int
-	ipHdrLen int
-	isV6     bool
-	gsoSize  int // per-segment UDP payload length
-	numSeg   int
-	totalPay int
-	// sealed closes the chain: set when a sub-gsoSize segment is appended
-	// (kernel UDP-GSO requires every segment but the last to be exactly
-	// gsoSize) or when limits are hit. No further appends after.
-	sealed  bool
-	payIovs [][]byte
-}
-
-// UDPCoalescer accumulates adjacent in-flow UDP datagrams across multiple
-// concurrent flows and emits each flow's run as a single GSO_UDP_L4
-// superpacket via tio.GSOWriter. Falls back to per-packet writes when the
-// underlying writer doesn't support USO.
-//
-// All output — coalesced or not — is deferred until Flush so per-flow
-// arrival order is preserved on the wire. Cross-flow order is NOT preserved
-// across the TCP/UDP/passthrough split when this coalescer runs alongside
-// others — see multi_coalesce.go. Per-flow order is preserved because a
-// single 5-tuple only ever lands in one lane and each lane preserves its
-// own slot order.
-//
-// Owns no locks; one coalescer per TUN write queue.
-type UDPCoalescer struct {
-	plainW io.Writer
-	gsoW   tio.GSOWriter // nil when the queue can't accept GSO_UDP_L4
-
-	slots     []*udpSlot
-	openSlots map[flowKey]*udpSlot
-	pool      []*udpSlot
-	reserver  Reserver
-	resetter  Resetter
-}
-
-// NewUDPCoalescer wraps w. The caller is responsible for only constructing
-// this when the underlying Queue's Capabilities advertise USO; otherwise
-// the kernel may reject GSO_UDP_L4 writes. If w does not implement
-// tio.GSOWriter at all (single-packet Queue), the coalescer degrades to
-// plain Writes — same defensive shape as the TCP coalescer.
-func NewUDPCoalescer(w io.Writer, reserver Reserver, resetter Resetter) *UDPCoalescer {
-	c := &UDPCoalescer{
-		plainW:    w,
-		slots:     make([]*udpSlot, 0, initialSlots),
-		openSlots: make(map[flowKey]*udpSlot, initialSlots),
-		pool:      make([]*udpSlot, 0, initialSlots),
-		reserver:  reserver,
-		resetter:  resetter,
-	}
-	if gw, ok := tio.SupportsGSO(w, tio.GSOProtoUDP); ok {
-		c.gsoW = gw
-	}
-	return c
-}
-
-// parsedUDP holds the fields extracted from a single parse so later steps
-// (admission, slot lookup, canAppend) don't re-walk the header.
-type parsedUDP struct {
-	fk       flowKey
-	ipHdrLen int
-	hdrLen   int // ipHdrLen + 8
-	payLen   int
-}
-
-// parseUDP extracts the flow key and IP/UDP offsets for a UDP packet.
-// Returns ok=false for non-UDP, malformed, or unsupported header shapes
-// (IPv4 with options/fragmentation, IPv6 with extension headers).
-func parseUDP(pkt []byte) (parsedUDP, bool) {
-	var p parsedUDP
-	ip, ok := parseIPPrologue(pkt, ipProtoUDP)
-	if !ok {
-		return p, false
-	}
-	pkt = ip.pkt
-	p.fk = ip.fk
-	p.ipHdrLen = ip.ipHdrLen
-
-	if len(pkt) < p.ipHdrLen+8 {
-		return p, false
-	}
-	p.hdrLen = p.ipHdrLen + 8
-	// UDP `length` field: must equal IP-derived length-of-UDP-header-plus-payload.
-	udpLen := int(binary.BigEndian.Uint16(pkt[p.ipHdrLen+4 : p.ipHdrLen+6]))
-	if udpLen < 8 || udpLen > len(pkt)-p.ipHdrLen {
-		return p, false
-	}
-	p.payLen = udpLen - 8
-	p.fk.sport = binary.BigEndian.Uint16(pkt[p.ipHdrLen : p.ipHdrLen+2])
-	p.fk.dport = binary.BigEndian.Uint16(pkt[p.ipHdrLen+2 : p.ipHdrLen+4])
-	return p, true
-}
-
-func (c *UDPCoalescer) Reserve(sz int) []byte {
-	return c.reserver(sz)
-}
-
-// Commit borrows pkt. The caller must keep pkt valid until the next Flush.
-func (c *UDPCoalescer) Commit(pkt []byte) error {
-	if c.gsoW == nil {
-		c.addPassthrough(pkt)
-		return nil
-	}
-	info, ok := parseUDP(pkt)
-	if !ok {
-		c.addPassthrough(pkt)
-		return nil
-	}
-	return c.commitParsed(pkt, info)
-}
-
-// commitParsed is the post-parse half of Commit. The caller must have
-// already verified parseUDP succeeded. Used by MultiCoalescer.Commit to
-// avoid re-walking the IP/UDP header.
-func (c *UDPCoalescer) commitParsed(pkt []byte, info parsedUDP) error {
-	if c.gsoW == nil {
-		c.addPassthrough(pkt)
-		return nil
-	}
-	// A zero-length UDP datagram (UDP `length` == 8) is legal and must still
-	// reach the TUN, but it can't be coalesced: a GSO slot would store an
-	// empty payload iovec and the kernel has nothing to segment. Seal any
-	// open chain for this flow (so a later, non-empty datagram seeds fresh
-	// *after* this one and per-flow arrival order is preserved) and deliver
-	// it as a plain single datagram.
-	if info.payLen == 0 {
-		delete(c.openSlots, info.fk)
-		c.addPassthrough(pkt)
-		return nil
-	}
-	if open := c.openSlots[info.fk]; open != nil {
-		if c.canAppend(open, pkt, info) {
-			c.appendPayload(open, pkt, info)
-			if open.sealed {
-				delete(c.openSlots, info.fk)
-			}
-			return nil
-		}
-		// Can't extend — seal it and fall through to seed a fresh slot.
-		delete(c.openSlots, info.fk)
-	}
-	c.seed(pkt, info)
-	return nil
-}
-
-// Flush drains every queued slot and calls the configured Resetter.
-func (c *UDPCoalescer) Flush() error {
-	first := c.drain()
-	if c.resetter != nil {
-		c.resetter()
-	}
-	return first
-}
-
-// drain emits every queued slot in arrival order and clears the slot state.
-// It does NOT reset the arena: borrowed payload slices stay valid until the
-// arena's owner recycles it.
-func (c *UDPCoalescer) drain() error {
-	var first error
-	for _, s := range c.slots {
-		var err error
-		if s.passthrough {
-			_, err = c.plainW.Write(s.rawPkt)
-		} else {
-			err = c.flushSlot(s)
-		}
-		if err != nil && first == nil {
-			first = err
-		}
-		c.release(s)
-	}
-	clear(c.slots)
-	c.slots = c.slots[:0]
-	clear(c.openSlots)
-	return first
-}
-
-func (c *UDPCoalescer) addPassthrough(pkt []byte) {
-	s := c.take()
-	s.passthrough = true
-	s.rawPkt = pkt
-	c.slots = append(c.slots, s)
-}
-
-func (c *UDPCoalescer) seed(pkt []byte, info parsedUDP) {
-	if info.hdrLen > udpCoalesceHdrCap || info.hdrLen+info.payLen > udpCoalesceBufSize {
-		c.addPassthrough(pkt)
-		return
-	}
-	s := c.take()
-	s.passthrough = false
-	s.rawPkt = nil
-	copy(s.hdrBuf[:], pkt[:info.hdrLen])
-	s.hdrLen = info.hdrLen
-	s.ipHdrLen = info.ipHdrLen
-	s.isV6 = info.fk.isV6
-	s.fk = info.fk
-	s.gsoSize = info.payLen
-	s.numSeg = 1
-	s.totalPay = info.payLen
-	s.sealed = false
-	s.payIovs = append(s.payIovs[:0], pkt[info.hdrLen:info.hdrLen+info.payLen])
-	c.slots = append(c.slots, s)
-	c.openSlots[info.fk] = s
-}
-
-// canAppend reports whether info's packet extends the slot's seed.
-// Kernel UDP-GSO requires every segment except possibly the last to be
-// exactly gsoSize, and the last may be shorter (≤ gsoSize).
-func (c *UDPCoalescer) canAppend(s *udpSlot, pkt []byte, info parsedUDP) bool {
-	if s.sealed {
-		return false
-	}
-	if info.hdrLen != s.hdrLen {
-		return false
-	}
-	if s.numSeg >= udpCoalesceMaxSegs {
-		return false
-	}
-	if info.payLen > s.gsoSize {
-		return false
-	}
-	if s.hdrLen+s.totalPay+info.payLen > udpCoalesceBufSize {
-		return false
-	}
-	if !udpHeadersMatch(s.hdrBuf[:s.hdrLen], pkt[:info.hdrLen], s.isV6, s.ipHdrLen) {
-		return false
-	}
-	return true
-}
-
-func (c *UDPCoalescer) appendPayload(s *udpSlot, pkt []byte, info parsedUDP) {
-	s.payIovs = append(s.payIovs, pkt[info.hdrLen:info.hdrLen+info.payLen])
-	s.numSeg++
-	s.totalPay += info.payLen
-	if info.payLen < s.gsoSize {
-		// Last-segment-can-be-shorter: this seals the chain.
-		s.sealed = true
-	}
-}
-
-func (c *UDPCoalescer) take() *udpSlot {
-	if n := len(c.pool); n > 0 {
-		s := c.pool[n-1]
-		c.pool[n-1] = nil
-		c.pool = c.pool[:n-1]
-		return s
-	}
-	return &udpSlot{}
-}
-
-func (c *UDPCoalescer) release(s *udpSlot) {
-	s.passthrough = false
-	s.rawPkt = nil
-	clear(s.payIovs)
-	s.payIovs = s.payIovs[:0]
-	s.numSeg = 0
-	s.totalPay = 0
-	s.sealed = false
-	c.pool = append(c.pool, s)
-}
-
-// flushSlot patches the IP header total length / IPv6 payload length and
-// the UDP length to the *total* across all coalesced segments, then seeds
-// the UDP checksum field with the pseudo-header partial (single-fold, not
-// inverted) per virtio NEEDS_CSUM. The kernel's ip_rcv_core (v4) and
-// ip6_rcv_core (v6) trim the skb to those length fields, so per-segment
-// values would silently drop everything but the first segment. The kernel
-// then walks each segment in __udp_gso_segment, recomputing per-segment
-// uh->len / iph->tot_len / IPv6 plen and adjusting the checksum via
-// `check = csum16_add(csum16_sub(uh->check, uh->len), newlen)` — meaning
-// our seed's uh->check must be consistent with the seed's uh->len, which
-// is what passing the total to both pseudoSum and the UDP length field
-// guarantees.
-func (c *UDPCoalescer) flushSlot(s *udpSlot) error {
-	hdr := s.hdrBuf[:s.hdrLen]
-	total := s.hdrLen + s.totalPay // full IP+UDP+all_payloads bytes
-	l4Len := total - s.ipHdrLen    // total UDP (8 + sum of payloads)
-
-	if s.isV6 {
-		binary.BigEndian.PutUint16(hdr[4:6], uint16(l4Len))
-	} else {
-		binary.BigEndian.PutUint16(hdr[2:4], uint16(total))
-		hdr[10] = 0
-		hdr[11] = 0
-		binary.BigEndian.PutUint16(hdr[10:12], ipv4HdrChecksum(hdr[:s.ipHdrLen]))
-	}
-
-	// UDP length field (offset 4 inside the UDP header) = total UDP size.
-	binary.BigEndian.PutUint16(hdr[s.ipHdrLen+4:s.ipHdrLen+6], uint16(l4Len))
-
-	var psum uint32
-	if s.isV6 {
-		psum = pseudoSumIPv6(hdr[8:24], hdr[24:40], ipProtoUDP, l4Len)
-	} else {
-		psum = pseudoSumIPv4(hdr[12:16], hdr[16:20], ipProtoUDP, l4Len)
-	}
-	udpCsumOff := s.ipHdrLen + 6
-	binary.BigEndian.PutUint16(hdr[udpCsumOff:udpCsumOff+2], foldOnceNoInvert(psum))
-
-	return c.gsoW.WriteGSO(hdr[:s.ipHdrLen], hdr[s.ipHdrLen:], s.payIovs, tio.GSOProtoUDP)
-}
-
-// udpHeadersMatch compares two IP+UDP header prefixes for byte-equality on
-// every field that must be identical across coalesced segments. Length
-// fields are masked out (flushSlot rewrites them), but the IP-level ECN
-// codepoint is compared (via ipHeadersMatch) so segments with differing ECN
-// don't coalesce, matching kernel GRO.
-func udpHeadersMatch(a, b []byte, isV6 bool, ipHdrLen int) bool {
-	if len(a) != len(b) {
-		return false
-	}
-	if !ipHeadersMatch(a, b, isV6) {
-		return false
-	}
-	// UDP: compare sport+dport ([0:4]). Skip length [4:6] and checksum [6:8] —
-	// length varies (we rewrite at flush) and the checksum will be redone.
-	udp := ipHdrLen
-	if a[udp] != b[udp] || a[udp+1] != b[udp+1] || a[udp+2] != b[udp+2] || a[udp+3] != b[udp+3] {
-		return false
-	}
-	return true
-}

overlay/batch/udp_coalesce_test.go

@@ -1,472 +0,0 @@
-package batch
-
-import (
-	"encoding/binary"
-	"testing"
-)
-
-// buildUDPv4 builds a minimal IPv4+UDP packet with the given payload and ports.
-func buildUDPv4(sport, dport uint16, payload []byte) []byte {
-	const ipHdrLen = 20
-	const udpHdrLen = 8
-	total := ipHdrLen + udpHdrLen + len(payload)
-	pkt := make([]byte, total)
-
-	pkt[0] = 0x45
-	pkt[1] = 0x00
-	binary.BigEndian.PutUint16(pkt[2:4], uint16(total))
-	binary.BigEndian.PutUint16(pkt[4:6], 0)
-	binary.BigEndian.PutUint16(pkt[6:8], 0x4000)
-	pkt[8] = 64
-	pkt[9] = ipProtoUDP
-	copy(pkt[12:16], []byte{10, 0, 0, 1})
-	copy(pkt[16:20], []byte{10, 0, 0, 2})
-
-	binary.BigEndian.PutUint16(pkt[20:22], sport)
-	binary.BigEndian.PutUint16(pkt[22:24], dport)
-	binary.BigEndian.PutUint16(pkt[24:26], uint16(udpHdrLen+len(payload)))
-	binary.BigEndian.PutUint16(pkt[26:28], 0)
-
-	copy(pkt[28:], payload)
-	return pkt
-}
-
-// buildUDPv6 builds a minimal IPv6+UDP packet.
-func buildUDPv6(sport, dport uint16, payload []byte) []byte {
-	const ipHdrLen = 40
-	const udpHdrLen = 8
-	total := ipHdrLen + udpHdrLen + len(payload)
-	pkt := make([]byte, total)
-
-	pkt[0] = 0x60
-	binary.BigEndian.PutUint16(pkt[4:6], uint16(udpHdrLen+len(payload)))
-	pkt[6] = ipProtoUDP
-	pkt[7] = 64
-	pkt[8] = 0xfe
-	pkt[9] = 0x80
-	pkt[23] = 1
-	pkt[24] = 0xfe
-	pkt[25] = 0x80
-	pkt[39] = 2
-
-	binary.BigEndian.PutUint16(pkt[40:42], sport)
-	binary.BigEndian.PutUint16(pkt[42:44], dport)
-	binary.BigEndian.PutUint16(pkt[44:46], uint16(udpHdrLen+len(payload)))
-	binary.BigEndian.PutUint16(pkt[46:48], 0)
-
-	copy(pkt[48:], payload)
-	return pkt
-}
-
-func TestUDPCoalescerPassthroughWhenGSOUnavailable(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: false}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pkt := buildUDPv4(1000, 53, make([]byte, 100))
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 0 || len(w.gsoWrites) != 0 {
-		t.Fatalf("no Add-time writes: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
-		t.Fatalf("want single plain write, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-}
-
-func TestUDPCoalescerNonUDPPassthrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	// ICMP packet
-	pkt := make([]byte, 28)
-	pkt[0] = 0x45
-	binary.BigEndian.PutUint16(pkt[2:4], 28)
-	pkt[9] = 1
-	copy(pkt[12:16], []byte{10, 0, 0, 1})
-	copy(pkt[16:20], []byte{10, 0, 0, 2})
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
-		t.Fatalf("ICMP must pass through unchanged: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-}
-
-func TestUDPCoalescerSeedThenFlushAlone(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pkt := buildUDPv4(1000, 53, make([]byte, 800))
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	// Single-segment flush goes through WriteGSO; the writer infers GSO_NONE
-	// from len(pays)==1 and the kernel fills in the UDP csum (NEEDS_CSUM).
-	if len(w.gsoWrites) != 1 || len(w.writes) != 0 {
-		t.Fatalf("single-seg flush: writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-}
-
-func TestUDPCoalescerCoalescesEqualSized(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pay := make([]byte, 1200)
-	for i := 0; i < 3; i++ {
-		if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
-			t.Fatal(err)
-		}
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 1 {
-		t.Fatalf("want 1 gso write, got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
-	}
-	g := w.gsoWrites[0]
-	if g.gsoSize != 1200 {
-		t.Errorf("gsoSize=%d want 1200", g.gsoSize)
-	}
-	if len(g.pays) != 3 {
-		t.Errorf("pay count=%d want 3", len(g.pays))
-	}
-	if g.csumStart != 20 {
-		t.Errorf("csumStart=%d want 20", g.csumStart)
-	}
-	// IP totalLen and UDP length must be the TOTAL across all segments —
-	// the kernel's ip_rcv_core trims skbs to iph->tot_len, so a per-segment
-	// value would silently drop everything but the first segment. Total =
-	// IP(20) + UDP(8) + 3*1200 = 3628.
-	gotTotalLen := binary.BigEndian.Uint16(g.hdr[2:4])
-	if gotTotalLen != 3628 {
-		t.Errorf("ipv4 total_len=%d want 3628 (must be total across segments)", gotTotalLen)
-	}
-	gotUDPLen := binary.BigEndian.Uint16(g.hdr[20+4 : 20+6])
-	if gotUDPLen != 8+3*1200 {
-		t.Errorf("udp len=%d want %d", gotUDPLen, 8+3*1200)
-	}
-}
-
-// Last segment may be shorter, sealing the chain.
-func TestUDPCoalescerShortLastSegmentSeals(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	full := make([]byte, 1200)
-	tail := make([]byte, 600)
-	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(1000, 53, tail)); err != nil {
-		t.Fatal(err)
-	}
-	// A 4th packet, even same-sized, must NOT join — chain is sealed.
-	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 2 {
-		t.Fatalf("want 2 gso writes (sealed + new seed), got %d", len(w.gsoWrites))
-	}
-	if len(w.gsoWrites[0].pays) != 3 {
-		t.Errorf("first super: want 3 pays, got %d", len(w.gsoWrites[0].pays))
-	}
-	if len(w.gsoWrites[1].pays) != 1 {
-		t.Errorf("second super: want 1 pay (re-seed), got %d", len(w.gsoWrites[1].pays))
-	}
-}
-
-// A larger-than-gsoSize packet cannot extend the slot — it reseeds.
-func TestUDPCoalescerLargerThanSeedReseeds(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	if err := c.Commit(buildUDPv4(1000, 53, make([]byte, 800))); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(1000, 53, make([]byte, 1200))); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 2 {
-		t.Fatalf("want 2 separate seeds, got %d", len(w.gsoWrites))
-	}
-}
-
-// Different 5-tuples must not coalesce.
-func TestUDPCoalescerDifferentFlowsKeepSeparate(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pay := make([]byte, 800)
-	if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(2000, 53, pay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(2000, 53, pay)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	// Two flows × 2 datagrams each = 2 superpackets of 2 segments.
-	if len(w.gsoWrites) != 2 {
-		t.Fatalf("want 2 gso writes (one per flow), got %d", len(w.gsoWrites))
-	}
-	for i, g := range w.gsoWrites {
-		if len(g.pays) != 2 {
-			t.Errorf("super %d: want 2 pays, got %d", i, len(g.pays))
-		}
-	}
-}
-
-// Caps at udpCoalesceMaxSegs.
-func TestUDPCoalescerCapsAtMaxSegs(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pay := make([]byte, 100)
-	for i := 0; i < udpCoalesceMaxSegs+5; i++ {
-		if err := c.Commit(buildUDPv4(1000, 53, pay)); err != nil {
-			t.Fatal(err)
-		}
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	// First superpacket holds udpCoalesceMaxSegs segments; the spillover
-	// reseeds a new one.
-	if len(w.gsoWrites) != 2 {
-		t.Fatalf("want 2 gso writes (cap then reseed), got %d", len(w.gsoWrites))
-	}
-	if len(w.gsoWrites[0].pays) != udpCoalesceMaxSegs {
-		t.Errorf("first super: pays=%d want %d", len(w.gsoWrites[0].pays), udpCoalesceMaxSegs)
-	}
-	if len(w.gsoWrites[1].pays) != 5 {
-		t.Errorf("second super: pays=%d want 5", len(w.gsoWrites[1].pays))
-	}
-}
-
-// Differing IP ECN codepoints must not coalesce: udpHeadersMatch compares
-// the full ToS byte (matching kernel GRO). A CE-marked datagram mid-run
-// seals the Not-ECT chain and seeds a fresh superpacket that keeps CE; the
-// trailing Not-ECT datagram seeds another.
-func TestUDPCoalescerDifferingECNReseeds(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pay := make([]byte, 800)
-	pkt0 := buildUDPv4(1000, 53, pay) // ECN=00 (Not-ECT)
-	pkt1 := buildUDPv4(1000, 53, pay)
-	pkt1[1] = 0x03                    // CE
-	pkt2 := buildUDPv4(1000, 53, pay) // ECN=00 again
-	for _, p := range [][]byte{pkt0, pkt1, pkt2} {
-		if err := c.Commit(p); err != nil {
-			t.Fatal(err)
-		}
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 3 {
-		t.Fatalf("want 3 separate seeds (differing ECN), got %d (plain=%d)", len(w.gsoWrites), len(w.writes))
-	}
-	wantECN := []byte{0x00, 0x03, 0x00}
-	for i, g := range w.gsoWrites {
-		if len(g.pays) != 1 {
-			t.Errorf("gso %d pay count=%d want 1", i, len(g.pays))
-		}
-		if got := g.hdr[1] & 0x03; got != wantECN[i] {
-			t.Errorf("gso %d ECN=%#x want %#x", i, got, wantECN[i])
-		}
-	}
-}
-
-// IPv6 path: same flow, equal-sized → coalesced.
-func TestUDPCoalescerIPv6Coalesces(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pay := make([]byte, 1200)
-	for i := 0; i < 3; i++ {
-		if err := c.Commit(buildUDPv6(1000, 53, pay)); err != nil {
-			t.Fatal(err)
-		}
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 1 {
-		t.Fatalf("want 1 gso write, got %d", len(w.gsoWrites))
-	}
-	g := w.gsoWrites[0]
-	if !g.isV6 {
-		t.Errorf("expected v6 write")
-	}
-	if g.csumStart != 40 {
-		t.Errorf("csumStart=%d want 40", g.csumStart)
-	}
-	// IPv6 payload_len and UDP length must be TOTAL — kernel's
-	// ip6_rcv_core trims to payload_len + ipv6 hdr size. Total UDP = 8 +
-	// 3*1200 = 3608.
-	gotPlen := binary.BigEndian.Uint16(g.hdr[4:6])
-	if gotPlen != 8+3*1200 {
-		t.Errorf("ipv6 payload_len=%d want %d (must be total)", gotPlen, 8+3*1200)
-	}
-	gotUDPLen := binary.BigEndian.Uint16(g.hdr[40+4 : 40+6])
-	if gotUDPLen != 8+3*1200 {
-		t.Errorf("udp len=%d want %d", gotUDPLen, 8+3*1200)
-	}
-}
-
-// DSCP differences must reseed: udpHeadersMatch compares the full ToS byte.
-func TestUDPCoalescerDSCPMismatchReseeds(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pay := make([]byte, 800)
-	pkt0 := buildUDPv4(1000, 53, pay)
-	pkt1 := buildUDPv4(1000, 53, pay)
-	pkt1[1] = 0xb8 // EF DSCP, ECN=0
-	if err := c.Commit(pkt0); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(pkt1); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.gsoWrites) != 2 {
-		t.Fatalf("want 2 separate seeds (different DSCP), got %d", len(w.gsoWrites))
-	}
-}
-
-// Fragmented IPv4 must not be coalesced.
-func TestUDPCoalescerFragmentedIPv4PassesThrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pkt := buildUDPv4(1000, 53, make([]byte, 200))
-	binary.BigEndian.PutUint16(pkt[6:8], 0x2000) // MF=1
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
-		t.Fatalf("frag must pass through plain, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-}
-
-// A zero-length UDP datagram (UDP length == 8, no payload) is legal and
-// must be delivered as a plain single datagram — never coalesced. Seeding
-// it into a GSO slot stores an empty payload iovec that panics WriteGSO
-// (index-out-of-range on &pay[0]); this is a remote DoS if we ever let it
-// reach the GSO path. Regression: must not panic and must be written.
-func TestUDPCoalescerZeroLengthPayloadPassesThrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pkt := buildUDPv4(1000, 53, nil) // UDP length 8, zero payload
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
-		t.Fatalf("zero-length UDP must pass through plain, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-	if len(w.writes[0]) != len(pkt) {
-		t.Errorf("delivered %d bytes, want the whole %d-byte datagram", len(w.writes[0]), len(pkt))
-	}
-}
-
-// IPv6 zero-length UDP datagram: same passthrough contract as v4.
-func TestUDPCoalescerZeroLengthPayloadIPv6PassesThrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pkt := buildUDPv6(1000, 53, nil) // UDP length 8, zero payload
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
-		t.Fatalf("zero-length IPv6 UDP must pass through plain, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-	if len(w.writes[0]) != len(pkt) {
-		t.Errorf("delivered %d bytes, want the whole %d-byte datagram", len(w.writes[0]), len(pkt))
-	}
-}
-
-// A zero-length datagram arriving mid-flow must seal the open chain so the
-// datagram after it seeds a fresh superpacket *after* the empty one on the
-// wire — per-flow arrival order (full, empty, full) must be preserved.
-func TestUDPCoalescerZeroLengthMidFlowSealsAndPreservesOrder(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	full := make([]byte, 800)
-	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(1000, 53, nil)); err != nil { // zero-length
-		t.Fatal(err)
-	}
-	if err := c.Commit(buildUDPv4(1000, 53, full)); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	// The empty datagram sealed the first slot, so the trailing full packet
-	// can't join it: two single-segment superpackets bracket one plain write.
-	if len(w.gsoWrites) != 2 || len(w.writes) != 1 {
-		t.Fatalf("want 2 gso writes + 1 plain, got gso=%d plain=%d", len(w.gsoWrites), len(w.writes))
-	}
-}
-
-// IPv4 with options is not admissible (we require IHL=5).
-func TestUDPCoalescerIPv4WithOptionsPassesThrough(t *testing.T) {
-	w := &fakeTunWriter{gsoEnabled: true}
-	arena := NewArena(0)
-	c := NewUDPCoalescer(w, arena.Reserve, arena.Reset)
-	pkt := buildUDPv4(1000, 53, make([]byte, 200))
-	pkt[0] = 0x46 // IHL = 6 (24-byte IPv4 header — has options)
-	if err := c.Commit(pkt); err != nil {
-		t.Fatal(err)
-	}
-	if err := c.Flush(); err != nil {
-		t.Fatal(err)
-	}
-	if len(w.writes) != 1 || len(w.gsoWrites) != 0 {
-		t.Fatalf("ipv4-with-options must pass through plain, got writes=%d gso=%d", len(w.writes), len(w.gsoWrites))
-	}
-}

overlay/checksum/checksum_amd64.go

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

overlay/checksum/checksum_amd64.s

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

overlay/checksum/checksum_arm64.go

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

overlay/checksum/checksum_arm64.s

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

overlay/checksum/checksum_fallback.go

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

overlay/checksum/checksum_test.go

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

overlay/device.go

@@ -4,25 +4,15 @@ import (
 	"io"
 	"net/netip"
 
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
-// defaultBatchBufSize is the per-Queue scratch size for Read on backends
-// that don't do TSO segmentation. 65535 covers any single IP packet.
-const defaultBatchBufSize = 65535
-
 type Device interface {
-	io.Closer
+	io.ReadWriteCloser
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
-	// Queues returns the device's packet queues, opening additional ones as
-	// needed until there are n. Platforms without multiqueue support return
-	// 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)
+	SupportsMultiqueue() bool
+	NewMultiQueueReader() (io.ReadWriteCloser, error)
 }

overlay/overlaytest/noop.go

@@ -3,9 +3,10 @@
 package overlaytest
 
 import (
+	"errors"
+	"io"
 	"net/netip"
 
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -30,16 +31,20 @@ func (NoopTun) Name() string {
 	return "noop"
 }
 
-func (NoopTun) Read() ([]tio.Packet, error) {
-	return nil, nil
+func (NoopTun) Read([]byte) (int, error) {
+	return 0, nil
 }
 
 func (NoopTun) Write([]byte) (int, error) {
 	return 0, nil
 }
 
-func (NoopTun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{NoopTun{}}, nil
+func (NoopTun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (NoopTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, errors.New("unsupported")
 }
 
 func (NoopTun) Close() error {

overlay/tio/blockon_linux.go

@@ -1,45 +0,0 @@
-//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_gso_linux.go

@@ -1,99 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-package tio
-
-import (
-	"encoding/binary"
-	"errors"
-	"fmt"
-	"sync/atomic"
-
-	"golang.org/x/sys/unix"
-)
-
-type offloadQueueSet struct {
-	pq []*Offload
-	// pqi is exactly the same as pq, but stored as the interface type
-	pqi        []Queue
-	shutdownFd int
-	// usoEnabled is true when newTun successfully negotiated TUN_F_USO4|6
-	// with the kernel. Queues created by Add inherit this and surface it
-	// via Offload.USOSupported so coalescers can gate USO emission.
-	usoEnabled bool
-	closed     atomic.Bool
-}
-
-// NewOffloadQueueSet creates a QueueSet that uses virtio_net_hdr to do
-// TSO segmentation in userspace. usoEnabled tells downstream queues whether
-// the kernel agreed to deliver/accept GSO_UDP_L4 superpackets — coalescers
-// should fall back to per-packet writes when this is false.
-func NewOffloadQueueSet(usoEnabled bool) (QueueSet, error) {
-	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create eventfd: %w", err)
-	}
-
-	out := &offloadQueueSet{
-		pq:         []*Offload{},
-		pqi:        []Queue{},
-		shutdownFd: shutdownFd,
-		usoEnabled: usoEnabled,
-	}
-
-	return out, nil
-}
-
-func (c *offloadQueueSet) Queues() []Queue {
-	return c.pqi
-}
-
-func (c *offloadQueueSet) Add(fd int) error {
-	x, err := newOffload(fd, c.shutdownFd, c.usoEnabled)
-	if err != nil {
-		return err
-	}
-	c.pq = append(c.pq, x)
-	c.pqi = append(c.pqi, x)
-
-	return nil
-}
-
-func (c *offloadQueueSet) wakeForShutdown() error {
-	var buf [8]byte
-	binary.NativeEndian.PutUint64(buf[:], 1)
-	_, err := unix.Write(c.shutdownFd, buf[:])
-	return err
-}
-
-func (c *offloadQueueSet) Close() error {
-	if c.closed.Swap(true) {
-		return nil
-	}
-
-	errs := []error{}
-
-	// Signal all readers blocked in poll to wake up and exit. They observe
-	// POLLIN on the shutdown eventfd and return 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)
-		}
-	}
-
-	// Close the shutdown eventfd last: every reader's pollfd set references
-	// it, so it must outlive the wake + per-queue teardown above.
-	if err := unix.Close(c.shutdownFd); err != nil {
-		errs = append(errs, err)
-	}
-	c.shutdownFd = -1
-
-	return errors.Join(errs...)
-}

overlay/tio/queueset_poll_linux.go

@@ -1,90 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-package tio
-
-import (
-	"encoding/binary"
-	"errors"
-	"fmt"
-	"sync/atomic"
-
-	"golang.org/x/sys/unix"
-)
-
-type pollQueueSet struct {
-	pq []*Poll
-	// pqi is exactly the same as pq, but stored as the interface type
-	pqi        []Queue
-	shutdownFd int
-	closed     atomic.Bool
-}
-
-func NewPollQueueSet() (QueueSet, error) {
-	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create eventfd: %w", err)
-	}
-
-	out := &pollQueueSet{
-		pq:         []*Poll{},
-		pqi:        []Queue{},
-		shutdownFd: shutdownFd,
-	}
-
-	return out, nil
-}
-
-func (c *pollQueueSet) Queues() []Queue {
-	return c.pqi
-}
-
-func (c *pollQueueSet) Add(fd int) error {
-	x, err := newPoll(fd, c.shutdownFd)
-	if err != nil {
-		return err
-	}
-	c.pq = append(c.pq, x)
-	c.pqi = append(c.pqi, x)
-
-	return nil
-}
-
-func (c *pollQueueSet) wakeForShutdown() error {
-	var buf [8]byte
-	binary.NativeEndian.PutUint64(buf[:], 1)
-	_, err := unix.Write(int(c.shutdownFd), buf[:])
-	return err
-}
-
-func (c *pollQueueSet) Close() error {
-	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)
-		}
-	}
-
-	// Close the shutdown eventfd last: every reader's pollfd set references
-	// it, so it must outlive the wake + per-queue teardown above.
-	if err := unix.Close(c.shutdownFd); err != nil {
-		errs = append(errs, err)
-	}
-	c.shutdownFd = -1
-
-	return errors.Join(errs...)
-}

overlay/tio/segment_bench_test.go

@@ -1,65 +0,0 @@
-//go:build linux && !android && !e2e_testing
-
-package tio
-
-import "testing"
-
-// fakeBatch stands in for batch.TxBatcher inside the bench — same shape
-// of pointer-capturing closure that sendInsideMessage builds.
-type fakeBatch struct{ buf [65536]byte }
-
-func (b *fakeBatch) Reserve(sz int) []byte { return b.buf[:sz] }
-func (b *fakeBatch) Commit([]byte)         {}
-
-type fakeHostInfo struct {
-	remoteIndexId uint32
-	counter       uint64
-}
-type fakeIface struct {
-	rebindCount uint8
-	hi          *fakeHostInfo
-}
-
-// BenchmarkSegmentSuperpacketAllocsTSO measures allocation per
-// SegmentSuperpacket call when a closure captures pointer-bearing
-// receivers — the realistic shape of sendInsideMessage's closure.
-func BenchmarkSegmentSuperpacketAllocsTSO(b *testing.B) {
-	const mss = 1400
-	const numSeg = 32
-	pkt := buildTSOv6(mss*numSeg, mss)
-	gso := GSOInfo{
-		Size:      mss,
-		HdrLen:    60, // 40 (IPv6) + 20 (TCP)
-		CsumStart: 40,
-		Proto:     GSOProtoTCP,
-	}
-	p := Packet{Bytes: pkt, GSO: gso}
-
-	hi := &fakeHostInfo{remoteIndexId: 0xdeadbeef}
-	f := &fakeIface{rebindCount: 7, hi: hi}
-	fb := &fakeBatch{}
-
-	// SegmentSuperpacket consumes pkt destructively; refresh from a master
-	// copy each iter (matches the production pattern where every TUN read
-	// hands the segmenter a fresh kernel-supplied buffer).
-	master := append([]byte(nil), pkt...)
-	work := make([]byte, len(pkt))
-	p.Bytes = work
-
-	b.ReportAllocs()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		copy(work, master)
-		err := SegmentSuperpacket(p, func(seg []byte) error {
-			out := fb.Reserve(16 + len(seg) + 16)
-			out[0] = byte(f.rebindCount)
-			out[1] = byte(hi.counter)
-			hi.counter++
-			fb.Commit(out)
-			return nil
-		})
-		if err != nil {
-			b.Fatalf("SegmentSuperpacket: %v", err)
-		}
-	}
-}

overlay/tio/segment_other.go

@@ -1,22 +0,0 @@
-//go:build !linux || android || e2e_testing
-
-package tio
-
-import "fmt"
-
-func protoFromGSOType(_ uint8) (GSOProto, error) {
-	return 0, fmt.Errorf("GSO unsupported")
-}
-
-// SegmentSuperpacket invokes fn once per segment of pkt. On non-Linux
-// builds (and Android/e2e_testing) this package does not provide a Queue
-// implementation, so any caller that does construct a Packet here can only
-// be operating on non-superpacket bytes and the stub forwards them
-// directly. A non-zero GSO field is a programming error from the caller
-// and returns an explicit error rather than silently misbehaving.
-func SegmentSuperpacket(pkt Packet, fn func(seg []byte) error) error {
-	if pkt.GSO.IsSuperpacket() {
-		return fmt.Errorf("tio: GSO superpacket on platform without segmentation support")
-	}
-	return fn(pkt.Bytes)
-}

overlay/tio/single.go

@@ -1,50 +0,0 @@
-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

@@ -1,172 +0,0 @@
-package tio
-
-import (
-	"io"
-)
-
-// QueueSet holds one or many Queue objects and helps close them in an orderly way.
-type QueueSet interface {
-	io.Closer
-	Queues() []Queue
-
-	// Add takes a tun fd, adds it to the set, and prepares it for use as a Queue.
-	Add(fd int) error
-}
-
-// Capabilities advertises which kernel offload features a Queue
-// successfully negotiated. Callers consult this to decide which coalescers
-// to wire onto the write path — a Queue without TSO can't usefully accept a
-// TCPCoalescer, and a Queue without USO can't accept a UDPCoalescer.
-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. Concurrency contract: a single
-// read goroutine drives Read; plain Write is safe for concurrent callers;
-// WriteGSO (on Queues that implement GSOWriter) is single-writer per queue.
-type Queue interface {
-	io.Closer
-
-	// Read returns one or more packets. The returned Packet.Bytes slices
-	// are borrowed from the Queue's internal buffer and are only valid
-	// until the next Read or Close on this Queue - callers must encrypt
-	// or copy each slice before the next call. A Packet may carry a
-	// GSO/USO superpacket (see GSOInfo); when GSO.IsSuperpacket() is
-	// true the caller must segment Bytes before treating it as a single
-	// IP datagram. Single-reader only: not 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. Safe for concurrent use.
-	Write(p []byte) (int, error)
-}
-
-// Packet is the unit Queue.Read returns. Bytes points into the queue's
-// internal buffer and is only valid until the next Read or Close on the
-// queue that produced it. GSO is the zero value for an already-segmented
-// IP datagram; when non-zero it describes a kernel-supplied TSO/USO
-// superpacket the caller must segment before consuming.
-type Packet struct {
-	Bytes []byte
-	GSO   GSOInfo
-}
-
-// GSOInfo describes a kernel-supplied superpacket sitting in Packet.Bytes.
-// The zero value means "not a superpacket" — Bytes is one regular IP
-// datagram and no segmentation is required.
-type GSOInfo struct {
-	// Size is the GSO segment size: max payload bytes per segment
-	// (== 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
-}
-
-// IsSuperpacket reports whether g describes a multi-segment GSO/USO
-// superpacket that needs segmentation before its bytes can be encrypted
-// and sent on the wire.
-func (g GSOInfo) IsSuperpacket() bool { return g.Size > 0 }
-
-// Clone returns a Packet whose Bytes is a freshly allocated copy of p.Bytes,
-// safe to retain past the next Read or Close on the originating Queue.
-// GSO metadata is copied verbatim. Use this only when a caller genuinely
-// needs to outlive the borrowed-slice contract — the hot path reads should
-// continue to consume the borrow synchronously to avoid the allocation.
-func (p Packet) Clone() Packet {
-	if p.Bytes == nil {
-		return p
-	}
-	cp := make([]byte, len(p.Bytes))
-	copy(cp, p.Bytes)
-	return Packet{Bytes: cp, GSO: p.GSO}
-}
-
-// CapsProvider is an optional interface implemented by Queues that
-// successfully negotiated kernel offload features at open time. Callers
-// pick a write-path coalescer based on the result. Queues that don't
-// implement it are treated as having no offload capability — callers must
-// fall back to plain per-packet writes.
-type CapsProvider interface {
-	Capabilities() Capabilities
-}
-
-// QueueCapabilities returns q's negotiated offload capabilities, or the
-// zero value when q does not advertise any.
-func QueueCapabilities(q Queue) Capabilities {
-	if cp, ok := q.(CapsProvider); ok {
-		return cp.Capabilities()
-	}
-	return Capabilities{}
-}
-
-// GSOProto selects the L4 protocol for a GSO superpacket. Determines which
-// VIRTIO_NET_HDR_GSO_* type the writer stamps and which checksum offset
-// inside the transport header virtio NEEDS_CSUM expects.
-type GSOProto uint8
-
-const (
-	GSOProtoTCP GSOProto = iota
-	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 any, want GSOProto) (GSOWriter, bool) {
-	gw, ok := w.(GSOWriter)
-	if !ok {
-		return nil, false
-	}
-	cp, ok := w.(CapsProvider)
-	if !ok {
-		return gw, true
-	}
-	caps := cp.Capabilities()
-	switch want {
-	case GSOProtoTCP:
-		return gw, caps.TSO
-	case GSOProtoUDP:
-		return gw, caps.USO
-	}
-	return gw, false
-}

overlay/tio/tio_gso_linux.go

@@ -1,414 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-package tio
-
-import (
-	"fmt"
-	"io"
-	"log/slog"
-	"os"
-	"sync/atomic"
-	"syscall"
-	"unsafe"
-
-	"golang.org/x/sys/unix"
-
-	"github.com/slackhq/nebula/overlay/tio/virtio"
-)
-
-// tunRxBufSize is the per-Read worst-case footprint inside rxBuf: one
-// kernel-supplied packet body, which is at most ~64 KiB (tunReadBufSize).
-// Segmentation happens at encrypt time on a per-routine MTU-sized scratch
-// (see SegmentSuperpacket), so rxBuf only holds raw kernel-supplied bytes.
-// We round up to give comfortable margin for the drain headroom check
-// below.
-const tunRxBufSize = 64 * 1024
-
-// tunRxBufCap is the total size we allocate for the per-reader rx
-// buffer. With reads landing directly in rxBuf, each drain iteration
-// consumes up to tunRxBufSize of headroom for the kernel-supplied bytes.
-// Sized to eight such iterations so a single poll wake can drain several
-// TSO/USO superpackets under bulk load, amortizing the wake and giving
-// the sendmmsg planner longer same-destination runs. Hold latency stays
-// bounded because listenIn flushes its send batch incrementally rather
-// than only at end-of-drain.
-const tunRxBufCap = tunRxBufSize * 8
-
-// tunDrainCap caps how many packets a single Read will accumulate via
-// the post-wake drain loop. Sized to soak up a burst of small ACKs while
-// bounding how much work a single caller holds before handing off.
-const tunDrainCap = 64
-
-// gsoMaxIovs caps the iovec budget WriteGSO assembles per call: 3 fixed
-// entries (virtio_net_hdr, IP hdr, transport hdr) plus up to gsoMaxIovs-3
-// payload fragments. Sized comfortably above the typical kernel GSO
-// segment cap (Linux UDP_GRO is 64) so realistic coalesced bursts never
-// touch the limit. iovecs are tiny (16 bytes), so the entire scratch is
-// 4 KiB — fine to keep resident on every queue. WriteGSO returns an error
-// rather than reallocating when a caller exceeds this budget.
-const gsoMaxIovs = 256
-
-// validVnetHdr is the 10-byte virtio_net_hdr we prepend to every non-GSO TUN
-// write. Only flag set is VIRTIO_NET_HDR_F_DATA_VALID, which marks the skb
-// CHECKSUM_UNNECESSARY so the receiving network stack skips L4 checksum
-// verification. All packets that reach the plain Write paths already carry
-// a valid L4 checksum (either supplied by a remote peer whose ciphertext we
-// AEAD-authenticated, produced by segmentTCPYield/segmentUDPYield during
-// superpacket segmentation, or built locally by CreateRejectPacket), so
-// trusting them is safe.
-var validVnetHdr = [virtio.Size]byte{unix.VIRTIO_NET_HDR_F_DATA_VALID}
-
-// Offload wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
-// A shared eventfd allows Close to wake all readers blocked in poll.
-type Offload struct {
-	fd         int
-	shutdownFd int
-	closed     atomic.Bool
-	rxBuf      []byte   // backing store for kernel-handed packets read this drain
-	rxOff      int      // cursor into rxBuf for the current Read drain
-	pending    []Packet // packets returned from the most recent Read
-
-	// readVnetScratch holds the 10-byte virtio_net_hdr split off the front of
-	// every TUN read via readv(2). Decoupling the header from the packet body
-	// lets us read the body directly into rxBuf at the current rxOff with
-	// no userspace copy on the GSO_NONE fast path.
-	readVnetScratch [virtio.Size]byte
-	// readIovs is the readv(2) iovec scratch wired once at construction —
-	// iovec[0] points at readVnetScratch; iovec[1].Base/Len is updated per
-	// read to address the current rxBuf slot.
-	readIovs [2]unix.Iovec
-
-	// usoEnabled records whether the kernel agreed to TUN_F_USO* on this FD,
-	// so writers can decide whether emitting GSO_UDP_L4 superpackets is safe.
-	usoEnabled bool
-
-	// gsoHdrBuf is a per-queue 10-byte scratch for the virtio_net_hdr emitted
-	// by WriteGSO. Kept separate from the read-only package-level validVnetHdr
-	// so non-GSO Writes can ship that constant directly while WriteGSO
-	// rewrites this scratch on every call.
-	gsoHdrBuf [virtio.Size]byte
-	// gsoIovs is the writev iovec scratch for WriteGSO. Pre-sized to
-	// gsoMaxIovs at construction; never grown. WriteGSO returns an error
-	// (and drops the call) if a caller hands it more fragments than fit.
-	gsoIovs []unix.Iovec
-}
-
-func newOffload(fd int, shutdownFd int, usoEnabled bool) (*Offload, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-
-	out := &Offload{
-		fd:         fd,
-		shutdownFd: shutdownFd,
-		usoEnabled: usoEnabled,
-		closed:     atomic.Bool{},
-
-		rxBuf:   make([]byte, tunRxBufCap),
-		gsoIovs: make([]unix.Iovec, 2, gsoMaxIovs),
-	}
-
-	out.gsoIovs[0].Base = &out.gsoHdrBuf[0]
-	out.gsoIovs[0].SetLen(virtio.Size)
-
-	// readIovs[0] is wired once to the virtio_net_hdr scratch; per-read we
-	// only repoint readIovs[1] at the next rxBuf slot (see readPacket).
-	out.readIovs[0].Base = &out.readVnetScratch[0]
-	out.readIovs[0].SetLen(virtio.Size)
-
-	return out, nil
-}
-
-func (r *Offload) blockOnRead() error {
-	return blockOn(int32(r.fd), int32(r.shutdownFd), unix.POLLIN)
-}
-
-func (r *Offload) blockOnWrite() error {
-	return blockOn(int32(r.fd), int32(r.shutdownFd), unix.POLLOUT)
-}
-
-// readPacket issues a single readv(2) splitting the virtio_net_hdr off
-// into readVnetScratch and reading the packet body directly into rxBuf at
-// the current rxOff. Returns the body length (zero virtio header bytes,
-// just the IP packet/superpacket). block controls whether EAGAIN is
-// retried via poll: the initial read of a drain blocks; subsequent drain
-// reads do not.
-//
-// The body iovec capacity is always tunReadBufSize; callers (the Read
-// drain loop) gate entry on tunRxBufCap-rxOff >= tunRxBufSize, sized to
-// hold one worst-case kernel-supplied packet body. Without that gate the
-// body iovec could be smaller than the next inbound packet and the
-// kernel would truncate.
-func (r *Offload) readPacket(block bool) (int, error) {
-	for {
-		r.readIovs[1].Base = &r.rxBuf[r.rxOff]
-		r.readIovs[1].SetLen(tunReadBufSize)
-		n, _, errno := syscall.Syscall(unix.SYS_READV, uintptr(r.fd), uintptr(unsafe.Pointer(&r.readIovs[0])), uintptr(len(r.readIovs)))
-		if errno == 0 {
-			if int(n) < virtio.Size {
-				return 0, io.ErrShortWrite
-			}
-			return int(n) - virtio.Size, nil
-		}
-		if errno == unix.EAGAIN {
-			if !block {
-				return 0, errno
-			}
-			if err := r.blockOnRead(); err != nil {
-				return 0, err
-			}
-			continue
-		}
-		if errno == unix.EINTR {
-			continue
-		}
-		if errno == unix.EBADF {
-			return 0, os.ErrClosed
-		}
-		return 0, errno
-	}
-}
-
-// Read returns one or more packets from the tun. Each Packet either
-// carries a single ready-to-use IP datagram (GSO zero) or a TSO/USO
-// superpacket plus the GSOInfo a caller needs to segment it (see
-// SegmentSuperpacket). The first read blocks via poll; once the fd is
-// known readable we drain additional packets non-blocking until the
-// kernel queue is empty (EAGAIN), we've collected tunDrainCap packets,
-// or we're out of rxBuf headroom. This amortizes the poll wake over
-// bursts of small packets (e.g. TCP ACKs). Packet.Bytes slices point
-// into the Offload's internal buffer and are only valid until the next
-// Read or Close on this Queue.
-func (r *Offload) Read() ([]Packet, error) {
-	r.pending = r.pending[:0]
-	r.rxOff = 0
-
-	// Initial (blocking) read. Retry on decode errors so a single bad
-	// packet does not stall the reader.
-	for {
-		n, err := r.readPacket(true)
-		if err != nil {
-			return nil, err
-		}
-		if err := r.decodeRead(n); err != nil {
-			// Drop and read again — a bad packet should not kill the reader.
-			continue
-		}
-		break
-	}
-
-	// Drain: non-blocking reads until the kernel queue is empty, the drain
-	// cap is reached, or rxBuf no longer has room for another worst-case
-	// kernel-supplied packet (tunRxBufSize).
-	for len(r.pending) < tunDrainCap && tunRxBufCap-r.rxOff >= tunRxBufSize {
-		n, err := r.readPacket(false)
-		if err != nil {
-			// EAGAIN / EINTR / anything else: stop draining. We already
-			// have a valid batch from the first read.
-			break
-		}
-		if n <= 0 {
-			break
-		}
-		if err := r.decodeRead(n); err != nil {
-			// Drop this packet and stop the drain; we'd rather hand off
-			// what we have than keep spinning here.
-			break
-		}
-	}
-
-	return r.pending, nil
-}
-
-// decodeRead processes the packet sitting in rxBuf at rxOff (length
-// pktLen). The bytes stay in rxBuf — for GSO_NONE we slice them as a
-// regular IP datagram (running finishChecksum if NEEDS_CSUM is set);
-// for TSO/USO superpackets we attach the corrected GSO metadata so the
-// caller can segment lazily at encrypt time. rxOff advances past the
-// kernel-supplied body and nothing else, since segmentation no longer
-// writes back into rxBuf.
-func (r *Offload) decodeRead(pktLen int) error {
-	if pktLen <= 0 {
-		return fmt.Errorf("short tun read: %d", pktLen)
-	}
-	var hdr virtio.Hdr
-	hdr.Decode(r.readVnetScratch[:])
-
-	body := r.rxBuf[r.rxOff : r.rxOff+pktLen]
-
-	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_NONE {
-		if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
-			if err := virtio.FinishChecksum(body, hdr); err != nil {
-				return err
-			}
-		}
-		r.pending = append(r.pending, Packet{Bytes: body})
-		r.rxOff += pktLen
-		return nil
-	}
-
-	// GSO superpacket: validate, fix the kernel-supplied HdrLen on the
-	// FORWARD path (CorrectHdrLen), pick the L4 protocol, and attach
-	// the metadata. The bytes stay in rxBuf untouched, segmentation
-	// happens in SegmentSuperpacket at encrypt time.
-	if err := virtio.CheckValid(body, hdr); err != nil {
-		return err
-	}
-	if err := virtio.CorrectHdrLen(body, &hdr); err != nil {
-		return err
-	}
-	proto, err := protoFromGSOType(hdr.GSOType)
-	if err != nil {
-		return err
-	}
-	r.pending = append(r.pending, Packet{
-		Bytes: body,
-		GSO: GSOInfo{
-			Size:      hdr.GSOSize,
-			HdrLen:    hdr.HdrLen,
-			CsumStart: hdr.CsumStart,
-			Proto:     proto,
-		},
-	})
-	r.rxOff += pktLen
-	return nil
-}
-
-func (r *Offload) Write(buf []byte) (int, error) {
-	iovs := [2]unix.Iovec{
-		{Base: &validVnetHdr[0]},
-		{Base: &buf[0]},
-	}
-	iovs[0].SetLen(virtio.Size)
-	iovs[1].SetLen(len(buf))
-	return r.writeWithScratch(buf, &iovs)
-}
-
-func (r *Offload) writeWithScratch(buf []byte, iovs *[2]unix.Iovec) (int, error) {
-	if len(buf) == 0 {
-		return 0, nil
-	}
-	iovs[1].Base = &buf[0]
-	iovs[1].SetLen(len(buf))
-	return r.rawWrite(unsafe.Slice(&iovs[0], len(iovs)))
-}
-
-func (r *Offload) rawWrite(iovs []unix.Iovec) (int, error) {
-	for {
-		n, _, errno := syscall.Syscall(unix.SYS_WRITEV, uintptr(r.fd), uintptr(unsafe.Pointer(&iovs[0])), uintptr(len(iovs)))
-		if errno == 0 {
-			if int(n) < virtio.Size {
-				return 0, io.ErrShortWrite
-			}
-			return int(n) - virtio.Size, nil
-		}
-		if errno == unix.EAGAIN {
-			if err := r.blockOnWrite(); err != nil {
-				return 0, err
-			}
-			continue
-		}
-		if errno == unix.EINTR {
-			continue
-		}
-		if errno == unix.EBADF {
-			return 0, os.ErrClosed
-		}
-		return 0, errno
-	}
-}
-
-// Capabilities reports the offload features negotiated for this Queue. TSO
-// is always true for Offload (we only construct it on IFF_VNET_HDR FDs);
-// USO is true only when the kernel agreed to TUN_F_USO4|6 at open time
-// (Linux ≥ 6.2).
-func (r *Offload) Capabilities() Capabilities {
-	return Capabilities{TSO: true, USO: r.usoEnabled}
-}
-
-func (r *Offload) WriteGSO(hdr []byte, transportHdr []byte, pays [][]byte, proto GSOProto) error {
-	if len(hdr) == 0 || len(pays) == 0 || len(transportHdr) == 0 {
-		return nil
-	}
-	// L4 checksum offset inside transportHdr: TCP=16 (the `check` field after
-	// seq/ack/dataoff/flags/window), UDP=6 (after sport/dport/length).
-	var csumOff uint16
-	switch proto {
-	case GSOProtoUDP:
-		csumOff = 6
-	default:
-		csumOff = 16
-	}
-	vhdr := virtio.Hdr{
-		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-		HdrLen:     uint16(len(hdr) + len(transportHdr)),
-		GSOSize:    uint16(len(pays[0])),
-		CsumStart:  uint16(len(hdr)),
-		CsumOffset: csumOff,
-	}
-	if len(pays) > 1 {
-		ipVer := hdr[0] >> 4
-		switch {
-		case proto == GSOProtoUDP && (ipVer == 4 || ipVer == 6):
-			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_UDP_L4
-		case ipVer == 6:
-			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV6
-		case ipVer == 4:
-			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_TCPV4
-		default:
-			vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
-			vhdr.GSOSize = 0
-		}
-	} else {
-		vhdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
-		vhdr.GSOSize = 0
-	}
-	vhdr.Encode(r.gsoHdrBuf[:])
-
-	// Build the iovec array: [virtio_hdr, hdr, transportHdr, pays...]. r.gsoIovs[0] is
-	// wired to gsoHdrBuf at construction and never changes.
-	need := 3 + len(pays)
-	if need > cap(r.gsoIovs) {
-		slog.Default().Warn("tio: WriteGSO iovec budget exceeded; dropping superpacket",
-			"need", need, "cap", cap(r.gsoIovs), "segments", len(pays))
-		return fmt.Errorf("tio: WriteGSO needs %d iovecs but cap is %d", need, cap(r.gsoIovs))
-	}
-	r.gsoIovs = r.gsoIovs[:need]
-	r.gsoIovs[1].Base = &hdr[0]
-	r.gsoIovs[1].SetLen(len(hdr))
-	r.gsoIovs[2].Base = &transportHdr[0]
-	r.gsoIovs[2].SetLen(len(transportHdr))
-	// Defense in depth: an empty payload fragment can't be a valid GSO
-	// segment and &p[0] would panic on it. Callers route zero-length
-	// datagrams through the plain path (see UDPCoalescer.commitParsed), so
-	// this should never fire, but skip empties rather than index into one.
-	// `n` tracks where the next payload iovec lands, since skips make it
-	// drift from 3+i.
-	n := 3
-	for _, p := range pays {
-		if len(p) == 0 {
-			continue
-		}
-		r.gsoIovs[n].Base = &p[0]
-		r.gsoIovs[n].SetLen(len(p))
-		n++
-	}
-	r.gsoIovs = r.gsoIovs[:n]
-
-	_, err := r.rawWrite(r.gsoIovs)
-	return err
-}
-
-func (r *Offload) Close() error {
-	if r.closed.Swap(true) {
-		return nil
-	}
-
-	//shutdownFd is owned by the container, so we should not close it
-	// Close the underlying fd but do NOT null r.fd: a reader may still be
-	// loading it in readOne, and mutating the field would race that load.
-	// It gets EBADF -> os.ErrClosed (or wakes via the shutdown eventfd's
-	// ppoll first). closed.Swap already guarantees we only close once.
-	return unix.Close(r.fd)
-}

overlay/tio/tio_poll_linux.go

@@ -1,117 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-package tio
-
-import (
-	"fmt"
-	"os"
-	"sync/atomic"
-
-	"golang.org/x/sys/unix"
-)
-
-// Maximum size we accept for a single read from a TUN with IFF_VNET_HDR. A
-// TSO superpacket can be up to 64KiB of payload plus a single L2/L3/L4 header
-// prefix plus the virtio header.
-const tunReadBufSize = 65535
-
-type Poll struct {
-	fd         int
-	shutdownFd int
-	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 matches the
-// newOffload convention and keeps closes at exactly one on every path.
-func newPoll(fd int, shutdownFd int) (*Poll, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set Poll device as nonblocking: %w", err)
-	}
-
-	out := &Poll{
-		fd:         fd,
-		shutdownFd: shutdownFd,
-		readBuf:    make([]byte, tunReadBufSize),
-	}
-	return out, nil
-}
-
-// blockOnRead waits until the Poll fd is readable or shutdown has been signaled.
-// Returns os.ErrClosed if Close was called.
-func (t *Poll) blockOnRead() error {
-	return blockOn(int32(t.fd), int32(t.shutdownFd), unix.POLLIN)
-}
-
-func (t *Poll) blockOnWrite() error {
-	return blockOn(int32(t.fd), int32(t.shutdownFd), unix.POLLOUT)
-}
-
-func (t *Poll) Read() ([]Packet, error) {
-	n, err := t.readOne(t.readBuf)
-	if err != nil {
-		return nil, err
-	}
-	t.batchRet[0] = Packet{Bytes: t.readBuf[:n]}
-	return t.batchRet[:], nil
-}
-
-func (t *Poll) readOne(to []byte) (int, error) {
-	for {
-		n, errno := unix.Read(t.fd, to)
-		if errno == nil {
-			return n, nil
-		}
-		switch errno {
-		case unix.EAGAIN:
-			if err := t.blockOnRead(); err != nil {
-				return 0, err
-			}
-		case unix.EINTR:
-			// retry
-		case unix.EBADF:
-			return 0, os.ErrClosed
-		default:
-			return 0, errno
-		}
-	}
-}
-
-// Write is safe for concurrent use
-func (t *Poll) Write(from []byte) (int, error) {
-	for {
-		n, errno := unix.Write(t.fd, from)
-		if errno == nil {
-			return n, nil
-		}
-		switch errno {
-		case unix.EAGAIN:
-			if err := t.blockOnWrite(); err != nil {
-				return 0, err
-			}
-		case unix.EINTR:
-			// retry
-		case unix.EBADF:
-			return 0, os.ErrClosed
-		default:
-			return 0, errno
-		}
-	}
-}
-
-func (t *Poll) Close() error {
-	if t.closed.Swap(true) {
-		return nil
-	}
-	//shutdownFd is owned by the container, so we should not close it
-	// Close the underlying fd but do NOT null t.fd: a reader may still be
-	// loading it in readOne, and mutating the field would race that load.
-	// It gets EBADF -> os.ErrClosed (or wakes via the shutdown eventfd's
-	// ppoll first). closed.Swap already guarantees we only close once.
-	return unix.Close(t.fd)
-}

overlay/tio/tun_file_linux_test.go

@@ -1,227 +0,0 @@
-//go:build linux && !android && !e2e_testing
-// +build linux,!android,!e2e_testing
-
-package tio
-
-import (
-	"errors"
-	"os"
-	"sync"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/require"
-	"golang.org/x/sys/unix"
-)
-
-// newReadPipe returns a read fd. The matching write fd is registered for cleanup.
-// The caller takes ownership of the read fd (pass it into a QueueSet).
-func newReadPipe(t *testing.T) int {
-	t.Helper()
-	var fds [2]int
-	if err := unix.Pipe2(fds[:], unix.O_CLOEXEC); err != nil {
-		t.Fatalf("pipe2: %v", err)
-	}
-	t.Cleanup(func() { _ = unix.Close(fds[1]) })
-	return fds[0]
-}
-
-func TestPoll_WakeForShutdown_WakesFriends(t *testing.T) {
-	pipe1 := newReadPipe(t)
-	pipe2 := newReadPipe(t)
-	parent, err := NewPollQueueSet()
-	require.NoError(t, err)
-	require.NoError(t, parent.Add(pipe1))
-	require.NoError(t, parent.Add(pipe2))
-	t.Cleanup(func() {
-		_ = unix.Close(pipe1)
-		_ = unix.Close(pipe2)
-	})
-
-	readers := parent.Queues()
-	errs := make([]error, len(readers))
-	var wg sync.WaitGroup
-	for i, r := range readers {
-		wg.Add(1)
-		go func(i int, r Queue) {
-			defer wg.Done()
-			_, errs[i] = r.Read()
-		}(i, r)
-	}
-
-	time.Sleep(50 * time.Millisecond)
-
-	if err := parent.Close(); err != nil {
-		t.Fatalf("Close: %v", err)
-	}
-
-	done := make(chan struct{})
-	go func() { wg.Wait(); close(done) }()
-	select {
-	case <-done:
-	case <-time.After(2 * time.Second):
-		t.Fatal("readers did not wake")
-	}
-
-	for i, err := range errs {
-		if !errors.Is(err, os.ErrClosed) {
-			t.Errorf("reader %d: expected os.ErrClosed, got %v", i, err)
-		}
-	}
-}
-
-// 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) })
-
-	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)
-	}
-	if err := tf.Close(); err != nil {
-		t.Fatalf("second Close should be a no-op, got %v", err)
-	}
-}
-
-// 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)))
-
-	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")
-
-	// Second Close must not touch fds (shutdownFd is now -1) and must return nil.
-	require.NoError(t, qs.Close())
-}
-
-// TestOffloadQueueSet_Close_ClosesShutdownFd mirrors the poll regression test
-// for the GSO/offload queueset.
-func TestOffloadQueueSet_Close_ClosesShutdownFd(t *testing.T) {
-	qs, err := NewOffloadQueueSet(false)
-	require.NoError(t, err)
-	c, ok := qs.(*offloadQueueSet)
-	require.True(t, ok)
-	require.NoError(t, qs.Add(newReadPipe(t)))
-
-	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")
-
-	// Second Close must not touch fds (shutdownFd is now -1) and must return nil.
-	require.NoError(t, qs.Close())
-}

overlay/tio/tun_linux_offload.go

@@ -1,51 +0,0 @@
-//go:build linux && !android && !e2e_testing
-// +build linux,!android,!e2e_testing
-
-package tio
-
-import (
-	"fmt"
-
-	"golang.org/x/sys/unix"
-
-	"github.com/slackhq/nebula/overlay/tio/virtio"
-)
-
-// protoFromGSOType maps a virtio_net_hdr GSOType to the GSOProto value the
-// segment-time helpers use. Returns an error for GSO_NONE or any unknown
-// value — the caller should only invoke this on a confirmed superpacket.
-func protoFromGSOType(t uint8) (GSOProto, error) {
-	switch t {
-	case unix.VIRTIO_NET_HDR_GSO_TCPV4, unix.VIRTIO_NET_HDR_GSO_TCPV6:
-		return GSOProtoTCP, nil
-	case unix.VIRTIO_NET_HDR_GSO_UDP_L4:
-		return GSOProtoUDP, nil
-	default:
-		return 0, fmt.Errorf("unsupported virtio gso type: %d", t)
-	}
-}
-
-// SegmentSuperpacket invokes fn once per segment of pkt. For non-GSO pkts
-// fn is called once with pkt.Bytes (no segmentation, no copy). For GSO/USO
-// superpackets fn is called once per segment with a slice of pkt.Bytes
-// holding that segment's plaintext (a freshly-patched L3+L4 header sliced
-// in front of the original payload chunk). The slide is destructive: pkt is
-// consumed by this call and its bytes are in an undefined state when
-// SegmentSuperpacket returns. Callers must not retain pkt or any earlier
-// seg slice past fn's return for that segment. The scratch parameter is
-// unused on the destructive path and kept only for cross-platform
-// signature compatibility. Aborts and returns the first error from fn or
-// from per-segment construction.
-func SegmentSuperpacket(pkt Packet, fn func(seg []byte) error) error {
-	if !pkt.GSO.IsSuperpacket() {
-		return fn(pkt.Bytes)
-	}
-	switch pkt.GSO.Proto {
-	case GSOProtoTCP:
-		return virtio.SegmentTCP(pkt.Bytes, pkt.GSO.HdrLen, pkt.GSO.CsumStart, pkt.GSO.Size, fn)
-	case GSOProtoUDP:
-		return virtio.SegmentUDP(pkt.Bytes, pkt.GSO.HdrLen, pkt.GSO.CsumStart, pkt.GSO.Size, fn)
-	default:
-		return fmt.Errorf("unsupported gso proto: %d", pkt.GSO.Proto)
-	}
-}

overlay/tio/tun_linux_offload_test.go

@@ -1,826 +0,0 @@
-//go:build linux && !android && !e2e_testing
-// +build linux,!android,!e2e_testing
-
-package tio
-
-import (
-	"encoding/binary"
-	"os"
-	"testing"
-
-	"golang.org/x/sys/unix"
-	"gvisor.dev/gvisor/pkg/tcpip/checksum"
-
-	"github.com/slackhq/nebula/overlay/tio/virtio"
-)
-
-// testSegScratchSize is a generous segmentation scratch sized to fit any
-// of the synthetic TSO/USO superpackets these tests generate (one
-// worst-case 64 KiB superpacket plus replicated per-segment headers).
-const testSegScratchSize = 192 * 1024
-
-// verifyChecksum confirms that the one's-complement sum across `b`, seeded
-// with a folded pseudo-header sum, equals all-ones (valid).
-func verifyChecksum(b []byte, pseudo uint16) bool {
-	return checksum.Checksum(b, pseudo) == 0xffff
-}
-
-// segmentForTest is the test-only counterpart to the production
-// SegmentSuperpacket path. It handles GSO_NONE (with optional
-// finishChecksum) inline and dispatches GSO superpackets through
-// SegmentSuperpacket, draining each yielded segment into a
-// freshly-copied [][]byte slot so callers can iterate after the call
-// returns. Tests pre-set hdr.HdrLen correctly, so correctHdrLen is not
-// invoked here.
-func segmentForTest(pkt []byte, hdr virtio.Hdr, out *[][]byte, scratch []byte) error {
-	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_NONE {
-		cp := append([]byte(nil), pkt...)
-		if hdr.Flags&unix.VIRTIO_NET_HDR_F_NEEDS_CSUM != 0 {
-			if err := virtio.FinishChecksum(cp, hdr); err != nil {
-				return err
-			}
-		}
-		*out = append(*out, cp)
-		return nil
-	}
-	proto, err := protoFromGSOType(hdr.GSOType)
-	if err != nil {
-		return err
-	}
-	gso := GSOInfo{
-		Size:      hdr.GSOSize,
-		HdrLen:    hdr.HdrLen,
-		CsumStart: hdr.CsumStart,
-		Proto:     proto,
-	}
-	return SegmentSuperpacket(Packet{Bytes: pkt, GSO: gso}, func(seg []byte) error {
-		*out = append(*out, append([]byte(nil), seg...))
-		return nil
-	})
-}
-
-// pseudoHeaderIPv4 returns the folded pseudo-header sum used to verify a
-// TCP/UDP segment's checksum in tests. src/dst are 4 bytes each.
-func pseudoHeaderIPv4(src, dst []byte, proto byte, l4Len int) uint16 {
-	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
-	s += uint32(proto) + uint32(l4Len)
-	s = (s & 0xffff) + (s >> 16)
-	s = (s & 0xffff) + (s >> 16)
-	return uint16(s)
-}
-
-// pseudoHeaderIPv6 returns the folded pseudo-header sum used to verify a
-// TCP/UDP segment's checksum in tests. src/dst are 16 bytes each.
-func pseudoHeaderIPv6(src, dst []byte, proto byte, l4Len int) uint16 {
-	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
-	s += uint32(l4Len>>16) + uint32(l4Len&0xffff) + uint32(proto)
-	s = (s & 0xffff) + (s >> 16)
-	s = (s & 0xffff) + (s >> 16)
-	return uint16(s)
-}
-
-// buildTSOv4 builds a synthetic IPv4/TCP TSO superpacket with a payload of
-// `payLen` bytes split at `mss`.
-func buildTSOv4(t *testing.T, payLen, mss int) ([]byte, virtio.Hdr) {
-	t.Helper()
-	const ipLen = 20
-	const tcpLen = 20
-	pkt := make([]byte, ipLen+tcpLen+payLen)
-
-	// IPv4 header
-	pkt[0] = 0x45 // version 4, IHL 5
-	// total length is meaningless for TSO but set it anyway
-	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+payLen))
-	binary.BigEndian.PutUint16(pkt[4:6], 0x4242) // original ID
-	pkt[8] = 64                                  // TTL
-	pkt[9] = unix.IPPROTO_TCP
-	copy(pkt[12:16], []byte{10, 0, 0, 1}) // src
-	copy(pkt[16:20], []byte{10, 0, 0, 2}) // dst
-
-	// TCP header
-	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
-	binary.BigEndian.PutUint16(pkt[22:24], 80)    // dport
-	binary.BigEndian.PutUint32(pkt[24:28], 10000) // seq
-	binary.BigEndian.PutUint32(pkt[28:32], 20000) // ack
-	pkt[32] = 0x50                                // data offset 5 words
-	pkt[33] = 0x18                                // ACK | PSH
-	binary.BigEndian.PutUint16(pkt[34:36], 65535) // window
-
-	// payload
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+tcpLen+i] = byte(i & 0xff)
-	}
-
-	return pkt, virtio.Hdr{
-		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
-		HdrLen:     uint16(ipLen + tcpLen),
-		GSOSize:    uint16(mss),
-		CsumStart:  uint16(ipLen),
-		CsumOffset: 16,
-	}
-}
-
-func TestSegmentTCPv4(t *testing.T) {
-	const mss = 100
-	const numSeg = 3
-	pkt, hdr := buildTSOv4(t, mss*numSeg, mss)
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != numSeg {
-		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
-	}
-
-	for i, seg := range out {
-		if len(seg) != 40+mss {
-			t.Errorf("seg %d: unexpected len %d", i, len(seg))
-		}
-		totalLen := binary.BigEndian.Uint16(seg[2:4])
-		if totalLen != uint16(40+mss) {
-			t.Errorf("seg %d: total_len=%d want %d", i, totalLen, 40+mss)
-		}
-		id := binary.BigEndian.Uint16(seg[4:6])
-		if id != 0x4242+uint16(i) {
-			t.Errorf("seg %d: ip id=%#x want %#x", i, id, 0x4242+uint16(i))
-		}
-		seq := binary.BigEndian.Uint32(seg[24:28])
-		wantSeq := uint32(10000 + i*mss)
-		if seq != wantSeq {
-			t.Errorf("seg %d: seq=%d want %d", i, seq, wantSeq)
-		}
-		flags := seg[33]
-		wantFlags := byte(0x10) // ACK only, PSH cleared
-		if i == numSeg-1 {
-			wantFlags = 0x18 // ACK | PSH preserved on last
-		}
-		if flags != wantFlags {
-			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
-		}
-		// IPv4 header checksum must verify against itself.
-		if !verifyChecksum(seg[:20], 0) {
-			t.Errorf("seg %d: bad IPv4 header checksum", i)
-		}
-		// TCP checksum must verify against the pseudo-header.
-		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+mss)
-		if !verifyChecksum(seg[20:], psum) {
-			t.Errorf("seg %d: bad TCP checksum", i)
-		}
-	}
-}
-
-func TestSegmentTCPv4OddTail(t *testing.T) {
-	// Payload of 250 bytes with MSS 100 → segments of 100, 100, 50.
-	pkt, hdr := buildTSOv4(t, 250, 100)
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != 3 {
-		t.Fatalf("want 3 segments, got %d", len(out))
-	}
-	wantPayLens := []int{100, 100, 50}
-	for i, seg := range out {
-		if len(seg)-40 != wantPayLens[i] {
-			t.Errorf("seg %d: pay len %d want %d", i, len(seg)-40, wantPayLens[i])
-		}
-		if !verifyChecksum(seg[:20], 0) {
-			t.Errorf("seg %d: bad IPv4 header checksum", i)
-		}
-		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+wantPayLens[i])
-		if !verifyChecksum(seg[20:], psum) {
-			t.Errorf("seg %d: bad TCP checksum", i)
-		}
-	}
-}
-
-func TestSegmentTCPv6(t *testing.T) {
-	const ipLen = 40
-	const tcpLen = 20
-	const mss = 120
-	const numSeg = 2
-	payLen := mss * numSeg
-	pkt := make([]byte, ipLen+tcpLen+payLen)
-
-	// IPv6 header
-	pkt[0] = 0x60 // version 6
-	binary.BigEndian.PutUint16(pkt[4:6], uint16(tcpLen+payLen))
-	pkt[6] = unix.IPPROTO_TCP
-	pkt[7] = 64
-	// src/dst fe80::1 / fe80::2
-	pkt[8] = 0xfe
-	pkt[9] = 0x80
-	pkt[23] = 1
-	pkt[24] = 0xfe
-	pkt[25] = 0x80
-	pkt[39] = 2
-
-	// TCP header
-	binary.BigEndian.PutUint16(pkt[40:42], 12345)
-	binary.BigEndian.PutUint16(pkt[42:44], 80)
-	binary.BigEndian.PutUint32(pkt[44:48], 7)
-	binary.BigEndian.PutUint32(pkt[48:52], 99)
-	pkt[52] = 0x50
-	pkt[53] = 0x19 // FIN | ACK | PSH — exercise FIN clearing too
-	binary.BigEndian.PutUint16(pkt[54:56], 65535)
-
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+tcpLen+i] = byte(i)
-	}
-
-	hdr := virtio.Hdr{
-		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV6,
-		HdrLen:     uint16(ipLen + tcpLen),
-		GSOSize:    uint16(mss),
-		CsumStart:  uint16(ipLen),
-		CsumOffset: 16,
-	}
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != numSeg {
-		t.Fatalf("want %d segments, got %d", numSeg, len(out))
-	}
-
-	for i, seg := range out {
-		if len(seg) != ipLen+tcpLen+mss {
-			t.Errorf("seg %d: len %d want %d", i, len(seg), ipLen+tcpLen+mss)
-		}
-		pl := binary.BigEndian.Uint16(seg[4:6])
-		if pl != uint16(tcpLen+mss) {
-			t.Errorf("seg %d: payload_length=%d want %d", i, pl, tcpLen+mss)
-		}
-		seq := binary.BigEndian.Uint32(seg[44:48])
-		if seq != uint32(7+i*mss) {
-			t.Errorf("seg %d: seq=%d want %d", i, seq, 7+i*mss)
-		}
-		flags := seg[53]
-		// Original flags = 0x19 (FIN|ACK|PSH). FIN(0x01)+PSH(0x08) should be
-		// cleared on all but the last; ACK(0x10) always preserved.
-		wantFlags := byte(0x10)
-		if i == numSeg-1 {
-			wantFlags = 0x19
-		}
-		if flags != wantFlags {
-			t.Errorf("seg %d: flags=%#x want %#x", i, flags, wantFlags)
-		}
-		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_TCP, tcpLen+mss)
-		if !verifyChecksum(seg[ipLen:], psum) {
-			t.Errorf("seg %d: bad TCP checksum", i)
-		}
-	}
-}
-
-func TestSegmentGSONonePassesThrough(t *testing.T) {
-	pkt, hdr := buildTSOv4(t, 100, 100)
-	hdr.GSOType = unix.VIRTIO_NET_HDR_GSO_NONE
-	hdr.Flags = 0 // no NEEDS_CSUM, leave packet untouched
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != 1 {
-		t.Fatalf("want 1 segment, got %d", len(out))
-	}
-	if len(out[0]) != len(pkt) {
-		t.Fatalf("unexpected length: %d vs %d", len(out[0]), len(pkt))
-	}
-}
-
-// TestSegmentRejectsLegacyUDPGSO ensures the legacy GSO_UDP (UFO) marker is
-// still rejected; only modern GSO_UDP_L4 (USO) is supported.
-func TestSegmentRejectsLegacyUDPGSO(t *testing.T) {
-	hdr := virtio.Hdr{GSOType: unix.VIRTIO_NET_HDR_GSO_UDP}
-	var out [][]byte
-	if err := segmentForTest(nil, hdr, &out, nil); err == nil {
-		t.Fatalf("expected rejection for legacy UDP GSO")
-	}
-}
-
-// buildUSOv4 builds a synthetic IPv4/UDP USO superpacket with payload of
-// payLen bytes, segmented at gsoSize.
-func buildUSOv4(t *testing.T, payLen, gsoSize int) ([]byte, virtio.Hdr) {
-	t.Helper()
-	const ipLen = 20
-	const udpLen = 8
-	pkt := make([]byte, ipLen+udpLen+payLen)
-
-	// IPv4 header
-	pkt[0] = 0x45 // version 4, IHL 5
-	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+udpLen+payLen))
-	binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
-	pkt[8] = 64
-	pkt[9] = unix.IPPROTO_UDP
-	copy(pkt[12:16], []byte{10, 0, 0, 1})
-	copy(pkt[16:20], []byte{10, 0, 0, 2})
-
-	// UDP header (length + checksum filled in per segment by segmentUDPYield)
-	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
-	binary.BigEndian.PutUint16(pkt[22:24], 53)    // dport
-
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+udpLen+i] = byte(i & 0xff)
-	}
-
-	return pkt, virtio.Hdr{
-		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-		GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
-		HdrLen:     uint16(ipLen + udpLen),
-		GSOSize:    uint16(gsoSize),
-		CsumStart:  uint16(ipLen),
-		CsumOffset: 6,
-	}
-}
-
-func TestSegmentUDPv4(t *testing.T) {
-	const gso = 100
-	const numSeg = 3
-	pkt, hdr := buildUSOv4(t, gso*numSeg, gso)
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != numSeg {
-		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
-	}
-
-	for i, seg := range out {
-		if len(seg) != 28+gso {
-			t.Errorf("seg %d: len %d want %d", i, len(seg), 28+gso)
-		}
-		totalLen := binary.BigEndian.Uint16(seg[2:4])
-		if totalLen != uint16(28+gso) {
-			t.Errorf("seg %d: total_len=%d want %d", i, totalLen, 28+gso)
-		}
-		// kernel UDP-GSO does NOT bump the IPv4 ID across segments; every
-		// segment carries the same ID as the seed.
-		id := binary.BigEndian.Uint16(seg[4:6])
-		if id != 0x4242 {
-			t.Errorf("seg %d: ip id=%#x want %#x", i, id, 0x4242)
-		}
-		udpLen := binary.BigEndian.Uint16(seg[24:26])
-		if udpLen != uint16(8+gso) {
-			t.Errorf("seg %d: udp len=%d want %d", i, udpLen, 8+gso)
-		}
-		if !verifyChecksum(seg[:20], 0) {
-			t.Errorf("seg %d: bad IPv4 header checksum", i)
-		}
-		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_UDP, 8+gso)
-		if !verifyChecksum(seg[20:], psum) {
-			t.Errorf("seg %d: bad UDP checksum", i)
-		}
-	}
-}
-
-func TestSegmentUDPv4OddTail(t *testing.T) {
-	// 250 bytes payload, gsoSize=100 → segments of 100, 100, 50.
-	pkt, hdr := buildUSOv4(t, 250, 100)
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != 3 {
-		t.Fatalf("want 3 segments, got %d", len(out))
-	}
-	wantPay := []int{100, 100, 50}
-	for i, seg := range out {
-		if len(seg)-28 != wantPay[i] {
-			t.Errorf("seg %d: pay len %d want %d", i, len(seg)-28, wantPay[i])
-		}
-		udpLen := binary.BigEndian.Uint16(seg[24:26])
-		if udpLen != uint16(8+wantPay[i]) {
-			t.Errorf("seg %d: udp len=%d want %d", i, udpLen, 8+wantPay[i])
-		}
-		if !verifyChecksum(seg[:20], 0) {
-			t.Errorf("seg %d: bad IPv4 header checksum", i)
-		}
-		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_UDP, 8+wantPay[i])
-		if !verifyChecksum(seg[20:], psum) {
-			t.Errorf("seg %d: bad UDP checksum", i)
-		}
-	}
-}
-
-func TestSegmentUDPv6(t *testing.T) {
-	const ipLen = 40
-	const udpLen = 8
-	const gso = 120
-	const numSeg = 2
-	payLen := gso * numSeg
-	pkt := make([]byte, ipLen+udpLen+payLen)
-
-	// IPv6 header
-	pkt[0] = 0x60
-	binary.BigEndian.PutUint16(pkt[4:6], uint16(udpLen+payLen))
-	pkt[6] = unix.IPPROTO_UDP
-	pkt[7] = 64
-	pkt[8] = 0xfe
-	pkt[9] = 0x80
-	pkt[23] = 1
-	pkt[24] = 0xfe
-	pkt[25] = 0x80
-	pkt[39] = 2
-
-	binary.BigEndian.PutUint16(pkt[40:42], 12345)
-	binary.BigEndian.PutUint16(pkt[42:44], 53)
-
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+udpLen+i] = byte(i)
-	}
-
-	hdr := virtio.Hdr{
-		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-		GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
-		HdrLen:     uint16(ipLen + udpLen),
-		GSOSize:    uint16(gso),
-		CsumStart:  uint16(ipLen),
-		CsumOffset: 6,
-	}
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != numSeg {
-		t.Fatalf("want %d segments, got %d", numSeg, len(out))
-	}
-
-	for i, seg := range out {
-		if len(seg) != ipLen+udpLen+gso {
-			t.Errorf("seg %d: len %d want %d", i, len(seg), ipLen+udpLen+gso)
-		}
-		pl := binary.BigEndian.Uint16(seg[4:6])
-		if pl != uint16(udpLen+gso) {
-			t.Errorf("seg %d: payload_length=%d want %d", i, pl, udpLen+gso)
-		}
-		ul := binary.BigEndian.Uint16(seg[ipLen+4 : ipLen+6])
-		if ul != uint16(udpLen+gso) {
-			t.Errorf("seg %d: udp len=%d want %d", i, ul, udpLen+gso)
-		}
-		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_UDP, udpLen+gso)
-		if !verifyChecksum(seg[ipLen:], psum) {
-			t.Errorf("seg %d: bad UDP checksum", i)
-		}
-	}
-}
-
-// TestSegmentUDPCEPropagates confirms IP-level CE marks on the seed appear on
-// every segment. UDP has no transport-level CWR/ECE: the IP TOS/TC byte is
-// copied verbatim into every segment by the segment-prefix copy.
-func TestSegmentUDPCEPropagates(t *testing.T) {
-	pkt, hdr := buildUSOv4(t, 200, 100)
-	pkt[1] = 0x03 // CE codepoint in IP-ECN
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != 2 {
-		t.Fatalf("want 2 segments, got %d", len(out))
-	}
-	for i, seg := range out {
-		if seg[1]&0x03 != 0x03 {
-			t.Errorf("seg %d: CE missing (tos=%#x)", i, seg[1])
-		}
-		if !verifyChecksum(seg[:20], 0) {
-			t.Errorf("seg %d: bad IPv4 header checksum", i)
-		}
-	}
-}
-
-// TestSegmentTCPCwrFirstSegmentOnly confirms RFC 3168 §6.1.2: when a TSO
-// burst's seed has CWR set, only the first emitted segment carries CWR.
-// ECE is preserved on every segment (different signal, persistent state).
-func TestSegmentTCPCwrFirstSegmentOnly(t *testing.T) {
-	const mss = 100
-	const numSeg = 3
-	pkt, hdr := buildTSOv4(t, mss*numSeg, mss)
-	// Seed flags: CWR | ECE | ACK | PSH.
-	pkt[33] = 0x80 | 0x40 | 0x10 | 0x08
-
-	scratch := make([]byte, testSegScratchSize)
-	var out [][]byte
-	if err := segmentForTest(pkt, hdr, &out, scratch); err != nil {
-		t.Fatalf("segmentForTest: %v", err)
-	}
-	if len(out) != numSeg {
-		t.Fatalf("expected %d segments, got %d", numSeg, len(out))
-	}
-	for i, seg := range out {
-		flags := seg[33]
-		hasCwr := flags&0x80 != 0
-		hasEce := flags&0x40 != 0
-		hasPsh := flags&0x08 != 0
-		wantCwr := i == 0
-		wantPsh := i == numSeg-1
-		if hasCwr != wantCwr {
-			t.Errorf("seg %d: CWR=%v want %v (flags=%#x)", i, hasCwr, wantCwr, flags)
-		}
-		if !hasEce {
-			t.Errorf("seg %d: ECE missing (flags=%#x)", i, flags)
-		}
-		if hasPsh != wantPsh {
-			t.Errorf("seg %d: PSH=%v want %v (flags=%#x)", i, hasPsh, wantPsh, flags)
-		}
-		// IP and TCP checksums must still verify after the flag rewrite.
-		if !verifyChecksum(seg[:20], 0) {
-			t.Errorf("seg %d: bad IPv4 header checksum", i)
-		}
-		psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, 20+mss)
-		if !verifyChecksum(seg[20:], psum) {
-			t.Errorf("seg %d: bad TCP checksum", i)
-		}
-	}
-}
-
-func BenchmarkSegmentTCPv4(b *testing.B) {
-	sizes := []struct {
-		name   string
-		payLen int
-		mss    int
-	}{
-		{"64KiB_MSS1460", 65000, 1460},
-		{"16KiB_MSS1460", 16384, 1460},
-		{"4KiB_MSS1460", 4096, 1460},
-	}
-	for _, sz := range sizes {
-		b.Run(sz.name, func(b *testing.B) {
-			const ipLen = 20
-			const tcpLen = 20
-			pkt := make([]byte, ipLen+tcpLen+sz.payLen)
-			pkt[0] = 0x45
-			binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+sz.payLen))
-			binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
-			pkt[8] = 64
-			pkt[9] = unix.IPPROTO_TCP
-			copy(pkt[12:16], []byte{10, 0, 0, 1})
-			copy(pkt[16:20], []byte{10, 0, 0, 2})
-			binary.BigEndian.PutUint16(pkt[20:22], 12345)
-			binary.BigEndian.PutUint16(pkt[22:24], 80)
-			binary.BigEndian.PutUint32(pkt[24:28], 10000)
-			binary.BigEndian.PutUint32(pkt[28:32], 20000)
-			pkt[32] = 0x50
-			pkt[33] = 0x18
-			binary.BigEndian.PutUint16(pkt[34:36], 65535)
-			for i := 0; i < sz.payLen; i++ {
-				pkt[ipLen+tcpLen+i] = byte(i)
-			}
-			hdr := virtio.Hdr{
-				Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-				GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV4,
-				HdrLen:     uint16(ipLen + tcpLen),
-				GSOSize:    uint16(sz.mss),
-				CsumStart:  uint16(ipLen),
-				CsumOffset: 16,
-			}
-
-			scratch := make([]byte, testSegScratchSize)
-			out := make([][]byte, 0, 64)
-
-			// SegmentSuperpacket consumes its input destructively; restore
-			// pkt from a master copy each iteration. The restore mirrors the
-			// kernel→userspace copy that hands a fresh GSO blob to the
-			// segmenter in production, so it's representative cost rather
-			// than bench overhead.
-			master := append([]byte(nil), pkt...)
-			work := make([]byte, len(pkt))
-
-			b.SetBytes(int64(len(pkt)))
-			b.ResetTimer()
-			for i := 0; i < b.N; i++ {
-				copy(work, master)
-				out = out[:0]
-				if err := segmentForTest(work, hdr, &out, scratch); err != nil {
-					b.Fatal(err)
-				}
-			}
-		})
-	}
-}
-
-// TestTunFileWriteVnetHdrNoAlloc verifies the IFF_VNET_HDR fast-path write is
-// allocation-free. We write to /dev/null so every call succeeds synchronously.
-func TestTunFileWriteVnetHdrNoAlloc(t *testing.T) {
-	fd, err := unix.Open("/dev/null", os.O_WRONLY, 0)
-	if err != nil {
-		t.Fatalf("open /dev/null: %v", err)
-	}
-	t.Cleanup(func() { _ = unix.Close(fd) })
-
-	tf := &Offload{fd: fd}
-
-	payload := make([]byte, 1400)
-	// Warm up (first call may trigger one-time internal allocations elsewhere).
-	if _, err := tf.Write(payload); err != nil {
-		t.Fatalf("Write: %v", err)
-	}
-
-	allocs := testing.AllocsPerRun(1000, func() {
-		if _, err := tf.Write(payload); err != nil {
-			t.Fatalf("Write: %v", err)
-		}
-	})
-	if allocs != 0 {
-		t.Fatalf("Write allocated %.1f times per call, want 0", allocs)
-	}
-}
-
-// TestWriteGSOSkipsEmptyPayloads is the defense-in-depth guard for the
-// zero-length UDP DoS: a payload fragment of length zero would make &p[0]
-// panic (index-out-of-range) when building the iovec array. WriteGSO must
-// skip empties instead. We write to /dev/null so the writev always succeeds
-// synchronously; the point is simply that neither call panics.
-func TestWriteGSOSkipsEmptyPayloads(t *testing.T) {
-	fd, err := unix.Open("/dev/null", os.O_WRONLY, 0)
-	if err != nil {
-		t.Fatalf("open /dev/null: %v", err)
-	}
-	t.Cleanup(func() { _ = unix.Close(fd) })
-
-	o := &Offload{fd: fd, gsoIovs: make([]unix.Iovec, 2, gsoMaxIovs)}
-	o.gsoIovs[0].Base = &o.gsoHdrBuf[0]
-	o.gsoIovs[0].SetLen(virtio.Size)
-
-	ipHdr := make([]byte, 20)
-	ipHdr[0] = 0x45 // IPv4, IHL 5
-	udpHdr := make([]byte, 8)
-
-	// Sole payload empty: exercises the all-empty skip (n stays at 3).
-	if err := o.WriteGSO(ipHdr, udpHdr, [][]byte{{}}, GSOProtoUDP); err != nil {
-		t.Fatalf("WriteGSO with a single empty payload: %v", err)
-	}
-	// Empty mixed with a real fragment: exercises the index-drift skip so a
-	// later non-empty payload still lands in the right iovec slot.
-	real := make([]byte, 1200)
-	if err := o.WriteGSO(ipHdr, udpHdr, [][]byte{real, {}}, GSOProtoUDP); err != nil {
-		t.Fatalf("WriteGSO with a trailing empty payload: %v", err)
-	}
-}
-
-// buildTSOv6 builds a synthetic IPv6/TCP TSO superpacket with payLen bytes
-// of payload, segmented at gso. Returns the packet bytes only; the
-// virtio_net_hdr is the caller's responsibility.
-func buildTSOv6(payLen, gso int) []byte {
-	const ipLen = 40
-	const tcpLen = 20
-	pkt := make([]byte, ipLen+tcpLen+payLen)
-
-	pkt[0] = 0x60 // version 6
-	binary.BigEndian.PutUint16(pkt[4:6], uint16(tcpLen+payLen))
-	pkt[6] = unix.IPPROTO_TCP
-	pkt[7] = 64
-	pkt[8] = 0xfe
-	pkt[9] = 0x80
-	pkt[23] = 1
-	pkt[24] = 0xfe
-	pkt[25] = 0x80
-	pkt[39] = 2
-
-	binary.BigEndian.PutUint16(pkt[40:42], 12345)
-	binary.BigEndian.PutUint16(pkt[42:44], 80)
-	binary.BigEndian.PutUint32(pkt[44:48], 7)
-	binary.BigEndian.PutUint32(pkt[48:52], 99)
-	pkt[52] = 0x50
-	pkt[53] = 0x10 // ACK only
-	binary.BigEndian.PutUint16(pkt[54:56], 65535)
-
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+tcpLen+i] = byte(i)
-	}
-	return pkt
-}
-
-// TestDecodeReadFitsMaxTSOAtDrainThreshold proves the rxBuf sizing is
-// correct: when rxOff is at the maximum value the drain headroom check
-// allows, decodeRead must still be able to absorb a worst-case 64KiB
-// TSO superpacket without dropping the burst. With segmentation deferred
-// to encrypt time, decodeRead writes only the kernel-supplied bytes into
-// rxBuf, so the size requirement is just "fit one worst-case input."
-//
-// Regression history: in a prior layout the rx buffer doubled as the
-// segmentation output, a near-threshold drain read returned "scratch too
-// small", the whole 45-segment TSO burst was dropped, and the remote's TCP
-// fast-retransmit collapsed cwnd. Keeping this test in the new layout
-// guards against re-introducing a drain headroom shortfall.
-func TestDecodeReadFitsMaxTSOAtDrainThreshold(t *testing.T) {
-	const ipv6HdrLen = 40
-	const tcpHdrLen = 20
-	const headerLen = ipv6HdrLen + tcpHdrLen
-	// Maximum TUN read body. The tunReadBufSize cap on readv's body iovec
-	// is what bounds the kernel's superpacket length.
-	pktLen := tunReadBufSize
-	payLen := pktLen - headerLen
-	const targetSegs = 64
-	gsoSize := (payLen + targetSegs - 1) / targetSegs
-
-	pkt := buildTSOv6(payLen, gsoSize)
-	if len(pkt) != pktLen {
-		t.Fatalf("buildTSOv6 produced %d bytes, want %d", len(pkt), pktLen)
-	}
-
-	o := &Offload{
-		rxBuf: make([]byte, tunRxBufCap),
-	}
-	// rxOff at the maximum value the drain headroom check permits before
-	// it would refuse another read. Any drain-time read up to this
-	// threshold MUST still process correctly.
-	o.rxOff = tunRxBufCap - tunRxBufSize
-
-	// Stage the body in rxBuf as if readv(2) just placed it there.
-	copy(o.rxBuf[o.rxOff:], pkt)
-
-	// Encode the matching virtio_net_hdr.
-	hdr := virtio.Hdr{
-		Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-		GSOType:    unix.VIRTIO_NET_HDR_GSO_TCPV6,
-		HdrLen:     uint16(headerLen),
-		GSOSize:    uint16(gsoSize),
-		CsumStart:  uint16(ipv6HdrLen),
-		CsumOffset: 16,
-	}
-	hdr.Encode(o.readVnetScratch[:])
-
-	startRxOff := o.rxOff
-	if err := o.decodeRead(pktLen); err != nil {
-		t.Fatalf("decodeRead at drain threshold returned %v — rxBuf sizing regression: "+
-			"tunRxBufSize=%d must hold one worst-case input (%d)",
-			err, tunRxBufSize, pktLen)
-	}
-
-	if len(o.pending) != 1 {
-		t.Fatalf("got %d packets, want 1 superpacket entry", len(o.pending))
-	}
-	got := o.pending[0]
-	if !got.GSO.IsSuperpacket() {
-		t.Fatalf("expected superpacket GSO metadata, got %+v", got.GSO)
-	}
-	if got.GSO.Proto != GSOProtoTCP {
-		t.Errorf("GSO.Proto=%d want TCP", got.GSO.Proto)
-	}
-	if got.GSO.Size != uint16(gsoSize) {
-		t.Errorf("GSO.Size=%d want %d", got.GSO.Size, gsoSize)
-	}
-	if got.GSO.HdrLen != uint16(headerLen) {
-		t.Errorf("GSO.HdrLen=%d want %d", got.GSO.HdrLen, headerLen)
-	}
-	if got.GSO.CsumStart != uint16(ipv6HdrLen) {
-		t.Errorf("GSO.CsumStart=%d want %d", got.GSO.CsumStart, ipv6HdrLen)
-	}
-	if len(got.Bytes) != pktLen {
-		t.Errorf("len(Bytes)=%d want %d", len(got.Bytes), pktLen)
-	}
-
-	// rxOff advances exactly by the kernel-supplied body length — no
-	// segmentation output to account for any more.
-	if o.rxOff != startRxOff+pktLen {
-		t.Errorf("rxOff=%d want %d", o.rxOff, startRxOff+pktLen)
-	}
-	if o.rxOff > tunRxBufCap {
-		t.Fatalf("rxOff=%d overran rxBuf (cap=%d)", o.rxOff, tunRxBufCap)
-	}
-
-	// Validate that segmenting the returned superpacket reproduces the
-	// expected per-segment IPv6 payload length and TCP checksum.
-	wantSegs := (payLen + gsoSize - 1) / gsoSize
-	gotSegs := 0
-	if err := SegmentSuperpacket(got, func(seg []byte) error {
-		defer func() { gotSegs++ }()
-		if len(seg) < headerLen+1 {
-			t.Errorf("seg %d too short: %d", gotSegs, len(seg))
-			return nil
-		}
-		if seg[0]>>4 != 6 {
-			t.Errorf("seg %d: bad IP version %#x", gotSegs, seg[0])
-		}
-		segPay := len(seg) - headerLen
-		gotPL := binary.BigEndian.Uint16(seg[4:6])
-		if gotPL != uint16(tcpHdrLen+segPay) {
-			t.Errorf("seg %d: payload_len=%d want %d", gotSegs, gotPL, tcpHdrLen+segPay)
-		}
-		psum := pseudoHeaderIPv6(seg[8:24], seg[24:40], unix.IPPROTO_TCP, tcpHdrLen+segPay)
-		if !verifyChecksum(seg[ipv6HdrLen:], psum) {
-			t.Errorf("seg %d: bad TCP checksum", gotSegs)
-		}
-		return nil
-	}); err != nil {
-		t.Fatalf("SegmentSuperpacket: %v", err)
-	}
-	if gotSegs != wantSegs {
-		t.Fatalf("got %d segments, want %d", gotSegs, wantSegs)
-	}
-}

overlay/tio/virtio/header_linux.go

@@ -1,43 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-package virtio
-
-import "encoding/binary"
-
-// Size is the on-wire length of struct virtio_net_hdr the kernel
-// prepends/expects on a TUN opened with IFF_VNET_HDR (TUNSETVNETHDRSZ
-// not set).
-const Size = 10
-
-// Hdr is the Go view of the legacy virtio_net_hdr.
-type Hdr struct {
-	Flags      uint8
-	GSOType    uint8
-	HdrLen     uint16
-	GSOSize    uint16
-	CsumStart  uint16
-	CsumOffset uint16
-}
-
-// Decode reads a virtio_net_hdr in host byte order (TUN default; we never
-// call TUNSETVNETLE so the kernel matches our endianness).
-func (h *Hdr) Decode(b []byte) {
-	h.Flags = b[0]
-	h.GSOType = b[1]
-	h.HdrLen = binary.NativeEndian.Uint16(b[2:4])
-	h.GSOSize = binary.NativeEndian.Uint16(b[4:6])
-	h.CsumStart = binary.NativeEndian.Uint16(b[6:8])
-	h.CsumOffset = binary.NativeEndian.Uint16(b[8:10])
-}
-
-// Encode is the inverse of Decode: writes the virtio_net_hdr fields into b
-// (must be at least Size bytes). Used to emit a TSO superpacket on egress.
-func (h *Hdr) Encode(b []byte) {
-	b[0] = h.Flags
-	b[1] = h.GSOType
-	binary.NativeEndian.PutUint16(b[2:4], h.HdrLen)
-	binary.NativeEndian.PutUint16(b[4:6], h.GSOSize)
-	binary.NativeEndian.PutUint16(b[6:8], h.CsumStart)
-	binary.NativeEndian.PutUint16(b[8:10], h.CsumOffset)
-}

overlay/tio/virtio/header_other.go

@@ -1,3 +0,0 @@
-//go:build !linux || android
-
-package virtio

overlay/tio/virtio/segment_linux.go

@@ -1,434 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-// Package virtio implements the pure validation, header-correction, and
-// per-segment slicing logic for kernel-supplied TSO/USO superpackets on
-// IFF_VNET_HDR TUN devices. It is FD-free and depends only on the byte
-// layout of the virtio_net_hdr and the IP/TCP/UDP headers it describes,
-// so it can be unit-tested in isolation from the tio Queue runtime.
-package virtio
-
-import (
-	"encoding/binary"
-	"errors"
-	"fmt"
-
-	"golang.org/x/sys/unix"
-
-	"github.com/slackhq/nebula/overlay/checksum"
-)
-
-// Protocol header size bounds used to validate / cap kernel-supplied offsets.
-const (
-	ipv4HeaderMinLen = 20 // IHL=5, no options
-	ipv4HeaderMaxLen = 60 // IHL=15, max options
-	ipv6FixedLen     = 40 // IPv6 base header; extensions would extend this
-	tcpHeaderMinLen  = 20 // data-offset=5, no options
-	tcpHeaderMaxLen  = 60 // data-offset=15, max options
-)
-
-// maxSegHdrLen bounds the L3+L4 header we snapshot before stamping each
-// segment. The largest header the segmenter supports is IPv4 (max IHL 60)
-// plus TCP (max data-offset 60) = 120 bytes; the array is sized to that
-// worst case so the snapshot lives on the stack with no per-call heap
-// allocation.
-const maxSegHdrLen = ipv4HeaderMaxLen + tcpHeaderMaxLen // 120
-
-// Byte offsets inside an IPv4 header.
-const (
-	ipv4TotalLenOff = 2
-	ipv4IDOff       = 4
-	ipv4ChecksumOff = 10
-	ipv4SrcOff      = 12
-	ipv4AddrsEnd    = 20 // end of dst address (ipv4SrcOff + 2*4)
-)
-
-// Byte offsets inside an IPv6 header.
-const (
-	ipv6PayloadLenOff = 4
-	ipv6SrcOff        = 8
-	ipv6AddrsEnd      = 40 // end of dst address (ipv6SrcOff + 2*16)
-)
-
-// Byte offsets inside a TCP header (relative to its start, i.e. csumStart).
-const (
-	tcpSeqOff      = 4
-	tcpDataOffOff  = 12 // upper nibble is header len in 32-bit words
-	tcpFlagsOff    = 13
-	tcpChecksumOff = 16
-)
-
-// UDP header is fixed at 8 bytes: {sport, dport, length, checksum}.
-const (
-	udpHeaderLen   = 8
-	udpLengthOff   = 4
-	udpChecksumOff = 6
-)
-
-// tcpFinPshMask is cleared on every segment except the last of a TSO burst.
-const tcpFinPshMask = 0x09 // FIN(0x01) | PSH(0x08)
-
-// tcpCwrFlag is cleared on every segment except the first. Per RFC 3168
-// §6.1.2 the CWR bit signals a one-shot transition (the sender just halved
-// its window) and must appear on the first segment of a TSO burst only.
-const tcpCwrFlag = 0x80
-
-// CheckValid rejects packets whose virtio_net_hdr/IP combination would
-// cause a downstream miscompute. The TUN should never emit RSC_INFO and
-// the GSO type must agree with the IP version nibble.
-func CheckValid(pkt []byte, hdr Hdr) error {
-	// When RSC_INFO is set the csum_start/csum_offset fields are repurposed to
-	// carry coalescing info rather than checksum offsets. A TUN writing via
-	// IFF_VNET_HDR should never emit this, but if it did we would silently
-	// miscompute the segment checksums — refuse the packet instead.
-	if hdr.Flags&unix.VIRTIO_NET_HDR_F_RSC_INFO != 0 {
-		return fmt.Errorf("virtio RSC_INFO flag not supported on TUN reads")
-	}
-	if len(pkt) < ipv4HeaderMinLen {
-		return fmt.Errorf("packet too short")
-	}
-	ipVersion := pkt[0] >> 4
-	switch hdr.GSOType {
-	case unix.VIRTIO_NET_HDR_GSO_TCPV4:
-		if ipVersion != 4 {
-			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
-		}
-	case unix.VIRTIO_NET_HDR_GSO_TCPV6:
-		if ipVersion != 6 {
-			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
-		}
-	case unix.VIRTIO_NET_HDR_GSO_UDP_L4:
-		// USO carries either v4 or v6; the leading nibble disambiguates.
-		if !(ipVersion == 4 || ipVersion == 6) {
-			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
-		}
-	default:
-		if !(ipVersion == 6 || ipVersion == 4) {
-			return fmt.Errorf("invalid IP version %d for GSO type %d", ipVersion, hdr.GSOType)
-		}
-	}
-
-	return nil
-}
-
-// CorrectHdrLen rewrites hdr.HdrLen based on the actual transport header
-// length read out of pkt. The kernel's hdr.HdrLen on the FORWARD path can
-// be the length of the entire first packet, so we don't trust it.
-func CorrectHdrLen(pkt []byte, hdr *Hdr) error {
-	// Thank you wireguard-go for documenting these edge-cases
-	// Don't trust hdr.hdrLen from the kernel as it can be equal to the length
-	// of the entire first packet when the kernel is handling it as part of a
-	// FORWARD path. Instead, parse the transport header length and add it onto
-	// csumStart, which is synonymous for IP header length.
-
-	if hdr.GSOType == unix.VIRTIO_NET_HDR_GSO_UDP_L4 {
-		hdr.HdrLen = hdr.CsumStart + 8
-	} else {
-		if len(pkt) <= int(hdr.CsumStart+tcpDataOffOff) {
-			return errors.New("packet is too short")
-		}
-
-		tcpHLen := uint16(pkt[hdr.CsumStart+tcpDataOffOff] >> 4 * 4)
-		if tcpHLen < 20 || tcpHLen > 60 {
-			// A TCP header must be between 20 and 60 bytes in length.
-			return fmt.Errorf("tcp header len is invalid: %d", tcpHLen)
-		}
-		hdr.HdrLen = hdr.CsumStart + tcpHLen
-	}
-
-	if len(pkt) < int(hdr.HdrLen) {
-		return fmt.Errorf("length of packet (%d) < virtioNetHdr.HdrLen (%d)", len(pkt), hdr.HdrLen)
-	}
-
-	if hdr.HdrLen < hdr.CsumStart {
-		return fmt.Errorf("virtioNetHdr.HdrLen (%d) < virtioNetHdr.CsumStart (%d)", hdr.HdrLen, hdr.CsumStart)
-	}
-	cSumAt := int(hdr.CsumStart + hdr.CsumOffset)
-	if cSumAt+1 >= len(pkt) {
-		return fmt.Errorf("end of checksum offset (%d) exceeds packet length (%d)", cSumAt+1, len(pkt))
-	}
-	return nil
-}
-
-// SegmentTCP walks a TSO superpacket pkt, yielding each segment as a
-// slice into pkt itself. Per-segment plaintext is laid out by stamping a
-// copy of the original L3+L4 header into pkt at offset i*gsoSize, where it
-// sits immediately before that segment's payload chunk in the original
-// buffer. The stamp is destructive but harmless: iter i's header write lands
-// on pkt[i*G : i*G+hdrLen], which is the tail of seg_{i-1}'s payload (already
-// consumed) and ends exactly where seg_i's payload begins, so it never clobbers
-// live payload — this holds even when gsoSize < hdrLen. The header bytes are
-// sourced from a pristine snapshot taken before the loop (savedHdr), NOT from
-// pkt[:hdrLen], because when gsoSize < hdrLen the stamps would otherwise
-// overwrite the leading header in place and every stamp after the first would
-// copy corrupted bytes. pkt is consumed by this call and must not be inspected
-// by the caller after the final yield.
-func SegmentTCP(pkt []byte, hdrLenU, csumStartU, gsoSizeU uint16, yield func(seg []byte) error) error {
-	if gsoSizeU == 0 {
-		return fmt.Errorf("gso_size is zero")
-	}
-	if csumStartU == 0 {
-		return fmt.Errorf("csum_start is zero")
-	}
-
-	headerLen := int(hdrLenU)
-	csumStart := int(csumStartU)
-	if headerLen > maxSegHdrLen {
-		return fmt.Errorf("header len %d exceeds max %d", headerLen, maxSegHdrLen)
-	}
-	isV4 := pkt[0]>>4 == 4
-
-	tcpHdrLen := int(pkt[csumStart+tcpDataOffOff]>>4) * 4
-	payLen := len(pkt) - headerLen
-	gsoSize := int(gsoSizeU)
-	numSeg := (payLen + gsoSize - 1) / gsoSize
-	if numSeg == 0 {
-		numSeg = 1
-	}
-
-	origSeq := binary.BigEndian.Uint32(pkt[csumStart+tcpSeqOff : csumStart+tcpSeqOff+4])
-	origFlags := pkt[csumStart+tcpFlagsOff]
-
-	var tmp [tcpHeaderMaxLen]byte
-	copy(tmp[:tcpHdrLen], pkt[csumStart:headerLen])
-	tmp[tcpSeqOff], tmp[tcpSeqOff+1], tmp[tcpSeqOff+2], tmp[tcpSeqOff+3] = 0, 0, 0, 0
-	tmp[tcpFlagsOff] = 0
-	tmp[tcpChecksumOff], tmp[tcpChecksumOff+1] = 0, 0
-	baseTcpHdrSum := uint32(checksum.Checksum(tmp[:tcpHdrLen], 0))
-
-	var baseProtoSum uint32
-	if isV4 {
-		baseProtoSum = uint32(checksum.Checksum(pkt[ipv4SrcOff:ipv4AddrsEnd], 0))
-	} else {
-		baseProtoSum = uint32(checksum.Checksum(pkt[ipv6SrcOff:ipv6AddrsEnd], 0))
-	}
-	baseProtoSum += uint32(unix.IPPROTO_TCP)
-
-	var origIPID uint16
-	var baseIPHdrSum uint32
-	if isV4 {
-		origIPID = binary.BigEndian.Uint16(pkt[ipv4IDOff : ipv4IDOff+2])
-		ihl := int(pkt[0]&0x0f) * 4
-		if ihl < ipv4HeaderMinLen || ihl > csumStart {
-			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
-		}
-		var ipTmp [ipv4HeaderMaxLen]byte
-		copy(ipTmp[:ihl], pkt[:ihl])
-		ipTmp[ipv4TotalLenOff], ipTmp[ipv4TotalLenOff+1] = 0, 0
-		ipTmp[ipv4IDOff], ipTmp[ipv4IDOff+1] = 0, 0
-		ipTmp[ipv4ChecksumOff], ipTmp[ipv4ChecksumOff+1] = 0, 0
-		baseIPHdrSum = uint32(checksum.Checksum(ipTmp[:ihl], 0))
-	}
-
-	// Snapshot the pristine L3+L4 header once. Every segment's header is
-	// stamped from this copy, so overlapping stamps (gsoSize < headerLen)
-	// can never corrupt the source. The variable fields (seq/flags/cksum/
-	// totalLen/id) captured here are stale but are overwritten per segment.
-	var savedHdr [maxSegHdrLen]byte
-	copy(savedHdr[:headerLen], pkt[:headerLen])
-
-	for i := 0; i < numSeg; i++ {
-		segStart := i * gsoSize
-		segEnd := segStart + gsoSize
-		if segEnd > payLen {
-			segEnd = payLen
-		}
-		segPayLen := segEnd - segStart
-		segLen := headerLen + segPayLen
-		headerOff := i * gsoSize
-
-		// Stamp the header into place immediately before this segment's
-		// payload, sourced from the pristine snapshot. Iter 0's header is
-		// already at pkt[:headerLen] (identical to savedHdr), so only i ≥ 1
-		// needs the stamp. The per-segment patches below overwrite the
-		// variable fields.
-		if i > 0 {
-			copy(pkt[headerOff:headerOff+headerLen], savedHdr[:headerLen])
-		}
-		seg := pkt[headerOff : headerOff+segLen]
-
-		segSeq := origSeq + uint32(segStart)
-		segFlags := origFlags
-		if i != 0 {
-			segFlags &^= tcpCwrFlag
-		}
-		if i != numSeg-1 {
-			segFlags &^= tcpFinPshMask
-		}
-		totalLen := segLen
-
-		if isV4 {
-			segID := origIPID + uint16(i)
-			binary.BigEndian.PutUint16(seg[ipv4TotalLenOff:ipv4TotalLenOff+2], uint16(totalLen))
-			binary.BigEndian.PutUint16(seg[ipv4IDOff:ipv4IDOff+2], segID)
-			ipSum := baseIPHdrSum + uint32(totalLen) + uint32(segID)
-			binary.BigEndian.PutUint16(seg[ipv4ChecksumOff:ipv4ChecksumOff+2], foldComplement(ipSum))
-		} else {
-			binary.BigEndian.PutUint16(seg[ipv6PayloadLenOff:ipv6PayloadLenOff+2], uint16(headerLen-ipv6FixedLen+segPayLen))
-		}
-
-		binary.BigEndian.PutUint32(seg[csumStart+tcpSeqOff:csumStart+tcpSeqOff+4], segSeq)
-		seg[csumStart+tcpFlagsOff] = segFlags
-
-		tcpLen := tcpHdrLen + segPayLen
-		// Payload bytes still live at their original offset in pkt. The
-		// header slide above only writes into pkt[i*G : i*G+H], which is
-		// the tail of seg_{i-1}'s payload (already consumed) and never
-		// overlaps seg_i's own payload at pkt[H+i*G : H+(i+1)*G].
-		paySum := uint32(checksum.Checksum(pkt[headerLen+segStart:headerLen+segEnd], 0))
-		wide := uint64(baseTcpHdrSum) + uint64(paySum) + uint64(baseProtoSum)
-		wide += uint64(segSeq) + uint64(segFlags) + uint64(tcpLen)
-		wide = (wide & 0xffffffff) + (wide >> 32)
-		wide = (wide & 0xffffffff) + (wide >> 32)
-		binary.BigEndian.PutUint16(seg[csumStart+tcpChecksumOff:csumStart+tcpChecksumOff+2], foldComplement(uint32(wide)))
-
-		if err := yield(seg); err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
-
-// SegmentUDP walks a USO superpacket, stamping a per-segment-patched copy of
-// the original L3+L4 header into pkt at offset i*gsoSize and yielding
-// pkt[i*G:i*G+segLen] to the caller. Per-segment patches are total_len +
-// IPv4 csum (or IPv6 payload_len) plus the UDP length and checksum. pkt is
-// consumed destructively; see SegmentTCP for the layout reasoning, including
-// why the header is stamped from a pristine snapshot rather than pkt[:hdrLen]
-// (correctness when gsoSize < hdrLen).
-//
-// UDP-GSO leaves the IPv4 ID identical across segments (the kernel does not
-// bump it), which is why the IP-level per-segment work is limited to
-// total_len + IPv4 header checksum (v4) or payload_len (v6).
-func SegmentUDP(pkt []byte, hdrLenU, csumStartU, gsoSizeU uint16, yield func(seg []byte) error) error {
-	if gsoSizeU == 0 {
-		return fmt.Errorf("gso_size is zero")
-	}
-	if csumStartU == 0 {
-		return fmt.Errorf("csum_start is zero")
-	}
-
-	isV4 := pkt[0]>>4 == 4
-	headerLen := int(hdrLenU)
-	csumStart := int(csumStartU)
-	if headerLen > maxSegHdrLen {
-		return fmt.Errorf("header len %d exceeds max %d", headerLen, maxSegHdrLen)
-	}
-	if headerLen-csumStart != udpHeaderLen {
-		return fmt.Errorf("udp header len mismatch: %d", headerLen-csumStart)
-	}
-
-	payLen := len(pkt) - headerLen
-	gsoSize := int(gsoSizeU)
-	numSeg := (payLen + gsoSize - 1) / gsoSize
-	if numSeg == 0 {
-		numSeg = 1
-	}
-
-	var udpTmp [udpHeaderLen]byte
-	copy(udpTmp[:], pkt[csumStart:headerLen])
-	udpTmp[udpLengthOff], udpTmp[udpLengthOff+1] = 0, 0
-	udpTmp[udpChecksumOff], udpTmp[udpChecksumOff+1] = 0, 0
-	baseUDPHdrSum := uint32(checksum.Checksum(udpTmp[:], 0))
-
-	var baseProtoSum uint32
-	if isV4 {
-		baseProtoSum = uint32(checksum.Checksum(pkt[ipv4SrcOff:ipv4AddrsEnd], 0))
-	} else {
-		baseProtoSum = uint32(checksum.Checksum(pkt[ipv6SrcOff:ipv6AddrsEnd], 0))
-	}
-	baseProtoSum += uint32(unix.IPPROTO_UDP)
-
-	var baseIPHdrSum uint32
-	if isV4 {
-		ihl := int(pkt[0]&0x0f) * 4
-		if ihl < ipv4HeaderMinLen || ihl > csumStart {
-			return fmt.Errorf("bad IPv4 IHL: %d", ihl)
-		}
-		var ipTmp [ipv4HeaderMaxLen]byte
-		copy(ipTmp[:ihl], pkt[:ihl])
-		ipTmp[ipv4TotalLenOff], ipTmp[ipv4TotalLenOff+1] = 0, 0
-		ipTmp[ipv4ChecksumOff], ipTmp[ipv4ChecksumOff+1] = 0, 0
-		baseIPHdrSum = uint32(checksum.Checksum(ipTmp[:ihl], 0))
-	}
-
-	// Snapshot the pristine L3+L4 header once and stamp every segment from
-	// it; see SegmentTCP for why sourcing from pkt[:headerLen] corrupts
-	// segments when gsoSize < headerLen.
-	var savedHdr [maxSegHdrLen]byte
-	copy(savedHdr[:headerLen], pkt[:headerLen])
-
-	for i := 0; i < numSeg; i++ {
-		segStart := i * gsoSize
-		segEnd := segStart + gsoSize
-		if segEnd > payLen {
-			segEnd = payLen
-		}
-		segPayLen := segEnd - segStart
-		segLen := headerLen + segPayLen
-		headerOff := i * gsoSize
-
-		if i > 0 {
-			copy(pkt[headerOff:headerOff+headerLen], savedHdr[:headerLen])
-		}
-		seg := pkt[headerOff : headerOff+segLen]
-
-		totalLen := segLen
-		udpLen := udpHeaderLen + segPayLen
-
-		if isV4 {
-			binary.BigEndian.PutUint16(seg[ipv4TotalLenOff:ipv4TotalLenOff+2], uint16(totalLen))
-			ipSum := baseIPHdrSum + uint32(totalLen)
-			binary.BigEndian.PutUint16(seg[ipv4ChecksumOff:ipv4ChecksumOff+2], foldComplement(ipSum))
-		} else {
-			binary.BigEndian.PutUint16(seg[ipv6PayloadLenOff:ipv6PayloadLenOff+2], uint16(headerLen-ipv6FixedLen+segPayLen))
-		}
-
-		binary.BigEndian.PutUint16(seg[csumStart+udpLengthOff:csumStart+udpLengthOff+2], uint16(udpLen))
-
-		paySum := uint32(checksum.Checksum(pkt[headerLen+segStart:headerLen+segEnd], 0))
-		wide := uint64(baseUDPHdrSum) + uint64(paySum) + uint64(baseProtoSum)
-		wide += uint64(udpLen) + uint64(udpLen)
-		wide = (wide & 0xffffffff) + (wide >> 32)
-		wide = (wide & 0xffffffff) + (wide >> 32)
-		csum := foldComplement(uint32(wide))
-		if csum == 0 {
-			csum = 0xffff
-		}
-		binary.BigEndian.PutUint16(seg[csumStart+udpChecksumOff:csumStart+udpChecksumOff+2], csum)
-
-		if err := yield(seg); err != nil {
-			return err
-		}
-	}
-
-	return nil
-}
-
-// FinishChecksum computes the L4 checksum for a non-GSO packet that the kernel
-// handed us with NEEDS_CSUM set. csum_start / csum_offset point at the 16-bit
-// checksum field; we zero it, fold a full sum (the field was pre-loaded with
-// the pseudo-header partial sum by the kernel), and store the result.
-func FinishChecksum(seg []byte, hdr Hdr) error {
-	cs := int(hdr.CsumStart)
-	co := int(hdr.CsumOffset)
-	if cs+co+2 > len(seg) {
-		return fmt.Errorf("csum offsets out of range: start=%d offset=%d len=%d", cs, co, len(seg))
-	}
-	// The kernel stores a partial pseudo-header sum at [cs+co:]; sum over the
-	// L4 region starting at cs, folding the prior partial in as the seed.
-	partial := binary.BigEndian.Uint16(seg[cs+co : cs+co+2])
-	seg[cs+co] = 0
-	seg[cs+co+1] = 0
-	binary.BigEndian.PutUint16(seg[cs+co:cs+co+2], ^checksum.Checksum(seg[cs:], partial))
-	return nil
-}
-
-// foldComplement folds a 32-bit one's-complement partial sum to 16 bits and
-// complements it, yielding the on-wire Internet checksum value.
-func foldComplement(sum uint32) uint16 {
-	sum = (sum & 0xffff) + (sum >> 16)
-	sum = (sum & 0xffff) + (sum >> 16)
-	return ^uint16(sum)
-}

overlay/tio/virtio/segment_linux_test.go

@@ -1,335 +0,0 @@
-//go:build linux && !android
-// +build linux,!android
-
-package virtio
-
-import (
-	"bytes"
-	"encoding/binary"
-	"testing"
-
-	"golang.org/x/sys/unix"
-
-	"github.com/slackhq/nebula/overlay/checksum"
-)
-
-// verifyChecksum confirms that the one's-complement sum across b, seeded with
-// a folded pseudo-header sum, equals all-ones (a valid on-wire checksum).
-// A corrupted header stamped into a segment makes this fail even when the
-// checksum field itself was computed from the (pristine) base sums, because
-// the bytes the receiver would sum no longer match what was checksummed.
-func verifyChecksum(b []byte, pseudo uint16) bool {
-	return checksum.Checksum(b, pseudo) == 0xffff
-}
-
-// pseudoHeaderIPv4 folds the TCP/UDP pseudo-header sum from a segment's own
-// address and length fields, used to independently verify its L4 checksum.
-func pseudoHeaderIPv4(src, dst []byte, proto byte, l4Len int) uint16 {
-	s := uint32(checksum.Checksum(src, 0)) + uint32(checksum.Checksum(dst, 0))
-	s += uint32(proto) + uint32(l4Len)
-	s = (s & 0xffff) + (s >> 16)
-	s = (s & 0xffff) + (s >> 16)
-	return uint16(s)
-}
-
-// buildTCPv4Super constructs a synthetic IPv4/TCP TSO superpacket with a
-// payload of payLen bytes and returns it alongside the header fields the
-// segmenter needs. The header is a fixed 40 bytes (20 IPv4 + 20 TCP).
-func buildTCPv4Super(payLen int) (pkt []byte, hdrLen, csumStart uint16) {
-	const ipLen = 20
-	const tcpLen = 20
-	pkt = make([]byte, ipLen+tcpLen+payLen)
-
-	// IPv4 header.
-	pkt[0] = 0x45 // version 4, IHL 5
-	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+tcpLen+payLen))
-	binary.BigEndian.PutUint16(pkt[4:6], 0x4242) // ID
-	pkt[8] = 64                                  // TTL
-	pkt[9] = unix.IPPROTO_TCP
-	copy(pkt[12:16], []byte{10, 0, 0, 1}) // src
-	copy(pkt[16:20], []byte{10, 0, 0, 2}) // dst
-
-	// TCP header.
-	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
-	binary.BigEndian.PutUint16(pkt[22:24], 80)    // dport
-	binary.BigEndian.PutUint32(pkt[24:28], 10000) // seq
-	binary.BigEndian.PutUint32(pkt[28:32], 20000) // ack
-	pkt[32] = 0x50                                // data offset 5 words
-	pkt[33] = 0x18                                // ACK | PSH
-	binary.BigEndian.PutUint16(pkt[34:36], 65535) // window
-
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+tcpLen+i] = byte(i & 0xff)
-	}
-	return pkt, ipLen + tcpLen, ipLen
-}
-
-// buildUDPv4Super constructs a synthetic IPv4/UDP USO superpacket with a
-// payload of payLen bytes. Header is a fixed 28 bytes (20 IPv4 + 8 UDP).
-func buildUDPv4Super(payLen int) (pkt []byte, hdrLen, csumStart uint16) {
-	const ipLen = 20
-	const udpLen = 8
-	pkt = make([]byte, ipLen+udpLen+payLen)
-
-	pkt[0] = 0x45
-	binary.BigEndian.PutUint16(pkt[2:4], uint16(ipLen+udpLen+payLen))
-	binary.BigEndian.PutUint16(pkt[4:6], 0x4242)
-	pkt[8] = 64
-	pkt[9] = unix.IPPROTO_UDP
-	copy(pkt[12:16], []byte{10, 0, 0, 1})
-	copy(pkt[16:20], []byte{10, 0, 0, 2})
-
-	binary.BigEndian.PutUint16(pkt[20:22], 12345) // sport
-	binary.BigEndian.PutUint16(pkt[22:24], 53)    // dport
-
-	for i := 0; i < payLen; i++ {
-		pkt[ipLen+udpLen+i] = byte(i & 0xff)
-	}
-	return pkt, ipLen + udpLen, ipLen
-}
-
-// collectTCP segments a fresh copy of pkt and returns each segment as an
-// independent slice so assertions can run after segmentation completes.
-func collectTCP(t *testing.T, pkt []byte, hdrLen, csumStart, gsoSize uint16) [][]byte {
-	t.Helper()
-	work := append([]byte(nil), pkt...)
-	var out [][]byte
-	err := SegmentTCP(work, hdrLen, csumStart, gsoSize, func(seg []byte) error {
-		out = append(out, append([]byte(nil), seg...))
-		return nil
-	})
-	if err != nil {
-		t.Fatalf("SegmentTCP: %v", err)
-	}
-	return out
-}
-
-func collectUDP(t *testing.T, pkt []byte, hdrLen, csumStart, gsoSize uint16) [][]byte {
-	t.Helper()
-	work := append([]byte(nil), pkt...)
-	var out [][]byte
-	err := SegmentUDP(work, hdrLen, csumStart, gsoSize, func(seg []byte) error {
-		out = append(out, append([]byte(nil), seg...))
-		return nil
-	})
-	if err != nil {
-		t.Fatalf("SegmentUDP: %v", err)
-	}
-	return out
-}
-
-// TestSegmentTCPHeaderNotCorrupted is the regression test for the in-place
-// header-slide bug: when gsoSize < headerLen the old code stamped each
-// segment's header from pkt[:headerLen], which had already been overwritten
-// by the previous segment's overlapping stamp, so segments 2..n carried a
-// corrupted header (garbage src/dst/ports/seq). Every segment must instead
-// carry the ORIGINAL constant header fields with correct per-segment seq.
-func TestSegmentTCPHeaderNotCorrupted(t *testing.T) {
-	const origSeq = 10000
-	cases := []struct {
-		name    string
-		payLen  int
-		gsoSize uint16
-	}{
-		// gsoSize (8) < headerLen (40): the bug's trigger. Even split.
-		{"small-gso-even", 40, 8},
-		// gsoSize (8) < headerLen (40) with a short final segment.
-		{"small-gso-odd-tail", 44, 8},
-		// gsoSize (100) >= headerLen (40): the normal path, must still work.
-		{"normal-gso", 250, 100},
-	}
-
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			pkt, hdrLen, csumStart := buildTCPv4Super(tc.payLen)
-			gso := int(tc.gsoSize)
-			wantSeg := (tc.payLen + gso - 1) / gso
-			segs := collectTCP(t, pkt, hdrLen, csumStart, tc.gsoSize)
-			if len(segs) != wantSeg {
-				t.Fatalf("got %d segments, want %d", len(segs), wantSeg)
-			}
-
-			off := 0
-			for i, seg := range segs {
-				// Constant header fields must be identical to the original in
-				// EVERY segment. These are exactly the bytes the old code
-				// corrupted in segments 2..n.
-				if got := seg[0]; got != 0x45 {
-					t.Errorf("seg %d: version/IHL byte=%#x want 0x45", i, got)
-				}
-				if seg[9] != unix.IPPROTO_TCP {
-					t.Errorf("seg %d: proto=%d want %d", i, seg[9], unix.IPPROTO_TCP)
-				}
-				if !bytes.Equal(seg[12:16], []byte{10, 0, 0, 1}) {
-					t.Errorf("seg %d: src=%v want [10 0 0 1]", i, seg[12:16])
-				}
-				if !bytes.Equal(seg[16:20], []byte{10, 0, 0, 2}) {
-					t.Errorf("seg %d: dst=%v want [10 0 0 2]", i, seg[16:20])
-				}
-				if sport := binary.BigEndian.Uint16(seg[20:22]); sport != 12345 {
-					t.Errorf("seg %d: sport=%d want 12345", i, sport)
-				}
-				if dport := binary.BigEndian.Uint16(seg[22:24]); dport != 80 {
-					t.Errorf("seg %d: dport=%d want 80", i, dport)
-				}
-				if ack := binary.BigEndian.Uint32(seg[28:32]); ack != 20000 {
-					t.Errorf("seg %d: ack=%d want 20000", i, ack)
-				}
-				if seg[32] != 0x50 {
-					t.Errorf("seg %d: data-offset byte=%#x want 0x50", i, seg[32])
-				}
-
-				// Per-segment seq must advance by the payload offset.
-				segStart := i * gso
-				if seq := binary.BigEndian.Uint32(seg[24:28]); seq != uint32(origSeq+segStart) {
-					t.Errorf("seg %d: seq=%d want %d", i, seq, origSeq+segStart)
-				}
-
-				// Payload bytes must be the original contiguous slice.
-				segPayLen := len(seg) - int(hdrLen)
-				wantPay := make([]byte, segPayLen)
-				for k := 0; k < segPayLen; k++ {
-					wantPay[k] = byte((off + k) & 0xff)
-				}
-				if !bytes.Equal(seg[hdrLen:], wantPay) {
-					t.Errorf("seg %d: payload mismatch", i)
-				}
-				off += segPayLen
-
-				// End-to-end: the stamped header must checksum-verify. A
-				// corrupted header fails here because the written checksum was
-				// derived from the pristine header.
-				if !verifyChecksum(seg[:20], 0) {
-					t.Errorf("seg %d: bad IPv4 header checksum", i)
-				}
-				psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_TCP, len(seg)-20)
-				if !verifyChecksum(seg[20:], psum) {
-					t.Errorf("seg %d: bad TCP checksum", i)
-				}
-			}
-		})
-	}
-}
-
-// TestCorrectHdrLenChecksumBound guards the checksum-field bounds check in
-// CorrectHdrLen. The checksum field sits at CsumStart+CsumOffset, so the check
-// must be computed from CsumStart+CsumOffset — NOT CsumStart+CsumStart, a
-// regression that doubled CsumStart and thus over-tightened the bound (since
-// CsumOffset, 6 for UDP / 16 for TCP, is always < CsumStart >= 20). That bogus
-// bound spuriously rejected valid small USO superpackets in decodeRead.
-func TestCorrectHdrLenChecksumBound(t *testing.T) {
-	// A valid IPv4 USO superpacket: 20B IPv4 + 8B UDP + two 6-byte segments
-	// (payload 12) = 40 bytes total. CsumStart=20, CsumOffset=6, so the UDP
-	// checksum field lives at bytes 26..27, comfortably inside the 40-byte
-	// packet. The OLD formula computed cSumAt = CsumStart+CsumStart = 40 and
-	// rejected on cSumAt+1 (41) >= len(pkt) (40); the fix (CsumStart+CsumOffset
-	// = 26) accepts. This case FAILS against the CsumStart+CsumStart regression.
-	t.Run("valid-small-uso-accepted", func(t *testing.T) {
-		pkt, _, csumStart := buildUDPv4Super(12) // total len 40
-		hdr := Hdr{
-			Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-			GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
-			GSOSize:    6, // two 6-byte segments
-			CsumStart:  csumStart,
-			CsumOffset: 6,
-		}
-		if err := CorrectHdrLen(pkt, &hdr); err != nil {
-			t.Fatalf("CorrectHdrLen rejected a valid 40-byte USO superpacket: %v", err)
-		}
-		if hdr.HdrLen != csumStart+udpHeaderLen {
-			t.Errorf("HdrLen = %d, want %d", hdr.HdrLen, csumStart+udpHeaderLen)
-		}
-	})
-
-	// A genuinely-too-short packet: CsumStart=20, CsumOffset=6 means the
-	// checksum field would end at byte 27, but the packet is only 25 bytes
-	// (CsumStart+CsumOffset+2 = 28 > 25). CorrectHdrLen must still reject it.
-	t.Run("too-short-rejected", func(t *testing.T) {
-		pkt := make([]byte, 25)
-		pkt[0] = 0x45 // IPv4, IHL 5
-		hdr := Hdr{
-			Flags:      unix.VIRTIO_NET_HDR_F_NEEDS_CSUM,
-			GSOType:    unix.VIRTIO_NET_HDR_GSO_UDP_L4,
-			GSOSize:    6,
-			CsumStart:  20,
-			CsumOffset: 6,
-		}
-		if err := CorrectHdrLen(pkt, &hdr); err == nil {
-			t.Fatalf("CorrectHdrLen accepted a too-short (25-byte) packet")
-		}
-	})
-}
-
-// TestSegmentUDPHeaderNotCorrupted is the USO counterpart: SegmentUDP performs
-// the same header stamp and must be correct when gsoSize < headerLen.
-func TestSegmentUDPHeaderNotCorrupted(t *testing.T) {
-	cases := []struct {
-		name    string
-		payLen  int
-		gsoSize uint16
-	}{
-		{"small-gso-even", 40, 8},
-		{"small-gso-odd-tail", 44, 8},
-		{"normal-gso", 250, 100},
-	}
-
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			pkt, hdrLen, csumStart := buildUDPv4Super(tc.payLen)
-			gso := int(tc.gsoSize)
-			wantSeg := (tc.payLen + gso - 1) / gso
-			segs := collectUDP(t, pkt, hdrLen, csumStart, tc.gsoSize)
-			if len(segs) != wantSeg {
-				t.Fatalf("got %d segments, want %d", len(segs), wantSeg)
-			}
-
-			off := 0
-			for i, seg := range segs {
-				if got := seg[0]; got != 0x45 {
-					t.Errorf("seg %d: version/IHL byte=%#x want 0x45", i, got)
-				}
-				if seg[9] != unix.IPPROTO_UDP {
-					t.Errorf("seg %d: proto=%d want %d", i, seg[9], unix.IPPROTO_UDP)
-				}
-				if !bytes.Equal(seg[12:16], []byte{10, 0, 0, 1}) {
-					t.Errorf("seg %d: src=%v want [10 0 0 1]", i, seg[12:16])
-				}
-				if !bytes.Equal(seg[16:20], []byte{10, 0, 0, 2}) {
-					t.Errorf("seg %d: dst=%v want [10 0 0 2]", i, seg[16:20])
-				}
-				if sport := binary.BigEndian.Uint16(seg[20:22]); sport != 12345 {
-					t.Errorf("seg %d: sport=%d want 12345", i, sport)
-				}
-				if dport := binary.BigEndian.Uint16(seg[22:24]); dport != 53 {
-					t.Errorf("seg %d: dport=%d want 53", i, dport)
-				}
-				// UDP-GSO keeps the same IPv4 ID across every segment.
-				if id := binary.BigEndian.Uint16(seg[4:6]); id != 0x4242 {
-					t.Errorf("seg %d: ip id=%#x want 0x4242", i, id)
-				}
-
-				segPayLen := len(seg) - int(hdrLen)
-				if udpLen := binary.BigEndian.Uint16(seg[24:26]); udpLen != uint16(8+segPayLen) {
-					t.Errorf("seg %d: udp len=%d want %d", i, udpLen, 8+segPayLen)
-				}
-
-				wantPay := make([]byte, segPayLen)
-				for k := 0; k < segPayLen; k++ {
-					wantPay[k] = byte((off + k) & 0xff)
-				}
-				if !bytes.Equal(seg[hdrLen:], wantPay) {
-					t.Errorf("seg %d: payload mismatch", i)
-				}
-				off += segPayLen
-
-				if !verifyChecksum(seg[:20], 0) {
-					t.Errorf("seg %d: bad IPv4 header checksum", i)
-				}
-				psum := pseudoHeaderIPv4(seg[12:16], seg[16:20], unix.IPPROTO_UDP, len(seg)-20)
-				if !verifyChecksum(seg[20:], psum) {
-					t.Errorf("seg %d: bad UDP checksum", i)
-				}
-			}
-		})
-	}
-}

overlay/tun_android.go

@@ -13,7 +13,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 )
@@ -64,7 +63,7 @@ func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
 	return r
 }
 
-func (t *tun) Activate() error {
+func (t tun) Activate() error {
 	return nil
 }
 
@@ -97,6 +96,10 @@ func (t *tun) Name() string {
 	return "android"
 }
 
-func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *tun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for android")
 }

overlay/tun_darwin.go

@@ -6,6 +6,7 @@ package overlay
 import (
 	"errors"
 	"fmt"
+	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -15,7 +16,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -550,7 +550,7 @@ func (t *tun) Read(to []byte) (int, error) {
 	return n - 4, nil
 }
 
-// Write pushes one IP packet onto the utun device. Only valid for single threaded use.
+// Write pushes one IP packet onto the utun device.
 func (t *tun) Write(from []byte) (int, error) {
 	if len(from) == 0 {
 		return 0, syscall.EIO
@@ -606,6 +606,10 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *tun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
 }

overlay/tun_disabled.go

@@ -10,7 +10,6 @@ import (
 
 	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/iputil"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -24,23 +23,6 @@ type disabledTun struct {
 	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 {
 	tun := &disabledTun{
 		vpnNetworks: vpnNetworks,
@@ -75,6 +57,24 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
+func (t *disabledTun) Read(b []byte) (int, error) {
+	r, ok := <-t.read
+	if !ok {
+		return 0, io.EOF
+	}
+
+	if len(r) > len(b) {
+		return 0, fmt.Errorf("packet larger than mtu: %d > %d bytes", len(r), len(b))
+	}
+
+	t.tx.Inc(1)
+	if t.l.Enabled(context.Background(), slog.LevelDebug) {
+		t.l.Debug("Write payload", "raw", prettyPacket(r))
+	}
+
+	return copy(b, r), nil
+}
+
 func (t *disabledTun) handleICMPEchoRequest(b []byte) bool {
 	out := make([]byte, len(b))
 	out = iputil.CreateICMPEchoResponse(b, out)
@@ -106,14 +106,12 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
-func (t *disabledTun) Queues(n int) ([]tio.Queue, error) {
-	out := make([]tio.Queue, n)
-	for i := range out {
-		// NoClose: the shared channel and metrics are owned by the
-		// disabledTun; Close on the device tears them down once for everybody.
-		out[i] = tio.NewSingleQueueNoClose(t, defaultBatchBufSize)
-	}
-	return out, nil
+func (t *disabledTun) SupportsMultiqueue() bool {
+	return true
+}
+
+func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return t, nil
 }
 
 func (t *disabledTun) Close() error {

overlay/tun_file_linux_test.go

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

overlay/tun_freebsd.go

@@ -7,6 +7,7 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
+	"io"
 	"io/fs"
 	"log/slog"
 	"net/netip"
@@ -19,7 +20,7 @@ import (
 	"github.com/gaissmai/bart"
 
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
+
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -560,8 +561,12 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *tun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for freebsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_ios.go

@@ -16,7 +16,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"golang.org/x/sys/unix"
@@ -160,6 +159,10 @@ func (t *tun) Name() string {
 	return "iOS"
 }
 
-func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *tun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for ios")
 }

overlay/tun_linux.go

@@ -4,7 +4,9 @@
 package overlay
 
 import (
+	"encoding/binary"
 	"fmt"
+	"io"
 	"log/slog"
 	"net"
 	"net/netip"
@@ -17,15 +19,180 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/vishvananda/netlink"
 	"golang.org/x/sys/unix"
 )
 
+// tunFile wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
+// A shared eventfd allows Close to wake all readers blocked in poll.
+type tunFile struct {
+	fd         int
+	shutdownFd int
+	lastOne    bool
+	readPoll   [2]unix.PollFd
+	writePoll  [2]unix.PollFd
+	closed     bool
+}
+
+// newFriend makes a tunFile for a MultiQueueReader that copies the shutdown eventfd from the parent tun
+func (r *tunFile) newFriend(fd int) (*tunFile, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+	return &tunFile{
+		fd:         fd,
+		shutdownFd: r.shutdownFd,
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
+		},
+	}, nil
+}
+
+func newTunFd(fd int) (*tunFile, error) {
+	if err := unix.SetNonblock(fd, true); err != nil {
+		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
+	}
+
+	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create eventfd: %w", err)
+	}
+
+	out := &tunFile{
+		fd:         fd,
+		shutdownFd: shutdownFd,
+		lastOne:    true,
+		readPoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLIN},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+		writePoll: [2]unix.PollFd{
+			{Fd: int32(fd), Events: unix.POLLOUT},
+			{Fd: int32(shutdownFd), Events: unix.POLLIN},
+		},
+	}
+
+	return out, nil
+}
+
+func (r *tunFile) blockOnRead() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.readPoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.readPoll[0].Revents
+	shutdownEvents := r.readPoll[1].Revents
+	r.readPoll[0].Revents = 0
+	r.readPoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *tunFile) blockOnWrite() error {
+	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
+	var err error
+	for {
+		_, err = unix.Poll(r.writePoll[:], -1)
+		if err != unix.EINTR {
+			break
+		}
+	}
+	//always reset these!
+	tunEvents := r.writePoll[0].Revents
+	shutdownEvents := r.writePoll[1].Revents
+	r.writePoll[0].Revents = 0
+	r.writePoll[1].Revents = 0
+	//do the err check before trusting the potentially bogus bits we just got
+	if err != nil {
+		return err
+	}
+	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
+		return os.ErrClosed
+	} else if tunEvents&problemFlags != 0 {
+		return os.ErrClosed
+	}
+	return nil
+}
+
+func (r *tunFile) Read(buf []byte) (int, error) {
+	for {
+		if n, err := unix.Read(r.fd, buf); err == nil {
+			return n, nil
+		} else if err == unix.EAGAIN {
+			if err = r.blockOnRead(); err != nil {
+				return 0, err
+			}
+			continue
+		} else if err == unix.EINTR {
+			continue
+		} else if err == unix.EBADF {
+			return 0, os.ErrClosed
+		} else {
+			return 0, err
+		}
+	}
+}
+
+func (r *tunFile) Write(buf []byte) (int, error) {
+	for {
+		if n, err := unix.Write(r.fd, buf); err == nil {
+			return n, nil
+		} else if err == unix.EAGAIN {
+			if err = r.blockOnWrite(); err != nil {
+				return 0, err
+			}
+			continue
+		} else if err == unix.EINTR {
+			continue
+		} else if err == unix.EBADF {
+			return 0, os.ErrClosed
+		} else {
+			return 0, err
+		}
+	}
+}
+
+func (r *tunFile) wakeForShutdown() error {
+	var buf [8]byte
+	binary.NativeEndian.PutUint64(buf[:], 1)
+	_, err := unix.Write(int(r.readPoll[1].Fd), buf[:])
+	return err
+}
+
+func (r *tunFile) Close() error {
+	if r.closed { // avoid closing more than once. Technically a fd could get re-used, which would be a problem
+		return nil
+	}
+	r.closed = true
+	if r.lastOne {
+		_ = unix.Close(r.shutdownFd)
+	}
+	return unix.Close(r.fd)
+}
+
 type tun struct {
-	readers     tio.QueueSet
+	*tunFile
+	readers     []*tunFile
 	closeLock   sync.Mutex
 	Device      string
 	vpnNetworks []netip.Prefix
@@ -34,23 +201,6 @@ type tun struct {
 	TXQueueLen  int
 	deviceIndex int
 	ioctlFd     uintptr
-	vnetHdr     bool
-	// offloadFlags is the exact TUN_F_* offload mask newTun negotiated with
-	// the kernel: usoOffloadFlags when USO was accepted, tsoOffloadFlags on
-	// the TSO-only fallback, or 0 when vnetHdr is off. TUNSETOFFLOAD is
-	// device-wide (drivers/net/tun.c set_offload updates tun->set_features
-	// for the whole netdev), so addQueue must replay this exact
-	// mask on every added queue — issuing a narrower mask there would
-	// silently downgrade offloads (e.g. disable USO) for all queues while
-	// they still advertise the stale capability.
-	offloadFlags uint
-	// routeFeatureECN, when true, sets RTAX_FEATURE_ECN on every route we
-	// install for the tun. The kernel then actively negotiates ECN for
-	// connections destined to those prefixes (equivalent to `ip route
-	// change ... features ecn`) regardless of net.ipv4.tcp_ecn, so flows
-	// across the nebula mesh use ECN even when the host default is the
-	// passive setting (=2). Disable via tunnels.ecn=false.
-	routeFeatureECN bool
 
 	Routes                    atomic.Pointer[[]Route]
 	routeTree                 atomic.Pointer[bart.Table[routing.Gateways]]
@@ -89,9 +239,7 @@ 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
-	// on IFF_VNET_HDR after TUNSETIFF, so skip offload on inherited fds.
-	t, err := newTunGeneric(c, l, deviceFd, false, 0, vpnNetworks)
+	t, err := newTunGeneric(c, l, deviceFd, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -101,118 +249,46 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	return t, nil
 }
 
-// openTunDev opens /dev/net/tun, creating the device node first if it's
-// missing (docker containers occasionally omit it).
-func openTunDev() (int, error) {
-	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
-	if err == nil {
-		return fd, nil
-	}
-	if !os.IsNotExist(err) {
-		return -1, err
-	}
-	if err = os.MkdirAll("/dev/net", 0755); err != nil {
-		return -1, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
-	}
-	if err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200))); err != nil {
-		return -1, fmt.Errorf("failed to create /dev/net/tun: %w", err)
-	}
-	fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
-	if err != nil {
-		return -1, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
-	}
-	return fd, nil
-}
-
-// tunSetIff runs TUNSETIFF with the given flags and returns the kernel-chosen
-// device name on success.
-func tunSetIff(fd int, name string, flags uint16) (string, error) {
-	var req ifReq
-	req.Flags = flags
-	copy(req.Name[:], name)
-	if err := ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		return "", err
-	}
-	return strings.Trim(string(req.Name[:]), "\x00"), nil
-}
-
-// tsoOffloadFlags are the TUN_F_* bits we ask the kernel to enable when a
-// TSO-capable TUN is available. CSUM is required as a prerequisite for TSO.
-// TSO_ECN tells the kernel we propagate ECN correctly through coalesce and
-// segmentation, so it can deliver superpackets whose seed has CWR/ECE set
-// or whose IP-level codepoint is CE.
-const tsoOffloadFlags = unix.TUN_F_CSUM | unix.TUN_F_TSO4 | unix.TUN_F_TSO6 | unix.TUN_F_TSO_ECN
-
-// usoOffloadFlags adds UDP Segmentation Offload to tsoOffloadFlags. Requires
-// Linux ≥ 6.2; older kernels reject it and we fall back to TCP-only TSO via
-// tsoOffloadFlags.
-const usoOffloadFlags = tsoOffloadFlags | unix.TUN_F_USO4 | unix.TUN_F_USO6
-
-// offloadUSOEnabled reports whether the negotiated offload mask includes UDP
-// Segmentation Offload. It is the single source of truth for the usoEnabled
-// capability surfaced by each queue, so the mask stored on the tun and the USO
-// bit reported to coalescers can never drift apart.
-func offloadUSOEnabled(offloadFlags uint) bool {
-	return offloadFlags&(unix.TUN_F_USO4|unix.TUN_F_USO6) != 0
-}
-
 func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
-	baseFlags := uint16(unix.IFF_TUN | unix.IFF_NO_PI)
-	if multiqueue {
-		baseFlags |= unix.IFF_MULTI_QUEUE
-	}
-	nameStr := c.GetString("tun.dev", "")
+	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
+	if err != nil {
+		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
+		if os.IsNotExist(err) {
+			err = os.MkdirAll("/dev/net", 0755)
+			if err != nil {
+				return nil, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
+			}
+			err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200)))
+			if err != nil {
+				return nil, fmt.Errorf("failed to create /dev/net/tun: %w", err)
+			}
 
-	// First try to enable IFF_VNET_HDR via TUNSETIFF and negotiate TUN_F_*
-	// offloads via TUNSETOFFLOAD so we can receive TSO/USO superpackets.
-	// We try TSO+USO first, fall back to TSO-only on kernels without USO
-	// (Linux < 6.2), and finally give up on virtio headers entirely and
-	// reopen as a plain TUN if neither offload mask is accepted.
-	fd, err := openTunDev()
-	if err != nil {
-		return nil, err
-	}
-	vnetHdr := true
-	// offloadFlags is the exact TUN_F_* mask the kernel accepted. We remember
-	// it (rather than a plain bool) so addQueue can replay the
-	// identical device-wide mask on added queues instead of downgrading them.
-	var offloadFlags uint
-	name, err := tunSetIff(fd, nameStr, baseFlags|unix.IFF_VNET_HDR)
-	if err != nil {
-		_ = unix.Close(fd)
-		vnetHdr = false
-	} else {
-		// Try TSO+USO first. On kernels without USO support (Linux < 6.2)
-		// the ioctl returns EINVAL; fall back to the TCP-only mask before
-		// giving up on VNET_HDR entirely.
-		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(usoOffloadFlags)); err == nil {
-			offloadFlags = usoOffloadFlags
-		} else if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(tsoOffloadFlags)); err == nil {
-			offloadFlags = tsoOffloadFlags
+			fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
+			if err != nil {
+				return nil, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
+			}
 		} else {
-			l.Warn("Failed to enable TUN offload (TSO); proceeding without virtio headers", "error", err)
-			_ = unix.Close(fd)
-			vnetHdr = false
-		}
-	}
-
-	if !vnetHdr {
-		fd, err = openTunDev()
-		if err != nil {
 			return nil, err
 		}
-		name, err = tunSetIff(fd, nameStr, baseFlags)
-		if err != nil {
-			_ = unix.Close(fd)
-			return nil, &NameError{Name: nameStr, Underlying: err}
+	}
+
+	var req ifReq
+	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
+	if multiqueue {
+		req.Flags |= unix.IFF_MULTI_QUEUE
+	}
+	nameStr := c.GetString("tun.dev", "")
+	copy(req.Name[:], nameStr)
+	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+		_ = unix.Close(fd)
+		return nil, &NameError{
+			Name:       nameStr,
+			Underlying: err,
 		}
 	}
+	name := strings.Trim(string(req.Name[:]), "\x00")
 
-	if vnetHdr {
-		l.Info("TUN offload enabled", "tso", true, "uso", offloadUSOEnabled(offloadFlags))
-	}
-
-	t, err := newTunGeneric(c, l, fd, vnetHdr, offloadFlags, vpnNetworks)
+	t, err := newTunGeneric(c, l, fd, vpnNetworks)
 	if err != nil {
 		return nil, err
 	}
@@ -222,40 +298,21 @@ 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. offloadFlags is the TUN_F_* mask
-// newTun negotiated (0 when vnetHdr is off); the queues' USO capability is
-// derived from it so it can never disagree with the mask we replay on added
-// multiqueue readers.
-func newTunGeneric(c *config.C, l *slog.Logger, fd int, vnetHdr bool, offloadFlags uint, vpnNetworks []netip.Prefix) (*tun, error) {
-	var qs tio.QueueSet
-	var err error
-	if vnetHdr {
-		qs, err = tio.NewOffloadQueueSet(offloadUSOEnabled(offloadFlags))
-	} else {
-		qs, err = tio.NewPollQueueSet()
-	}
-
+// newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
+func newTunGeneric(c *config.C, l *slog.Logger, fd int, vpnNetworks []netip.Prefix) (*tun, error) {
+	tfd, err := newTunFd(fd)
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, err
 	}
-	err = qs.Add(fd)
-	if err != nil {
-		_ = unix.Close(fd)
-		return nil, err
-	}
-
 	t := &tun{
-		readers:                   qs,
+		tunFile:                   tfd,
+		readers:                   []*tunFile{tfd},
 		closeLock:                 sync.Mutex{},
-		vnetHdr:                   vnetHdr,
-		offloadFlags:              offloadFlags,
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
 		useSystemRoutes:           c.GetBool("tun.use_system_route_table", false),
 		useSystemRoutesBufferSize: c.GetInt("tun.use_system_route_table_buffer_size", 0),
-		routeFeatureECN:           c.GetBool("tunnels.ecn", true),
 		routesFromSystem:          map[netip.Prefix]routing.Gateways{},
 		l:                         l,
 	}
@@ -349,56 +406,36 @@ func (t *tun) reload(c *config.C, initial bool) error {
 	return nil
 }
 
-// Queues opens additional kernel multiqueue fds until the device has n
-// queues, then returns them all. The first queue was opened by newTun; each
-// extra fd replays the negotiated offload state (see addQueue).
-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) SupportsMultiqueue() bool {
+	return true
 }
 
-// addQueue opens one more IFF_MULTI_QUEUE fd on the device and adds it to
-// the queue set.
-func (t *tun) addQueue() error {
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
 
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
-		return err
+		return nil, err
 	}
 
-	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
-	if t.vnetHdr {
-		flags |= unix.IFF_VNET_HDR
-	}
-	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
+	var req ifReq
+	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	copy(req.Name[:], t.Device)
+	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
 		_ = unix.Close(fd)
-		return err
+		return nil, err
 	}
 
-	if t.vnetHdr {
-		// Replay the exact mask newTun negotiated. TUNSETOFFLOAD is
-		// device-wide, so issuing the TSO-only mask here would disable USO
-		// for every queue (including queue 0) on kernels where newTun
-		// successfully enabled it, while the queues keep advertising USO.
-		if err = ioctl(uintptr(fd), unix.TUNSETOFFLOAD, uintptr(t.offloadFlags)); err != nil {
-			_ = unix.Close(fd)
-			return fmt.Errorf("failed to enable offload on multiqueue tun fd: %w", err)
-		}
-	}
-
-	err = t.readers.Add(fd)
+	out, err := t.tunFile.newFriend(fd)
 	if err != nil {
 		_ = unix.Close(fd)
-		return err
+		return nil, err
 	}
 
-	return nil
+	t.readers = append(t.readers, out)
+
+	return out, nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -566,18 +603,6 @@ 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
-	}
-	if t.routeFeatureECN {
-		nr.Features |= unix.RTAX_FEATURE_ECN
-	}
 	err := netlink.RouteReplace(&nr)
 	if err != nil {
 		t.l.Warn("Failed to set default route MTU, retrying", "error", err, "cidr", cidr)
@@ -627,9 +652,6 @@ func (t *tun) addRoutes(logErrors bool) error {
 		if r.Metric > 0 {
 			nr.Priority = r.Metric
 		}
-		if t.routeFeatureECN {
-			nr.Features |= unix.RTAX_FEATURE_ECN
-		}
 
 		err := netlink.RouteReplace(&nr)
 		if err != nil {
@@ -856,10 +878,32 @@ func (t *tun) Close() error {
 		t.routeChan = nil
 	}
 
+	// Signal all readers blocked in poll to wake up and exit
+	_ = t.tunFile.wakeForShutdown()
+
 	if t.ioctlFd > 0 {
 		_ = unix.Close(int(t.ioctlFd))
 		t.ioctlFd = 0
 	}
 
-	return t.readers.Close()
+	for i := range t.readers {
+		if i == 0 {
+			continue //we want to close the zeroth reader last
+		}
+		err := t.readers[i].Close()
+		if err != nil {
+			t.l.Error("error closing tun reader", "reader", i, "error", err)
+		} else {
+			t.l.Info("closed tun reader", "reader", i)
+		}
+	}
+
+	//this is t.readers[0] too
+	err := t.tunFile.Close()
+	if err != nil {
+		t.l.Error("error closing tun reader", "reader", 0, "error", err)
+	} else {
+		t.l.Info("closed tun reader", "reader", 0)
+	}
+	return err
 }

overlay/tun_linux_test.go

@@ -3,9 +3,7 @@
 
 package overlay
 
-import (
-	"testing"
-)
+import "testing"
 
 var runAdvMSSTests = []struct {
 	name     string
@@ -34,65 +32,3 @@ func TestTunAdvMSS(t *testing.T) {
 		})
 	}
 }
-
-// TestOffloadUSOEnabled pins the single source of truth for the per-queue USO
-// capability: it is derived from the negotiated offload mask, so the mask
-// stored on the tun and the capability reported to coalescers cannot drift.
-func TestOffloadUSOEnabled(t *testing.T) {
-	// usoOffloadFlags must be a strict superset of tsoOffloadFlags. Otherwise
-	// the TSO-only fallback (and the historic hardcoded-mask bug in
-	// addQueue) would not actually be a downgrade.
-	if usoOffloadFlags&tsoOffloadFlags != tsoOffloadFlags {
-		t.Fatalf("usoOffloadFlags (%#x) is not a superset of tsoOffloadFlags (%#x)", usoOffloadFlags, tsoOffloadFlags)
-	}
-	if usoOffloadFlags == tsoOffloadFlags {
-		t.Fatal("usoOffloadFlags must add bits beyond tsoOffloadFlags")
-	}
-
-	cases := []struct {
-		name         string
-		offloadFlags uint
-		wantUSO      bool
-	}{
-		{"uso-negotiated", usoOffloadFlags, true},
-		{"tso-fallback", tsoOffloadFlags, false},
-		{"no-vnet-hdr", 0, false},
-	}
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			if got := offloadUSOEnabled(tc.offloadFlags); got != tc.wantUSO {
-				t.Fatalf("offloadUSOEnabled(%#x) = %v, want %v", tc.offloadFlags, got, tc.wantUSO)
-			}
-		})
-	}
-}
-
-// TestAddQueueReplaysNegotiatedMask guards the device-wide TUNSETOFFLOAD
-// downgrade bug: addQueue must issue the exact mask newTun negotiated
-// (t.offloadFlags), not a hardcoded TSO-only mask. Because TUNSETOFFLOAD is
-// per-netdev, a narrower mask on an added queue silently disables USO for
-// every queue on a USO-capable kernel while the queues keep advertising it.
-//
-// A full multi-queue exercise needs /dev/net/tun and CAP_NET_ADMIN, which are
-// not available in CI/sandbox, so this asserts on the struct field that the
-// TUNSETOFFLOAD argument is read from.
-func TestAddQueueReplaysNegotiatedMask(t *testing.T) {
-	t.Run("uso-negotiated", func(t *testing.T) {
-		tn := &tun{vnetHdr: true, offloadFlags: usoOffloadFlags}
-		// The ioctl argument in addQueue is uintptr(t.offloadFlags);
-		// it must equal the negotiated USO mask, and must NOT be the TSO-only
-		// mask (the original bug).
-		if tn.offloadFlags != usoOffloadFlags {
-			t.Fatalf("offloadFlags = %#x, want %#x", tn.offloadFlags, usoOffloadFlags)
-		}
-		if tn.offloadFlags == tsoOffloadFlags {
-			t.Fatal("added queue would downgrade USO: offloadFlags must not be the TSO-only mask when USO was negotiated")
-		}
-	})
-	t.Run("tso-fallback", func(t *testing.T) {
-		tn := &tun{vnetHdr: true, offloadFlags: tsoOffloadFlags}
-		if tn.offloadFlags != tsoOffloadFlags {
-			t.Fatalf("offloadFlags = %#x, want %#x", tn.offloadFlags, tsoOffloadFlags)
-		}
-	})
-}

overlay/tun_netbsd.go

@@ -6,6 +6,7 @@ package overlay
 import (
 	"errors"
 	"fmt"
+	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -16,7 +17,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -390,8 +390,12 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *tun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for netbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_openbsd.go

@@ -6,6 +6,7 @@ package overlay
 import (
 	"errors"
 	"fmt"
+	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -16,7 +17,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	netroute "golang.org/x/net/route"
@@ -138,8 +138,8 @@ func tunWritev(fd int, iovecs []unix.Iovec) (n int, err error)
 //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.
+// 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
 
@@ -369,8 +369,12 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *tun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for openbsd")
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_tester.go

@@ -14,7 +14,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/udp"
 )
@@ -178,6 +177,10 @@ func (t *TestTun) Read(b []byte) (int, error) {
 	return n, nil
 }
 
-func (t *TestTun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, udp.MTU)}, nil
+func (t *TestTun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *TestTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented")
 }

overlay/tun_windows.go

@@ -6,6 +6,7 @@ package overlay
 import (
 	"crypto"
 	"fmt"
+	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -17,7 +18,6 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/slackhq/nebula/wintun"
@@ -47,10 +47,6 @@ type winTun struct {
 	tun *wintun.NativeTun
 }
 
-func (t *winTun) Read(b []byte) (int, error) {
-	return t.tun.Read(b, 0)
-}
-
 func newTunFromFd(_ *config.C, _ *slog.Logger, _ int, _ []netip.Prefix) (Device, error) {
 	return nil, fmt.Errorf("newTunFromFd not supported in Windows")
 }
@@ -259,12 +255,20 @@ func (t *winTun) Name() string {
 	return t.Device
 }
 
+func (t *winTun) Read(b []byte) (int, error) {
+	return t.tun.Read(b, 0)
+}
+
 func (t *winTun) Write(b []byte) (int, error) {
 	return t.tun.Write(b, 0)
 }
 
-func (t *winTun) Queues(int) ([]tio.Queue, error) {
-	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
+func (t *winTun) SupportsMultiqueue() bool {
+	return false
+}
+
+func (t *winTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return nil, fmt.Errorf("TODO: multiqueue not implemented for windows")
 }
 
 func (t *winTun) Close() error {

overlay/user.go

@@ -6,7 +6,6 @@ import (
 	"net/netip"
 
 	"github.com/slackhq/nebula/config"
-	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -47,16 +46,12 @@ func (d *UserDevice) RoutesFor(ip netip.Addr) routing.Gateways {
 	return routing.Gateways{routing.NewGateway(ip, 1)}
 }
 
-func (d *UserDevice) Queues(n int) ([]tio.Queue, error) {
-	out := make([]tio.Queue, n)
-	for i := range out {
-		// All queues share the underlying pipes (the io.Pipe serializes
-		// concurrent callers) but each owns a private scratch buffer so
-		// concurrent Reads across queues never alias. NoClose: the pipes are
-		// owned by the UserDevice and torn down once by UserDevice.Close.
-		out[i] = tio.NewSingleQueueNoClose(d, defaultBatchBufSize)
-	}
-	return out, nil
+func (d *UserDevice) SupportsMultiqueue() bool {
+	return true
+}
+
+func (d *UserDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+	return d, nil
 }
 
 func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
@@ -66,11 +61,9 @@ func (d *UserDevice) Pipe() (*io.PipeReader, *io.PipeWriter) {
 func (d *UserDevice) Read(p []byte) (n int, err error) {
 	return d.outboundReader.Read(p)
 }
-
 func (d *UserDevice) Write(p []byte) (n int, err error) {
 	return d.inboundWriter.Write(p)
 }
-
 func (d *UserDevice) Close() error {
 	d.inboundWriter.Close()
 	d.outboundWriter.Close()

overlay/user_test.go

@@ -1,163 +0,0 @@
-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
-}

pki.go

@@ -1,6 +1,7 @@
 package nebula
 
 import (
+	"crypto/fips140"
 	"encoding/binary"
 	"encoding/json"
 	"errors"
@@ -237,10 +238,17 @@ func (cs *CertState) getCertificate(v cert.Version) cert.Certificate {
 	return nil
 }
 
-func newCipherSuite(curve cert.Curve, pkcs11backed bool, cipher string) (noise.CipherSuite, error) {
+// newCipherSuite builds the noise.CipherSuite for the given curve and cipher.
+// When fips140Enforced is true (FIPS 140-only mode), non-approved algorithms
+// (Curve25519 and ChaChaPoly) are rejected with an error. Callers pass
+// fips140.Enforced() for fips140Enforced.
+func newCipherSuite(curve cert.Curve, pkcs11backed bool, cipher string, fips140Enforced bool) (noise.CipherSuite, error) {
 	var dhFunc noise.DHFunc
 	switch curve {
 	case cert.Curve_CURVE25519:
+		if fips140Enforced {
+			return nil, errors.New("pki: use of Curve25519 is not allowed in FIPS 140-only mode")
+		}
 		dhFunc = noise.DH25519
 	case cert.Curve_P256:
 		if pkcs11backed {
@@ -253,6 +261,9 @@ func newCipherSuite(curve cert.Curve, pkcs11backed bool, cipher string) (noise.C
 	}
 
 	if cipher == "chachapoly" {
+		if fips140Enforced {
+			return nil, errors.New("pki: use of ChaChaPoly is not allowed in FIPS 140-only mode")
+		}
 		return noise.NewCipherSuite(dhFunc, noise.CipherChaChaPoly, noise.HashSHA256), nil
 	}
 	return noise.NewCipherSuite(dhFunc, noiseutil.CipherAESGCM, noise.HashSHA256), nil
@@ -326,6 +337,10 @@ func newCertStateFromConfig(c *config.C, cipher string) (*CertState, error) {
 			return nil, err
 		}
 
+		if fips140.Enforced() && crt.Curve() != cert.Curve_P256 {
+			return nil, fmt.Errorf("pki: use of %s is not allowed in FIPS 140-only mode", crt.Curve())
+		}
+
 		switch crt.Version() {
 		case cert.Version1:
 			if v1 != nil {
@@ -405,7 +420,7 @@ func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, p
 			//NOTE: We do not currently have a method to verify a public private key pair when the private key is in an hsm
 		} else {
 			if err := v1.VerifyPrivateKey(privateKeyCurve, privateKey); err != nil {
-				return nil, fmt.Errorf("private key is not a pair with public key in nebula cert")
+				return nil, fmt.Errorf("private key is not a pair with public key in nebula cert: %w", err)
 			}
 		}
 
@@ -413,7 +428,7 @@ func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, p
 		if err != nil {
 			return nil, fmt.Errorf("error marshalling v1 certificate for handshake: %w", err)
 		}
-		ncs, err := newCipherSuite(v1.Curve(), pkcs11backed, cipher)
+		ncs, err := newCipherSuite(v1.Curve(), pkcs11backed, cipher, fips140.Enforced())
 		if err != nil {
 			return nil, err
 		}
@@ -430,7 +445,7 @@ func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, p
 			//NOTE: We do not currently have a method to verify a public private key pair when the private key is in an hsm
 		} else {
 			if err := v2.VerifyPrivateKey(privateKeyCurve, privateKey); err != nil {
-				return nil, fmt.Errorf("private key is not a pair with public key in nebula cert")
+				return nil, fmt.Errorf("private key is not a pair with public key in nebula cert: %w", err)
 			}
 		}
 
@@ -438,7 +453,7 @@ func newCertState(dv cert.Version, v1, v2 cert.Certificate, pkcs11backed bool, p
 		if err != nil {
 			return nil, fmt.Errorf("error marshalling v2 certificate for handshake: %w", err)
 		}
-		ncs, err := newCipherSuite(v2.Curve(), pkcs11backed, cipher)
+		ncs, err := newCipherSuite(v2.Curve(), pkcs11backed, cipher, fips140.Enforced())
 		if err != nil {
 			return nil, err
 		}

pki_test.go

@@ -0,0 +1,94 @@
+package nebula
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/slackhq/nebula/cert"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestNewCipherSuite(t *testing.T) {
+	tests := []struct {
+		name            string
+		curve           cert.Curve
+		cipher          string
+		fips140Enforced bool
+		wantErr         string
+		// wantName is the full expected CipherSuite name (<DH>_<Cipher>_<Hash>),
+		// only checked when wantErr is empty. Asserting the whole name makes both
+		// the curve and cipher selection load-bearing.
+		wantName string
+	}{
+		{
+			name:     "curve25519 aesgcm, not enforced",
+			curve:    cert.Curve_CURVE25519,
+			cipher:   "aesgcm",
+			wantName: "25519_AESGCM_SHA256",
+		},
+		{
+			name:     "curve25519 chachapoly, not enforced",
+			curve:    cert.Curve_CURVE25519,
+			cipher:   "chachapoly",
+			wantName: "25519_ChaChaPoly_SHA256",
+		},
+		{
+			name:     "p256 aesgcm, not enforced",
+			curve:    cert.Curve_P256,
+			cipher:   "aesgcm",
+			wantName: "P256_AESGCM_SHA256",
+		},
+		{
+			name:            "p256 aesgcm, enforced is allowed",
+			curve:           cert.Curve_P256,
+			cipher:          "aesgcm",
+			fips140Enforced: true,
+			wantName:        "P256_AESGCM_SHA256",
+		},
+		{
+			name:            "curve25519 rejected when enforced",
+			curve:           cert.Curve_CURVE25519,
+			cipher:          "aesgcm",
+			fips140Enforced: true,
+			wantErr:         "pki: use of Curve25519 is not allowed in FIPS 140-only mode",
+		},
+		{
+			name:            "chachapoly rejected when enforced",
+			curve:           cert.Curve_P256,
+			cipher:          "chachapoly",
+			fips140Enforced: true,
+			wantErr:         "pki: use of ChaChaPoly is not allowed in FIPS 140-only mode",
+		},
+		{
+			// Curve is checked before cipher, so a Curve25519+ChaChaPoly
+			// request reports the Curve25519 rejection.
+			name:            "curve25519 chachapoly rejected on curve when enforced",
+			curve:           cert.Curve_CURVE25519,
+			cipher:          "chachapoly",
+			fips140Enforced: true,
+			wantErr:         "pki: use of Curve25519 is not allowed in FIPS 140-only mode",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cs, err := newCipherSuite(tt.curve, false, tt.cipher, tt.fips140Enforced)
+			if tt.wantErr != "" {
+				require.EqualError(t, err, tt.wantErr)
+				assert.Nil(t, cs)
+				return
+			}
+			require.NoError(t, err)
+			require.NotNil(t, cs)
+			assert.Equal(t, tt.wantName, string(cs.Name()))
+		})
+	}
+}
+
+func TestNewCipherSuiteUnsupportedCurve(t *testing.T) {
+	cs, err := newCipherSuite(cert.Curve(99), false, "aesgcm", false)
+	require.Error(t, err)
+	assert.True(t, strings.HasPrefix(err.Error(), "unsupported curve:"), "got: %v", err)
+	assert.Nil(t, cs)
+}