Add support of rootless containers

* support podmand rootless containers
* support docker rootless containers

Fixes #311

PiperOrigin-RevId: 458573377

.buildkite/pipeline.yaml

@@ -267,6 +267,12 @@ steps:
     agents:
       cgroup: "v2"
       os: "ubuntu"
+  - <<: *common
+    label: ":podman: Podman"
+    command: sudo ./test/podman/run.sh
+    agents:
+      cgroup: "v2"
+      os: "ubuntu"
 
   # Check the website builds.
   - <<: *common

runsc/cmd/do.go

@@ -48,6 +48,8 @@ type Do struct {
 	ip      string
 	quiet   bool
 	overlay bool
+	uidMap  idMapSlice
+	gidMap  idMapSlice
 }
 
 // Name implements subcommands.Command.Name.
@@ -72,13 +74,53 @@ used for testing only.
 `
 }
 
+type idMapSlice []specs.LinuxIDMapping
+
+// String implements flag.Value.String.
+func (is *idMapSlice) String() string {
+	return fmt.Sprintf("%#v", is)
+}
+
+// Get implements flag.Value.Get.
+func (is *idMapSlice) Get() interface{} {
+	return is
+}
+
+// Set implements flag.Value.Set.
+func (is *idMapSlice) Set(s string) error {
+	fs := strings.Fields(s)
+	if len(fs) != 3 {
+		return fmt.Errorf("invalid mapping: %s", s)
+	}
+	var cid, hid, size int
+	var err error
+	if cid, err = strconv.Atoi(fs[0]); err != nil {
+		return fmt.Errorf("invalid mapping: %s", s)
+	}
+	if hid, err = strconv.Atoi(fs[1]); err != nil {
+		return fmt.Errorf("invalid mapping: %s", s)
+	}
+	if size, err = strconv.Atoi(fs[2]); err != nil {
+		return fmt.Errorf("invalid mapping: %s", s)
+	}
+	m := specs.LinuxIDMapping{
+		ContainerID: uint32(cid),
+		HostID:      uint32(hid),
+		Size:        uint32(size),
+	}
+	*is = append(*is, m)
+	return nil
+}
+
 // SetFlags implements subcommands.Command.SetFlags.
 func (c *Do) SetFlags(f *flag.FlagSet) {
 	f.StringVar(&c.root, "root", "/", `path to the root directory, defaults to "/"`)
 	f.StringVar(&c.cwd, "cwd", ".", "path to the current directory, defaults to the current directory")
 	f.StringVar(&c.ip, "ip", "192.168.10.2", "IPv4 address for the sandbox")
 	f.BoolVar(&c.quiet, "quiet", false, "suppress runsc messages to stdout. Application output is still sent to stdout and stderr")
 	f.BoolVar(&c.overlay, "force-overlay", true, "use an overlay. WARNING: disabling gives the command write access to the host")
+	f.Var(&c.uidMap, "uid-map", "Add a user id mapping [ContainerID, HostID, Size]")
+	f.Var(&c.gidMap, "gid-map", "Add a group id mapping [ContainerID, HostID, Size]")
 }
 
 // Execute implements subcommands.Command.Execute.
@@ -129,6 +171,12 @@ func (c *Do) Execute(_ context.Context, f *flag.FlagSet, args ...interface{}) su
 
 	cid := fmt.Sprintf("runsc-%06d", rand.Int31n(1000000))
 
+	if c.uidMap != nil {
+		addNamespace(spec, specs.LinuxNamespace{Type: specs.UserNamespace})
+		spec.Linux.UIDMappings = c.uidMap
+		spec.Linux.GIDMappings = c.gidMap
+	}
+
 	if conf.Network == config.NetworkNone {
 		addNamespace(spec, specs.LinuxNamespace{Type: specs.NetworkNamespace})
 	} else if conf.Rootless {

runsc/cmd/gofer.go

@@ -18,8 +18,10 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
 	"os"
 	"path/filepath"
+	"runtime"
 	"runtime/debug"
 	"strings"
 
@@ -63,8 +65,9 @@ type Gofer struct {
 	applyCaps bool
 	setUpRoot bool
 
-	specFD   int
-	mountsFD int
+	specFD       int
+	mountsFD     int
+	syncUsernsFD int
 }
 
 // Name implements subcommands.Command.
@@ -92,6 +95,7 @@ func (g *Gofer) SetFlags(f *flag.FlagSet) {
 	f.Var(&g.ioFDs, "io-fds", "list of FDs to connect gofer servers. They must follow this order: root first, then mounts as defined in the spec")
 	f.IntVar(&g.specFD, "spec-fd", -1, "required fd with the container spec")
 	f.IntVar(&g.mountsFD, "mounts-fd", -1, "mountsFD is the file descriptor to write list of mounts after they have been resolved (direct paths, no symlinks).")
+	f.IntVar(&g.syncUsernsFD, "sync-userns-fd", -1, "file descriptor used to synchronize rootless user namespace initialization.")
 }
 
 // Execute implements subcommands.Command.
@@ -113,6 +117,26 @@ func (g *Gofer) Execute(_ context.Context, f *flag.FlagSet, args ...interface{})
 		util.Fatalf("reading spec: %v", err)
 	}
 
+	if g.syncUsernsFD >= 0 {
+		f := os.NewFile(uintptr(g.syncUsernsFD), "sync FD")
+		defer f.Close()
+		var b [1]byte
+		if n, err := f.Read(b[:]); n != 0 || err != io.EOF {
+			util.Fatalf("failed to sync: %v: %v", n, err)
+		}
+
+		f.Close()
+		// SETUID changes UID on the current system thread, so we have
+		// to re-execute current binary.
+		runtime.LockOSThread()
+		if _, _, errno := unix.RawSyscall(unix.SYS_SETUID, 0, 0, 0); errno != 0 {
+			util.Fatalf("failed to set UID: %v", errno)
+		}
+		if _, _, errno := unix.RawSyscall(unix.SYS_SETGID, 0, 0, 0); errno != 0 {
+			util.Fatalf("failed to set GID: %v", errno)
+		}
+	}
+
 	if g.setUpRoot {
 		if err := setupRootFS(spec, conf); err != nil {
 			util.Fatalf("Error setting up root FS: %v", err)
@@ -122,7 +146,7 @@ func (g *Gofer) Execute(_ context.Context, f *flag.FlagSet, args ...interface{})
 		// Disable caps when calling myself again.
 		// Note: minimal argument handling for the default case to keep it simple.
 		args := os.Args
-		args = append(args, "--apply-caps=false", "--setup-root=false")
+		args = append(args, "--apply-caps=false", "--setup-root=false", "--sync-userns-fd=-1")
 		util.Fatalf("setCapsAndCallSelf(%v, %v): %v", args, goferCaps, setCapsAndCallSelf(args, goferCaps))
 		panic("unreachable")
 	}

runsc/container/container.go

@@ -971,15 +971,49 @@ func (c *Container) createGoferProcess(spec *specs.Spec, conf *config.Config, bu
 		{Type: specs.UTSNamespace},
 	}
 
+	rootlessEUID := unix.Getuid() != 0
+	var syncFile *os.File
 	// Setup any uid/gid mappings, and create or join the configured user
 	// namespace so the gofer's view of the filesystem aligns with the
 	// users in the sandbox.
-	userNS := specutils.FilterNS([]specs.LinuxNamespaceType{specs.UserNamespace}, spec)
-	nss = append(nss, userNS...)
-	specutils.SetUIDGIDMappings(cmd, spec)
-	if len(userNS) != 0 {
-		// We need to set UID and GID to have capabilities in a new user namespace.
-		cmd.SysProcAttr.Credential = &syscall.Credential{Uid: 0, Gid: 0}
+	if !rootlessEUID {
+		userNS := specutils.FilterNS([]specs.LinuxNamespaceType{specs.UserNamespace}, spec)
+		nss = append(nss, userNS...)
+		specutils.SetUIDGIDMappings(cmd, spec)
+		if len(userNS) != 0 {
+			// We need to set UID and GID to have capabilities in a new user namespace.
+			cmd.SysProcAttr.Credential = &syscall.Credential{Uid: 0, Gid: 0}
+		}
+	} else {
+		userNS := specutils.FilterNS([]specs.LinuxNamespaceType{specs.UserNamespace}, spec)
+		if len(userNS) == 0 {
+			return nil, nil, fmt.Errorf("unable to run a rootless container without userns")
+		}
+		fds, err := unix.Socketpair(unix.AF_UNIX, unix.SOCK_STREAM|unix.SOCK_CLOEXEC, 0)
+		if err != nil {
+			return nil, nil, err
+		}
+		syncFile = os.NewFile(uintptr(fds[0]), "sync FD")
+		defer syncFile.Close()
+
+		f := os.NewFile(uintptr(fds[1]), "sync other FD")
+		donations.DonateAndClose("sync-userns-fd", f)
+		if cmd.SysProcAttr == nil {
+			cmd.SysProcAttr = &unix.SysProcAttr{}
+		}
+		cmd.SysProcAttr.AmbientCaps = []uintptr{
+			unix.CAP_CHOWN,
+			unix.CAP_DAC_OVERRIDE,
+			unix.CAP_DAC_READ_SEARCH,
+			unix.CAP_FOWNER,
+			unix.CAP_FSETID,
+			unix.CAP_SYS_CHROOT,
+			unix.CAP_SETUID,
+			unix.CAP_SETGID,
+			unix.CAP_SYS_ADMIN,
+			unix.CAP_SETPCAP,
+		}
+		nss = append(nss, specs.LinuxNamespace{Type: specs.UserNamespace})
 	}
 
 	donations.Transfer(cmd, nextFD)
@@ -990,6 +1024,43 @@ func (c *Container) createGoferProcess(spec *specs.Spec, conf *config.Config, bu
 	if err := specutils.StartInNS(cmd, nss); err != nil {
 		return nil, nil, fmt.Errorf("gofer: %v", err)
 	}
+
+	if rootlessEUID {
+		log.Debugf("Setting user mappings")
+		args := []string{strconv.Itoa(cmd.Process.Pid)}
+		for _, idMap := range spec.Linux.UIDMappings {
+			log.Infof("Mapping host uid %d to container uid %d (size=%d)",
+				idMap.HostID, idMap.ContainerID, idMap.Size)
+			args = append(args,
+				strconv.Itoa(int(idMap.ContainerID)),
+				strconv.Itoa(int(idMap.HostID)),
+				strconv.Itoa(int(idMap.Size)),
+			)
+		}
+
+		out, err := exec.Command("newuidmap", args...).CombinedOutput()
+		log.Debugf("newuidmap: %#v\n%s", args, out)
+		if err != nil {
+			return nil, nil, fmt.Errorf("newuidmap failed: %w", err)
+		}
+
+		args = []string{strconv.Itoa(cmd.Process.Pid)}
+		for _, idMap := range spec.Linux.GIDMappings {
+			log.Infof("Mapping host uid %d to container uid %d (size=%d)",
+				idMap.HostID, idMap.ContainerID, idMap.Size)
+			args = append(args,
+				strconv.Itoa(int(idMap.ContainerID)),
+				strconv.Itoa(int(idMap.HostID)),
+				strconv.Itoa(int(idMap.Size)),
+			)
+		}
+		out, err = exec.Command("newgidmap", args...).CombinedOutput()
+		log.Debugf("newgidmap: %#v\n%s", args, out)
+		if err != nil {
+			return nil, nil, fmt.Errorf("newgidmap failed: %w", err)
+		}
+	}
+
 	log.Infof("Gofer started, PID: %d", cmd.Process.Pid)
 	c.GoferPid = cmd.Process.Pid
 	c.goferIsChild = true

runsc/sandbox/sandbox.go

@@ -18,6 +18,7 @@ package sandbox
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"math"
@@ -536,6 +537,7 @@ func (s *Sandbox) createSandboxProcess(conf *config.Config, args *Args, startSyn
 	donations := donation.Agency{}
 	defer donations.Close()
 
+	rootlessEUID := unix.Getuid() != 0
 	//
 	// These flags must come BEFORE the "boot" command in cmd.Args.
 	//
@@ -722,13 +724,13 @@ func (s *Sandbox) createSandboxProcess(conf *config.Config, args *Args, startSyn
 		if conf.TestOnlyAllowRunAsCurrentUserWithoutChroot {
 			log.Warningf("Running sandbox in test mode as current user (uid=%d gid=%d). This is only safe in tests!", os.Getuid(), os.Getgid())
 			log.Warningf("Running sandbox in test mode without chroot. This is only safe in tests!")
-		} else if specutils.HasCapabilities(capability.CAP_SETUID, capability.CAP_SETGID) {
+		} else if rootlessEUID || specutils.HasCapabilities(capability.CAP_SETUID, capability.CAP_SETGID) {
 			log.Infof("Sandbox will be started in new user namespace")
 			nss = append(nss, specs.LinuxNamespace{Type: specs.UserNamespace})
 			cmd.Args = append(cmd.Args, "--setup-root")
 
 			const nobody = 65534
-			if conf.Rootless {
+			if rootlessEUID || conf.Rootless {
 				log.Infof("Rootless mode: sandbox will run as nobody inside user namespace, mapped to the current user, uid: %d, gid: %d", os.Getuid(), os.Getgid())
 			} else {
 				// Map nobody in the new namespace to nobody in the parent namespace.
@@ -1419,6 +1421,10 @@ func (s *Sandbox) configureStdios(conf *config.Config, stdios []*os.File) error
 	for _, file := range stdios {
 		log.Debugf("Changing %q ownership to %d/%d", file.Name(), s.UID, s.GID)
 		if err := file.Chown(s.UID, s.GID); err != nil {
+			if errors.Is(err, unix.EINVAL) || errors.Is(err, unix.EPERM) || errors.Is(err, unix.EROFS) {
+				log.Warningf("can't change an owner of %s: %s", file.Name(), err)
+				continue
+			}
 			return err
 		}
 	}

runsc/specutils/namespace.go

@@ -119,7 +119,7 @@ func setNS(fd, nsType uintptr) error {
 // that will restore the namespace to the original value.
 //
 // Preconditions: Must be called with os thread locked.
-func ApplyNS(ns specs.LinuxNamespace) (func(), error) {
+func ApplyNS(ns specs.LinuxNamespace) (func() error, error) {
 	log.Infof("Applying namespace %v at path %q", ns.Type, ns.Path)
 	newNS, err := os.Open(ns.Path)
 	if err != nil {
@@ -140,27 +140,49 @@ func ApplyNS(ns specs.LinuxNamespace) (func(), error) {
 		oldNS.Close()
 		return nil, fmt.Errorf("error setting namespace of type %v and path %q: %v", ns.Type, ns.Path, err)
 	}
-	return func() {
+	return func() error {
 		log.Infof("Restoring namespace %v", ns.Type)
 		defer oldNS.Close()
 		if err := setNS(oldNS.Fd(), flag); err != nil {
-			panic(fmt.Sprintf("error restoring namespace: of type %v: %v", ns.Type, err))
+			return fmt.Errorf("error restoring namespace: of type %v: %v", ns.Type, err)
 		}
+		return nil
 	}, nil
 }
 
 // StartInNS joins or creates the given namespaces and calls cmd.Start before
 // restoring the namespaces to the original values.
 func StartInNS(cmd *exec.Cmd, nss []specs.LinuxNamespace) error {
-	// We are about to setup namespaces, which requires the os thread being
-	// locked so that Go doesn't change the thread out from under us.
-	runtime.LockOSThread()
-	defer runtime.UnlockOSThread()
+	errChan := make(chan error)
+	go func() {
+		runtime.LockOSThread()
+		defer runtime.UnlockOSThread()
+
+		rstFuncs, err := startInNS(cmd, nss)
+		errChan <- err
+		for _, rstFunc := range rstFuncs {
+			err := rstFunc()
+			if err == nil {
+				continue
+			}
+
+			// One or more namespaces have not been restored, but
+			// we can't destroy the current system thread, because
+			// a child process is execited with Pdeathsig.
+			log.Debugf("Block the current system thread due to: %s", err)
+			c := make(chan interface{})
+			<-c
+		}
+	}()
+	return <-errChan
+}
 
+func startInNS(cmd *exec.Cmd, nss []specs.LinuxNamespace) ([]func() error, error) {
 	if cmd.SysProcAttr == nil {
 		cmd.SysProcAttr = &unix.SysProcAttr{}
 	}
 
+	var deferFuncs []func() error
 	for _, ns := range nss {
 		if ns.Path == "" {
 			// No path.  Just set a flag to create a new namespace.
@@ -171,12 +193,12 @@ func StartInNS(cmd *exec.Cmd, nss []specs.LinuxNamespace) error {
 		// before exiting.
 		restoreNS, err := ApplyNS(ns)
 		if err != nil {
-			return err
+			return deferFuncs, err
 		}
-		defer restoreNS()
+		deferFuncs = append(deferFuncs, restoreNS)
 	}
 
-	return cmd.Start()
+	return deferFuncs, cmd.Start()
 }
 
 // SetUIDGIDMappings sets the given uid/gid mappings from the spec on the cmd.