Add a runtime config

This adds support for a runtime configuration file that can do:
- `mkdir` to make a directory at runtime, eg in `/var` or `/tmp`, to avoid workarounds
- `interface` that can create network interfaces in a container or move them
- `bindNS` that can bind mount namespaces of an `onboot` container to a file so a service can be started in that namespace.

It merges the `service` and `onboot` tools (in `init`) to avoid duplication. This also saves some size for
eg LCOW which did not use the `onboot` code in `runc`.

Signed-off-by: Justin Cormack <justin.cormack@docker.com>
This commit is contained in:
Justin Cormack
2017-08-16 16:40:58 +01:00
parent df16f6fb9e
commit 0c81ce19e8
14 changed files with 420 additions and 280 deletions

View File

@@ -0,0 +1,96 @@
package main
import (
"flag"
"fmt"
"os"
"path/filepath"
"strings"
log "github.com/sirupsen/logrus"
)
const (
defaultSocket = "/run/containerd/containerd.sock"
defaultPath = "/containers/services"
defaultContainerd = "/usr/bin/containerd"
installPath = "/usr/bin/service"
onbootPath = "/containers/onboot"
shutdownPath = "/containers/onshutdown"
)
var (
defaultLogFormatter = &log.TextFormatter{}
)
// infoFormatter overrides the default format for Info() log events to
// provide an easier to read output
type infoFormatter struct {
}
func (f *infoFormatter) Format(entry *log.Entry) ([]byte, error) {
if entry.Level == log.InfoLevel {
return append([]byte(entry.Message), '\n'), nil
}
return defaultLogFormatter.Format(entry)
}
func main() {
flag.Usage = func() {
fmt.Printf("USAGE: %s [options] COMMAND\n\n", filepath.Base(os.Args[0]))
fmt.Printf("Commands:\n")
fmt.Printf(" system-init Prepare the system at start of day\n")
fmt.Printf(" start Start a service\n")
fmt.Printf(" help Print this message\n")
fmt.Printf("\n")
fmt.Printf("Run '%s COMMAND --help' for more information on the command\n", filepath.Base(os.Args[0]))
fmt.Printf("\n")
fmt.Printf("Options:\n")
flag.PrintDefaults()
}
flagQuiet := flag.Bool("q", false, "Quiet execution")
flagVerbose := flag.Bool("v", false, "Verbose execution")
// Set up logging
log.SetFormatter(new(infoFormatter))
log.SetLevel(log.InfoLevel)
flag.Parse()
if *flagQuiet && *flagVerbose {
fmt.Printf("Can't set quiet and verbose flag at the same time\n")
os.Exit(1)
}
if *flagQuiet {
log.SetLevel(log.ErrorLevel)
}
if *flagVerbose {
// Switch back to the standard formatter
log.SetFormatter(defaultLogFormatter)
log.SetLevel(log.DebugLevel)
}
args := flag.Args()
if len(args) < 1 {
// check if called form startup scripts
command := os.Args[0]
switch {
case strings.Contains(command, "onboot"):
os.Exit(runcInit(onbootPath))
case strings.Contains(command, "onshutdown"):
os.Exit(runcInit(shutdownPath))
case strings.Contains(command, "containerd"):
systemInitCmd([]string{})
os.Exit(0)
}
}
switch args[0] {
case "start":
startCmd(args[1:])
case "system-init":
systemInitCmd(args[1:])
default:
fmt.Printf("%q is not valid command.\n\n", args[0])
flag.Usage()
os.Exit(1)
}
}

View File

@@ -0,0 +1,255 @@
package main
import (
"encoding/json"
"fmt"
"io/ioutil"
"log"
"os"
"path/filepath"
"syscall"
"github.com/vishvananda/netlink"
)
const (
wgPath = "/usr/bin/wg"
nsenterPath = "/usr/bin/nsenter-net"
)
// Note these definitions are from moby/tool/src/moby/config.go and should be kept in sync
// Runtime is the type of config processed at runtime, not used to build the OCI spec
type Runtime struct {
Mkdir []string `yaml:"mkdir" json:"mkdir,omitempty"`
Interfaces []Interface `yaml:"interfaces" json:"interfaces,omitempty"`
BindNS *Namespaces `yaml:"bindNS" json:"bindNS,omitempty"`
}
// Namespaces is the type for configuring paths to bind namespaces
type Namespaces struct {
Cgroup string `yaml:"cgroup" json:"cgroup,omitempty"`
Ipc string `yaml:"ipc" json:"ipc,omitempty"`
Mnt string `yaml:"mnt" json:"mnt,omitempty"`
Net string `yaml:"net" json:"net,omitempty"`
Pid string `yaml:"pid" json:"pid,omitempty"`
User string `yaml:"user" json:"user,omitempty"`
Uts string `yaml:"uts" json:"uts,omitempty"`
}
// Interface is the runtime config for network interfaces
type Interface struct {
Name string `yaml:"name" json:"name,omitempty"`
Add string `yaml:"add" json:"add,omitempty"`
Peer string `yaml:"peer" json:"peer,omitempty"`
CreateInRoot bool `yaml:"createInRoot" json:"createInRoot"`
}
func getRuntimeConfig(path string) Runtime {
var runtime Runtime
conf, err := ioutil.ReadFile(filepath.Join(path, "runtime.json"))
if err != nil {
// if it does not exist it is fine to return an empty runtime, to not do anything
if os.IsNotExist(err) {
return runtime
}
log.Fatalf("Cannot read runtime config: %v", err)
}
if err := json.Unmarshal(conf, &runtime); err != nil {
log.Fatalf("Cannot parse runtime config: %v", err)
}
return runtime
}
// prepareFilesystem sets up the mounts, before the container is created
func prepareFilesystem(path string, runtime Runtime) error {
// execute the runtime config that should be done up front
for _, dir := range runtime.Mkdir {
// in future we may need to change the structure to set mode, ownership
var mode os.FileMode = 0755
err := os.MkdirAll(dir, mode)
if err != nil {
return fmt.Errorf("Cannot create directory %s: %v", dir, err)
}
}
// see if we are dealing with a read only or read write container
if _, err := os.Stat(filepath.Join(path, "lower")); err != nil {
if os.IsNotExist(err) {
return prepareRO(path)
}
return err
}
return prepareRW(path)
}
func prepareRO(path string) error {
// make rootfs a mount point, as runc doesn't like it much otherwise
rootfs := filepath.Join(path, "rootfs")
if err := syscall.Mount(rootfs, rootfs, "", syscall.MS_BIND, ""); err != nil {
return err
}
return nil
}
func prepareRW(path string) error {
// mount a tmpfs on tmp for upper and workdirs
// make it private as nothing else should be using this
tmp := filepath.Join(path, "tmp")
if err := syscall.Mount("tmpfs", tmp, "tmpfs", 0, "size=10%"); err != nil {
return err
}
// make it private as nothing else should be using this
if err := syscall.Mount("", tmp, "", syscall.MS_REMOUNT|syscall.MS_PRIVATE, ""); err != nil {
return err
}
upper := filepath.Join(tmp, "upper")
// make the mount points
if err := os.Mkdir(upper, 0755); err != nil {
return err
}
work := filepath.Join(tmp, "work")
if err := os.Mkdir(work, 0755); err != nil {
return err
}
lower := filepath.Join(path, "lower")
rootfs := filepath.Join(path, "rootfs")
opt := fmt.Sprintf("lowerdir=%s,upperdir=%s,workdir=%s", lower, upper, work)
if err := syscall.Mount("overlay", rootfs, "overlay", 0, opt); err != nil {
return err
}
return nil
}
// bind mount a namespace file
func bindNS(ns string, path string, pid int) error {
if path == "" {
return nil
}
// the path and file need to exist for the bind to succeed, so try to create
dir := filepath.Dir(path)
if err := os.MkdirAll(dir, 0755); err != nil {
return fmt.Errorf("Cannot create leading directories %s for bind mount destination: %v", dir, err)
}
fi, err := os.Create(path)
if err != nil {
return fmt.Errorf("Cannot create a mount point for namespace bind at %s: %v", path, err)
}
if err := fi.Close(); err != nil {
return err
}
if err := syscall.Mount(fmt.Sprintf("/proc/%d/ns/%s", pid, ns), path, "", syscall.MS_BIND, ""); err != nil {
return fmt.Errorf("Failed to bind %s namespace at %s: %v", ns, path, err)
}
return nil
}
// prepareProcess sets up anything that needs to be done after the container process is created, but before it runs
// for example networking
func prepareProcess(pid int, runtime Runtime) error {
for _, iface := range runtime.Interfaces {
if iface.Name == "" {
return fmt.Errorf("Interface requires a name")
}
var link netlink.Link
var ns interface{} = netlink.NsPid(pid)
var move bool
var err error
if iface.Peer != "" && iface.Add == "" {
// must be a veth if specify peer
iface.Add = "veth"
}
// if create in root is set, create in root namespace first, then move
// also do the same for a veth pair
if iface.CreateInRoot || iface.Add == "veth" {
ns = nil
move = true
}
if iface.Add != "" {
switch iface.Add {
case "veth":
if iface.Peer == "" {
return fmt.Errorf("Creating a veth pair %s requires a peer to be set", iface.Name)
}
la := netlink.LinkAttrs{Name: iface.Name, Namespace: ns}
link = &netlink.Veth{LinkAttrs: la, PeerName: iface.Peer}
default:
// no special creation options needed
la := netlink.LinkAttrs{Name: iface.Name, Namespace: ns}
link = &netlink.GenericLink{la, iface.Add}
}
if err := netlink.LinkAdd(link); err != nil {
return fmt.Errorf("Link add %s of type %s failed: %v", iface.Name, iface.Add, err)
}
fmt.Fprintf(os.Stderr, "Created interface %s type %s\n", iface.Name, iface.Add)
} else {
// find existing interface
link, err = netlink.LinkByName(iface.Name)
if err != nil {
return fmt.Errorf("Cannot find interface %s: %v", iface.Name, err)
}
// then move into namespace
move = true
}
if move {
if err := netlink.LinkSetNsPid(link, int(pid)); err != nil {
return fmt.Errorf("Cannot move interface %s into namespace: %v", iface.Name, err)
}
fmt.Fprintf(os.Stderr, "Moved interface %s to pid %d\n", iface.Name, pid)
}
}
if runtime.BindNS != nil {
binds := []struct {
ns string
path string
}{
{"cgroup", runtime.BindNS.Cgroup},
{"ipc", runtime.BindNS.Ipc},
{"mnt", runtime.BindNS.Mnt},
{"net", runtime.BindNS.Net},
{"pid", runtime.BindNS.Pid},
{"user", runtime.BindNS.User},
{"uts", runtime.BindNS.Uts},
}
for _, b := range binds {
if err := bindNS(b.ns, b.path, pid); err != nil {
return err
}
}
}
return nil
}
// cleanup functions are best efforts only, mainly for rw onboot containers
func cleanup(path string) {
// see if we are dealing with a read only or read write container
if _, err := os.Stat(filepath.Join(path, "lower")); err != nil {
cleanupRO(path)
} else {
cleanupRW(path)
}
}
func cleanupRO(path string) {
// remove the bind mount
rootfs := filepath.Join(path, "rootfs")
_ = syscall.Unmount(rootfs, 0)
}
func cleanupRW(path string) {
// remove the overlay mount
rootfs := filepath.Join(path, "rootfs")
_ = os.RemoveAll(rootfs)
_ = syscall.Unmount(rootfs, 0)
// remove the tmpfs
tmp := filepath.Join(path, "tmp")
_ = os.RemoveAll(tmp)
_ = syscall.Unmount(tmp, 0)
}

View File

@@ -0,0 +1,111 @@
package main
import (
"io/ioutil"
"log"
"os"
"os/exec"
"path/filepath"
"strconv"
"github.com/containerd/containerd/sys"
)
const (
runcBinary = "/usr/bin/runc"
)
func runcInit(rootPath string) int {
// do nothing if the path does not exist
if _, err := os.Stat(rootPath); err != nil && os.IsNotExist(err) {
return 0
}
// get files; note ReadDir already sorts them
files, err := ioutil.ReadDir(rootPath)
if err != nil {
log.Fatalf("Cannot read files in %s: %v", rootPath, err)
}
tmpdir, err := ioutil.TempDir("", filepath.Base(rootPath))
if err != nil {
log.Fatalf("Cannot create temporary directory: %v", err)
}
// need to set ourselves as a child subreaper or we cannot wait for runc as reparents to init
if err := sys.SetSubreaper(1); err != nil {
log.Fatalf("Cannot set as subreaper: %v", err)
}
status := 0
for _, file := range files {
name := file.Name()
path := filepath.Join(rootPath, name)
runtimeConfig := getRuntimeConfig(path)
if err := prepareFilesystem(path, runtimeConfig); err != nil {
log.Printf("Error preparing %s: %v", name, err)
status = 1
continue
}
pidfile := filepath.Join(tmpdir, name)
cmd := exec.Command(runcBinary, "create", "--bundle", path, "--pid-file", pidfile, name)
cmd.Stdout = os.Stdout
cmd.Stderr = os.Stderr
if err := cmd.Run(); err != nil {
log.Printf("Error creating %s: %v", name, err)
status = 1
// skip cleanup on error for debug
continue
}
pf, err := ioutil.ReadFile(pidfile)
if err != nil {
log.Printf("Cannot read pidfile: %v", err)
status = 1
continue
}
pid, err := strconv.Atoi(string(pf))
if err != nil {
log.Printf("Cannot parse pid from pidfile: %v", err)
status = 1
continue
}
if err := prepareProcess(pid, runtimeConfig); err != nil {
log.Printf("Cannot prepare process: %v", err)
status = 1
continue
}
waitFor := make(chan *os.ProcessState)
go func() {
// never errors in Unix
p, _ := os.FindProcess(pid)
state, err := p.Wait()
if err != nil {
log.Printf("Process wait error: %v", err)
}
waitFor <- state
}()
cmd = exec.Command(runcBinary, "start", name)
cmd.Stdout = os.Stdout
cmd.Stderr = os.Stderr
if err := cmd.Run(); err != nil {
log.Printf("Error starting %s: %v", name, err)
status = 1
continue
}
_ = <-waitFor
cleanup(path)
_ = os.Remove(pidfile)
}
_ = os.RemoveAll(tmpdir)
return status
}

View File

@@ -0,0 +1,136 @@
package main
import (
"context"
"encoding/json"
"flag"
"fmt"
"io/ioutil"
"os"
"path/filepath"
"github.com/containerd/containerd"
"github.com/containerd/containerd/namespaces"
specs "github.com/opencontainers/runtime-spec/specs-go"
log "github.com/sirupsen/logrus"
)
func startCmd(args []string) {
invoked := filepath.Base(os.Args[0])
flags := flag.NewFlagSet("start", flag.ExitOnError)
flags.Usage = func() {
fmt.Printf("USAGE: %s start [service]\n\n", invoked)
fmt.Printf("Options:\n")
flags.PrintDefaults()
}
sock := flags.String("sock", defaultSocket, "Path to containerd socket")
path := flags.String("path", defaultPath, "Path to service configs")
dumpSpec := flags.String("dump-spec", "", "Dump container spec to file before start")
if err := flags.Parse(args); err != nil {
log.Fatal("Unable to parse args")
}
args = flags.Args()
if len(args) != 1 {
fmt.Println("Please specify the service")
flags.Usage()
os.Exit(1)
}
service := args[0]
log.Infof("Starting service: %q", service)
log := log.WithFields(log.Fields{
"service": service,
})
id, pid, msg, err := start(service, *sock, *path, *dumpSpec)
if err != nil {
log.WithError(err).Fatal(msg)
}
log.Debugf("Started %s pid %d", id, pid)
}
func start(service, sock, basePath, dumpSpec string) (string, uint32, string, error) {
path := filepath.Join(basePath, service)
runtimeConfig := getRuntimeConfig(path)
rootfs := filepath.Join(path, "rootfs")
if err := prepareFilesystem(path, runtimeConfig); err != nil {
return "", 0, "preparing filesystem", err
}
client, err := containerd.New(sock)
if err != nil {
return "", 0, "creating containerd client", err
}
ctx := namespaces.WithNamespace(context.Background(), "default")
var spec *specs.Spec
specf, err := os.Open(filepath.Join(path, "config.json"))
if err != nil {
return "", 0, "failed to read service spec", err
}
if err := json.NewDecoder(specf).Decode(&spec); err != nil {
return "", 0, "failed to parse service spec", err
}
spec.Root.Path = rootfs
if dumpSpec != "" {
d, err := os.Create(dumpSpec)
if err != nil {
return "", 0, "failed to open file for spec dump", err
}
enc := json.NewEncoder(d)
enc.SetIndent("", " ")
if err := enc.Encode(&spec); err != nil {
return "", 0, "failed to write spec dump", err
}
}
ctr, err := client.NewContainer(ctx, service, containerd.WithSpec(spec))
if err != nil {
return "", 0, "failed to create container", err
}
io := func(id string) (*containerd.IO, error) {
logfile := filepath.Join("/var/log", service+".log")
// We just need this to exist.
if err := ioutil.WriteFile(logfile, []byte{}, 0600); err != nil {
// if we cannot write to log, discard output
logfile = "/dev/null"
}
return &containerd.IO{
Stdin: "/dev/null",
Stdout: logfile,
Stderr: logfile,
Terminal: false,
}, nil
}
task, err := ctr.NewTask(ctx, io)
if err != nil {
// Don't bother to destroy the container here.
return "", 0, "failed to create task", err
}
if err := prepareProcess(int(task.Pid()), runtimeConfig); err != nil {
return "", 0, "preparing process", err
}
if err := task.Start(ctx); err != nil {
// Don't destroy the container here so it can be inspected for debugging.
return "", 0, "failed to start task", err
}
return ctr.ID(), task.Pid(), "", nil
}

View File

@@ -0,0 +1,149 @@
package main
import (
"context"
"flag"
"fmt"
"io/ioutil"
"os"
"os/exec"
"path/filepath"
"syscall"
"time"
"github.com/containerd/containerd"
"github.com/containerd/containerd/errdefs"
"github.com/containerd/containerd/namespaces"
"github.com/pkg/errors"
log "github.com/sirupsen/logrus"
)
func cleanupTask(ctx context.Context, ctr containerd.Container) error {
task, err := ctr.Task(ctx, nil)
if err != nil {
if errdefs.IsNotFound(err) {
return nil
}
return errors.Wrap(err, "getting task")
}
deleteErr := make(chan error, 1)
deleteCtx, deleteCancel := context.WithCancel(ctx)
defer deleteCancel()
go func(ctx context.Context, ch chan error) {
_, err := task.Delete(ctx)
if err != nil {
ch <- errors.Wrap(err, "killing task")
}
ch <- nil
}(deleteCtx, deleteErr)
sig := syscall.SIGKILL
if err := task.Kill(ctx, sig); err != nil && !errdefs.IsNotFound(err) {
return errors.Wrapf(err, "killing task with %q", sig)
}
select {
case err := <-deleteErr:
return err
case <-ctx.Done():
return ctx.Err()
}
}
func systemInitCmd(args []string) {
invoked := filepath.Base(os.Args[0])
flags := flag.NewFlagSet("system-init", flag.ExitOnError)
flags.Usage = func() {
fmt.Printf("USAGE: %s system-init\n\n", invoked)
fmt.Printf("Options:\n")
flags.PrintDefaults()
}
sock := flags.String("sock", defaultSocket, "Path to containerd socket")
path := flags.String("path", defaultPath, "Path to service configs")
binary := flags.String("containerd", defaultContainerd, "Path to containerd")
if err := flags.Parse(args); err != nil {
log.Fatal("Unable to parse args")
}
args = flags.Args()
if len(args) != 0 {
fmt.Println("Unexpected argument")
flags.Usage()
os.Exit(1)
}
// remove (unlikely) old containerd socket
_ = os.Remove(*sock)
// start up containerd
cmd := exec.Command(*binary)
cmd.Stdout = os.Stdout
cmd.Stderr = os.Stderr
if err := cmd.Start(); err != nil {
log.WithError(err).Fatal("cannot start containerd")
}
// wait for containerd socket to appear
for {
_, err := os.Stat(*sock)
if err == nil {
break
}
err = cmd.Process.Signal(syscall.Signal(0))
if err != nil {
// process not there, wait() to find error
err = cmd.Wait()
log.WithError(err).Fatal("containerd process exited")
}
time.Sleep(100 * time.Millisecond)
}
// connect to containerd
client, err := containerd.New(*sock)
if err != nil {
log.WithError(err).Fatal("creating containerd client")
}
ctx := namespaces.WithNamespace(context.Background(), "default")
ctrs, err := client.Containers(ctx)
if err != nil {
log.WithError(err).Fatal("listing containers")
}
// Clean up any old containers
// None of the errors in this loop are fatal since we want to
// keep trying.
for _, ctr := range ctrs {
log.Infof("Cleaning up stale service: %q", ctr.ID())
log := log.WithFields(log.Fields{
"service": ctr.ID(),
})
if err := cleanupTask(ctx, ctr); err != nil {
log.WithError(err).Error("cleaning up task")
}
if err := ctr.Delete(ctx); err != nil {
log.WithError(err).Error("deleting container")
}
}
// Start up containers
files, err := ioutil.ReadDir(*path)
// just skip if there is an error, eg no such path
if err != nil {
return
}
for _, file := range files {
if id, pid, msg, err := start(file.Name(), *sock, *path, ""); err != nil {
log.WithError(err).Error(msg)
} else {
log.Debugf("Started %s pid %d", id, pid)
}
}
}