fix(*): empty cpuset.mems will lead to failure of tmo

Signed-off-by: linzhecheng <[email protected]>

pkg/agent/qrm-plugins/memory/dynamicpolicy/policy_advisor_handler.go

@@ -736,13 +736,9 @@ func (p *DynamicPolicy) handleAdvisorMemoryOffloading(_ *config.Configuration,
 		}
 	}
 
-	cpuSetStats, err := cgroupmgr.GetCPUSetWithAbsolutePath(absCGPath)
+	_, mems, err := cgroupmgr.GetEffectiveCPUSetWithAbsolutePath(absCGPath)
 	if err != nil {
-		return fmt.Errorf("GetCPUSetWithAbsolutePath failed with error: %v", err)
-	}
-	mems, err := machine.Parse(cpuSetStats.Mems)
-	if err != nil {
-		return fmt.Errorf("parse cpuSetStats failed with error: %v", err)
+		return fmt.Errorf("GetEffectiveCPUSetWithAbsolutePath failed with error: %v", err)
 	}
 
 	// start a asynchronous work to execute memory offloading

pkg/agent/sysadvisor/plugin/qosaware/reporter/nodemetric_reporter.go

@@ -540,7 +540,11 @@ func (p *nodeMetricsReporterPlugin) getGroupUsage(pods []*v1.Pod, qosLevel strin
 		resourceMetric.CPU = aggCPU
 	}
 
-	for numaID, resourceUsages := range numaUsages {
+	for numaID := 0; numaID < p.metaServer.NumNUMANodes; numaID++ {
+		resourceUsages, ok := numaUsages[numaID]
+		if !ok {
+			continue
+		}
 		resourceNUMAMetric := nodeapis.ResourceMetric{}
 
 		cpuUsage := resourceUsages[v1.ResourceCPU]

pkg/util/cgroup/common/types.go

@@ -184,8 +184,10 @@ type CPUStats struct {
 
 // CPUSetStats get cgroup cpuset data
 type CPUSetStats struct {
-	CPUs string
-	Mems string
+	CPUs          string
+	EffectiveCPUs string
+	Mems          string
+	EffectiveMems string
 }
 
 // MemoryMetrics get memory cgroup metrics

pkg/util/cgroup/manager/cgroup.go

@@ -21,11 +21,15 @@ import (
 	"fmt"
 	"io/fs"
 	"math"
+	"os"
 	"os/exec"
 	"path/filepath"
 	"strconv"
+	"syscall"
 	"time"
 
+	"golang.org/x/sys/unix"
+
 	"github.com/kubewharf/katalyst-core/pkg/consts"
 	"github.com/kubewharf/katalyst-core/pkg/metrics"
 	"github.com/kubewharf/katalyst-core/pkg/util/asyncworker"
@@ -514,3 +518,50 @@ func MemoryOffloadingWithAbsolutePath(ctx context.Context, absCgroupPath string,
 
 	return err
 }
+
+func IsCgroupPath(path string) bool {
+	var fstat syscall.Statfs_t
+	err := syscall.Statfs(path, &fstat)
+	if err != nil {
+		general.ErrorS(err, "failed to Statfs", "path", path)
+		return false
+	}
+	return fstat.Type == unix.CGROUP2_SUPER_MAGIC || fstat.Type == unix.CGROUP_SUPER_MAGIC
+}
+
+func GetEffectiveCPUSetWithAbsolutePath(absCgroupPath string) (machine.CPUSet, machine.CPUSet, error) {
+	if !IsCgroupPath(absCgroupPath) {
+		return machine.CPUSet{}, machine.CPUSet{}, fmt.Errorf("path %s is not a cgroup", absCgroupPath)
+	}
+
+	cpusetStat, err := GetCPUSetWithAbsolutePath(absCgroupPath)
+	if err != nil {
+		// if controller is disabled, we should walk the parent's dir.
+		if os.IsNotExist(err) {
+			return GetEffectiveCPUSetWithAbsolutePath(filepath.Dir(absCgroupPath))
+		}
+		return machine.CPUSet{}, machine.CPUSet{}, err
+	}
+	// if the cpus or mems is empty, they will inherit the parent's mask.
+	cpus, err := machine.Parse(cpusetStat.EffectiveCPUs)
+	if err != nil {
+		return machine.CPUSet{}, machine.CPUSet{}, err
+	}
+	if cpus.IsEmpty() {
+		cpus, _, err = GetEffectiveCPUSetWithAbsolutePath(filepath.Dir(absCgroupPath))
+		if err != nil {
+			return machine.CPUSet{}, machine.CPUSet{}, err
+		}
+	}
+	mems, err := machine.Parse(cpusetStat.EffectiveMems)
+	if err != nil {
+		return machine.CPUSet{}, machine.CPUSet{}, err
+	}
+	if mems.IsEmpty() {
+		_, mems, err = GetEffectiveCPUSetWithAbsolutePath(filepath.Dir(absCgroupPath))
+		if err != nil {
+			return machine.CPUSet{}, machine.CPUSet{}, err
+		}
+	}
+	return cpus, mems, nil
+}