func (m *Manager) Apply(pid int) error {
	var (
		c          = m.Cgroups
		unitName   = getUnitName(c)
		slice      = "system.slice"
		properties []systemdDbus.Property
	)

	if c.Slice != "" {
		slice = c.Slice
	}

	properties = append(properties,
		systemdDbus.PropSlice(slice),
		systemdDbus.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, nil); 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/opencontainers/runc/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
}
示例#2
0
func (m *Manager) Apply(pid int) error {
	var (
		c          = m.Cgroups
		unitName   = getUnitName(c)
		slice      = "system.slice"
		properties []systemdDbus.Property
	)

	if c.Paths != nil {
		paths := make(map[string]string)
		for name, path := range c.Paths {
			_, err := getSubsystemPath(m.Cgroups, name)
			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[name] = path
		}
		m.Paths = paths
		return cgroups.EnterPid(m.Paths, pid)
	}

	if c.Parent != "" {
		slice = c.Parent
	}

	properties = append(properties,
		systemdDbus.PropSlice(slice),
		systemdDbus.PropDescription("docker container "+c.Name),
		newProp("PIDs", []uint32{uint32(pid)}),
		// This is only supported on systemd versions 218 and above.
		newProp("Delegate", true),
	)

	// 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.Resources.Memory != 0 {
		properties = append(properties,
			newProp("MemoryLimit", uint64(c.Resources.Memory)))
	}

	if c.Resources.CpuShares != 0 {
		properties = append(properties,
			newProp("CPUShares", uint64(c.Resources.CpuShares)))
	}

	if c.Resources.BlkioWeight != 0 {
		properties = append(properties,
			newProp("BlockIOWeight", uint64(c.Resources.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.Resources.KernelMemory > 0 {
		if err := setKernelMemory(c); err != nil {
			return err
		}
	}

	if _, err := theConn.StartTransientUnit(unitName, "replace", properties, nil); err != nil {
		return err
	}

	if err := joinCgroups(c, pid); err != nil {
		return err
	}

	paths := make(map[string]string)
	for _, s := range subsystems {
		subsystemPath, err := getSubsystemPath(m.Cgroups, s.Name())
		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[s.Name()] = subsystemPath
	}
	m.Paths = paths
	return nil
}
示例#3
0
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)))
	}

	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
	}

	// -1 disables memorySwap
	if c.MemorySwap >= 0 && c.Memory != 0 {
		if err := joinMemory(c, pid); err != nil {
			return 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 err
	}

	if err := joinCpuset(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

	return nil
}
示例#4
0
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
}
示例#5
0
func (m *Manager) Apply(pid int) error {
	var (
		c          = m.Cgroups
		unitName   = getUnitName(c)
		slice      = "system.slice"
		properties []systemdDbus.Property
	)

	if c.Paths != nil {
		paths := make(map[string]string)
		for name, path := range c.Paths {
			_, err := getSubsystemPath(m.Cgroups, name)
			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[name] = path
		}
		m.Paths = paths
		return cgroups.EnterPid(m.Paths, pid)
	}

	if c.Parent != "" {
		slice = c.Parent
	}

	properties = append(properties, systemdDbus.PropDescription("libcontainer container "+c.Name))

	// if we create a slice, the parent is defined via a Wants=
	if strings.HasSuffix(unitName, ".slice") {
		// This was broken until systemd v229, but has been back-ported on RHEL environments >= 219
		if !hasStartTransientSliceUnit {
			return fmt.Errorf("systemd version does not support ability to start a slice as transient unit")
		}
		properties = append(properties, systemdDbus.PropWants(slice))
	} else {
		// otherwise, we use Slice=
		properties = append(properties, systemdDbus.PropSlice(slice))
	}

	// only add pid if its valid, -1 is used w/ general slice creation.
	if pid != -1 {
		properties = append(properties, newProp("PIDs", []uint32{uint32(pid)}))
	}

	if hasDelegate {
		// This is only supported on systemd versions 218 and above.
		properties = append(properties, newProp("Delegate", true))
	}

	// 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.Resources.Memory != 0 {
		properties = append(properties,
			newProp("MemoryLimit", uint64(c.Resources.Memory)))
	}

	if c.Resources.CpuShares != 0 {
		properties = append(properties,
			newProp("CPUShares", uint64(c.Resources.CpuShares)))
	}

	if c.Resources.BlkioWeight != 0 {
		properties = append(properties,
			newProp("BlockIOWeight", uint64(c.Resources.BlkioWeight)))
	}

	// We have to set kernel memory here, as we can't change it once
	// processes have been attached to the cgroup.
	if c.Resources.KernelMemory != 0 {
		if err := setKernelMemory(c); err != nil {
			return err
		}
	}

	if _, err := theConn.StartTransientUnit(unitName, "replace", properties, nil); err != nil {
		return err
	}

	if err := joinCgroups(c, pid); err != nil {
		return err
	}

	paths := make(map[string]string)
	for _, s := range subsystems {
		subsystemPath, err := getSubsystemPath(m.Cgroups, s.Name())
		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[s.Name()] = subsystemPath
	}
	m.Paths = paths
	return nil
}