rootless: use a single user namespace

simplify the rootless implementation to use a single user namespace
for all the running containers.

This makes the rootless implementation behave more like root Podman,
where each container is created in the host environment.

There are multiple advantages to it: 1) much simpler implementation as
there is only one namespace to join.  2) we can join namespaces owned
by different containers.  3) commands like ps won't be limited to what
container they can access as previously we either had access to the
storage from a new namespace or access to /proc when running from the
host.  4) rootless varlink works.  5) there are only two ways to enter
in a namespace, either by creating a new one if no containers are
running or joining the existing one from any container.

Containers created by older Podman versions must be restarted.

Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>

4 files changed, 114 insertions, 235 deletions

diff --git a/pkg/rootless/rootless.go b/pkg/rootless/rootless.go
deleted file mode 100644
index a531e43ce..000000000
--- a/pkg/rootless/rootless.go
+++ /dev/null
@@ -1,9 +0,0 @@
-package rootless
-
-// Opts allows to customize how re-execing to a rootless process is done
-type Opts struct {
-	// Argument overrides the arguments on the command line
-	// for the re-execed process.  The process in the namespace
-	// must use rootless.Argument() to read its value.
-	Argument string
-}
diff --git a/pkg/rootless/rootless_linux.c b/pkg/rootless/rootless_linux.c
index 2e2c3acac..9cb79ed4d 100644
--- a/pkg/rootless/rootless_linux.c
+++ b/pkg/rootless/rootless_linux.c
@@ -13,10 +13,36 @@
 #include <sys/wait.h>
 #include <string.h>
 #include <stdbool.h>
+#include <sys/types.h>
+#include <sys/prctl.h>
+#include <dirent.h>
 
 static const char *_max_user_namespaces = "/proc/sys/user/max_user_namespaces";
 static const char *_unprivileged_user_namespaces = "/proc/sys/kernel/unprivileged_userns_clone";
 
+static int n_files;
+
+static void __attribute__((constructor)) init()
+{
+  DIR *d;
+
+  /* Store how many FDs were open before the Go runtime kicked in.  */
+  d = opendir ("/proc/self/fd");
+  if (d)
+    {
+      struct dirent *ent;
+
+      for (ent = readdir (d); ent; ent = readdir (d))
+        {
+          int fd = atoi (ent->d_name);
+          if (fd > n_files && fd != dirfd (d))
+            n_files = fd;
+        }
+      closedir (d);
+    }
+}
+
+
 static int
 syscall_setresuid (uid_t ruid, uid_t euid, uid_t suid)
 {
@@ -133,12 +159,25 @@ reexec_userns_join (int userns, int mountns)
   pid = fork ();
   if (pid < 0)
     fprintf (stderr, "cannot fork: %s\n", strerror (errno));
+
   if (pid)
-    return pid;
+    {
+      /* We passed down these fds, close them.  */
+      int f;
+      for (f = 3; f < n_files; f++)
+        close (f);
+      return pid;
+    }
 
   setenv ("_CONTAINERS_USERNS_CONFIGURED", "init", 1);
   setenv ("_CONTAINERS_ROOTLESS_UID", uid, 1);
 
+  if (prctl (PR_SET_PDEATHSIG, SIGTERM, 0, 0, 0) < 0)
+    {
+      fprintf (stderr, "cannot prctl(PR_SET_PDEATHSIG): %s\n", strerror (errno));
+      _exit (EXIT_FAILURE);
+    }
+
   if (setns (userns, 0) < 0)
     {
       fprintf (stderr, "cannot setns: %s\n", strerror (errno));
diff --git a/pkg/rootless/rootless_linux.go b/pkg/rootless/rootless_linux.go
index c753228f1..1d1b1713d 100644
--- a/pkg/rootless/rootless_linux.go
+++ b/pkg/rootless/rootless_linux.go
@@ -46,11 +46,6 @@ func IsRootless() bool {
 	return isRootless
 }
 
-// Argument returns the argument that was set for the rootless session.
-func Argument() string {
-	return os.Getenv("_CONTAINERS_ROOTLESS_ARG")
-}
-
 // GetRootlessUID returns the UID of the user in the parent userNS
 func GetRootlessUID() int {
 	uidEnv := os.Getenv("_CONTAINERS_ROOTLESS_UID")
@@ -90,51 +85,86 @@ func tryMappingTool(tool string, pid int, hostID int, mappings []idtools.IDMap)
 	return nil
 }
 
-// JoinNS re-exec podman in a new userNS and join the user namespace of the specified
-// PID.
-func JoinNS(pid uint, preserveFDs int) (bool, int, error) {
-	if os.Geteuid() == 0 || os.Getenv("_CONTAINERS_USERNS_CONFIGURED") != "" {
-		return false, -1, nil
+func readUserNs(path string) (string, error) {
+	b := make([]byte, 256)
+	_, err := syscall.Readlink(path, b)
+	if err != nil {
+		return "", err
+	}
+	return string(b), nil
+}
+
+func readUserNsFd(fd uintptr) (string, error) {
+	return readUserNs(fmt.Sprintf("/proc/self/fd/%d", fd))
+}
+
+func getParentUserNs(fd uintptr) (uintptr, error) {
+	const nsGetParent = 0xb702
+	ret, _, errno := syscall.Syscall(syscall.SYS_IOCTL, fd, uintptr(nsGetParent), 0)
+	if errno != 0 {
+		return 0, errno
 	}
+	return (uintptr)(unsafe.Pointer(ret)), nil
+}
 
-	userNS, err := getUserNSForPid(pid)
+// getUserNSFirstChild returns an open FD for the first direct child user namespace that created the process
+// Each container creates a new user namespace where the runtime runs.  The current process in the container
+// might have created new user namespaces that are child of the initial namespace we created.
+// This function finds the initial namespace created for the container that is a child of the current namespace.
+//
+//                                     current ns
+//                                       /     \
+//                           TARGET ->  a   [other containers]
+//                                     /
+//                                    b
+//                                   /
+//        NS READ USING THE PID ->  c
+func getUserNSFirstChild(fd uintptr) (*os.File, error) {
+	currentNS, err := readUserNs("/proc/self/ns/user")
 	if err != nil {
-		return false, -1, err
+		return nil, err
 	}
-	defer userNS.Close()
 
-	pidC := C.reexec_userns_join(C.int(userNS.Fd()), -1)
-	if int(pidC) < 0 {
-		return false, -1, errors.Errorf("cannot re-exec process")
+	ns, err := readUserNsFd(fd)
+	if err != nil {
+		return nil, errors.Wrapf(err, "cannot read user namespace")
 	}
-	if preserveFDs > 0 {
-		for fd := 3; fd < 3+preserveFDs; fd++ {
-			// These fds were passed down to the runtime.  Close them
-			// and not interfere
-			os.NewFile(uintptr(fd), fmt.Sprintf("fd-%d", fd)).Close()
-		}
+	if ns == currentNS {
+		return nil, errors.New("process running in the same user namespace")
 	}
 
-	ret := C.reexec_in_user_namespace_wait(pidC)
-	if ret < 0 {
-		return false, -1, errors.New("error waiting for the re-exec process")
-	}
+	for {
+		nextFd, err := getParentUserNs(fd)
+		if err != nil {
+			return nil, errors.Wrapf(err, "cannot get parent user namespace")
+		}
 
-	return true, int(ret), nil
-}
+		ns, err = readUserNsFd(nextFd)
+		if err != nil {
+			return nil, errors.Wrapf(err, "cannot read user namespace")
+		}
 
-// JoinDirectUserAndMountNS re-exec podman in a new userNS and join the user and mount
-// namespace of the specified PID without looking up its parent.  Useful to join directly
-// the conmon process.  It is a convenience function for JoinDirectUserAndMountNSWithOpts
-// with a default configuration.
-func JoinDirectUserAndMountNS(pid uint) (bool, int, error) {
-	return JoinDirectUserAndMountNSWithOpts(pid, nil)
+		if ns == currentNS {
+			syscall.Close(int(nextFd))
+
+			// Drop O_CLOEXEC for the fd.
+			_, _, errno := syscall.Syscall(syscall.SYS_FCNTL, fd, syscall.F_SETFD, 0)
+			if errno != 0 {
+				syscall.Close(int(fd))
+				return nil, errno
+			}
+
+			return os.NewFile(fd, "userns child"), nil
+		}
+		syscall.Close(int(fd))
+		fd = nextFd
+	}
 }
 
-// JoinDirectUserAndMountNSWithOpts re-exec podman in a new userNS and join the user and
-// mount namespace of the specified PID without looking up its parent.  Useful to join
-// directly the conmon process.
-func JoinDirectUserAndMountNSWithOpts(pid uint, opts *Opts) (bool, int, error) {
+// JoinUserAndMountNS re-exec podman in a new userNS and join the user and mount
+// namespace of the specified PID without looking up its parent.  Useful to join directly
+// the conmon process.
+func JoinUserAndMountNS(pid uint) (bool, int, error) {
 	if os.Geteuid() == 0 || os.Getenv("_CONTAINERS_USERNS_CONFIGURED") != "" {
 		return false, -1, nil
 	}
@@ -151,39 +181,11 @@ func JoinDirectUserAndMountNSWithOpts(pid uint, opts *Opts) (bool, int, error) {
 	}
 	defer userNS.Close()
 
-	if opts != nil && opts.Argument != "" {
-		if err := os.Setenv("_CONTAINERS_ROOTLESS_ARG", opts.Argument); err != nil {
-			return false, -1, err
-		}
-	}
-
-	pidC := C.reexec_userns_join(C.int(userNS.Fd()), C.int(mountNS.Fd()))
-	if int(pidC) < 0 {
-		return false, -1, errors.Errorf("cannot re-exec process")
-	}
-
-	ret := C.reexec_in_user_namespace_wait(pidC)
-	if ret < 0 {
-		return false, -1, errors.New("error waiting for the re-exec process")
-	}
-
-	return true, int(ret), nil
-}
-
-// JoinNSPath re-exec podman in a new userNS and join the owner user namespace of the
-// specified path.
-func JoinNSPath(path string) (bool, int, error) {
-	if os.Geteuid() == 0 || os.Getenv("_CONTAINERS_USERNS_CONFIGURED") != "" {
-		return false, -1, nil
-	}
-
-	userNS, err := getUserNSForPath(path)
+	fd, err := getUserNSFirstChild(userNS.Fd())
 	if err != nil {
 		return false, -1, err
 	}
-	defer userNS.Close()
-
-	pidC := C.reexec_userns_join(C.int(userNS.Fd()), -1)
+	pidC := C.reexec_userns_join(C.int(fd.Fd()), C.int(mountNS.Fd()))
 	if int(pidC) < 0 {
 		return false, -1, errors.Errorf("cannot re-exec process")
 	}
@@ -199,16 +201,8 @@ func JoinNSPath(path string) (bool, int, error) {
 // BecomeRootInUserNS re-exec podman in a new userNS.  It returns whether podman was re-executed
 // into a new user namespace and the return code from the re-executed podman process.
 // If podman was re-executed the caller needs to propagate the error code returned by the child
-// process.  It is a convenience function for BecomeRootInUserNSWithOpts with a default configuration.
-func BecomeRootInUserNS() (bool, int, error) {
-	return BecomeRootInUserNSWithOpts(nil)
-}
-
-// BecomeRootInUserNSWithOpts re-exec podman in a new userNS.  It returns whether podman was
-// re-execute into a new user namespace and the return code from the re-executed podman process.
-// If podman was re-executed the caller needs to propagate the error code returned by the child
 // process.
-func BecomeRootInUserNSWithOpts(opts *Opts) (bool, int, error) {
+func BecomeRootInUserNS() (bool, int, error) {
 	if os.Geteuid() == 0 || os.Getenv("_CONTAINERS_USERNS_CONFIGURED") != "" {
 		if os.Getenv("_CONTAINERS_USERNS_CONFIGURED") == "init" {
 			return false, 0, runInUser()
@@ -227,12 +221,6 @@ func BecomeRootInUserNSWithOpts(opts *Opts) (bool, int, error) {
 	defer w.Close()
 	defer w.Write([]byte("0"))
 
-	if opts != nil && opts.Argument != "" {
-		if err := os.Setenv("_CONTAINERS_ROOTLESS_ARG", opts.Argument); err != nil {
-			return false, -1, err
-		}
-	}
-
 	pidC := C.reexec_in_user_namespace(C.int(r.Fd()))
 	pid := int(pidC)
 	if pid < 0 {
@@ -314,112 +302,3 @@ func BecomeRootInUserNSWithOpts(opts *Opts) (bool, int, error) {
 
 	return true, int(ret), nil
 }
-
-func readUserNs(path string) (string, error) {
-	b := make([]byte, 256)
-	_, err := syscall.Readlink(path, b)
-	if err != nil {
-		return "", err
-	}
-	return string(b), nil
-}
-
-func readUserNsFd(fd uintptr) (string, error) {
-	return readUserNs(fmt.Sprintf("/proc/self/fd/%d", fd))
-}
-
-func getOwner(fd uintptr) (uintptr, error) {
-	const nsGetUserns = 0xb701
-	ret, _, errno := syscall.Syscall(syscall.SYS_IOCTL, fd, uintptr(nsGetUserns), 0)
-	if errno != 0 {
-		return 0, errno
-	}
-	return (uintptr)(unsafe.Pointer(ret)), nil
-}
-
-func getParentUserNs(fd uintptr) (uintptr, error) {
-	const nsGetParent = 0xb702
-	ret, _, errno := syscall.Syscall(syscall.SYS_IOCTL, fd, uintptr(nsGetParent), 0)
-	if errno != 0 {
-		return 0, errno
-	}
-	return (uintptr)(unsafe.Pointer(ret)), nil
-}
-
-func getUserNSForPath(path string) (*os.File, error) {
-	u, err := os.Open(path)
-	if err != nil {
-		return nil, errors.Wrapf(err, "cannot open %s", path)
-	}
-	defer u.Close()
-	fd, err := getOwner(u.Fd())
-	if err != nil {
-		return nil, err
-	}
-
-	return getUserNSFirstChild(fd)
-}
-
-func getUserNSForPid(pid uint) (*os.File, error) {
-	path := fmt.Sprintf("/proc/%d/ns/user", pid)
-	u, err := os.Open(path)
-	if err != nil {
-		return nil, errors.Wrapf(err, "cannot open %s", path)
-	}
-
-	return getUserNSFirstChild(u.Fd())
-}
-
-// getUserNSFirstChild returns an open FD for the first direct child user namespace that created the process
-// Each container creates a new user namespace where the runtime runs.  The current process in the container
-// might have created new user namespaces that are child of the initial namespace we created.
-// This function finds the initial namespace created for the container that is a child of the current namespace.
-//
-//                                     current ns
-//                                       /     \
-//                           TARGET ->  a   [other containers]
-//                                     /
-//                                    b
-//                                   /
-//        NS READ USING THE PID ->  c
-func getUserNSFirstChild(fd uintptr) (*os.File, error) {
-	currentNS, err := readUserNs("/proc/self/ns/user")
-	if err != nil {
-		return nil, err
-	}
-
-	ns, err := readUserNsFd(fd)
-	if err != nil {
-		return nil, errors.Wrapf(err, "cannot read user namespace")
-	}
-	if ns == currentNS {
-		return nil, errors.New("process running in the same user namespace")
-	}
-
-	for {
-		nextFd, err := getParentUserNs(fd)
-		if err != nil {
-			return nil, errors.Wrapf(err, "cannot get parent user namespace")
-		}
-
-		ns, err = readUserNsFd(nextFd)
-		if err != nil {
-			return nil, errors.Wrapf(err, "cannot read user namespace")
-		}
-
-		if ns == currentNS {
-			syscall.Close(int(nextFd))
-
-			// Drop O_CLOEXEC for the fd.
-			_, _, errno := syscall.Syscall(syscall.SYS_FCNTL, fd, syscall.F_SETFD, 0)
-			if errno != 0 {
-				syscall.Close(int(fd))
-				return nil, errno
-			}
-
-			return os.NewFile(fd, "userns child"), nil
-		}
-		syscall.Close(int(fd))
-		fd = nextFd
-	}
-}
diff --git a/pkg/rootless/rootless_unsupported.go b/pkg/rootless/rootless_unsupported.go
index 24009610a..47b5dd7cc 100644
--- a/pkg/rootless/rootless_unsupported.go
+++ b/pkg/rootless/rootless_unsupported.go
@@ -19,45 +19,15 @@ func BecomeRootInUserNS() (bool, int, error) {
 	return false, -1, errors.New("this function is not supported on this os")
 }
 
-// BecomeRootInUserNS is a stub function that always returns false and an
-// error on unsupported OS's
-func BecomeRootInUserNSWithOpts(opts *Opts) (bool, int, error) {
-	return false, -1, errors.New("this function is not supported on this os")
-}
-
 // GetRootlessUID returns the UID of the user in the parent userNS
 func GetRootlessUID() int {
 	return -1
 }
 
-// JoinNS re-exec podman in a new userNS and join the user namespace of the specified
-// PID.
-func JoinNS(pid uint, preserveFDs int) (bool, int, error) {
-	return false, -1, errors.New("this function is not supported on this os")
-}
-
-// JoinNSPath re-exec podman in a new userNS and join the owner user namespace of the
-// specified path.
-func JoinNSPath(path string) (bool, int, error) {
-	return false, -1, errors.New("this function is not supported on this os")
-}
-
-// JoinDirectUserAndMountNSWithOpts re-exec podman in a new userNS and join the user and
-// mount namespace of the specified PID without looking up its parent.  Useful to join
-// directly the conmon process.
-func JoinDirectUserAndMountNSWithOpts(pid uint, opts *Opts) (bool, int, error) {
-	return false, -1, errors.New("this function is not supported on this os")
-}
-
-// JoinDirectUserAndMountNS re-exec podman in a new userNS and join the user and mount
+// JoinUserAndMountNS re-exec podman in a new userNS and join the user and mount
 // namespace of the specified PID without looking up its parent.  Useful to join directly
-// the conmon process.  It is a convenience function for JoinDirectUserAndMountNSWithOpts
+// the conmon process.  It is a convenience function for JoinUserAndMountNSWithOpts
 // with a default configuration.
-func JoinDirectUserAndMountNS(pid uint) (bool, int, error) {
+func JoinUserAndMountNS(pid uint) (bool, int, error) {
 	return false, -1, errors.New("this function is not supported on this os")
 }
-
-// Argument returns the argument that was set for the rootless session.
-func Argument() string {
-	return ""
-}