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

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

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

control_lifecycle_test.go

@@ -11,6 +11,7 @@ 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/test"
 	"github.com/slackhq/nebula/udp"
@@ -30,9 +31,9 @@ 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(p []byte) (int, error) {
+func (d *fakeDevice) Read() ([]tio.Packet, error) {
 	<-d.closedCh
-	return 0, io.EOF
+	return nil, io.EOF
 }
 
 func (d *fakeDevice) Write(p []byte) (int, error) { return len(p), nil }
@@ -49,10 +50,8 @@ func (d *fakeDevice) Activate() error                       { return nil }
 func (d *fakeDevice) Networks() []netip.Prefix              { return nil }
 func (d *fakeDevice) Name() string                          { return "fake" }
 func (d *fakeDevice) RoutesFor(netip.Addr) routing.Gateways { return nil }
-func (d *fakeDevice) SupportsMultiqueue() bool              { return false }
-func (d *fakeDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, errors.New("unsupported")
-}
+
+func (d *fakeDevice) Queues(int) ([]tio.Queue, error) { return []tio.Queue{d}, nil }
 
 // newReadyControl hand-builds the minimum Control that Main would have
 // produced right before Start, including the construction token NewInterface
@@ -78,7 +77,6 @@ func newReadyControl(t *testing.T) (*Control, *fakeDevice, *fakeConn) {
 		inside:     dev,
 		outside:    conn,
 		writers:    []udp.Conn{conn},
-		readers:    make([]io.ReadWriteCloser, 1),
 		routines:   1,
 		hostMap:    newHostMap(l),
 		lightHouse: lh,
@@ -155,7 +153,14 @@ type multiqueueDevice struct {
 	*fakeDevice
 }
 
-func (d *multiqueueDevice) SupportsMultiqueue() bool { return true }
+// Queues claims multiqueue support but fails to open the second queue,
+// exercising the activation error path.
+func (d *multiqueueDevice) Queues(n int) ([]tio.Queue, error) {
+	if n > 1 {
+		return nil, errors.New("second queue failed to open")
+	}
+	return d.fakeDevice.Queues(n)
+}
 
 func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
 	dev := &multiqueueDevice{fakeDevice: newFakeDevice()}
@@ -166,7 +171,6 @@ func TestControl_StartMultiqueueFailureReleases(t *testing.T) {
 		inside:   dev,
 		outside:  conn,
 		writers:  []udp.Conn{conn},
-		readers:  make([]io.ReadWriteCloser, 2),
 		routines: 2,
 		l:        test.NewLogger(),
 	}

examples/config.yml

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

inside.go

@@ -37,7 +37,7 @@ func (f *Interface) consumeInsidePacket(packet []byte, fwPacket *firewall.Packet
 		// routes packets from the Nebula addr to the Nebula addr through the Nebula
 		// TUN device.
 		if immediatelyForwardToSelf {
-			_, err := f.readers[q].Write(packet)
+			_, err := f.queues[q].Write(packet)
 			if err != nil {
 				f.l.Error("Failed to forward to tun", "error", err)
 			}
@@ -96,7 +96,7 @@ func (f *Interface) rejectInside(packet []byte, out []byte, q int) {
 		return
 	}
 
-	_, err := f.readers[q].Write(out)
+	_, err := f.queues[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}

interface.go

@@ -4,9 +4,9 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
+	"runtime"
 	"slices"
 	"sync"
 	"sync/atomic"
@@ -20,7 +20,9 @@ import (
 	"github.com/slackhq/nebula/firewall"
 	"github.com/slackhq/nebula/header"
 	"github.com/slackhq/nebula/overlay"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/udp"
+	"github.com/slackhq/nebula/util"
 )
 
 const mtu = 9001
@@ -49,7 +51,19 @@ type InterfaceConfig struct {
 	reQueryWait     time.Duration
 
 	ConntrackCacheTimeout time.Duration
-	l                     *slog.Logger
+
+	// CpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
+	// should pin to. Queue i pins to CpuAffinity[i % len(CpuAffinity)] —
+	// shorter lists than `routines` cycle. Empty list keeps the default
+	// pin-to-(i % NumCPU) behavior. Only consulted when PinThreads is true.
+	CpuAffinity []int
+	// PinThreads controls whether each TUN reader OS thread is pinned to a
+	// single CPU (via tun.pin_threads, default true). Pinning keeps each
+	// goroutine's UDP sends on one XPS-selected NIC TX ring so per-flow
+	// packets stay ordered on the wire.
+	PinThreads bool
+
+	l *slog.Logger
 }
 
 type Interface struct {
@@ -73,7 +87,16 @@ type Interface struct {
 	routines              int
 	disconnectInvalid     atomic.Bool
 	closed                atomic.Bool
-	relayManager          *relayManager
+	// cpuAffinity, when non-empty, names the CPUs each TUN reader goroutine
+	// should pin to. Queue i pins to cpuAffinity[i % len(cpuAffinity)].
+	// Empty falls back to the default pin-to-(allowed CPU) behavior.
+	// Only consulted when pinThreads is true.
+	cpuAffinity []int
+	// pinThreads controls whether listenIn pins each TUN reader OS thread to
+	// a CPU at all (tun.pin_threads, default true). When false, threads are
+	// left free to migrate as on stock nebula.
+	pinThreads   bool
+	relayManager *relayManager
 
 	tryPromoteEvery atomic.Uint32
 	reQueryEvery    atomic.Uint32
@@ -90,7 +113,7 @@ type Interface struct {
 
 	ctx     context.Context
 	writers []udp.Conn
-	readers []io.ReadWriteCloser
+	queues  []tio.Queue
 	wg      sync.WaitGroup
 
 	// fatalErr holds the first unexpected reader error that caused shutdown.
@@ -189,7 +212,6 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		routines:              c.routines,
 		version:               c.version,
 		writers:               make([]udp.Conn, c.routines),
-		readers:               make([]io.ReadWriteCloser, c.routines),
 		myVpnNetworks:         cs.myVpnNetworks,
 		myVpnNetworksTable:    cs.myVpnNetworksTable,
 		myVpnAddrs:            cs.myVpnAddrs,
@@ -198,6 +220,8 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
 		relayManager:          c.relayManager,
 		connectionManager:     c.connectionManager,
 		conntrackCacheTimeout: c.ConntrackCacheTimeout,
+		cpuAffinity:           c.CpuAffinity,
+		pinThreads:            c.PinThreads,
 
 		metricHandshakes: metrics.GetOrRegisterHistogram("handshakes", nil, metrics.NewExpDecaySample(1028, 0.015)),
 		messageMetrics:   c.MessageMetrics,
@@ -240,27 +264,27 @@ func (f *Interface) activate() error {
 		"boringcrypto", boringEnabled(),
 	)
 
-	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")
-		}
+	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")
 	}
 
+	// Prepare the tun queues. A device that can't open that many hands back
+	// fewer (a single queue on platforms without multiqueue support) and we
+	// size the reader routines to what we actually got.
+	queues, err := f.inside.Queues(f.routines)
+	if err != nil {
+		return err
+	}
+	if len(queues) < f.routines {
+		f.l.Warn("tun multiqueue is not supported on this platform, falling back to fewer routines",
+			"requested", f.routines, "opened", len(queues))
+		f.routines = len(queues)
+	}
+	f.queues = queues
+
 	metrics.GetOrRegisterGauge("routines", nil).Update(int64(f.routines))
 
-	// Prepare n tun queues
-	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
 	// before releasing our resources so a waiter never observes a live context
 	if err = f.inside.Activate(); err != nil {
@@ -281,7 +305,7 @@ 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.readers[i], i)
+			f.listenIn(f.queues[i], i)
 		})
 	}
 
@@ -336,8 +360,29 @@ func (f *Interface) listenOut(i int) {
 	f.l.Debug("underlay reader is done", "reader", i)
 }
 
-func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
-	packet := make([]byte, mtu)
+func (f *Interface) listenIn(queue tio.Queue, i int) {
+	// Pinning this thread (and goroutine) to a single CPU keeps every UDP send from this goroutine going through
+	// the same TX ring on the nic (XPS selects the ring by CPU), 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)
+		}
+	}
+
 	out := make([]byte, mtu)
 	fwPacket := &firewall.Packet{}
 	nb := make([]byte, 12, 12)
@@ -345,7 +390,7 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 	conntrackCache := firewall.NewConntrackCacheTicker(f.ctx, f.l, f.conntrackCacheTimeout)
 
 	for {
-		n, err := reader.Read(packet)
+		pkts, err := queue.Read()
 		if err != nil {
 			// Same shutdown noise handling as listenOut
 			if !f.closed.Load() && f.ctx.Err() == nil {
@@ -355,7 +400,11 @@ func (f *Interface) listenIn(reader io.ReadWriteCloser, i int) {
 			break
 		}
 
-		f.consumeInsidePacket(packet[:n], fwPacket, nb, out, i, conntrackCache.Get())
+		for _, pkt := range pkts {
+			// borrowed: pkt.Bytes is owned by the queue and only valid until
+			// the next Read; consumeInsidePacket reads it synchronously.
+			f.consumeInsidePacket(pkt.Bytes, fwPacket, nb, out, i, conntrackCache.Get())
+		}
 	}
 
 	f.l.Debug("overlay reader is done", "reader", i)

main.go

@@ -7,6 +7,7 @@ import (
 	"net"
 	"net/netip"
 	"runtime/debug"
+	"slices"
 	"strings"
 	"time"
 
@@ -231,6 +232,8 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 		relayManager:          NewRelayManager(ctx, l, hostMap, c),
 		punchy:                punchy,
 		ConntrackCacheTimeout: conntrackCacheTimeout,
+		CpuAffinity:           parseCpuAffinity(c, l, routines),
+		PinThreads:            c.GetBool("tun.pin_threads", true),
 		l:                     l,
 	}
 
@@ -282,6 +285,70 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
 	}, nil
 }
 
+// parseCpuAffinity reads `tun.cpu_affinity` from the config — a list of
+// integer CPU IDs, one per TUN reader goroutine. Empty / unset returns nil
+// (listenIn falls back to spreading queues across the allowed CPU set).
+// Length mismatch with `routines` is a warning, not an error: shorter lists
+// are modulo-cycled across queues, longer lists' tail is ignored. Invalid
+// entries (non-integer, or a CPU ID we're not allowed to run on) are also a
+// warning and disable the override entirely so we don't silently pin to the
+// wrong CPU. Entries are validated against the process's current affinity
+// mask (util.AllowedCPUs) rather than 0..NumCPU-1: under a cgroup cpuset or
+// taskset the runnable IDs are frequently not that contiguous range, and
+// pinning to an unrunnable ID always fails. If the allowed set can't be
+// determined we fall back to a plain non-negative check.
+func parseCpuAffinity(c *config.C, l *slog.Logger, routines int) []int {
+	raw := c.Get("tun.cpu_affinity")
+	if raw == nil {
+		return nil
+	}
+	rv, ok := raw.([]any)
+	if !ok {
+		l.Warn("tun.cpu_affinity must be a list of integers; ignoring", "value", raw)
+		return nil
+	}
+	// allowed is the set of CPU IDs we're actually permitted to run on. A nil
+	// slice (unsupported platform or lookup error) means "can't tell", so we
+	// only apply the weaker non-negative check in that case.
+	allowed, err := util.AllowedCPUs()
+	if err != nil {
+		l.Warn("could not determine allowed CPUs; validating tun.cpu_affinity against non-negative only", "error", err)
+		allowed = nil
+	}
+	cpus := make([]int, 0, len(rv))
+	for i, e := range rv {
+		var cpu int
+		switch v := e.(type) {
+		case int:
+			cpu = v
+		case int64:
+			cpu = int(v)
+		case float64:
+			cpu = int(v)
+		default:
+			l.Warn("tun.cpu_affinity entry not an integer; ignoring affinity",
+				"index", i, "value", e)
+			return nil
+		}
+		if cpu < 0 {
+			l.Warn("tun.cpu_affinity entry out of range; ignoring affinity",
+				"index", i, "cpu", cpu)
+			return nil
+		}
+		if len(allowed) > 0 && !slices.Contains(allowed, cpu) {
+			l.Warn("tun.cpu_affinity entry not in allowed CPU set; ignoring affinity",
+				"index", i, "cpu", cpu, "allowed", allowed)
+			return nil
+		}
+		cpus = append(cpus, cpu)
+	}
+	if len(cpus) != routines {
+		l.Warn("tun.cpu_affinity length doesn't match routines; queues will modulo-cycle through the list",
+			"affinity_len", len(cpus), "routines", routines)
+	}
+	return cpus
+}
+
 func moduleVersion() string {
 	info, ok := debug.ReadBuildInfo()
 	if !ok {

main_test.go

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

outside.go

@@ -542,7 +542,7 @@ func (f *Interface) handleOutsideMessagePacket(hostinfo *HostInfo, out []byte, p
 		return
 	}
 
-	_, err = f.readers[q].Write(out)
+	_, err = f.queues[q].Write(out)
 	if err != nil {
 		f.l.Error("Failed to write to tun", "error", err)
 	}

overlay/device.go

@@ -4,15 +4,25 @@ import (
 	"io"
 	"net/netip"
 
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
+// defaultBatchBufSize is the per-Queue scratch size for Read. 65535 covers
+// any single IP packet.
+const defaultBatchBufSize = 65535
+
 type Device interface {
-	io.ReadWriteCloser
+	io.Closer
 	Activate() error
 	Networks() []netip.Prefix
 	Name() string
 	RoutesFor(netip.Addr) routing.Gateways
-	SupportsMultiqueue() bool
-	NewMultiQueueReader() (io.ReadWriteCloser, error)
+	// 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)
 }

overlay/overlaytest/noop.go

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

overlay/tio/blockon_linux.go

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

overlay/tio/queueset_poll_linux.go

@@ -0,0 +1,90 @@
+//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/single.go

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

overlay/tio/tio.go

@@ -0,0 +1,52 @@
+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
+}
+
+// Queue is a readable/writable packet queue. Concurrency contract: a single
+// read goroutine drives Read; plain Write is safe for concurrent callers.
+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. 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.
+type Packet struct {
+	Bytes []byte
+}
+
+// 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.
+// 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}
+}

overlay/tio/tio_poll_linux.go

@@ -0,0 +1,116 @@
+//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. 65535 covers any
+// single IP packet.
+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 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

@@ -0,0 +1,208 @@
+//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())
+}

overlay/tun_android.go

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

overlay/tun_darwin.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -16,6 +15,7 @@ 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"
@@ -606,10 +606,6 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-func (t *tun) SupportsMultiqueue() bool {
-	return false
-}
-
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for darwin")
+func (t *tun) Queues(int) ([]tio.Queue, error) {
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
 }

overlay/tun_disabled.go

@@ -10,6 +10,7 @@ import (
 
 	"github.com/rcrowley/go-metrics"
 	"github.com/slackhq/nebula/iputil"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -23,6 +24,23 @@ 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,
@@ -57,24 +75,6 @@ func (*disabledTun) Name() string {
 	return "disabled"
 }
 
-func (t *disabledTun) Read(b []byte) (int, error) {
-	r, ok := <-t.read
-	if !ok {
-		return 0, io.EOF
-	}
-
-	if len(r) > len(b) {
-		return 0, fmt.Errorf("packet larger than mtu: %d > %d bytes", len(r), len(b))
-	}
-
-	t.tx.Inc(1)
-	if t.l.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)
@@ -106,12 +106,14 @@ func (t *disabledTun) Write(b []byte) (int, error) {
 	return len(b), nil
 }
 
-func (t *disabledTun) SupportsMultiqueue() bool {
-	return true
-}
-
-func (t *disabledTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return t, 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) Close() error {

overlay/tun_file_linux_test.go

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

overlay/tun_freebsd.go

@@ -7,7 +7,6 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
-	"io"
 	"io/fs"
 	"log/slog"
 	"net/netip"
@@ -20,7 +19,7 @@ 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"
@@ -561,12 +560,8 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-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) Queues(int) ([]tio.Queue, error) {
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_ios.go

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

overlay/tun_linux.go

@@ -4,9 +4,7 @@
 package overlay
 
 import (
-	"encoding/binary"
 	"fmt"
-	"io"
 	"log/slog"
 	"net"
 	"net/netip"
@@ -19,180 +17,15 @@ import (
 
 	"github.com/gaissmai/bart"
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 	"github.com/slackhq/nebula/util"
 	"github.com/vishvananda/netlink"
 	"golang.org/x/sys/unix"
 )
 
-// tunFile wraps a TUN file descriptor with poll-based reads. The FD provided will be changed to non-blocking.
-// A shared eventfd allows Close to wake all readers blocked in poll.
-type tunFile struct {
-	fd         int
-	shutdownFd int
-	lastOne    bool
-	readPoll   [2]unix.PollFd
-	writePoll  [2]unix.PollFd
-	closed     bool
-}
-
-// newFriend makes a tunFile for a MultiQueueReader that copies the shutdown eventfd from the parent tun
-func (r *tunFile) newFriend(fd int) (*tunFile, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-	return &tunFile{
-		fd:         fd,
-		shutdownFd: r.shutdownFd,
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(r.shutdownFd), Events: unix.POLLIN},
-		},
-	}, nil
-}
-
-func newTunFd(fd int) (*tunFile, error) {
-	if err := unix.SetNonblock(fd, true); err != nil {
-		return nil, fmt.Errorf("failed to set tun fd non-blocking: %w", err)
-	}
-
-	shutdownFd, err := unix.Eventfd(0, unix.EFD_NONBLOCK|unix.EFD_CLOEXEC)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create eventfd: %w", err)
-	}
-
-	out := &tunFile{
-		fd:         fd,
-		shutdownFd: shutdownFd,
-		lastOne:    true,
-		readPoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLIN},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-		writePoll: [2]unix.PollFd{
-			{Fd: int32(fd), Events: unix.POLLOUT},
-			{Fd: int32(shutdownFd), Events: unix.POLLIN},
-		},
-	}
-
-	return out, nil
-}
-
-func (r *tunFile) blockOnRead() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(r.readPoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	//always reset these!
-	tunEvents := r.readPoll[0].Revents
-	shutdownEvents := r.readPoll[1].Revents
-	r.readPoll[0].Revents = 0
-	r.readPoll[1].Revents = 0
-	//do the err check before trusting the potentially bogus bits we just got
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	} else if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
-}
-
-func (r *tunFile) blockOnWrite() error {
-	const problemFlags = unix.POLLHUP | unix.POLLNVAL | unix.POLLERR
-	var err error
-	for {
-		_, err = unix.Poll(r.writePoll[:], -1)
-		if err != unix.EINTR {
-			break
-		}
-	}
-	//always reset these!
-	tunEvents := r.writePoll[0].Revents
-	shutdownEvents := r.writePoll[1].Revents
-	r.writePoll[0].Revents = 0
-	r.writePoll[1].Revents = 0
-	//do the err check before trusting the potentially bogus bits we just got
-	if err != nil {
-		return err
-	}
-	if shutdownEvents&(unix.POLLIN|problemFlags) != 0 {
-		return os.ErrClosed
-	} else if tunEvents&problemFlags != 0 {
-		return os.ErrClosed
-	}
-	return nil
-}
-
-func (r *tunFile) Read(buf []byte) (int, error) {
-	for {
-		if n, err := unix.Read(r.fd, buf); err == nil {
-			return n, nil
-		} else if err == unix.EAGAIN {
-			if err = r.blockOnRead(); err != nil {
-				return 0, err
-			}
-			continue
-		} else if err == unix.EINTR {
-			continue
-		} else if err == unix.EBADF {
-			return 0, os.ErrClosed
-		} else {
-			return 0, err
-		}
-	}
-}
-
-func (r *tunFile) Write(buf []byte) (int, error) {
-	for {
-		if n, err := unix.Write(r.fd, buf); err == nil {
-			return n, nil
-		} else if err == unix.EAGAIN {
-			if err = r.blockOnWrite(); err != nil {
-				return 0, err
-			}
-			continue
-		} else if err == unix.EINTR {
-			continue
-		} else if err == unix.EBADF {
-			return 0, os.ErrClosed
-		} else {
-			return 0, err
-		}
-	}
-}
-
-func (r *tunFile) wakeForShutdown() error {
-	var buf [8]byte
-	binary.NativeEndian.PutUint64(buf[:], 1)
-	_, err := unix.Write(int(r.readPoll[1].Fd), buf[:])
-	return err
-}
-
-func (r *tunFile) Close() error {
-	if r.closed { // avoid closing more than once. Technically a fd could get re-used, which would be a problem
-		return nil
-	}
-	r.closed = true
-	if r.lastOne {
-		_ = unix.Close(r.shutdownFd)
-	}
-	return unix.Close(r.fd)
-}
-
 type tun struct {
-	*tunFile
-	readers     []*tunFile
+	readers     tio.QueueSet
 	closeLock   sync.Mutex
 	Device      string
 	vpnNetworks []netip.Prefix
@@ -249,44 +82,57 @@ func newTunFromFd(c *config.C, l *slog.Logger, deviceFd int, vpnNetworks []netip
 	return t, nil
 }
 
-func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+// openTunDev opens /dev/net/tun, creating the device node first if it's
+// missing (docker containers occasionally omit it).
+func openTunDev() (int, error) {
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
-	if err != nil {
-		// If /dev/net/tun doesn't exist, try to create it (will happen in docker)
-		if os.IsNotExist(err) {
-			err = os.MkdirAll("/dev/net", 0755)
-			if err != nil {
-				return nil, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
-			}
-			err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200)))
-			if err != nil {
-				return nil, fmt.Errorf("failed to create /dev/net/tun: %w", err)
-			}
-
-			fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
-			if err != nil {
-				return nil, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
-			}
-		} else {
-			return nil, err
-		}
+	if err == nil {
+		return fd, nil
 	}
+	if !os.IsNotExist(err) {
+		return -1, err
+	}
+	if err = os.MkdirAll("/dev/net", 0755); err != nil {
+		return -1, fmt.Errorf("/dev/net/tun doesn't exist, failed to mkdir -p /dev/net: %w", err)
+	}
+	if err = unix.Mknod("/dev/net/tun", unix.S_IFCHR|0600, int(unix.Mkdev(10, 200))); err != nil {
+		return -1, fmt.Errorf("failed to create /dev/net/tun: %w", err)
+	}
+	fd, err = unix.Open("/dev/net/tun", os.O_RDWR, 0)
+	if err != nil {
+		return -1, fmt.Errorf("created /dev/net/tun, but still failed: %w", err)
+	}
+	return fd, nil
+}
 
+// tunSetIff runs TUNSETIFF with the given flags and returns the kernel-chosen
+// device name on success.
+func tunSetIff(fd int, name string, flags uint16) (string, error) {
 	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI)
+	req.Flags = flags
+	copy(req.Name[:], name)
+	if err := ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+		return "", err
+	}
+	return strings.Trim(string(req.Name[:]), "\x00"), nil
+}
+
+func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue bool) (*tun, error) {
+	baseFlags := uint16(unix.IFF_TUN | unix.IFF_NO_PI)
 	if multiqueue {
-		req.Flags |= unix.IFF_MULTI_QUEUE
+		baseFlags |= unix.IFF_MULTI_QUEUE
 	}
 	nameStr := c.GetString("tun.dev", "")
-	copy(req.Name[:], nameStr)
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
-		_ = unix.Close(fd)
-		return nil, &NameError{
-			Name:       nameStr,
-			Underlying: err,
-		}
+
+	fd, err := openTunDev()
+	if err != nil {
+		return nil, err
+	}
+	name, err := tunSetIff(fd, nameStr, baseFlags)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, &NameError{Name: nameStr, Underlying: err}
 	}
-	name := strings.Trim(string(req.Name[:]), "\x00")
 
 	t, err := newTunGeneric(c, l, fd, vpnNetworks)
 	if err != nil {
@@ -298,16 +144,22 @@ func newTun(c *config.C, l *slog.Logger, vpnNetworks []netip.Prefix, multiqueue
 	return t, nil
 }
 
-// newTunGeneric does all the stuff common to different tun initialization paths. It will close your files on error.
+// newTunGeneric does all the stuff common to different tun initialization
+// 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)
+	qs, err := tio.NewPollQueueSet()
 	if err != nil {
 		_ = unix.Close(fd)
 		return nil, err
 	}
+	err = qs.Add(fd)
+	if err != nil {
+		_ = unix.Close(fd)
+		return nil, err
+	}
+
 	t := &tun{
-		tunFile:                   tfd,
-		readers:                   []*tunFile{tfd},
+		readers:                   qs,
 		closeLock:                 sync.Mutex{},
 		vpnNetworks:               vpnNetworks,
 		TXQueueLen:                c.GetInt("tun.tx_queue", 500),
@@ -406,36 +258,41 @@ func (t *tun) reload(c *config.C, initial bool) error {
 	return nil
 }
 
-func (t *tun) SupportsMultiqueue() bool {
-	return true
+// Queues opens additional kernel multiqueue fds until the device has n
+// queues, then returns them all. The first queue was opened by newTun.
+func (t *tun) Queues(n int) ([]tio.Queue, error) {
+	for len(t.readers.Queues()) < n {
+		if err := t.addQueue(); err != nil {
+			return nil, err
+		}
+	}
+	return t.readers.Queues(), nil
 }
 
-func (t *tun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
+// addQueue opens one more IFF_MULTI_QUEUE fd on the device and adds it to
+// the queue set.
+func (t *tun) addQueue() error {
 	t.closeLock.Lock()
 	defer t.closeLock.Unlock()
 
 	fd, err := unix.Open("/dev/net/tun", os.O_RDWR, 0)
 	if err != nil {
-		return nil, err
+		return err
 	}
 
-	var req ifReq
-	req.Flags = uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
-	copy(req.Name[:], t.Device)
-	if err = ioctl(uintptr(fd), uintptr(unix.TUNSETIFF), uintptr(unsafe.Pointer(&req))); err != nil {
+	flags := uint16(unix.IFF_TUN | unix.IFF_NO_PI | unix.IFF_MULTI_QUEUE)
+	if _, err = tunSetIff(fd, t.Device, flags); err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	out, err := t.tunFile.newFriend(fd)
+	err = t.readers.Add(fd)
 	if err != nil {
 		_ = unix.Close(fd)
-		return nil, err
+		return err
 	}
 
-	t.readers = append(t.readers, out)
-
-	return out, nil
+	return nil
 }
 
 func (t *tun) RoutesFor(ip netip.Addr) routing.Gateways {
@@ -878,32 +735,10 @@ func (t *tun) Close() error {
 		t.routeChan = nil
 	}
 
-	// Signal all readers blocked in poll to wake up and exit
-	_ = t.tunFile.wakeForShutdown()
-
 	if t.ioctlFd > 0 {
 		_ = unix.Close(int(t.ioctlFd))
 		t.ioctlFd = 0
 	}
 
-	for i := range t.readers {
-		if i == 0 {
-			continue //we want to close the zeroth reader last
-		}
-		err := t.readers[i].Close()
-		if err != nil {
-			t.l.Error("error closing tun reader", "reader", i, "error", err)
-		} else {
-			t.l.Info("closed tun reader", "reader", i)
-		}
-	}
-
-	//this is t.readers[0] too
-	err := t.tunFile.Close()
-	if err != nil {
-		t.l.Error("error closing tun reader", "reader", 0, "error", err)
-	} else {
-		t.l.Info("closed tun reader", "reader", 0)
-	}
-	return err
+	return t.readers.Close()
 }

overlay/tun_netbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -17,6 +16,7 @@ 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"
@@ -390,12 +390,8 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-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) Queues(int) ([]tio.Queue, error) {
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_openbsd.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -17,6 +16,7 @@ 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"
@@ -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
 
@@ -369,12 +369,8 @@ func (t *tun) Name() string {
 	return t.Device
 }
 
-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) Queues(int) ([]tio.Queue, error) {
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
 }
 
 func (t *tun) addRoutes(logErrors bool) error {

overlay/tun_tester.go

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

overlay/tun_windows.go

@@ -6,7 +6,6 @@ package overlay
 import (
 	"crypto"
 	"fmt"
-	"io"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -18,6 +17,7 @@ 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"
 	"github.com/slackhq/nebula/wintun"
@@ -47,6 +47,10 @@ 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")
 }
@@ -255,20 +259,12 @@ 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) SupportsMultiqueue() bool {
-	return false
-}
-
-func (t *winTun) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return nil, fmt.Errorf("TODO: multiqueue not implemented for windows")
+func (t *winTun) Queues(int) ([]tio.Queue, error) {
+	return []tio.Queue{tio.NewSingleQueue(t, defaultBatchBufSize)}, nil
 }
 
 func (t *winTun) Close() error {

overlay/user.go

@@ -6,6 +6,7 @@ import (
 	"net/netip"
 
 	"github.com/slackhq/nebula/config"
+	"github.com/slackhq/nebula/overlay/tio"
 	"github.com/slackhq/nebula/routing"
 )
 
@@ -46,12 +47,16 @@ func (d *UserDevice) RoutesFor(ip netip.Addr) routing.Gateways {
 	return routing.Gateways{routing.NewGateway(ip, 1)}
 }
 
-func (d *UserDevice) SupportsMultiqueue() bool {
-	return true
-}
-
-func (d *UserDevice) NewMultiQueueReader() (io.ReadWriteCloser, error) {
-	return d, nil
+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) Pipe() (*io.PipeReader, *io.PipeWriter) {
@@ -61,9 +66,11 @@ 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

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

util/cpupin_linux.go

@@ -0,0 +1,43 @@
+//go:build linux && !android && !e2e_testing
+
+package util
+
+import (
+	"runtime"
+
+	"golang.org/x/sys/unix"
+)
+
+// PinThreadToCPU restricts the calling OS thread to the given CPU via
+// sched_setaffinity(2). Combined with runtime.LockOSThread on the
+// goroutine, this prevents the kernel from migrating us across CPUs and
+// in turn keeps every UDP send from this goroutine going through the
+// same XPS-selected TX ring, eliminating the wire-side reorder that
+// otherwise fragments one nebula flow across multiple rings.
+func PinThreadToCPU(cpu int) error {
+	runtime.LockOSThread()
+	var set unix.CPUSet
+	set.Zero()
+	set.Set(cpu)
+	return unix.SchedSetaffinity(0, &set)
+}
+
+// AllowedCPUs returns the CPU IDs the calling process is currently allowed to
+// run on, as reported by sched_getaffinity(2). Under a cgroup cpuset or a
+// `taskset` mask the allowed IDs are frequently not the contiguous range
+// 0..NumCPU-1 (e.g. pinned to CPUs 4-7: NumCPU reports 4 while the valid IDs
+// are 4,5,6,7). Callers that need a real CPU to pin to must choose from this
+// set rather than assuming i % NumCPU is runnable, or every pin fails.
+func AllowedCPUs() ([]int, error) {
+	var set unix.CPUSet
+	if err := unix.SchedGetaffinity(0, &set); err != nil {
+		return nil, err
+	}
+	cpus := make([]int, 0, set.Count())
+	for cpu := 0; cpu < len(set)*64; cpu++ {
+		if set.IsSet(cpu) {
+			cpus = append(cpus, cpu)
+		}
+	}
+	return cpus, nil
+}

util/cpupin_other.go

@@ -0,0 +1,18 @@
+//go:build !linux || android || e2e_testing
+
+package util
+
+// PinThreadToCPU is a no-op outside Linux: only Linux exposes a stable
+// per-thread CPU affinity API and only Linux has XPS-driven TX ring
+// selection in the first place. On every other platform there's nothing
+// to fix here.
+func PinThreadToCPU(_ int) error {
+	return nil
+}
+
+// AllowedCPUs has no meaningful answer off Linux (no sched_getaffinity), so it
+// reports "unknown" by returning a nil slice and nil error. Callers treat an
+// empty result as "fall back to the default CPU choice".
+func AllowedCPUs() ([]int, error) {
+	return nil, nil
+}