func (s *FreezerGroup) Set(d *data) error {
switch d.c.Freezer {
case cgroups.Frozen, cgroups.Thawed:
dir, err := d.path("freezer")
if err != nil {
return err
}
if err := writeFile(dir, "freezer.state", string(d.c.Freezer)); err != nil {
return err
}
for {
state, err := readFile(dir, "freezer.state")
if err != nil {
return err
}
if strings.TrimSpace(state) == string(d.c.Freezer) {
break
}
time.Sleep(1 * time.Millisecond)
}
default:
if _, err := d.join("freezer"); err != nil && !cgroups.IsNotFound(err) {
return err
}
}
return nil
}
func (s *CpusetGroup) Apply(d *data) error {
dir, err := d.path("cpuset")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
return s.ApplyDir(dir, d.c, d.pid)
}
func joinMemory(c *configs.Cgroup, pid int) error {
path, err := getSubsystemPath(c, "memory")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
// -1 disables memoryswap
if c.MemorySwap > 0 {
err = writeFile(path, "memory.memsw.limit_in_bytes", strconv.FormatInt(c.MemorySwap, 10))
if err != nil {
return err
}
}
if c.KernelMemory > 0 {
err = writeFile(path, "memory.kmem.limit_in_bytes", strconv.FormatInt(c.KernelMemory, 10))
if err != nil {
return err
}
}
if c.MemorySwappiness >= 0 && c.MemorySwappiness <= 100 {
err = writeFile(path, "memory.swappiness", strconv.FormatInt(c.MemorySwappiness, 10))
if err != nil {
return err
}
}
return nil
}
func joinFreezer(c *configs.Cgroup, pid int) error {
if _, err := join(c, "freezer", pid); err != nil && !cgroups.IsNotFound(err) {
return err
}
return nil
}
func GetStats(c *cgroups.Cgroup) (*cgroups.Stats, error) {
stats := cgroups.NewStats()
d, err := getCgroupData(c, 0)
if err != nil {
return nil, fmt.Errorf("getting CgroupData %s", err)
}
for sysname, sys := range subsystems {
path, err := d.path(sysname)
if err != nil {
// Don't fail if a cgroup hierarchy was not found, just skip this subsystem
if cgroups.IsNotFound(err) {
continue
}
return nil, err
}
if err := sys.GetStats(path, stats); err != nil {
return nil, err
}
}
return stats, nil
}
func (s *MemoryGroup) Apply(d *data) error {
path, err := d.path("memory")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if err := os.MkdirAll(path, 0755); err != nil && !os.IsExist(err) {
return err
}
if err := s.Set(path, d.c); err != nil {
return err
}
// We need to join memory cgroup after set memory limits, because
// kmem.limit_in_bytes can only be set when the cgroup is empty.
_, err = d.join("memory")
if err != nil {
return err
}
defer func() {
if err != nil {
os.RemoveAll(path)
}
}()
return nil
}
func joinMemory(c *configs.Cgroup, pid int) error {
path, err := getSubsystemPath(c, "memory")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
// -1 disables memoryswap
if c.Memory != 0 && c.MemorySwap >= 0 {
memorySwap := c.MemorySwap
if memorySwap == 0 {
// By default, MemorySwap is set to twice the size of RAM.
memorySwap = c.Memory * 2
}
err = writeFile(path, "memory.memsw.limit_in_bytes", strconv.FormatInt(memorySwap, 10))
if err != nil {
return err
}
}
if c.KernelMemory > 0 {
err = writeFile(path, "memory.kmem.limit_in_bytes", strconv.FormatInt(c.KernelMemory, 10))
if err != nil {
return err
}
}
return nil
}
func Apply(c *cgroups.Cgroup, pid int) (map[string]string, error) {
d, err := getCgroupData(c, pid)
if err != nil {
return nil, err
}
paths := make(map[string]string)
defer func() {
if err != nil {
cgroups.RemovePaths(paths)
}
}()
for name, sys := range subsystems {
if err := sys.Set(d); err != nil {
return nil, err
}
// FIXME: Apply should, ideally, be reentrant or be broken up into a separate
// create and join phase so that the cgroup hierarchy for a container can be
// created then join consists of writing the process pids to cgroup.procs
p, err := d.path(name)
if err != nil {
if cgroups.IsNotFound(err) {
continue
}
return nil, err
}
paths[name] = p
}
return paths, nil
}
func joinCpu(c *configs.Cgroup, pid int) error {
path, err := getSubsystemPath(c, "cpu")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if c.CpuQuota != 0 {
if err = writeFile(path, "cpu.cfs_quota_us", strconv.FormatInt(c.CpuQuota, 10)); err != nil {
return err
}
}
if c.CpuPeriod != 0 {
if err = writeFile(path, "cpu.cfs_period_us", strconv.FormatInt(c.CpuPeriod, 10)); err != nil {
return err
}
}
if c.CpuRtPeriod != 0 {
if err = writeFile(path, "cpu.rt_period_us", strconv.FormatInt(c.CpuRtPeriod, 10)); err != nil {
return err
}
}
if c.CpuRtRuntime != 0 {
if err = writeFile(path, "cpu.rt_runtime_us", strconv.FormatInt(c.CpuRtRuntime, 10)); err != nil {
return err
}
}
return nil
}
func (s *PerfEventGroup) Apply(d *data) error {
// we just want to join this group even though we don't set anything
if _, err := d.join("perf_event"); err != nil && !cgroups.IsNotFound(err) {
return err
}
return nil
}
func (s *BlkioGroup) Set(d *data) error {
// we just want to join this group even though we don't set anything
if _, err := d.join("blkio"); err != nil && !cgroups.IsNotFound(err) {
return err
}
return nil
}
func joinNetPrio(c *configs.Cgroup, pid int) error {
path, err := join(c, "net_prio", pid)
if err != nil && !cgroups.IsNotFound(err) {
return err
}
netPrio := subsystems["net_prio"]
return netPrio.Set(path, c)
}
func joinNetCls(c *configs.Cgroup, pid int) error {
path, err := join(c, "net_cls", pid)
if err != nil && !cgroups.IsNotFound(err) {
return err
}
netcls := subsystems["net_cls"]
return netcls.Set(path, c)
}
func joinFreezer(c *configs.Cgroup, pid int) error {
path, err := join(c, "freezer", pid)
if err != nil && !cgroups.IsNotFound(err) {
return err
}
freezer := subsystems["freezer"]
return freezer.Set(path, c)
}
func joinHugetlb(c *configs.Cgroup, pid int) error {
path, err := join(c, "hugetlb", pid)
if err != nil && !cgroups.IsNotFound(err) {
return err
}
hugetlb := subsystems["hugetlb"]
return hugetlb.Set(path, c)
}
// systemd does not atm set up the cpuset controller, so we must manually
// join it. Additionally that is a very finicky controller where each
// level must have a full setup as the default for a new directory is "no cpus"
func joinCpuset(c *configs.Cgroup, pid int) error {
path, err := getSubsystemPath(c, "cpuset")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
s := &fs.CpusetGroup{}
return s.ApplyDir(path, c, pid)
}
func (s *NetClsGroup) Apply(d *data) error {
dir, err := d.join("net_cls")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if err := s.Set(dir, d.c); err != nil {
return err
}
return nil
}
func (s *FreezerGroup) Apply(d *data) error {
dir, err := d.join("freezer")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if err := s.Set(dir, d.c); err != nil {
return err
}
return nil
}
func (s *HugetlbGroup) Apply(d *data) error {
dir, err := d.join("hugetlb")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if err := s.Set(dir, d.c); err != nil {
return err
}
return nil
}
func (s *BlkioGroup) Set(d *data) error {
dir, err := d.join("blkio")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if d.c.BlkioWeight != 0 {
if err := writeFile(dir, "blkio.weight", strconv.FormatInt(d.c.BlkioWeight, 10)); err != nil {
return err
}
}
return nil
}
func (s *CpuGroup) Apply(d *data) error {
// We always want to join the cpu group, to allow fair cpu scheduling
// on a container basis
dir, err := d.join("cpu")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if err := s.Set(dir, d.c); err != nil {
return err
}
return nil
}
func joinMemory(c *configs.Cgroup, pid int) error {
memorySwap := c.MemorySwap
if memorySwap == 0 {
// By default, MemorySwap is set to twice the size of RAM.
memorySwap = c.Memory * 2
}
path, err := getSubsystemPath(c, "memory")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
return writeFile(path, "memory.memsw.limit_in_bytes", strconv.FormatInt(memorySwap, 10))
}
func (s *BlkioGroup) Apply(d *data) error {
dir, err := d.join("blkio")
if err != nil {
if cgroups.IsNotFound(err) {
return nil
} else {
return err
}
}
if err := s.Set(dir, d.c); err != nil {
return err
}
return nil
}
func (s *MemoryGroup) Apply(d *data) error {
dir, err := d.join("memory")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
defer func() {
if err != nil {
os.RemoveAll(dir)
}
}()
if err := s.Set(dir, d.c); err != nil {
return err
}
return nil
}
func (m *Manager) Apply(pid int) error {
if m.Cgroups == nil {
return nil
}
var c = m.Cgroups
d, err := getCgroupData(m.Cgroups, pid)
if err != nil {
return err
}
paths := make(map[string]string)
defer func() {
if err != nil {
cgroups.RemovePaths(paths)
}
}()
for name, sys := range subsystems {
if err := sys.Apply(d); err != nil {
return err
}
// TODO: Apply should, ideally, be reentrant or be broken up into a separate
// create and join phase so that the cgroup hierarchy for a container can be
// created then join consists of writing the process pids to cgroup.procs
p, err := d.path(name)
if err != nil {
if cgroups.IsNotFound(err) {
continue
}
return err
}
paths[name] = p
}
m.Paths = paths
if paths["cpu"] != "" {
if err := CheckCpushares(paths["cpu"], c.CpuShares); err != nil {
return err
}
}
return nil
}
func setKernelMemory(c *configs.Cgroup) error {
path, err := getSubsystemPath(c, "memory")
if err != nil && !cgroups.IsNotFound(err) {
return err
}
if err := os.MkdirAll(path, 0755); err != nil && !os.IsExist(err) {
return err
}
if c.KernelMemory > 0 {
err = writeFile(path, "memory.kmem.limit_in_bytes", strconv.FormatInt(c.KernelMemory, 10))
if err != nil {
return err
}
}
return nil
}
func (raw *data) Paths() (map[string]string, error) {
paths := make(map[string]string)
for sysname := range subsystems {
path, err := raw.path(sysname)
if err != nil {
// Don't fail if a cgroup hierarchy was not found, just skip this subsystem
if cgroups.IsNotFound(err) {
continue
}
return nil, err
}
paths[sysname] = path
}
return paths, nil
}
/*
* This would be nicer to get from the systemd API when accounting
* is enabled, but sadly there is no way to do that yet.
* The lack of this functionality in the API & the approach taken
* is guided by
* http://www.freedesktop.org/wiki/Software/systemd/ControlGroupInterface/#readingaccountinginformation.
*/
func GetStats(c *cgroups.Cgroup) (*cgroups.Stats, error) {
stats := cgroups.NewStats()
for sysname, sys := range subsystems {
subsystemPath, err := getSubsystemPath(c, sysname)
if err != nil {
// Don't fail if a cgroup hierarchy was not found, just skip this subsystem
if cgroups.IsNotFound(err) {
continue
}
return nil, err
}
if err := sys.GetStats(subsystemPath, stats); err != nil {
return nil, err
}
}
return stats, nil
}
func Apply(c *cgroups.Cgroup, pid int) (map[string]string, error) {
var (
unitName = getUnitName(c)
slice = "system.slice"
properties []systemd.Property
res = &systemdCgroup{}
)
res.cgroup = c
if c.Slice != "" {
slice = c.Slice
}
properties = append(properties,
systemd.PropSlice(slice),
systemd.PropDescription("docker container "+c.Name),
newProp("PIDs", []uint32{uint32(pid)}),
)
// Always enable accounting, this gets us the same behaviour as the fs implementation,
// plus the kernel has some problems with joining the memory cgroup at a later time.
properties = append(properties,
newProp("MemoryAccounting", true),
newProp("CPUAccounting", true),
newProp("BlockIOAccounting", true))
if hasTransientDefaultDependencies {
properties = append(properties,
newProp("DefaultDependencies", false))
}
if c.Memory != 0 {
properties = append(properties,
newProp("MemoryLimit", uint64(c.Memory)))
}
// TODO: MemoryReservation and MemorySwap not available in systemd
if c.CpuShares != 0 {
properties = append(properties,
newProp("CPUShares", uint64(c.CpuShares)))
}
if c.BlkioWeight != 0 {
properties = append(properties,
newProp("BlockIOWeight", uint64(c.BlkioWeight)))
}
if _, err := theConn.StartTransientUnit(unitName, "replace", properties...); err != nil {
return nil, err
}
if err := joinDevices(c, pid); err != nil {
return nil, err
}
// -1 disables memorySwap
if c.MemorySwap >= 0 && c.Memory != 0 {
if err := joinMemory(c, pid); err != nil {
return nil, err
}
}
// we need to manually join the freezer and cpuset cgroup in systemd
// because it does not currently support it via the dbus api.
if err := joinFreezer(c, pid); err != nil {
return nil, err
}
if err := joinCpuset(c, pid); err != nil {
return nil, err
}
paths := make(map[string]string)
for _, sysname := range []string{
"devices",
"memory",
"cpu",
"cpuset",
"cpuacct",
"blkio",
"perf_event",
"freezer",
} {
subsystemPath, err := getSubsystemPath(res.cgroup, sysname)
if err != nil {
// Don't fail if a cgroup hierarchy was not found, just skip this subsystem
if cgroups.IsNotFound(err) {
continue
}
return nil, err
}
paths[sysname] = subsystemPath
}
return paths, nil
}
func (m *Manager) Apply(pid int) error {
var (
c = m.Cgroups
unitName = getUnitName(c)
slice = "system.slice"
properties []systemd.Property
)
if c.Slice != "" {
slice = c.Slice
}
properties = append(properties,
systemd.PropSlice(slice),
systemd.PropDescription("docker container "+c.Name),
newProp("PIDs", []uint32{uint32(pid)}),
)
// Always enable accounting, this gets us the same behaviour as the fs implementation,
// plus the kernel has some problems with joining the memory cgroup at a later time.
properties = append(properties,
newProp("MemoryAccounting", true),
newProp("CPUAccounting", true),
newProp("BlockIOAccounting", true))
if hasTransientDefaultDependencies {
properties = append(properties,
newProp("DefaultDependencies", false))
}
if c.Memory != 0 {
properties = append(properties,
newProp("MemoryLimit", uint64(c.Memory)))
}
// TODO: MemoryReservation and MemorySwap not available in systemd
if c.CpuShares != 0 {
properties = append(properties,
newProp("CPUShares", uint64(c.CpuShares)))
}
if c.BlkioWeight != 0 {
properties = append(properties,
newProp("BlockIOWeight", uint64(c.BlkioWeight)))
}
// We need to set kernel memory before processes join cgroup because
// kmem.limit_in_bytes can only be set when the cgroup is empty.
// And swap memory limit needs to be set after memory limit, only
// memory limit is handled by systemd, so it's kind of ugly here.
if c.KernelMemory > 0 {
if err := setKernelMemory(c); err != nil {
return err
}
}
if _, err := theConn.StartTransientUnit(unitName, "replace", properties...); err != nil {
return err
}
if err := joinDevices(c, pid); err != nil {
return err
}
// TODO: CpuQuota and CpuPeriod not available in systemd
// we need to manually join the cpu.cfs_quota_us and cpu.cfs_period_us
if err := joinCpu(c, pid); err != nil {
return err
}
if err := joinMemory(c, pid); err != nil {
return err
}
// we need to manually join the freezer, net_cls, net_prio and cpuset cgroup in systemd
// because it does not currently support it via the dbus api.
if err := joinFreezer(c, pid); err != nil {
return err
}
if err := joinNetPrio(c, pid); err != nil {
return err
}
if err := joinNetCls(c, pid); err != nil {
return err
}
if err := joinCpuset(c, pid); err != nil {
return err
}
if err := joinHugetlb(c, pid); err != nil {
return err
}
// FIXME: Systemd does have `BlockIODeviceWeight` property, but we got problem
// using that (at least on systemd 208, see https://github.com/docker/libcontainer/pull/354),
// so use fs work around for now.
if err := joinBlkio(c, pid); err != nil {
return err
}
paths := make(map[string]string)
for sysname := range subsystems {
subsystemPath, err := getSubsystemPath(m.Cgroups, sysname)
if err != nil {
// Don't fail if a cgroup hierarchy was not found, just skip this subsystem
if cgroups.IsNotFound(err) {
continue
}
return err
}
paths[sysname] = subsystemPath
}
m.Paths = paths
if paths["cpu"] != "" {
if err := fs.CheckCpushares(paths["cpu"], c.CpuShares); err != nil {
return err
}
}
return nil
}
