Compare commits

...
Author SHA1 Message Date
JackDoan 0f59f895a4 cpupick: Keep single readers off tun.avoid_cpus without cpupick, warn and ignore bad lists 2026-10-15 12:21:59 -05:00
JackDoan df1c4d9364 cpupick: Make tun.avoid_cpus a pin selection rule instead of a process-wide mask 2026-10-15 12:13:21 -05:00
JackDoan 36579bae03 cpupick: Move AvoidCPUs from util into cpupick as Avoid 2026-10-15 12:00:28 -05:00
Ryan Huber 73d26dd87d tun: Allow excluding CPUs from process threads at startup
Set tun.avoid_cpus to a list of logical CPUs to remove from the affinity of
every thread of the process on Linux, before services start and reader CPUs
are chosen. Later threads inherit the restricted mask; enumeration repeats so
threads created meanwhile are covered. Inherited taskset and cgroup
restrictions are kept. Keeping Nebula off the underlay NIC's receive IRQ CPUs
and their SMT siblings is what this is for; the example configuration shows
how to find them.

Invalid IDs, a list that would leave a thread no CPU, and syscall failures
fail startup. Configuration tests only validate. Other platforms log and
ignore the setting. The exclusion is process-wide, including embedded use, and
not reloadable. It is independent of thread pinning.

New API: util.AvoidCPUs, next to the existing PinThreadToCPU and AllowedCPUs.

Tests run in a child process and check existing and new threads, preserved
inherited masks, validation without changes, and the unix.CPUSet bounds on
32- and 64-bit platforms. No isolated performance claim is made.
2026-10-15 14:40:45 +00:00
10 changed files with 359 additions and 18 deletions

No files matched your search

+75
View File
@@ -0,0 +1,75 @@
//go:build linux && !android && !e2e_testing
package cpupick
import (
"errors"
"fmt"
"os"
"strconv"
"unsafe"
"golang.org/x/sys/unix"
)
// Avoid removes cpus from the affinity of every thread of the process, for use at startup. New threads inherit
// their parent's mask, so it repeats until a pass finds nothing left to change; a thread that exits meanwhile is
// skipped. Every mask is checked before any is changed. With apply false it only validates, for configuration tests.
// This affects the entire process, including other users of an embedded Nebula. Not wired up yet: tun.avoid_cpus
// only steers Default for now.
func Avoid(cpus []int, apply bool) error {
const capacity = int(unsafe.Sizeof(unix.CPUSet{})) * 8
for _, cpu := range cpus {
if cpu < 0 || cpu >= capacity {
return fmt.Errorf("CPU %d is out of range", cpu)
}
}
if len(cpus) == 0 {
return nil
}
for {
masks, err := avoidingMasks(cpus)
if err != nil || !apply || len(masks) == 0 {
return err
}
for tid, mask := range masks {
if err := unix.SchedSetaffinity(tid, &mask); err != nil && !errors.Is(err, unix.ESRCH) {
return fmt.Errorf("set thread %d affinity: %w", tid, err)
}
}
}
}
// avoidingMasks returns the mask without cpus of every thread whose mask has any of them.
func avoidingMasks(cpus []int) (map[int]unix.CPUSet, error) {
entries, err := os.ReadDir("/proc/self/task")
if err != nil {
return nil, err
}
masks := make(map[int]unix.CPUSet, len(entries))
for _, entry := range entries {
tid, err := strconv.Atoi(entry.Name())
if err != nil {
return nil, err
}
var mask unix.CPUSet
err = unix.SchedGetaffinity(tid, &mask)
if errors.Is(err, unix.ESRCH) {
continue
}
if err != nil {
return nil, fmt.Errorf("read thread %d affinity: %w", tid, err)
}
next := mask
for _, cpu := range cpus {
next.Clear(cpu)
}
if next.Count() == 0 {
return nil, fmt.Errorf("avoid_cpus leaves thread %d with no allowed CPU", tid)
}
if next != mask {
masks[tid] = next
}
}
return masks, nil
}
+98
View File
@@ -0,0 +1,98 @@
//go:build linux && !android && !e2e_testing
package cpupick
import (
"fmt"
"os"
"os/exec"
"runtime"
"strconv"
"sync"
"testing"
"unsafe"
"github.com/slackhq/nebula/util"
"github.com/stretchr/testify/require"
"golang.org/x/sys/unix"
)
// TestAvoid changes the affinity of the whole process, so it runs in a child, leaving the test runner untouched.
func TestAvoid(t *testing.T) {
if os.Getenv("NEBULA_CPUPICK_AVOID_TEST") != "1" {
exe, err := os.Executable()
require.NoError(t, err)
cmd := exec.Command(exe, "-test.run=^TestAvoid$", "-test.v")
cmd.Env = append(os.Environ(), "NEBULA_CPUPICK_AVOID_TEST=1")
out, err := cmd.CombinedOutput()
require.NoError(t, err, "%s", out)
return
}
allowed, err := util.AllowedCPUs()
require.NoError(t, err)
if len(allowed) < 2 {
t.Skip("requires at least two allowed CPUs")
}
cpu := allowed[0]
release := make(chan struct{})
var wg sync.WaitGroup
// lockThreads occupies 8 OS threads until release, so the runtime has to start more
lockThreads := func() {
var ready sync.WaitGroup
for range 8 {
ready.Add(1)
wg.Go(func() {
runtime.LockOSThread()
ready.Done()
<-release
})
}
ready.Wait()
}
lockThreads()
require.NoError(t, Avoid([]int{cpu}, false))
after, err := util.AllowedCPUs()
require.NoError(t, err)
require.Equal(t, allowed, after, "configuration check must not mutate affinity")
require.Error(t, Avoid(allowed, true), "must reject an empty mask before applying")
require.NoError(t, Avoid([]int{cpu, cpu}, true))
requireAvoided(t, cpu, allowed)
// Existing threads stay locked while new ones inherit the restriction
lockThreads()
requireAvoided(t, cpu, allowed)
close(release)
wg.Wait()
}
// requireAvoided checks no thread may run on cpu, and every thread still may on some CPU of allowed only.
func requireAvoided(t *testing.T, cpu int, allowed []int) {
t.Helper()
entries, err := os.ReadDir("/proc/self/task")
require.NoError(t, err)
for _, entry := range entries {
tid, err := strconv.Atoi(entry.Name())
require.NoError(t, err)
var mask unix.CPUSet
err = unix.SchedGetaffinity(tid, &mask)
if err == unix.ESRCH {
continue
}
require.NoError(t, err)
require.False(t, mask.IsSet(cpu), "thread %d", tid)
require.Positive(t, mask.Count())
for id := range len(mask) * 64 {
if mask.IsSet(id) {
require.Contains(t, allowed, id, "must not widen the inherited mask")
}
}
}
}
// TestAvoidRange checks the bounds of unix.CPUSet, whose words are 32 bits wide on 32-bit platforms.
func TestAvoidRange(t *testing.T) {
capacity := int(unsafe.Sizeof(unix.CPUSet{})) * 8
require.NoError(t, Avoid([]int{capacity - 1}, false))
require.EqualError(t, Avoid([]int{capacity}, false), fmt.Sprintf("CPU %d is out of range", capacity))
require.EqualError(t, Avoid([]int{-1}, false), "CPU -1 is out of range")
}
+8
View File
@@ -0,0 +1,8 @@
//go:build !linux || android || e2e_testing
package cpupick
import "errors"
// Avoid is unsupported outside Linux: there is no thread affinity API to restrict.
func Avoid(_ []int, _ bool) error { return errors.ErrUnsupported }
+17 -6
View File
@@ -16,10 +16,15 @@
// across distinct physical cores when the topology permits, CPU 0's physical
// core demoted to last resort, and the order rotated by a stable per-instance
// key so co-located instances spread instead of stacking.
//
// CPUs the operator asked to keep readers off (tun.avoid_cpus, typically the
// underlay NIC's receive IRQ CPUs) are removed from the allowed set before
// any of that, so no rule can bring one back.
package cpupick
import (
"log/slog"
"slices"
"github.com/slackhq/nebula/util"
)
@@ -57,12 +62,17 @@ func flatTopology(cpus []int) topology {
// Default computes the pin order for `routines` tun readers. key is any
// stable per-instance value; the bound UDP port is ideal — distinct across
// co-located instances, stable across restarts so benchmark runs stay
// comparable. Returns nil when there is nothing useful to say (no affinity
// support on this platform, lookup failure); callers keep their existing
// fallback spread.
func Default(routines int, key uint64, l *slog.Logger) []int {
// comparable. avoid lists CPUs never to return, matched literally. Returns
// nil when there is nothing useful to say (no affinity support on this
// platform, lookup failure, avoid covering every allowed CPU); callers keep
// their existing fallback spread.
func Default(routines int, key uint64, avoid []int, l *slog.Logger) []int {
allowed, err := util.AllowedCPUs()
if err != nil || len(allowed) == 0 {
if err != nil {
return nil
}
allowed = slices.DeleteFunc(allowed, func(c int) bool { return slices.Contains(avoid, c) })
if len(allowed) == 0 {
return nil
}
perf, signal := perfCPUs(allowed)
@@ -77,7 +87,8 @@ func Default(routines int, key uint64, l *slog.Logger) []int {
if l != nil {
l.Info("chose default pin CPUs for tun readers",
"cpus", cpus[:min(routines, len(cpus))],
"perfSignal", signal)
"perfSignal", signal,
"avoided", avoid)
}
return cpus
}
+35
View File
@@ -0,0 +1,35 @@
//go:build linux && !android && !e2e_testing
package cpupick
import (
"slices"
"testing"
"github.com/slackhq/nebula/util"
"github.com/stretchr/testify/assert"
"github.com/stretchr/testify/require"
)
func TestDefaultSkipsAvoided(t *testing.T) {
allowed, err := util.AllowedCPUs()
require.NoError(t, err)
if len(allowed) < 2 {
t.Skip("requires at least two allowed CPUs")
}
avoid := allowed[:len(allowed)/2]
for key := range uint64(16) {
got := Default(len(allowed), key, avoid, nil)
require.NotEmpty(t, got)
for _, cpu := range got {
assert.NotContains(t, avoid, cpu, "key %d", key)
assert.Contains(t, allowed, cpu, "key %d", key)
}
}
assert.Nil(t, Default(2, 0, allowed, nil), "avoiding everything leaves nothing to suggest")
// Without avoid every allowed CPU stays in play
got := Default(len(allowed), 0, nil, nil)
slices.Sort(got)
assert.Equal(t, allowed, got)
}
+9
View File
@@ -295,6 +295,15 @@ tun:
# - 2
# - 4
# Linux only. avoid_cpus lists CPU IDs the tun readers are never pinned to, exactly as listed (SMT siblings are not
# added for you). The default pin selection skips them, and a cpu_affinity entry naming one is rejected like any
# other invalid entry. Other threads are not restricted, and the setting does nothing with pin_threads: false. An
# invalid list, or one covering every allowed CPU, is ignored with a warning. Not reloadable. To keep Nebula off the
# underlay NIC's receive IRQ CPUs, find them in /proc/interrupts and /proc/irq/<IRQ>/effective_affinity_list, add
# their SMT siblings from /sys/devices/system/cpu/cpu<N>/topology/thread_siblings_list if you want those kept clear
# too, and revisit the list when IRQ placement changes.
#avoid_cpus: []
# Route based MTU overrides, you have known vpn ip paths that can support larger MTUs you can increase/decrease them here
routes:
#- mtu: 8800
+9 -1
View File
@@ -60,6 +60,9 @@ type InterfaceConfig struct {
// shorter lists than `routines` cycle. Empty list keeps the default
// pin-to-(i % NumCPU) behavior. Only consulted when PinThreads is true.
CpuAffinity []int
// AvoidCPUs (tun.avoid_cpus) lists CPUs the default pin spread skips when
// CpuAffinity is empty. Only consulted when PinThreads is true.
AvoidCPUs []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
@@ -95,6 +98,8 @@ type Interface struct {
// Empty falls back to the default pin-to-(allowed CPU) behavior.
// Only consulted when pinThreads is true.
cpuAffinity []int
// avoidCPUs are removed from that fallback spread, unless they would empty it.
avoidCPUs []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.
@@ -247,6 +252,7 @@ func NewInterface(ctx context.Context, c *InterfaceConfig) (*Interface, error) {
connectionManager: c.connectionManager,
conntrackCacheTimeout: c.ConntrackCacheTimeout,
cpuAffinity: c.CpuAffinity,
avoidCPUs: c.AvoidCPUs,
pinThreads: c.PinThreads,
udpTxErrLimiter: udpTxErrLimiter,
udpRelayTxErrLimiter: udpRelayTxErrLimiter,
@@ -437,12 +443,14 @@ func (f *Interface) pinThisThread(i int) {
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.
// validated the entries against the allowed CPU set and tun.avoid_cpus.
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.
// parseAvoidCPUs already dropped an avoid list that covers them all.
allowed = slices.DeleteFunc(allowed, func(c int) bool { return slices.Contains(f.avoidCPUs, c) })
cpu = allowed[i%len(allowed)]
} else {
cpu = i % runtime.NumCPU()
+20
View File
@@ -7,6 +7,7 @@ import (
"github.com/slackhq/nebula/cert"
"github.com/slackhq/nebula/config"
"github.com/slackhq/nebula/test"
"github.com/slackhq/nebula/util"
"github.com/stretchr/testify/assert"
"github.com/stretchr/testify/require"
)
@@ -118,3 +119,22 @@ func TestReloadFirewall_NoChange(t *testing.T) {
assert.Same(t, fw, f.firewall, "firewall should not have been replaced")
}
// TestPinThisThreadFallbackSkipsAvoided covers the single reader, which never goes through cpupick: its fallback
// spread over the allowed set must still skip tun.avoid_cpus.
func TestPinThisThreadFallbackSkipsAvoided(t *testing.T) {
allowed, err := util.AllowedCPUs()
require.NoError(t, err)
if len(allowed) < 2 {
t.Skip("requires at least two allowed CPUs")
}
f := &Interface{l: test.NewLogger(), pinThreads: true, avoidCPUs: []int{allowed[0]}}
got := make(chan []int)
go func() {
// The pin leaves this goroutine locked to its thread, so the runtime discards the thread when it returns.
f.pinThisThread(0)
cpus, _ := util.AllowedCPUs()
got <- cpus
}()
assert.Equal(t, []int{allowed[1]}, <-got)
}
+52 -6
View File
@@ -243,11 +243,19 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
}
pinThreads := c.GetBool("tun.pin_threads", true)
cpuAffinity := parseCpuAffinity(c, l, routines)
avoidCPUs := parseAvoidCPUs(c, l)
if len(avoidCPUs) > 0 && !pinThreads {
l.Warn("tun.avoid_cpus has no effect with tun.pin_threads disabled")
}
// Keeping every thread of the process off avoidCPUs, not just the pinned readers, is parked for now:
//if err := cpupick.Avoid(avoidCPUs, !configTest); err != nil && !errors.Is(err, errors.ErrUnsupported) {
// return nil, fmt.Errorf("tun.avoid_cpus: %w", err)
//}
cpuAffinity := parseCpuAffinity(c, l, routines, avoidCPUs)
if pinThreads && routines > 1 && len(cpuAffinity) == 0 && !configTest {
// The operator didn't choose pin CPUs, so pick a default set that
// prefers performance cores and doesn't stack co-located instances
// onto allowed[0].
// prefers performance cores, skips tun.avoid_cpus and doesn't stack
// co-located instances onto allowed[0].
// key is used to seed the spreading of routines->cores.
// use PID if you want to ensure many different Nebulas in VMs or containers land on different cores
@@ -270,7 +278,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
l.Warn("tun.pin_threads_key is invalid, using PID")
}
cpuAffinity = cpupick.Default(routines, key, l)
cpuAffinity = cpupick.Default(routines, key, avoidCPUs, l)
}
ifConfig := &InterfaceConfig{
@@ -295,6 +303,7 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
punchy: punchy,
ConntrackCacheTimeout: conntrackCacheTimeout,
CpuAffinity: cpuAffinity,
AvoidCPUs: avoidCPUs,
PinThreads: pinThreads,
l: l,
}
@@ -363,8 +372,10 @@ func Main(c *config.C, configTest bool, buildVersion string, l *slog.Logger, dev
// 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 {
// determined we fall back to a plain non-negative check. An entry listed in
// avoid (tun.avoid_cpus) contradicts the operator's own config and is treated
// the same as any other invalid entry.
func parseCpuAffinity(c *config.C, l *slog.Logger, routines int, avoid []int) []int {
raw := c.Get("tun.cpu_affinity")
if raw == nil {
return nil
@@ -407,6 +418,11 @@ func parseCpuAffinity(c *config.C, l *slog.Logger, routines int) []int {
"index", i, "cpu", cpu, "allowed", allowed)
return nil
}
if slices.Contains(avoid, cpu) {
l.Warn("tun.cpu_affinity entry is listed in tun.avoid_cpus; ignoring affinity",
"index", i, "cpu", cpu)
return nil
}
cpus = append(cpus, cpu)
}
if len(cpus) != routines {
@@ -416,6 +432,36 @@ func parseCpuAffinity(c *config.C, l *slog.Logger, routines int) []int {
return cpus
}
// parseAvoidCPUs reads tun.avoid_cpus, a list of CPU IDs the tun readers must not be pinned to. IDs outside the
// allowed set are accepted, since one config may serve differently shaped hosts. Like tun.cpu_affinity, an invalid
// list is a warning and is ignored entirely, as is one covering every allowed CPU, which would leave nowhere to pin.
func parseAvoidCPUs(c *config.C, l *slog.Logger) []int {
raw := c.Get("tun.avoid_cpus")
if raw == nil {
return nil
}
values, ok := raw.([]any)
if !ok {
l.Warn("tun.avoid_cpus must be a list of integers; ignoring", "value", raw)
return nil
}
cpus := make([]int, 0, len(values))
for i, value := range values {
cpu, ok := value.(int)
if !ok || cpu < 0 {
l.Warn("tun.avoid_cpus entry not a non-negative integer; ignoring avoid_cpus", "index", i, "value", value)
return nil
}
cpus = append(cpus, cpu)
}
if allowed, err := util.AllowedCPUs(); err == nil && len(allowed) > 0 &&
!slices.ContainsFunc(allowed, func(cpu int) bool { return !slices.Contains(cpus, cpu) }) {
l.Warn("tun.avoid_cpus covers every allowed CPU; ignoring avoid_cpus", "cpus", cpus, "allowed", allowed)
return nil
}
return cpus
}
func moduleVersion() string {
info, ok := debug.ReadBuildInfo()
if !ok {
+36 -5
View File
@@ -32,7 +32,7 @@ func TestParseCpuAffinity(t *testing.T) {
}
// unset -> nil (listenIn falls back to spreading across the allowed set)
assert.Nil(t, parseCpuAffinity(newConfig(nil), l, 1))
assert.Nil(t, parseCpuAffinity(newConfig(nil), l, 1, nil))
// Pick a CPU we're actually allowed to run on so a valid list survives
// validation regardless of the host's affinity mask.
@@ -43,19 +43,50 @@ func TestParseCpuAffinity(t *testing.T) {
}
// valid list -> parsed through unchanged
assert.Equal(t, []int{validCPU, validCPU}, parseCpuAffinity(newConfig([]any{validCPU, validCPU}), l, 2))
assert.Equal(t, []int{validCPU, validCPU}, parseCpuAffinity(newConfig([]any{validCPU, validCPU}), l, 2, nil))
// a negative entry is out of range on every platform -> disables the override
assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU, -1}), l, 2))
assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU, -1}), l, 2, nil))
// a non-integer entry -> disables the override
assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU, "not-a-cpu"}), l, 2))
assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU, "not-a-cpu"}), l, 2, nil))
// 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))
assert.Nil(t, parseCpuAffinity(newConfig([]any{1 << 20}), l, 1, nil))
}
// an entry also listed in tun.avoid_cpus -> disables the override
assert.Nil(t, parseCpuAffinity(newConfig([]any{validCPU}), l, 1, []int{validCPU}))
}
func TestParseAvoidCPUs(t *testing.T) {
l := test.NewLogger()
for _, tc := range []struct {
value any
want []int
}{
{nil, nil}, {[]any{}, []int{}}, {[]any{1, 1 << 20}, []int{1, 1 << 20}},
{"auto", nil}, {"1,2", nil}, {[]any{-1}, nil}, {[]any{1.5}, nil}, {[]any{1, "2"}, nil},
} {
c := config.NewC(l)
if tc.value != nil {
c.Settings["tun"] = map[string]any{"avoid_cpus": tc.value}
}
assert.Equal(t, tc.want, parseAvoidCPUs(c, l), "%v", tc.value)
}
// avoiding every allowed CPU leaves nowhere to pin; only checkable where the allowed set is known
if allowed, _ := util.AllowedCPUs(); len(allowed) > 0 {
all := make([]any, len(allowed))
for i, cpu := range allowed {
all[i] = cpu
}
c := config.NewC(l)
c.Settings["tun"] = map[string]any{"avoid_cpus": all}
assert.Nil(t, parseAvoidCPUs(c, l))
}
}