Merge pull request #2883 from flouthoc/master

Add support for rdma cgroup introduced in Linux Kernel 4.11

libcontainer/cgroups/fs/fs.go

@@ -31,6 +31,7 @@ var (
 		&NetPrioGroup{},
 		&PerfEventGroup{},
 		&FreezerGroup{},
+		&RdmaGroup{},
 		&NameGroup{GroupName: "name=systemd", Join: true},
 	}
 	HugePageSizes, _ = cgroups.GetHugePageSize()

libcontainer/cgroups/fs/rdma.go

@@ -0,0 +1,25 @@
+package fs
+
+import (
+	"github.com/opencontainers/runc/libcontainer/cgroups"
+	"github.com/opencontainers/runc/libcontainer/cgroups/fscommon"
+	"github.com/opencontainers/runc/libcontainer/configs"
+)
+
+type RdmaGroup struct{}
+
+func (s *RdmaGroup) Name() string {
+	return "rdma"
+}
+
+func (s *RdmaGroup) Apply(path string, d *cgroupData) error {
+	return join(path, d.pid)
+}
+
+func (s *RdmaGroup) Set(path string, r *configs.Resources) error {
+	return fscommon.RdmaSet(path, r)
+}
+
+func (s *RdmaGroup) GetStats(path string, stats *cgroups.Stats) error {
+	return fscommon.RdmaGetStats(path, stats)
+}

libcontainer/cgroups/fs2/fs2.go

@@ -125,6 +125,10 @@ func (m *manager) GetStats() (*cgroups.Stats, error) {
 	if err := statHugeTlb(m.dirPath, st); err != nil && !os.IsNotExist(err) {
 		errs = append(errs, err)
 	}
+	// rdma (since kernel 4.11)
+	if err := fscommon.RdmaGetStats(m.dirPath, st); err != nil && !os.IsNotExist(err) {
+		errs = append(errs, err)
+	}
 	if len(errs) > 0 && !m.rootless {
 		return st, fmt.Errorf("error while statting cgroup v2: %+v", errs)
 	}
@@ -183,6 +187,10 @@ func (m *manager) Set(r *configs.Resources) error {
 	if err := setHugeTlb(m.dirPath, r); err != nil {
 		return err
 	}
+	// rdma (since kernel 4.11)
+	if err := fscommon.RdmaSet(m.dirPath, r); err != nil {
+		return err
+	}
 	// freezer (since kernel 5.2, pseudo-controller)
 	if err := setFreezer(m.dirPath, r.Freezer); err != nil {
 		return err

libcontainer/cgroups/fscommon/rdma.go

@@ -0,0 +1,121 @@
+package fscommon
+
+import (
+	"bufio"
+	"errors"
+	"math"
+	"os"
+	"strconv"
+	"strings"
+
+	"github.com/opencontainers/runc/libcontainer/cgroups"
+	"github.com/opencontainers/runc/libcontainer/configs"
+	"golang.org/x/sys/unix"
+)
+
+// parseRdmaKV parses raw string to RdmaEntry.
+func parseRdmaKV(raw string, entry *cgroups.RdmaEntry) error {
+	var value uint32
+
+	parts := strings.SplitN(raw, "=", 3)
+
+	if len(parts) != 2 {
+		return errors.New("Unable to parse RDMA entry")
+	}
+
+	k, v := parts[0], parts[1]
+
+	if v == "max" {
+		value = math.MaxUint32
+	} else {
+		val64, err := strconv.ParseUint(v, 10, 32)
+		if err != nil {
+			return err
+		}
+		value = uint32(val64)
+	}
+	if k == "hca_handle" {
+		entry.HcaHandles = value
+	} else if k == "hca_object" {
+		entry.HcaObjects = value
+	}
+
+	return nil
+}
+
+// readRdmaEntries reads and converts array of rawstrings to RdmaEntries from file.
+// example entry: mlx4_0 hca_handle=2 hca_object=2000
+func readRdmaEntries(dir, file string) ([]cgroups.RdmaEntry, error) {
+	rdmaEntries := make([]cgroups.RdmaEntry, 0)
+	fd, err := cgroups.OpenFile(dir, file, unix.O_RDONLY)
+	if err != nil {
+		return nil, err
+	}
+	defer fd.Close() //nolint:errorlint
+	scanner := bufio.NewScanner(fd)
+	for scanner.Scan() {
+		parts := strings.SplitN(scanner.Text(), " ", 4)
+		if len(parts) == 3 {
+			entry := new(cgroups.RdmaEntry)
+			entry.Device = parts[0]
+			err = parseRdmaKV(parts[1], entry)
+			if err != nil {
+				continue
+			}
+			err = parseRdmaKV(parts[2], entry)
+			if err != nil {
+				continue
+			}
+
+			rdmaEntries = append(rdmaEntries, *entry)
+		}
+	}
+	return rdmaEntries, scanner.Err()
+}
+
+// RdmaGetStats returns rdma stats such as totalLimit and current entries.
+func RdmaGetStats(path string, stats *cgroups.Stats) error {
+	currentEntries, err := readRdmaEntries(path, "rdma.current")
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			err = nil
+		}
+		return err
+	}
+	maxEntries, err := readRdmaEntries(path, "rdma.max")
+	if err != nil {
+		return err
+	}
+	// If device got removed between reading two files, ignore returning stats.
+	if len(currentEntries) != len(maxEntries) {
+		return nil
+	}
+
+	stats.RdmaStats = cgroups.RdmaStats{
+		RdmaLimit:   maxEntries,
+		RdmaCurrent: currentEntries,
+	}
+
+	return nil
+}
+
+func createCmdString(device string, limits configs.LinuxRdma) string {
+	cmdString := device
+	if limits.HcaHandles != nil {
+		cmdString += " hca_handle=" + strconv.FormatUint(uint64(*limits.HcaHandles), 10)
+	}
+	if limits.HcaObjects != nil {
+		cmdString += " hca_object=" + strconv.FormatUint(uint64(*limits.HcaObjects), 10)
+	}
+	return cmdString
+}
+
+// RdmaSet sets RDMA resources.
+func RdmaSet(path string, r *configs.Resources) error {
+	for device, limits := range r.Rdma {
+		if err := cgroups.WriteFile(path, "rdma.max", createCmdString(device, limits)); err != nil {
+			return err
+		}
+	}
+	return nil
+}

libcontainer/cgroups/fscommon/rdma_test.go

@@ -0,0 +1,57 @@
+package fscommon
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/opencontainers/runc/libcontainer/configs"
+)
+
+/* Roadmap for future */
+// (Low-priority)  TODO: Check if it is possible to virtually mimic an actual RDMA device.
+// TODO: Think of more edge-cases to add.
+
+// TestRdmaSet performs an E2E test of RdmaSet(), parseRdmaKV() using dummy device and a dummy cgroup file-system.
+// Note: Following test does not guarantees that your host supports RDMA since this mocks underlying infrastructure.
+func TestRdmaSet(t *testing.T) {
+	testCgroupPath := filepath.Join(t.TempDir(), "rdma")
+
+	// Ensure the full mock cgroup path exists.
+	err := os.Mkdir(testCgroupPath, 0o755)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	rdmaDevice := "mlx5_1"
+	maxHandles := uint32(100)
+	maxObjects := uint32(300)
+
+	rdmaStubResource := &configs.Resources{
+		Rdma: map[string]configs.LinuxRdma{
+			rdmaDevice: {
+				HcaHandles: &maxHandles,
+				HcaObjects: &maxObjects,
+			},
+		},
+	}
+
+	if err := RdmaSet(testCgroupPath, rdmaStubResource); err != nil {
+		t.Fatal(err)
+	}
+
+	// The default rdma.max must be written.
+	rdmaEntries, err := readRdmaEntries(testCgroupPath, "rdma.max")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(rdmaEntries) != 1 {
+		t.Fatal("rdma_test: Got the wrong values while parsing entries from rdma.max")
+	}
+	if rdmaEntries[0].HcaHandles != maxHandles {
+		t.Fatalf("rdma_test: Got the wrong value for hca_handles")
+	}
+	if rdmaEntries[0].HcaObjects != maxObjects {
+		t.Fatalf("rdma_test: Got the wrong value for hca_Objects")
+	}
+}

libcontainer/cgroups/stats.go

@@ -146,6 +146,17 @@ type HugetlbStats struct {
 	Failcnt uint64 `json:"failcnt"`
 }
 
+type RdmaEntry struct {
+	Device     string `json:"device,omitempty"`
+	HcaHandles uint32 `json:"hca_handles,omitempty"`
+	HcaObjects uint32 `json:"hca_objects,omitempty"`
+}
+
+type RdmaStats struct {
+	RdmaLimit   []RdmaEntry `json:"rdma_limit,omitempty"`
+	RdmaCurrent []RdmaEntry `json:"rdma_current,omitempty"`
+}
+
 type Stats struct {
 	CpuStats    CpuStats    `json:"cpu_stats,omitempty"`
 	CPUSetStats CPUSetStats `json:"cpuset_stats,omitempty"`
@@ -154,6 +165,7 @@ type Stats struct {
 	BlkioStats  BlkioStats  `json:"blkio_stats,omitempty"`
 	// the map is in the format "size of hugepage: stats of the hugepage"
 	HugetlbStats map[string]HugetlbStats `json:"hugetlb_stats,omitempty"`
+	RdmaStats    RdmaStats               `json:"rdma_stats,omitempty"`
 }
 
 func NewStats() *Stats {

libcontainer/cgroups/systemd/v1.go

@@ -59,6 +59,7 @@ var legacySubsystems = []subsystem{
 	&fs.NetPrioGroup{},
 	&fs.NetClsGroup{},
 	&fs.NameGroup{GroupName: "name=systemd"},
+	&fs.RdmaGroup{},
 }
 
 func genV1ResourcesProperties(r *configs.Resources, cm *dbusConnManager) ([]systemdDbus.Property, error) {

libcontainer/configs/cgroup_linux.go

@@ -113,6 +113,9 @@ type Resources struct {
 	// Set class identifier for container's network packets
 	NetClsClassid uint32 `json:"net_cls_classid_u"`
 
+	// Rdma resource restriction configuration
+	Rdma map[string]LinuxRdma `json:"rdma"`
+
 	// Used on cgroups v2:
 
 	// CpuWeight sets a proportional bandwidth limit.

libcontainer/configs/rdma.go

@@ -0,0 +1,9 @@
+package configs
+
+// LinuxRdma for Linux cgroup 'rdma' resource management (Linux 4.11)
+type LinuxRdma struct {
+	// Maximum number of HCA handles that can be opened. Default is "no limit".
+	HcaHandles *uint32 `json:"hca_handles,omitempty"`
+	// Maximum number of HCA objects that can be created. Default is "no limit".
+	HcaObjects *uint32 `json:"hca_objects,omitempty"`
+}

libcontainer/specconv/spec_linux.go

@@ -608,6 +608,15 @@ func CreateCgroupConfig(opts *CreateOpts, defaultDevs []*devices.Device) (*confi
 					Limit:    l.Limit,
 				})
 			}
+			if len(r.Rdma) > 0 {
+				c.Resources.Rdma = make(map[string]configs.LinuxRdma, len(r.Rdma))
+				for k, v := range r.Rdma {
+					c.Resources.Rdma[k] = configs.LinuxRdma{
+						HcaHandles: v.HcaHandles,
+						HcaObjects: v.HcaObjects,
+					}
+				}
+			}
 			if r.Network != nil {
 				if r.Network.ClassID != nil {
 					c.Resources.NetClsClassid = *r.Network.ClassID