Merge pull request #1529 from AkihiroSuda/allow-disabling-hugepages

allow disabling hugepages

pkg/config/config.go

@@ -236,6 +236,10 @@ type PluginConfig struct {
 	// container requests with huge page limits if the cgroup controller for hugepages is not present.
 	// This helps with supporting Kubernetes <=1.18 out of the box. (default is `true`)
 	TolerateMissingHugetlbController bool `toml:"tolerate_missing_hugetlb_controller" json:"tolerateMissingHugetlbController"`
+	// DisableHugetlbController indicates to silently disable the hugetlb controller, even when it is
+	// present in /sys/fs/cgroup/cgroup.controllers.
+	// This helps with running rootless mode + cgroup v2 + systemd but without hugetlb delegation.
+	DisableHugetlbController bool `toml:"disable_hugetlb_controller" json:"disableHugetlbController"`
 	// IgnoreImageDefinedVolumes ignores volumes defined by the image. Useful for better resource
 	// isolation, security and early detection of issues in the mount configuration when using
 	// ReadOnlyRootFilesystem since containers won't silently mount a temporary volume.

pkg/config/config_unix.go

@@ -68,6 +68,7 @@ func DefaultConfig() PluginConfig {
 		MaxConcurrentDownloads:           3,
 		DisableProcMount:                 false,
 		TolerateMissingHugetlbController: true,
+		DisableHugetlbController:         true,
 		IgnoreImageDefinedVolumes:        false,
 	}
 }

pkg/containerd/opts/spec_unix.go

@@ -408,7 +408,7 @@ func WithSelinuxLabels(process, mount string) oci.SpecOpts {
 }
 
 // WithResources sets the provided resource restrictions
-func WithResources(resources *runtime.LinuxContainerResources, tolerateMissingHugetlbController bool) oci.SpecOpts {
+func WithResources(resources *runtime.LinuxContainerResources, tolerateMissingHugetlbController, disableHugetlbController bool) oci.SpecOpts {
 	return func(ctx context.Context, client oci.Client, c *containers.Container, s *runtimespec.Spec) (err error) {
 		if resources == nil {
 			return nil
@@ -451,19 +451,21 @@ func WithResources(resources *runtime.LinuxContainerResources, tolerateMissingHu
 		if limit != 0 {
 			s.Linux.Resources.Memory.Limit = &limit
 		}
-		if isHugetlbControllerPresent() {
-			for _, limit := range hugepages {
-				s.Linux.Resources.HugepageLimits = append(s.Linux.Resources.HugepageLimits, runtimespec.LinuxHugepageLimit{
-					Pagesize: limit.PageSize,
-					Limit:    limit.Limit,
-				})
+		if !disableHugetlbController {
+			if isHugetlbControllerPresent() {
+				for _, limit := range hugepages {
+					s.Linux.Resources.HugepageLimits = append(s.Linux.Resources.HugepageLimits, runtimespec.LinuxHugepageLimit{
+						Pagesize: limit.PageSize,
+						Limit:    limit.Limit,
+					})
+				}
+			} else {
+				if !tolerateMissingHugetlbController {
+					return errors.Errorf("huge pages limits are specified but hugetlb cgroup controller is missing. " +
+						"Please set tolerate_missing_hugetlb_controller to `true` to ignore this error")
+				}
+				logrus.Warn("hugetlb cgroup controller is absent. skipping huge pages limits")
 			}
-		} else {
-			if !tolerateMissingHugetlbController {
-				return errors.Errorf("huge pages limits are specified but hugetlb cgroup controller is missing. " +
-					"Please set tolerate_missing_hugetlb_controller to `true` to ignore this error")
-			}
-			logrus.Warn("hugetlb cgroup controller is absent. skipping huge pages limits")
 		}
 		return nil
 	}

pkg/server/container_create_unix.go

@@ -225,7 +225,7 @@ func (c *criService) containerSpec(id string, sandboxID string, sandboxPid uint3
 	if c.config.DisableCgroup {
 		specOpts = append(specOpts, customopts.WithDisabledCgroups)
 	} else {
-		specOpts = append(specOpts, customopts.WithResources(config.GetLinux().GetResources(), c.config.TolerateMissingHugetlbController))
+		specOpts = append(specOpts, customopts.WithResources(config.GetLinux().GetResources(), c.config.TolerateMissingHugetlbController, c.config.DisableHugetlbController))
 		if sandboxConfig.GetLinux().GetCgroupParent() != "" {
 			cgroupsPath := getCgroupsPath(sandboxConfig.GetLinux().GetCgroupParent(), id)
 			specOpts = append(specOpts, oci.WithCgroup(cgroupsPath))

pkg/server/container_update_resources_unix.go

@@ -73,7 +73,7 @@ func (c *criService) updateContainerResources(ctx context.Context,
 		return errors.Wrap(err, "failed to get container spec")
 	}
 	newSpec, err := updateOCILinuxResource(ctx, oldSpec, resources,
-		c.config.TolerateMissingHugetlbController)
+		c.config.TolerateMissingHugetlbController, c.config.DisableHugetlbController)
 	if err != nil {
 		return errors.Wrap(err, "failed to update resource in spec")
 	}
@@ -134,7 +134,7 @@ func updateContainerSpec(ctx context.Context, cntr containerd.Container, spec *r
 
 // updateOCILinuxResource updates container resource limit.
 func updateOCILinuxResource(ctx context.Context, spec *runtimespec.Spec, new *runtime.LinuxContainerResources,
-	tolerateMissingHugetlbController bool) (*runtimespec.Spec, error) {
+	tolerateMissingHugetlbController, disableHugetlbController bool) (*runtimespec.Spec, error) {
 	// Copy to make sure old spec is not changed.
 	var cloned runtimespec.Spec
 	if err := util.DeepCopy(&cloned, spec); err != nil {
@@ -143,7 +143,7 @@ func updateOCILinuxResource(ctx context.Context, spec *runtimespec.Spec, new *ru
 	if cloned.Linux == nil {
 		cloned.Linux = &runtimespec.Linux{}
 	}
-	if err := opts.WithResources(new, tolerateMissingHugetlbController)(ctx, nil, nil, &cloned); err != nil {
+	if err := opts.WithResources(new, tolerateMissingHugetlbController, disableHugetlbController)(ctx, nil, nil, &cloned); err != nil {
 		return nil, errors.Wrap(err, "unable to set linux container resources")
 	}
 	return &cloned, nil

pkg/server/container_update_resources_unix_test.go

@@ -153,7 +153,7 @@ func TestUpdateOCILinuxResource(t *testing.T) {
 		},
 	} {
 		t.Logf("TestCase %q", desc)
-		got, err := updateOCILinuxResource(context.Background(), test.spec, test.resources, false)
+		got, err := updateOCILinuxResource(context.Background(), test.spec, test.resources, false, false)
 		if test.expectErr {
 			assert.Error(t, err)
 		} else {