func getCgroupMounts(m *configs.Mount) ([]*configs.Mount, error) {
mounts, err := cgroups.GetCgroupMounts()
if err != nil {
return nil, err
}
cgroupPaths, err := cgroups.ParseCgroupFile("/proc/self/cgroup")
if err != nil {
return nil, err
}
var binds []*configs.Mount
for _, mm := range mounts {
dir, err := mm.GetThisCgroupDir(cgroupPaths)
if err != nil {
return nil, err
}
relDir, err := filepath.Rel(mm.Root, dir)
if err != nil {
return nil, err
}
binds = append(binds, &configs.Mount{
Device: "bind",
Source: filepath.Join(mm.Mountpoint, relDir),
Destination: filepath.Join(m.Destination, strings.Join(mm.Subsystems, ",")),
Flags: syscall.MS_BIND | syscall.MS_REC | m.Flags,
PropagationFlags: m.PropagationFlags,
})
}
return binds, nil
}
// Gets the (CPU) container the specified pid is in.
func getContainer(pid int) (string, error) {
cgs, err := cgroups.ParseCgroupFile(fmt.Sprintf("/proc/%d/cgroup", pid))
if err != nil {
return "", err
}
cg, ok := cgs["cpu"]
if ok {
return cg, nil
}
return "", cgroups.NewNotFoundError("cpu")
}
// getContainer returns the cgroup associated with the specified pid.
// It enforces a unified hierarchy for memory and cpu cgroups.
// On systemd environments, it uses the name=systemd cgroup for the specified pid.
func getContainer(pid int) (string, error) {
cgs, err := cgroups.ParseCgroupFile(fmt.Sprintf("/proc/%d/cgroup", pid))
if err != nil {
return "", err
}
cpu, found := cgs["cpu"]
if !found {
return "", cgroups.NewNotFoundError("cpu")
}
memory, found := cgs["memory"]
if !found {
return "", cgroups.NewNotFoundError("memory")
}
// since we use this container for accounting, we need to ensure its a unified hierarchy.
if cpu != memory {
return "", fmt.Errorf("cpu and memory cgroup hierarchy not unified. cpu: %s, memory: %s", cpu, memory)
}
// on systemd, every pid is in a unified cgroup hierarchy (name=systemd as seen in systemd-cgls)
// cpu and memory accounting is off by default, users may choose to enable it per unit or globally.
// users could enable CPU and memory accounting globally via /etc/systemd/system.conf (DefaultCPUAccounting=true DefaultMemoryAccounting=true).
// users could also enable CPU and memory accounting per unit via CPUAccounting=true and MemoryAccounting=true
// we only warn if accounting is not enabled for CPU or memory so as to not break local development flows where kubelet is launched in a terminal.
// for example, the cgroup for the user session will be something like /user.slice/user-X.slice/session-X.scope, but the cpu and memory
// cgroup will be the closest ancestor where accounting is performed (most likely /) on systems that launch docker containers.
// as a result, on those systems, you will not get cpu or memory accounting statistics for kubelet.
// in addition, you would not get memory or cpu accounting for the runtime unless accounting was enabled on its unit (or globally).
if systemd, found := cgs["name=systemd"]; found {
if systemd != cpu {
glog.Warningf("CPUAccounting not enabled for pid: %d", pid)
}
if systemd != memory {
glog.Warningf("MemoryAccounting not enabled for pid: %d", pid)
}
return systemd, nil
}
return cpu, nil
}
func (c *linuxContainer) criuApplyCgroups(pid int, req *criurpc.CriuReq) error {
if err := c.cgroupManager.Apply(pid); err != nil {
return err
}
path := fmt.Sprintf("/proc/%d/cgroup", pid)
cgroupsPaths, err := cgroups.ParseCgroupFile(path)
if err != nil {
return err
}
for c, p := range cgroupsPaths {
cgroupRoot := &criurpc.CgroupRoot{
Ctrl: proto.String(c),
Path: proto.String(p),
}
req.Opts.CgRoot = append(req.Opts.CgRoot, cgroupRoot)
}
return nil
}