device-injector: implement memory policy adjustment.

Add support for adjusting container linux memory policy based
on annotations.

Signed-off-by: Krisztian Litkey <krisztian.litkey@intel.com>

Author: klihub

plugins/device-injector/README.md

@@ -158,6 +158,43 @@ instead of a single container. Also many of the settings are not namespaced at
 all. Trying to adjust such a setting will result in an error and a failure to
 create the container.
 
+### Linux Memory Policy Annotations
+
+Memory policies are annotated in a similar manner to devices, but using the
+`memory-policy.noderesource.dev` annotation key prefix. As with devices, the
+`memory-policy.nri.io` annotation key prefix is also supported.
+
+The annotation value syntax for memory policy adjustment is
+
+```
+  mode: <policy mode, for instance MPOL_INTERLEAVE>
+  nodes: <list of NUMA nodes as a string, for instance "2,3">
+  flags: <list of policy flags as strings, for instance [ MPOL_F_STATIC_NODES ] >
+```
+
+The supported modes are:
+
+- MPOL_DEFAULT
+- MPOL_PREFERRED
+- MPOL_BIND
+- MPOL_INTERLEAVE
+- MPOL_LOCAL
+- MPOL_PREFERRED_MANY
+- MPOL_WEIGHTED_INTERLEAVE
+
+You can omit the MPOL_ prefix and can use lowercase at will.
+
+The supported flags are:
+
+- MPOL_F_STATIC_NODES
+- MPOL_F_RELATIVE_NODES
+- MPOL_F_NUMA_BALANCING
+
+As with modes, you can omit the MPOL_F_ prefix and use lowercase at will.
+
+See man set_mempolicy(2) for a description of what the effects are of these
+settings.
+
 ## Deployment
 
 The NRI repository contains minimal kustomize overlays for this plugin at

plugins/device-injector/device-injector.go

@@ -57,6 +57,10 @@ const (
 	oldSchedulerKey = "scheduling-policy.nri.io"
 	// Prefix of the key used for sysctl adjustment.
 	sysctlKey = "sysctl.noderesource.dev"
+	// Prefix of the key used for memory policy adjustment.
+	memoryPolicyKey = "memory-policy.noderesource.dev"
+	// Deprecated: Prefix of the key used for memory policy adjustment.
+	oldMemoryPolicyKey = "memory-policy.nri.io"
 )
 
 var (
@@ -106,6 +110,13 @@ type scheduler struct {
 	Period   uint64   `json:"period"`
 }
 
+// memoryPolicy adjustment
+type memoryPolicy struct {
+	Mode  string   `json:"mode"`
+	Nodes string   `json:"nodes"`
+	Flags []string `json:"flags"`
+}
+
 // our injector plugin
 type plugin struct {
 	stub stub.Stub
@@ -146,6 +157,10 @@ func (p *plugin) CreateContainer(_ context.Context, pod *api.PodSandbox, ctr *ap
 		return nil, nil, err
 	}
 
+	if err := adjustMemoryPolicy(pod, ctr, adjust); err != nil {
+		return nil, nil, err
+	}
+
 	if verbose {
 		dump(containerName(pod, ctr), "ContainerAdjustment", adjust)
 	}
@@ -443,6 +458,53 @@ func parseSysctl(ctr string, annotations map[string]string) (map[string]string,
 	return sysctl, nil
 }
 
+func adjustMemoryPolicy(pod *api.PodSandbox, ctr *api.Container, a *api.ContainerAdjustment) error {
+	pol, err := parseMemoryPolicy(ctr.Name, pod.Annotations)
+	if err != nil {
+		log.Errorf("%s: invalid memory policy annotation: %v",
+			containerName(pod, ctr), err)
+		return err
+	}
+
+	if pol == nil {
+		log.Debugf("%s: no memory policy annotated...", containerName(pod, ctr))
+		return nil
+	}
+
+	if verbose {
+		dump(containerName(pod, ctr), "annotated memory policy", pol)
+	}
+
+	amp, err := pol.ToNRI()
+	if err != nil {
+		log.Errorf("%s: invalid memory policy annotation: %v",
+			containerName(pod, ctr), err)
+		return fmt.Errorf("invalid memory policy: %w", err)
+	}
+
+	a.SetLinuxMemoryPolicy(amp.Mode, amp.Nodes, amp.Flags...)
+	log.Infof("%s: adjusted memory policy to %s...", containerName(pod, ctr), amp)
+
+	return nil
+}
+
+func parseMemoryPolicy(ctr string, annotations map[string]string) (*memoryPolicy, error) {
+	var (
+		policy = &memoryPolicy{}
+	)
+
+	annotation := getAnnotation(annotations, memoryPolicyKey, oldMemoryPolicyKey, ctr)
+	if annotation == nil {
+		return nil, nil
+	}
+
+	if err := yaml.Unmarshal(annotation, policy); err != nil {
+		return nil, fmt.Errorf("invalid memory policy annotation %q: %w", string(annotation), err)
+	}
+
+	return policy, nil
+}
+
 func getAnnotation(annotations map[string]string, mainKey, oldKey, ctr string) []byte {
 	for _, key := range []string{
 		mainKey + "/container." + ctr,
@@ -570,6 +632,46 @@ func (sc *scheduler) String() string {
 	return s
 }
 
+// Convert memoryPolicy to the NRI API representation.
+func (mp *memoryPolicy) ToNRI() (*api.LinuxMemoryPolicy, error) {
+	apiPol := &api.LinuxMemoryPolicy{
+		Nodes: mp.Nodes,
+	}
+
+	mode, ok := api.MpolMode_value[mp.Mode]
+	if !ok {
+		return nil, fmt.Errorf("invalid memory policy mode %q", mp.Mode)
+	}
+	apiPol.Mode = api.MpolMode(mode)
+
+	for _, f := range mp.Flags {
+		flag, ok := api.MpolFlag_value[f]
+		if !ok {
+			return nil, fmt.Errorf("invalid memory policy flag %q", f)
+		}
+		apiPol.Flags = append(apiPol.Flags, api.MpolFlag(flag))
+	}
+
+	return apiPol, nil
+}
+
+func (mp *memoryPolicy) String() string {
+	if mp == nil {
+		return "<no memory policy>"
+	}
+
+	s := fmt.Sprintf("<memory policy mode=%s", mp.Mode)
+	if mp.Nodes != "" {
+		s += fmt.Sprintf(", nodes=%s", mp.Nodes)
+	}
+	if len(mp.Flags) > 0 {
+		s += fmt.Sprintf(", flags=%v", mp.Flags)
+	}
+	s += ">"
+
+	return s
+}
+
 // Construct a container name for log messages.
 func containerName(pod *api.PodSandbox, container *api.Container) string {
 	if pod != nil {