support cgroup2

* only shim v2 runc v2 ("io.containerd.runc.v2") is supported
* only PID metrics is implemented. Others should be implemented in separate PRs.
* lots of code duplication in v1 metrics and v2 metrics. Dedupe should be separate PR.

Signed-off-by: Akihiro Suda <akihiro.suda.cz@hco.ntt.co.jp>

vendor/github.com/cilium/ebpf/internal/cpu.go

@@ -0,0 +1,64 @@
+package internal
+
+import (
+	"fmt"
+	"os"
+	"sync"
+
+	"github.com/pkg/errors"
+)
+
+var sysCPU struct {
+	once sync.Once
+	err  error
+	num  int
+}
+
+// PossibleCPUs returns the max number of CPUs a system may possibly have
+// Logical CPU numbers must be of the form 0-n
+func PossibleCPUs() (int, error) {
+	sysCPU.once.Do(func() {
+		sysCPU.num, sysCPU.err = parseCPUs("/sys/devices/system/cpu/possible")
+	})
+
+	return sysCPU.num, sysCPU.err
+}
+
+var onlineCPU struct {
+	once sync.Once
+	err  error
+	num  int
+}
+
+// OnlineCPUs returns the number of currently online CPUs
+// Logical CPU numbers must be of the form 0-n
+func OnlineCPUs() (int, error) {
+	onlineCPU.once.Do(func() {
+		onlineCPU.num, onlineCPU.err = parseCPUs("/sys/devices/system/cpu/online")
+	})
+
+	return onlineCPU.num, onlineCPU.err
+}
+
+// parseCPUs parses the number of cpus from sysfs,
+// in the format of "/sys/devices/system/cpu/{possible,online,..}.
+// Logical CPU numbers must be of the form 0-n
+func parseCPUs(path string) (int, error) {
+	file, err := os.Open(path)
+	if err != nil {
+		return 0, err
+	}
+	defer file.Close()
+
+	var low, high int
+	n, _ := fmt.Fscanf(file, "%d-%d", &low, &high)
+	if n < 1 || low != 0 {
+		return 0, errors.Wrapf(err, "%s has unknown format", path)
+	}
+	if n == 1 {
+		high = low
+	}
+
+	// cpus is 0 indexed
+	return high + 1, nil
+}

vendor/github.com/cilium/ebpf/internal/endian.go

@@ -0,0 +1,24 @@
+package internal
+
+import (
+	"encoding/binary"
+	"unsafe"
+)
+
+// NativeEndian is set to either binary.BigEndian or binary.LittleEndian,
+// depending on the host's endianness.
+var NativeEndian binary.ByteOrder
+
+func init() {
+	if isBigEndian() {
+		NativeEndian = binary.BigEndian
+	} else {
+		NativeEndian = binary.LittleEndian
+	}
+}
+
+func isBigEndian() (ret bool) {
+	i := int(0x1)
+	bs := (*[int(unsafe.Sizeof(i))]byte)(unsafe.Pointer(&i))
+	return bs[0] == 0
+}

vendor/github.com/cilium/ebpf/internal/feature.go

@@ -0,0 +1,85 @@
+package internal
+
+import (
+	"fmt"
+	"sync"
+
+	"github.com/pkg/errors"
+)
+
+// UnsupportedFeatureError is returned by FeatureTest() functions.
+type UnsupportedFeatureError struct {
+	// The minimum Linux mainline version required for this feature.
+	// Used for the error string, and for sanity checking during testing.
+	MinimumVersion Version
+
+	// The name of the feature that isn't supported.
+	Name string
+}
+
+func (ufe *UnsupportedFeatureError) Error() string {
+	return fmt.Sprintf("%s not supported (requires >= %s)", ufe.Name, ufe.MinimumVersion)
+}
+
+// FeatureTest wraps a function so that it is run at most once.
+//
+// name should identify the tested feature, while version must be in the
+// form Major.Minor[.Patch].
+//
+// Returns a descriptive UnsupportedFeatureError if the feature is not available.
+func FeatureTest(name, version string, fn func() bool) func() error {
+	v, err := NewVersion(version)
+	if err != nil {
+		return func() error { return err }
+	}
+
+	var (
+		once   sync.Once
+		result error
+	)
+
+	return func() error {
+		once.Do(func() {
+			if !fn() {
+				result = &UnsupportedFeatureError{
+					MinimumVersion: v,
+					Name:           name,
+				}
+			}
+		})
+		return result
+	}
+}
+
+// A Version in the form Major.Minor.Patch.
+type Version [3]uint16
+
+// NewVersion creates a version from a string like "Major.Minor.Patch".
+//
+// Patch is optional.
+func NewVersion(ver string) (Version, error) {
+	var major, minor, patch uint16
+	n, _ := fmt.Sscanf(ver, "%d.%d.%d", &major, &minor, &patch)
+	if n < 2 {
+		return Version{}, errors.Errorf("invalid version: %s", ver)
+	}
+	return Version{major, minor, patch}, nil
+}
+
+func (v Version) String() string {
+	if v[2] == 0 {
+		return fmt.Sprintf("v%d.%d", v[0], v[1])
+	}
+	return fmt.Sprintf("v%d.%d.%d", v[0], v[1], v[2])
+}
+
+// Less returns true if the version is less than another version.
+func (v Version) Less(other Version) bool {
+	for i, a := range v {
+		if a == other[i] {
+			continue
+		}
+		return a < other[i]
+	}
+	return false
+}

vendor/github.com/cilium/ebpf/internal/unix/types_linux.go

@@ -0,0 +1,127 @@
+// +build linux
+
+package unix
+
+import (
+	"syscall"
+
+	linux "golang.org/x/sys/unix"
+)
+
+const (
+	ENOENT                   = linux.ENOENT
+	EAGAIN                   = linux.EAGAIN
+	ENOSPC                   = linux.ENOSPC
+	EINVAL                   = linux.EINVAL
+	EPOLLIN                  = linux.EPOLLIN
+	BPF_OBJ_NAME_LEN         = linux.BPF_OBJ_NAME_LEN
+	BPF_TAG_SIZE             = linux.BPF_TAG_SIZE
+	SYS_BPF                  = linux.SYS_BPF
+	F_DUPFD_CLOEXEC          = linux.F_DUPFD_CLOEXEC
+	EPOLL_CTL_ADD            = linux.EPOLL_CTL_ADD
+	EPOLL_CLOEXEC            = linux.EPOLL_CLOEXEC
+	O_CLOEXEC                = linux.O_CLOEXEC
+	O_NONBLOCK               = linux.O_NONBLOCK
+	PROT_READ                = linux.PROT_READ
+	PROT_WRITE               = linux.PROT_WRITE
+	MAP_SHARED               = linux.MAP_SHARED
+	PERF_TYPE_SOFTWARE       = linux.PERF_TYPE_SOFTWARE
+	PERF_COUNT_SW_BPF_OUTPUT = linux.PERF_COUNT_SW_BPF_OUTPUT
+	PerfBitWatermark         = linux.PerfBitWatermark
+	PERF_SAMPLE_RAW          = linux.PERF_SAMPLE_RAW
+	PERF_FLAG_FD_CLOEXEC     = linux.PERF_FLAG_FD_CLOEXEC
+	RLIM_INFINITY            = linux.RLIM_INFINITY
+)
+
+// Statfs_t is a wrapper
+type Statfs_t = linux.Statfs_t
+
+// Rlimit is a wrapper
+type Rlimit = linux.Rlimit
+
+// Setrlimit is a wrapper
+func Setrlimit(resource int, rlim *Rlimit) (err error) {
+	return linux.Setrlimit(resource, rlim)
+}
+
+// Syscall is a wrapper
+func Syscall(trap, a1, a2, a3 uintptr) (r1, r2 uintptr, err syscall.Errno) {
+	return linux.Syscall(trap, a1, a2, a3)
+}
+
+// FcntlInt is a wrapper
+func FcntlInt(fd uintptr, cmd, arg int) (int, error) {
+	return linux.FcntlInt(fd, cmd, arg)
+}
+
+// Statfs is a wrapper
+func Statfs(path string, buf *Statfs_t) (err error) {
+	return linux.Statfs(path, buf)
+}
+
+// Close is a wrapper
+func Close(fd int) (err error) {
+	return linux.Close(fd)
+}
+
+// EpollEvent is a wrapper
+type EpollEvent = linux.EpollEvent
+
+// EpollWait is a wrapper
+func EpollWait(epfd int, events []EpollEvent, msec int) (n int, err error) {
+	return linux.EpollWait(epfd, events, msec)
+}
+
+// EpollCtl is a wrapper
+func EpollCtl(epfd int, op int, fd int, event *EpollEvent) (err error) {
+	return linux.EpollCtl(epfd, op, fd, event)
+}
+
+// Eventfd is a wrapper
+func Eventfd(initval uint, flags int) (fd int, err error) {
+	return linux.Eventfd(initval, flags)
+}
+
+// Write is a wrapper
+func Write(fd int, p []byte) (n int, err error) {
+	return linux.Write(fd, p)
+}
+
+// EpollCreate1 is a wrapper
+func EpollCreate1(flag int) (fd int, err error) {
+	return linux.EpollCreate1(flag)
+}
+
+// PerfEventMmapPage is a wrapper
+type PerfEventMmapPage linux.PerfEventMmapPage
+
+// SetNonblock is a wrapper
+func SetNonblock(fd int, nonblocking bool) (err error) {
+	return linux.SetNonblock(fd, nonblocking)
+}
+
+// Mmap is a wrapper
+func Mmap(fd int, offset int64, length int, prot int, flags int) (data []byte, err error) {
+	return linux.Mmap(fd, offset, length, prot, flags)
+}
+
+// Munmap is a wrapper
+func Munmap(b []byte) (err error) {
+	return linux.Munmap(b)
+}
+
+// PerfEventAttr is a wrapper
+type PerfEventAttr = linux.PerfEventAttr
+
+// PerfEventOpen is a wrapper
+func PerfEventOpen(attr *PerfEventAttr, pid int, cpu int, groupFd int, flags int) (fd int, err error) {
+	return linux.PerfEventOpen(attr, pid, cpu, groupFd, flags)
+}
+
+// Utsname is a wrapper
+type Utsname = linux.Utsname
+
+// Uname is a wrapper
+func Uname(buf *Utsname) (err error) {
+	return linux.Uname(buf)
+}

vendor/github.com/cilium/ebpf/internal/unix/types_other.go

@@ -0,0 +1,193 @@
+// +build !linux
+
+package unix
+
+import (
+	"fmt"
+	"runtime"
+	"syscall"
+)
+
+var errNonLinux = fmt.Errorf("unsupported platform %s/%s", runtime.GOOS, runtime.GOARCH)
+
+const (
+	ENOENT                   = syscall.ENOENT
+	EAGAIN                   = syscall.EAGAIN
+	ENOSPC                   = syscall.ENOSPC
+	EINVAL                   = syscall.EINVAL
+	BPF_OBJ_NAME_LEN         = 0x10
+	BPF_TAG_SIZE             = 0x8
+	SYS_BPF                  = 321
+	F_DUPFD_CLOEXEC          = 0x406
+	EPOLLIN                  = 0x1
+	EPOLL_CTL_ADD            = 0x1
+	EPOLL_CLOEXEC            = 0x80000
+	O_CLOEXEC                = 0x80000
+	O_NONBLOCK               = 0x800
+	PROT_READ                = 0x1
+	PROT_WRITE               = 0x2
+	MAP_SHARED               = 0x1
+	PERF_TYPE_SOFTWARE       = 0x1
+	PERF_COUNT_SW_BPF_OUTPUT = 0xa
+	PerfBitWatermark         = 0x4000
+	PERF_SAMPLE_RAW          = 0x400
+	PERF_FLAG_FD_CLOEXEC     = 0x8
+)
+
+// Statfs_t is a wrapper
+type Statfs_t struct {
+	Type    int64
+	Bsize   int64
+	Blocks  uint64
+	Bfree   uint64
+	Bavail  uint64
+	Files   uint64
+	Ffree   uint64
+	Fsid    [2]int32
+	Namelen int64
+	Frsize  int64
+	Flags   int64
+	Spare   [4]int64
+}
+
+// Rlimit is a wrapper
+type Rlimit struct {
+	Cur uint64
+	Max uint64
+}
+
+// Setrlimit is a wrapper
+func Setrlimit(resource int, rlim *Rlimit) (err error) {
+	return errNonLinux
+}
+
+// Syscall is a wrapper
+func Syscall(trap, a1, a2, a3 uintptr) (r1, r2 uintptr, err syscall.Errno) {
+	return 0, 0, syscall.Errno(1)
+}
+
+// FcntlInt is a wrapper
+func FcntlInt(fd uintptr, cmd, arg int) (int, error) {
+	return -1, errNonLinux
+}
+
+// Statfs is a wrapper
+func Statfs(path string, buf *Statfs_t) error {
+	return errNonLinux
+}
+
+// Close is a wrapper
+func Close(fd int) (err error) {
+	return errNonLinux
+}
+
+// EpollEvent is a wrapper
+type EpollEvent struct {
+	Events uint32
+	Fd     int32
+	Pad    int32
+}
+
+// EpollWait is a wrapper
+func EpollWait(epfd int, events []EpollEvent, msec int) (n int, err error) {
+	return 0, errNonLinux
+}
+
+// EpollCtl is a wrapper
+func EpollCtl(epfd int, op int, fd int, event *EpollEvent) (err error) {
+	return errNonLinux
+}
+
+// Eventfd is a wrapper
+func Eventfd(initval uint, flags int) (fd int, err error) {
+	return 0, errNonLinux
+}
+
+// Write is a wrapper
+func Write(fd int, p []byte) (n int, err error) {
+	return 0, errNonLinux
+}
+
+// EpollCreate1 is a wrapper
+func EpollCreate1(flag int) (fd int, err error) {
+	return 0, errNonLinux
+}
+
+// PerfEventMmapPage is a wrapper
+type PerfEventMmapPage struct {
+	Version        uint32
+	Compat_version uint32
+	Lock           uint32
+	Index          uint32
+	Offset         int64
+	Time_enabled   uint64
+	Time_running   uint64
+	Capabilities   uint64
+	Pmc_width      uint16
+	Time_shift     uint16
+	Time_mult      uint32
+	Time_offset    uint64
+	Time_zero      uint64
+	Size           uint32
+
+	Data_head   uint64
+	Data_tail   uint64
+	Data_offset uint64
+	Data_size   uint64
+	Aux_head    uint64
+	Aux_tail    uint64
+	Aux_offset  uint64
+	Aux_size    uint64
+}
+
+// SetNonblock is a wrapper
+func SetNonblock(fd int, nonblocking bool) (err error) {
+	return errNonLinux
+}
+
+// Mmap is a wrapper
+func Mmap(fd int, offset int64, length int, prot int, flags int) (data []byte, err error) {
+	return []byte{}, errNonLinux
+}
+
+// Munmap is a wrapper
+func Munmap(b []byte) (err error) {
+	return errNonLinux
+}
+
+// PerfEventAttr is a wrapper
+type PerfEventAttr struct {
+	Type               uint32
+	Size               uint32
+	Config             uint64
+	Sample             uint64
+	Sample_type        uint64
+	Read_format        uint64
+	Bits               uint64
+	Wakeup             uint32
+	Bp_type            uint32
+	Ext1               uint64
+	Ext2               uint64
+	Branch_sample_type uint64
+	Sample_regs_user   uint64
+	Sample_stack_user  uint32
+	Clockid            int32
+	Sample_regs_intr   uint64
+	Aux_watermark      uint32
+	Sample_max_stack   uint16
+}
+
+// PerfEventOpen is a wrapper
+func PerfEventOpen(attr *PerfEventAttr, pid int, cpu int, groupFd int, flags int) (fd int, err error) {
+	return 0, errNonLinux
+}
+
+// Utsname is a wrapper
+type Utsname struct {
+    Release    [65]byte
+}
+
+// Uname is a wrapper
+func Uname(buf *Utsname) (err error) {
+	return errNonLinux
+}