nspawn: optionally fix up OS tree uid/gids for userns

This adds a new --private-userns-chown switch that may be used in combination
with --private-userns. If it is passed a recursive chmod() operation is run on
the OS tree, fixing all file owner UID/GIDs to the right ranges. This should
make user namespacing pretty workable, as the OS trees don't need to be
prepared manually anymore.

6 files changed, 560 insertions, 2 deletions

diff --git a/.gitignore b/.gitignore
index 02ba86ef6f..c7eb14452d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -240,6 +240,7 @@
 /test-ns
 /test-nss
 /test-parse-util
+/test-patch-uid
 /test-path
 /test-path-lookup
 /test-path-util
diff --git a/Makefile.am b/Makefile.am
index 0f475c6d09..b323de55c6 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -3021,6 +3021,8 @@ systemd_nspawn_SOURCES = \
 	src/nspawn/nspawn-setuid.h \
 	src/nspawn/nspawn-stub-pid1.c \
 	src/nspawn/nspawn-stub-pid1.h \
+	src/nspawn/nspawn-patch-uid.c \
+	src/nspawn/nspawn-patch-uid.h \
 	src/core/mount-setup.c \
 	src/core/mount-setup.h \
 	src/core/loopback-setup.c \
@@ -3048,6 +3050,17 @@ systemd_nspawn_LDADD += \
 	libfirewall.la
 endif
 
+test_patch_uid_SOURCES = \
+	src/nspawn/nspawn-patch-uid.c \
+	src/nspawn/nspawn-patch-uid.h \
+	src/nspawn/test-patch-uid.c
+
+test_patch_uid_LDADD = \
+	libshared.la
+
+manual_tests += \
+	test-patch-uid
+
 # ------------------------------------------------------------------------------
 systemd_run_SOURCES = \
 	src/run/run.c
diff --git a/src/nspawn/nspawn-patch-uid.c b/src/nspawn/nspawn-patch-uid.c
new file mode 100644
index 0000000000..f53164fbbf
--- /dev/null
+++ b/src/nspawn/nspawn-patch-uid.c
@@ -0,0 +1,417 @@
+/***
+  This file is part of systemd.
+
+  Copyright 2016 Lennart Poettering
+
+  systemd is free software; you can redistribute it and/or modify it
+  under the terms of the GNU Lesser General Public License as published by
+  the Free Software Foundation; either version 2.1 of the License, or
+  (at your option) any later version.
+
+  systemd is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+  Lesser General Public License for more details.
+
+  You should have received a copy of the GNU Lesser General Public License
+  along with systemd; If not, see <http://www.gnu.org/licenses/>.
+***/
+
+#include <fcntl.h>
+#ifdef HAVE_ACL
+#include <sys/acl.h>
+#endif
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "acl-util.h"
+#include "dirent-util.h"
+#include "fd-util.h"
+#include "nspawn-patch-uid.h"
+#include "stdio-util.h"
+#include "string-util.h"
+#include "strv.h"
+#include "user-util.h"
+
+#ifdef HAVE_ACL
+
+static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
+        char procfs_path[strlen("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
+        acl_t acl;
+
+        assert(fd >= 0);
+        assert(ret);
+
+        if (name) {
+                _cleanup_close_ int child_fd = -1;
+
+                child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
+                if (child_fd < 0)
+                        return -errno;
+
+                xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
+                acl = acl_get_file(procfs_path, type);
+        } else if (type == ACL_TYPE_ACCESS)
+                acl = acl_get_fd(fd);
+        else {
+                xsprintf(procfs_path, "/proc/self/fd/%i", fd);
+                acl = acl_get_file(procfs_path, type);
+        }
+        if (!acl)
+                return -errno;
+
+        *ret = acl;
+        return 0;
+}
+
+static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
+        char procfs_path[strlen("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
+        int r;
+
+        assert(fd >= 0);
+        assert(acl);
+
+        if (name) {
+                _cleanup_close_ int child_fd = -1;
+
+                child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
+                if (child_fd < 0)
+                        return -errno;
+
+                xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
+                r = acl_set_file(procfs_path, type, acl);
+        } else if (type == ACL_TYPE_ACCESS)
+                r = acl_set_fd(fd, acl);
+        else {
+                xsprintf(procfs_path, "/proc/self/fd/%i", fd);
+                r = acl_set_file(procfs_path, type, acl);
+        }
+        if (r < 0)
+                return -errno;
+
+        return 0;
+}
+
+static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
+        _cleanup_(acl_freep) acl_t copy = NULL;
+        acl_entry_t i;
+        int r;
+
+        assert(acl);
+        assert(ret);
+
+        r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
+        if (r < 0)
+                return -errno;
+        while (r > 0) {
+                uid_t *old_uid, new_uid;
+                bool modify = false;
+                acl_tag_t tag;
+
+                if (acl_get_tag_type(i, &tag) < 0)
+                        return -errno;
+
+                if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
+
+                        /* We don't distuingish here between uid_t and gid_t, let's make sure the compiler checks that
+                         * this is actually OK */
+                        assert_cc(sizeof(uid_t) == sizeof(gid_t));
+
+                        old_uid = acl_get_qualifier(i);
+                        if (!old_uid)
+                                return -errno;
+
+                        new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
+                        if (!uid_is_valid(new_uid))
+                                return -EINVAL;
+
+                        modify = new_uid != *old_uid;
+                        if (modify && !copy) {
+                                int n;
+
+                                /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
+                                 * beginning, so that we copy all entries, starting from the first, this time. */
+
+                                n = acl_entries(acl);
+                                if (n < 0)
+                                        return -errno;
+
+                                copy = acl_init(n);
+                                if (!copy)
+                                        return -errno;
+
+                                /* Seek back to the beginning */
+                                r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
+                                if (r < 0)
+                                        return -errno;
+                                continue;
+                        }
+                }
+
+                if (copy) {
+                        acl_entry_t new_entry;
+
+                        if (acl_create_entry(&copy, &new_entry) < 0)
+                                return -errno;
+
+                        if (acl_copy_entry(new_entry, i) < 0)
+                                return -errno;
+
+                        if (modify)
+                                if (acl_set_qualifier(new_entry, &new_uid) < 0)
+                                        return -errno;
+                }
+
+                r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
+                if (r < 0)
+                        return -errno;
+        }
+
+        *ret = copy;
+        copy = NULL;
+
+        return !!*ret;
+}
+
+static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
+        _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
+        bool changed = false;
+        int r;
+
+        assert(fd >= 0);
+        assert(st);
+
+        /* ACLs are not supported on symlinks, there's no point in trying */
+        if (S_ISLNK(st->st_mode))
+                return 0;
+
+        r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
+        if (r == -EOPNOTSUPP)
+                return 0;
+        if (r < 0)
+                return r;
+
+        r = shift_acl(acl, shift, &shifted);
+        if (r < 0)
+                return r;
+        if (r > 0) {
+                r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
+                if (r < 0)
+                        return r;
+
+                changed = true;
+        }
+
+        if (S_ISDIR(st->st_mode)) {
+                acl_free(acl);
+                acl_free(shifted);
+
+                acl = shifted = NULL;
+
+                r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
+                if (r < 0)
+                        return r;
+
+                r = shift_acl(acl, shift, &shifted);
+                if (r < 0)
+                        return r;
+                if (r > 0) {
+                        r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
+                        if (r < 0)
+                                return r;
+
+                        changed = true;
+                }
+        }
+
+        return changed;
+}
+
+#else
+
+static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
+        return 0;
+}
+
+#endif
+
+static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
+        uid_t new_uid;
+        gid_t new_gid;
+        bool changed = false;
+        int r;
+
+        assert(fd >= 0);
+        assert(st);
+
+        new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
+        new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
+
+        if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
+                return -EINVAL;
+
+        if (st->st_uid != new_uid || st->st_gid != new_gid) {
+                if (name)
+                        r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
+                else
+                        r = fchown(fd, new_uid, new_gid);
+                if (r < 0)
+                        return -errno;
+
+                /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
+                if (name && !S_ISLNK(st->st_mode))
+                        r = fchmodat(fd, name, st->st_mode, 0);
+                else
+                        r = fchmod(fd, st->st_mode);
+                if (r < 0)
+                        return -errno;
+
+                changed = true;
+        }
+
+        r = patch_acls(fd, name, st, shift);
+        if (r < 0)
+                return r;
+
+        return r > 0 || changed;
+}
+
+static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift) {
+        bool changed = false;
+        int r;
+
+        assert(fd >= 0);
+
+        r = patch_fd(fd, NULL, st, shift);
+        if (r < 0)
+                goto finish;
+
+        if (S_ISDIR(st->st_mode)) {
+                _cleanup_closedir_ DIR *d = NULL;
+                struct dirent *de;
+
+                if (!donate_fd) {
+                        int copy;
+
+                        copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
+                        if (copy < 0)
+                                return -errno;
+
+                        fd = copy;
+                        donate_fd = true;
+                }
+
+                d = fdopendir(fd);
+                if (!d) {
+                        r = -errno;
+                        goto finish;
+                }
+                fd = -1;
+
+                FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
+                        struct stat fst;
+
+                        if (STR_IN_SET(de->d_name, ".", ".."))
+                                continue;
+
+                        if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
+                                r = -errno;
+                                goto finish;
+                        }
+
+                        if (S_ISDIR(fst.st_mode)) {
+                                int subdir_fd;
+
+                                subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
+                                if (subdir_fd < 0) {
+                                        r = -errno;
+                                        goto finish;
+
+                                }
+
+                                r = recurse_fd(subdir_fd, true, &fst, shift);
+                                if (r < 0)
+                                        goto finish;
+                                if (r > 0)
+                                        changed = true;
+
+                        } else {
+                                r = patch_fd(dirfd(d), de->d_name, &fst, shift);
+                                if (r < 0)
+                                        goto finish;
+                                if (r > 0)
+                                        changed = true;
+                        }
+                }
+        }
+
+        r = changed;
+
+finish:
+        if (donate_fd)
+                safe_close(fd);
+
+        return r;
+}
+
+static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
+        struct stat st;
+        int r;
+
+        assert(fd >= 0);
+
+        /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
+         * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
+         * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
+         * UID within the container. */
+
+        if ((shift & 0xFFFF) != 0) {
+                /* We only support containers where the shift starts at a 2^16 boundary */
+                r = -EOPNOTSUPP;
+                goto finish;
+        }
+
+        if (range != 0x10000) {
+                /* We only support containers with 16bit UID ranges for the patching logic */
+                r = -EOPNOTSUPP;
+                goto finish;
+        }
+
+        if (fstat(fd, &st) < 0) {
+                r = -errno;
+                goto finish;
+        }
+
+        if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
+                /* We only support containers where the uid/gid container ID match */
+                r = -EBADE;
+                goto finish;
+        }
+
+        /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
+         * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
+        if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
+                return 0;
+
+        return recurse_fd(fd, donate_fd, &st, shift);
+
+finish:
+        if (donate_fd)
+                safe_close(fd);
+
+        return r;
+}
+
+int fd_patch_uid(int fd, uid_t shift, uid_t range) {
+        return fd_patch_uid_internal(fd, false, shift, range);
+}
+
+int path_patch_uid(const char *path, uid_t shift, uid_t range) {
+        int fd;
+
+        fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
+        if (fd < 0)
+                return -errno;
+
+        return fd_patch_uid_internal(fd, true, shift, range);
+}
diff --git a/src/nspawn/nspawn-patch-uid.h b/src/nspawn/nspawn-patch-uid.h
new file mode 100644
index 0000000000..55d0990016
--- /dev/null
+++ b/src/nspawn/nspawn-patch-uid.h
@@ -0,0 +1,23 @@
+/***
+  This file is part of systemd.
+
+  Copyright 2016 Lennart Poettering
+
+  systemd is free software; you can redistribute it and/or modify it
+  under the terms of the GNU Lesser General Public License as published by
+  the Free Software Foundation; either version 2.1 of the License, or
+  (at your option) any later version.
+
+  systemd is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+  Lesser General Public License for more details.
+
+  You should have received a copy of the GNU Lesser General Public License
+  along with systemd; If not, see <http://www.gnu.org/licenses/>.
+***/
+
+#include <sys/types.h>
+
+int fd_patch_uid(int fd, uid_t shift, uid_t range);
+int path_patch_uid(const char *path, uid_t shift, uid_t range);
diff --git a/src/nspawn/nspawn.c b/src/nspawn/nspawn.c
index e1d37d383a..a3190545fa 100644
--- a/src/nspawn/nspawn.c
+++ b/src/nspawn/nspawn.c
@@ -75,6 +75,7 @@
 #include "nspawn-expose-ports.h"
 #include "nspawn-mount.h"
 #include "nspawn-network.h"
+#include "nspawn-patch-uid.h"
 #include "nspawn-register.h"
 #include "nspawn-settings.h"
 #include "nspawn-setuid.h"
@@ -175,6 +176,7 @@ static ExposePort *arg_expose_ports = NULL;
 static char **arg_property = NULL;
 static uid_t arg_uid_shift = UID_INVALID, arg_uid_range = 0x10000U;
 static bool arg_userns = false;
+static bool arg_userns_chown = false;
 static int arg_kill_signal = 0;
 static bool arg_unified_cgroup_hierarchy = false;
 static SettingsMask arg_settings_mask = 0;
@@ -204,6 +206,7 @@ static void help(void) {
                "     --property=NAME=VALUE  Set scope unit property\n"
                "     --private-users[=UIDBASE[:NUIDS]]\n"
                "                            Run within user namespace\n"
+               "     --private-user-chown   Adjust OS tree file ownership for private user range\n"
                "     --private-network      Disable network in container\n"
                "     --network-interface=INTERFACE\n"
                "                            Assign an existing network interface to the\n"
@@ -349,6 +352,7 @@ static int parse_argv(int argc, char *argv[]) {
                 ARG_KILL_SIGNAL,
                 ARG_SETTINGS,
                 ARG_CHDIR,
+                ARG_PRIVATE_USERS_CHOWN,
         };
 
         static const struct option options[] = {
@@ -392,6 +396,7 @@ static int parse_argv(int argc, char *argv[]) {
                 { "port",                  required_argument, NULL, 'p'                   },
                 { "property",              required_argument, NULL, ARG_PROPERTY          },
                 { "private-users",         optional_argument, NULL, ARG_PRIVATE_USERS     },
+                { "private-users-chown",   optional_argument, NULL, ARG_PRIVATE_USERS_CHOWN},
                 { "kill-signal",           required_argument, NULL, ARG_KILL_SIGNAL       },
                 { "settings",              required_argument, NULL, ARG_SETTINGS          },
                 { "chdir",                 required_argument, NULL, ARG_CHDIR             },
@@ -825,6 +830,10 @@ static int parse_argv(int argc, char *argv[]) {
                         arg_userns = true;
                         break;
 
+                case ARG_PRIVATE_USERS_CHOWN:
+                        arg_userns_chown = true;
+                        break;
+
                 case ARG_KILL_SIGNAL:
                         arg_kill_signal = signal_from_string_try_harder(optarg);
                         if (arg_kill_signal < 0) {
@@ -933,8 +942,15 @@ static int parse_argv(int argc, char *argv[]) {
                 return -EINVAL;
         }
 
-        if (arg_userns && access("/proc/self/uid_map", F_OK) < 0)
-                return log_error_errno(EOPNOTSUPP, "--private-users= is not supported, kernel compiled without user namespace support.");
+        if (arg_userns && access("/proc/self/uid_map", F_OK) < 0) {
+                log_error("--private-users= is not supported, kernel compiled without user namespace support.");
+                return -EOPNOTSUPP;
+        }
+
+        if (arg_userns_chown && arg_read_only) {
+                log_error("--read-only and --private-users-chown may not be combined.");
+                return -EINVAL;
+        }
 
         if (argc > optind) {
                 arg_parameters = strv_copy(argv + optind);
@@ -2218,6 +2234,29 @@ static int setup_machine_id(const char *directory) {
         return 0;
 }
 
+static int recursive_chown(const char *directory, uid_t shift, uid_t range) {
+        int r;
+
+        assert(directory);
+
+        if (!arg_userns || !arg_userns_chown)
+                return 0;
+
+        r = path_patch_uid(directory, arg_uid_shift, arg_uid_range);
+        if (r == -EOPNOTSUPP)
+                return log_error_errno(r, "Automatic UID/GID adjusting is only supported for UID/GID ranges starting at multiples of 2^16 with a range of 2^16.");
+        if (r == -EBADE)
+                return log_error_errno(r, "Upper 16 bits of root directory UID and GID do not match.");
+        if (r < 0)
+                return log_error_errno(r, "Failed to adjust UID/GID shift of OS tree: %m");
+        if (r == 0)
+                log_debug("Root directory of image is already owned by the right UID/GID range, skipping recursive chown operation.");
+        else
+                log_debug("Patched directory tree to match UID/GID range.");
+
+        return r;
+}
+
 static int mount_devices(
                 const char *where,
                 const char *root_device, bool root_device_rw,
@@ -2763,6 +2802,10 @@ static int outer_child(
         if (mount(directory, directory, NULL, MS_BIND|MS_REC, NULL) < 0)
                 return log_error_errno(errno, "Failed to make bind mount: %m");
 
+        r = recursive_chown(directory, arg_uid_shift, arg_uid_range);
+        if (r < 0)
+                return r;
+
         r = setup_volatile(directory, arg_volatile_mode, arg_userns, arg_uid_shift, arg_uid_range, arg_selinux_context);
         if (r < 0)
                 return r;
diff --git a/src/nspawn/test-patch-uid.c b/src/nspawn/test-patch-uid.c
new file mode 100644
index 0000000000..11c5321788
--- /dev/null
+++ b/src/nspawn/test-patch-uid.c
@@ -0,0 +1,61 @@
+/***
+  This file is part of systemd.
+
+  Copyright 2016 Lennart Poettering
+
+  systemd is free software; you can redistribute it and/or modify it
+  under the terms of the GNU Lesser General Public License as published by
+  the Free Software Foundation; either version 2.1 of the License, or
+  (at your option) any later version.
+
+  systemd is distributed in the hope that it will be useful, but
+  WITHOUT ANY WARRANTY; without even the implied warranty of
+  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+  Lesser General Public License for more details.
+
+  You should have received a copy of the GNU Lesser General Public License
+  along with systemd; If not, see <http://www.gnu.org/licenses/>.
+***/
+
+#include <stdlib.h>
+
+#include "log.h"
+#include "nspawn-patch-uid.h"
+#include "user-util.h"
+#include "util.h"
+
+int main(int argc, char *argv[]) {
+        uid_t shift, range;
+        int r;
+
+        log_set_max_level(LOG_DEBUG);
+        log_parse_environment();
+        log_open();
+
+        if (argc != 4) {
+                log_error("Expected PATH SHIFT RANGE parameters.");
+                return EXIT_FAILURE;
+        }
+
+        r = parse_uid(argv[2], &shift);
+        if (r < 0) {
+                log_error_errno(r, "Failed to parse UID shift %s.", argv[2]);
+                return EXIT_FAILURE;
+        }
+
+        r = parse_gid(argv[3], &range);
+        if (r < 0) {
+                log_error_errno(r, "Failed to parse UID range %s.", argv[3]);
+                return EXIT_FAILURE;
+        }
+
+        r = path_patch_uid(argv[1], shift, range);
+        if (r < 0) {
+                log_error_errno(r, "Failed to patch directory tree: %m");
+                return EXIT_FAILURE;
+        }
+
+        log_info("Changed: %s", yes_no(r));
+
+        return EXIT_SUCCESS;
+}