[1.4] Add memory policy support

Implement support for Linux memory policy in OCI spec PR:
https://github.com/opencontainers/runtime-spec/pull/1282

Signed-off-by: Antti Kervinen <antti.kervinen@intel.com>
(cherry picked from commit eda7bdf80c)
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>

features.go

@@ -59,6 +59,10 @@ var featuresCommand = cli.Command{
 					Enabled:  &t,
 					Schemata: &t,
 				},
+				MemoryPolicy: &features.MemoryPolicy{
+					Modes: specconv.KnownMemoryPolicyModes(),
+					Flags: specconv.KnownMemoryPolicyFlags(),
+				},
 				MountExtensions: &features.MountExtensions{
 					IDMap: &features.IDMap{
 						Enabled: &t,

internal/linux/linux.go

@@ -2,6 +2,7 @@ package linux
 
 import (
 	"os"
+	"unsafe"
 
 	"golang.org/x/sys/unix"
 )
@@ -72,3 +73,15 @@ func Sendmsg(fd int, p, oob []byte, to unix.Sockaddr, flags int) error {
 	})
 	return os.NewSyscallError("sendmsg", err)
 }
+
+// SetMempolicy wraps set_mempolicy.
+func SetMempolicy(mode uint, mask *unix.CPUSet) error {
+	err := retryOnEINTR(func() error {
+		_, _, errno := unix.Syscall(unix.SYS_SET_MEMPOLICY, uintptr(mode), uintptr(unsafe.Pointer(mask)), unsafe.Sizeof(*mask)*8)
+		if errno != 0 {
+			return errno
+		}
+		return nil
+	})
+	return os.NewSyscallError("set_mempolicy", err)
+}

libcontainer/configs/config.go

@@ -214,6 +214,9 @@ type Config struct {
 	// to limit the resources (e.g., L3 cache, memory bandwidth) the container has available
 	IntelRdt *IntelRdt `json:"intel_rdt,omitempty"`
 
+	// MemoryPolicy specifies NUMA memory policy for the container.
+	MemoryPolicy *LinuxMemoryPolicy `json:"memory_policy,omitempty"`
+
 	// RootlessEUID is set when the runc was launched with non-zero EUID.
 	// Note that RootlessEUID is set to false when launched with EUID=0 in userns.
 	// When RootlessEUID is set, runc creates a new userns for the container.
@@ -305,7 +308,8 @@ type CPUAffinity struct {
 	Initial, Final *unix.CPUSet
 }
 
-func toCPUSet(str string) (*unix.CPUSet, error) {
+// ToCPUSet parses a string in list format into a unix.CPUSet, e.g. "0-3,5,7-9".
+func ToCPUSet(str string) (*unix.CPUSet, error) {
 	if str == "" {
 		return nil, nil
 	}
@@ -356,7 +360,7 @@ func toCPUSet(str string) (*unix.CPUSet, error) {
 		}
 	}
 	if s.Count() == 0 {
-		return nil, fmt.Errorf("no CPUs found in %q", str)
+		return nil, fmt.Errorf("no members found in set %q", str)
 	}
 
 	return s, nil
@@ -367,11 +371,11 @@ func ConvertCPUAffinity(sa *specs.CPUAffinity) (*CPUAffinity, error) {
 	if sa == nil {
 		return nil, nil
 	}
-	initial, err := toCPUSet(sa.Initial)
+	initial, err := ToCPUSet(sa.Initial)
 	if err != nil {
 		return nil, fmt.Errorf("bad CPUAffinity.Initial: %w", err)
 	}
-	final, err := toCPUSet(sa.Final)
+	final, err := ToCPUSet(sa.Final)
 	if err != nil {
 		return nil, fmt.Errorf("bad CPUAffinity.Final: %w", err)
 	}

libcontainer/configs/memorypolicy.go

@@ -0,0 +1,31 @@
+package configs
+
+import "golang.org/x/sys/unix"
+
+// Memory policy modes and flags as defined in /usr/include/linux/mempolicy.h
+
+//nolint:revive,staticcheck,nolintlint // ignore ALL_CAPS errors in consts from numaif.h, will match unix.* in the future
+const (
+	MPOL_DEFAULT             = 0
+	MPOL_PREFERRED           = 1
+	MPOL_BIND                = 2
+	MPOL_INTERLEAVE          = 3
+	MPOL_LOCAL               = 4
+	MPOL_PREFERRED_MANY      = 5
+	MPOL_WEIGHTED_INTERLEAVE = 6
+
+	MPOL_F_STATIC_NODES   = 1 << 15
+	MPOL_F_RELATIVE_NODES = 1 << 14
+	MPOL_F_NUMA_BALANCING = 1 << 13
+)
+
+// LinuxMemoryPolicy contains memory policy configuration.
+type LinuxMemoryPolicy struct {
+	// Mode specifies memory policy mode without mode flags. See
+	// set_mempolicy() documentation for details.
+	Mode uint `json:"mode,omitempty"`
+	// Flags contains mode flags.
+	Flags uint `json:"flags,omitempty"`
+	// Nodes contains NUMA nodes to which the mode applies.
+	Nodes *unix.CPUSet `json:"nodes,omitempty"`
+}

libcontainer/configs/tocpuset_test.go

@@ -58,8 +58,8 @@ func TestToCPUSet(t *testing.T) {
 
 	for _, tc := range testCases {
 		t.Run(tc.in, func(t *testing.T) {
-			out, err := toCPUSet(tc.in)
-			t.Logf("toCPUSet(%q) = %v (error: %v)", tc.in, out, err)
+			out, err := ToCPUSet(tc.in)
+			t.Logf("ToCPUSet(%q) = %v (error: %v)", tc.in, out, err)
 			// Check the error.
 			if tc.isErr {
 				if err == nil {

libcontainer/configs/validate/validator.go

@@ -33,6 +33,7 @@ func Validate(config *configs.Config) error {
 		mountsStrict,
 		scheduler,
 		ioPriority,
+		memoryPolicy,
 	}
 	for _, c := range checks {
 		if err := c(config); err != nil {
@@ -482,3 +483,26 @@ func ioPriority(config *configs.Config) error {
 
 	return nil
 }
+
+func memoryPolicy(config *configs.Config) error {
+	mpol := config.MemoryPolicy
+	if mpol == nil {
+		return nil
+	}
+	switch mpol.Mode {
+	case configs.MPOL_DEFAULT, configs.MPOL_LOCAL:
+		if mpol.Nodes != nil && mpol.Nodes.Count() != 0 {
+			return fmt.Errorf("memory policy mode requires 0 nodes but got %d", mpol.Nodes.Count())
+		}
+	case configs.MPOL_BIND, configs.MPOL_INTERLEAVE,
+		configs.MPOL_PREFERRED_MANY, configs.MPOL_WEIGHTED_INTERLEAVE:
+		if mpol.Nodes == nil || mpol.Nodes.Count() == 0 {
+			return fmt.Errorf("memory policy mode requires at least one node but got 0")
+		}
+	case configs.MPOL_PREFERRED:
+		// Zero or more nodes are allowed by the kernel.
+	default:
+		return fmt.Errorf("invalid memory policy mode: %d", mpol.Mode)
+	}
+	return nil
+}

libcontainer/init_linux.go

@@ -659,6 +659,14 @@ func setupIOPriority(config *initConfig) error {
 	return nil
 }
 
+func setupMemoryPolicy(config *configs.Config) error {
+	mpol := config.MemoryPolicy
+	if mpol == nil {
+		return nil
+	}
+	return linux.SetMempolicy(mpol.Mode|mpol.Flags, config.MemoryPolicy.Nodes)
+}
+
 func setupPersonality(config *configs.Config) error {
 	return system.SetLinuxPersonality(config.Personality.Domain)
 }

libcontainer/setns_init_linux.go

@@ -88,6 +88,10 @@ func (l *linuxSetnsInit) Init() error {
 		}
 	}
 
+	if err := setupMemoryPolicy(l.config.Config); err != nil {
+		return err
+	}
+
 	// Tell our parent that we're ready to exec. This must be done before the
 	// Seccomp rules have been applied, because we need to be able to read and
 	// write to a socket.

libcontainer/specconv/spec_linux.go

@@ -8,6 +8,7 @@ import (
 	"maps"
 	"os"
 	"path/filepath"
+	"slices"
 	"sort"
 	"strings"
 	"sync"
@@ -41,6 +42,8 @@ var (
 		flag  int
 	}
 	complexFlags map[string]func(*configs.Mount)
+	mpolModeMap  map[string]uint
+	mpolModeFMap map[string]uint
 )
 
 func initMaps() {
@@ -148,6 +151,22 @@ func initMaps() {
 				m.IDMapping.Recursive = true
 			},
 		}
+
+		mpolModeMap = map[string]uint{
+			string(specs.MpolDefault):            configs.MPOL_DEFAULT,
+			string(specs.MpolPreferred):          configs.MPOL_PREFERRED,
+			string(specs.MpolBind):               configs.MPOL_BIND,
+			string(specs.MpolInterleave):         configs.MPOL_INTERLEAVE,
+			string(specs.MpolLocal):              configs.MPOL_LOCAL,
+			string(specs.MpolPreferredMany):      configs.MPOL_PREFERRED_MANY,
+			string(specs.MpolWeightedInterleave): configs.MPOL_WEIGHTED_INTERLEAVE,
+		}
+
+		mpolModeFMap = map[string]uint{
+			string(specs.MpolFStaticNodes):   configs.MPOL_F_STATIC_NODES,
+			string(specs.MpolFRelativeNodes): configs.MPOL_F_RELATIVE_NODES,
+			string(specs.MpolFNumaBalancing): configs.MPOL_F_NUMA_BALANCING,
+		}
 	})
 }
 
@@ -184,6 +203,20 @@ func KnownMountOptions() []string {
 	return res
 }
 
+// KnownMemoryPolicyModes returns the list of the known memory policy modes.
+// Used by `runc features`.
+func KnownMemoryPolicyModes() []string {
+	initMaps()
+	return slices.Sorted(maps.Keys(mpolModeMap))
+}
+
+// KnownMemoryPolicyFlags returns the list of the known memory policy mode flags.
+// Used by `runc features`.
+func KnownMemoryPolicyFlags() []string {
+	initMaps()
+	return slices.Sorted(maps.Keys(mpolModeFMap))
+}
+
 // AllowedDevices is the set of devices which are automatically included for
 // all containers.
 //
@@ -468,6 +501,28 @@ func CreateLibcontainerConfig(opts *CreateOpts) (*configs.Config, error) {
 				MemBwSchema:   spec.Linux.IntelRdt.MemBwSchema,
 			}
 		}
+		if spec.Linux.MemoryPolicy != nil {
+			var ok bool
+			var err error
+			specMp := spec.Linux.MemoryPolicy
+			confMp := &configs.LinuxMemoryPolicy{}
+			confMp.Mode, ok = mpolModeMap[string(specMp.Mode)]
+			if !ok {
+				return nil, fmt.Errorf("invalid memory policy mode %q", specMp.Mode)
+			}
+			confMp.Nodes, err = configs.ToCPUSet(specMp.Nodes)
+			if err != nil {
+				return nil, fmt.Errorf("invalid memory policy nodes %q: %w", specMp.Nodes, err)
+			}
+			for _, specFlag := range specMp.Flags {
+				confFlag, ok := mpolModeFMap[string(specFlag)]
+				if !ok {
+					return nil, fmt.Errorf("invalid memory policy flag %q", specFlag)
+				}
+				confMp.Flags |= confFlag
+			}
+			config.MemoryPolicy = confMp
+		}
 		if spec.Linux.Personality != nil {
 			if len(spec.Linux.Personality.Flags) > 0 {
 				logrus.Warnf("ignoring unsupported personality flags: %+v because personality flag has not supported at this time", spec.Linux.Personality.Flags)

libcontainer/standard_init_linux.go

@@ -171,6 +171,10 @@ func (l *linuxStandardInit) Init() error {
 		}
 	}
 
+	if err := setupMemoryPolicy(l.config.Config); err != nil {
+		return err
+	}
+
 	// Tell our parent that we're ready to exec. This must be done before the
 	// Seccomp rules have been applied, because we need to be able to read and
 	// write to a socket.

tests/integration/memorypolicy.bats

@@ -0,0 +1,133 @@
+#!/usr/bin/env bats
+
+load helpers
+
+function setup() {
+	setup_busybox
+}
+
+function teardown() {
+	teardown_bundle
+}
+
+@test "runc run memory policy interleave without flags" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "MPOL_INTERLEAVE",
+		"nodes": "0"
+	}'
+	runc run test_busybox
+	[ "$status" -eq 0 ]
+	[[ "${lines[0]}" == "interleave:0" ]]
+}
+
+@test "runc run memory policy bind static" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "MPOL_BIND",
+		"nodes": "0",
+		"flags": ["MPOL_F_STATIC_NODES"]
+	}'
+	runc run test_busybox
+	[ "$status" -eq 0 ]
+	[[ "${lines[0]}" == "bind"*"static"*"0" ]]
+}
+
+@test "runc run and exec memory policy prefer relative" {
+	update_config '
+	.linux.memoryPolicy = {
+		"mode": "MPOL_PREFERRED",
+		"nodes": "0",
+		"flags": ["MPOL_F_RELATIVE_NODES"]
+	}'
+	runc run -d --console-socket "$CONSOLE_SOCKET" test_busybox
+	[ "$status" -eq 0 ]
+
+	runc exec test_busybox /bin/sh -c "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"
+	[ "$status" -eq 0 ]
+	[[ "${lines[0]}" == "prefer"*"relative"*"0" ]]
+}
+
+@test "runc run empty memory policy" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+	}'
+	runc run test_busybox
+	[ "$status" -eq 1 ]
+	[[ "${lines[0]}" == *"invalid memory policy"* ]]
+}
+
+@test "runc run memory policy with non-existing mode" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "INTERLEAVE",
+		"nodes": "0"
+	}'
+	runc run test_busybox
+	[ "$status" -eq 1 ]
+	[[ "${lines[0]}" == *"invalid memory policy"* ]]
+}
+
+@test "runc run memory policy with invalid flag" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "MPOL_PREFERRED",
+		"nodes": "0",
+		"flags": ["MPOL_F_RELATIVE_NODES", "badflag"]
+	}'
+	runc run test_busybox
+	[ "$status" -eq 1 ]
+	[[ "${lines[0]}" == *"invalid memory policy flag"* ]]
+}
+
+@test "runc run memory policy default with missing nodes" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "MPOL_DEFAULT"
+	}'
+	runc run test_busybox
+	[ "$status" -eq 0 ]
+	[[ "${lines[0]}" == *"default"* ]]
+}
+
+@test "runc run memory policy with missing mode" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"nodes": "0-7"
+	}'
+	runc run test_busybox
+	[ "$status" -eq 1 ]
+	[[ "${lines[0]}" == *"invalid memory policy mode"* ]]
+}
+
+@test "runc run memory policy calls syscall with invalid arguments" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "MPOL_DEFAULT",
+		"nodes": "0-7",
+	}'
+	runc run test_busybox
+	[ "$status" -eq 1 ]
+	[[ "${lines[*]}" == *"mode requires 0 nodes but got 8"* ]]
+}
+
+@test "runc run memory policy bind way too large a node number" {
+	update_config '
+	.process.args = ["/bin/sh", "-c", "head -n 1 /proc/self/numa_maps | cut -d \" \" -f 2"]
+	| .linux.memoryPolicy = {
+		"mode": "MPOL_BIND",
+		"nodes": "0-9876543210",
+		"flags": []
+	}'
+	runc run test_busybox
+	[ "$status" -eq 1 ]
+	[[ "${lines[0]}" == *"invalid memory policy node"* ]]
+}