/* SPDX-License-Identifier: LGPL-2.1+ */ #include #include #include #include #include #include "clean-ipc.h" #include "dynamic-user.h" #include "fd-util.h" #include "fileio.h" #include "format-util.h" #include "fs-util.h" #include "io-util.h" #include "nscd-flush.h" #include "parse-util.h" #include "random-util.h" #include "serialize.h" #include "socket-util.h" #include "stdio-util.h" #include "string-util.h" #include "user-util.h" /* Takes a value generated randomly or by hashing and turns it into a UID in the right range */ #define UID_CLAMP_INTO_RANGE(rnd) (((uid_t) (rnd) % (DYNAMIC_UID_MAX - DYNAMIC_UID_MIN + 1)) + DYNAMIC_UID_MIN) DEFINE_PRIVATE_TRIVIAL_REF_FUNC(DynamicUser, dynamic_user); static DynamicUser* dynamic_user_free(DynamicUser *d) { if (!d) return NULL; if (d->manager) (void) hashmap_remove(d->manager->dynamic_users, d->name); safe_close_pair(d->storage_socket); return mfree(d); } static int dynamic_user_add(Manager *m, const char *name, int storage_socket[static 2], DynamicUser **ret) { DynamicUser *d; int r; assert(m); assert(name); assert(storage_socket); r = hashmap_ensure_allocated(&m->dynamic_users, &string_hash_ops); if (r < 0) return r; d = malloc0(offsetof(DynamicUser, name) + strlen(name) + 1); if (!d) return -ENOMEM; strcpy(d->name, name); d->storage_socket[0] = storage_socket[0]; d->storage_socket[1] = storage_socket[1]; r = hashmap_put(m->dynamic_users, d->name, d); if (r < 0) { free(d); return r; } d->manager = m; if (ret) *ret = d; return 0; } static int dynamic_user_acquire(Manager *m, const char *name, DynamicUser** ret) { _cleanup_close_pair_ int storage_socket[2] = { -1, -1 }; DynamicUser *d; int r; assert(m); assert(name); /* Return the DynamicUser structure for a specific user name. Note that this won't actually allocate a UID for * it, but just prepare the data structure for it. The UID is allocated only on demand, when it's really * needed, and in the child process we fork off, since allocation involves NSS checks which are not OK to do * from PID 1. To allow the children and PID 1 share information about allocated UIDs we use an anonymous * AF_UNIX/SOCK_DGRAM socket (called the "storage socket") that contains at most one datagram with the * allocated UID number, plus an fd referencing the lock file for the UID * (i.e. /run/systemd/dynamic-uid/$UID). Why involve the socket pair? So that PID 1 and all its children can * share the same storage for the UID and lock fd, simply by inheriting the storage socket fds. The socket pair * may exist in three different states: * * a) no datagram stored. This is the initial state. In this case the dynamic user was never realized. * * b) a datagram containing a UID stored, but no lock fd attached to it. In this case there was already a * statically assigned UID by the same name, which we are reusing. * * c) a datagram containing a UID stored, and a lock fd is attached to it. In this case we allocated a dynamic * UID and locked it in the file system, using the lock fd. * * As PID 1 and various children might access the socket pair simultaneously, and pop the datagram or push it * back in any time, we also maintain a lock on the socket pair. Note one peculiarity regarding locking here: * the UID lock on disk is protected via a BSD file lock (i.e. an fd-bound lock), so that the lock is kept in * place as long as there's a reference to the fd open. The lock on the storage socket pair however is a POSIX * file lock (i.e. a process-bound lock), as all users share the same fd of this (after all it is anonymous, * nobody else could get any access to it except via our own fd) and we want to synchronize access between all * processes that have access to it. */ d = hashmap_get(m->dynamic_users, name); if (d) { if (ret) { /* We already have a structure for the dynamic user, let's increase the ref count and reuse it */ d->n_ref++; *ret = d; } return 0; } if (!valid_user_group_name_or_id(name)) return -EINVAL; if (socketpair(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0, storage_socket) < 0) return -errno; r = dynamic_user_add(m, name, storage_socket, &d); if (r < 0) return r; storage_socket[0] = storage_socket[1] = -1; if (ret) { d->n_ref++; *ret = d; } return 1; } static int make_uid_symlinks(uid_t uid, const char *name, bool b) { char path1[STRLEN("/run/systemd/dynamic-uid/direct:") + DECIMAL_STR_MAX(uid_t) + 1]; const char *path2; int r = 0, k; /* Add direct additional symlinks for direct lookups of dynamic UIDs and their names by userspace code. The * only reason we have this is because dbus-daemon cannot use D-Bus for resolving users and groups (since it * would be its own client then). We hence keep these world-readable symlinks in place, so that the * unprivileged dbus user can read the mappings when it needs them via these symlinks instead of having to go * via the bus. Ideally, we'd use the lock files we keep for this anyway, but we can't since we use BSD locks * on them and as those may be taken by any user with read access we can't make them world-readable. */ xsprintf(path1, "/run/systemd/dynamic-uid/direct:" UID_FMT, uid); if (unlink(path1) < 0 && errno != ENOENT) r = -errno; if (b && symlink(name, path1) < 0) { k = log_warning_errno(errno, "Failed to symlink \"%s\": %m", path1); if (r == 0) r = k; } path2 = strjoina("/run/systemd/dynamic-uid/direct:", name); if (unlink(path2) < 0 && errno != ENOENT) { k = -errno; if (r == 0) r = k; } if (b && symlink(path1 + STRLEN("/run/systemd/dynamic-uid/direct:"), path2) < 0) { k = log_warning_errno(errno, "Failed to symlink \"%s\": %m", path2); if (r == 0) r = k; } return r; } static int pick_uid(char **suggested_paths, const char *name, uid_t *ret_uid) { /* Find a suitable free UID. We use the following strategy to find a suitable UID: * * 1. Initially, we try to read the UID of a number of specified paths. If any of these UIDs works, we use * them. We use in order to increase the chance of UID reuse, if StateDirectory=, CacheDirectory= or * LogsDirectory= are used, as reusing the UID these directories are owned by saves us from having to * recursively chown() them to new users. * * 2. If that didn't yield a currently unused UID, we hash the user name, and try to use that. This should be * pretty good, as the use ris by default derived from the unit name, and hence the same service and same * user should usually get the same UID as long as our hashing doesn't clash. * * 3. Finally, if that didn't work, we randomly pick UIDs, until we find one that is empty. * * Since the dynamic UID space is relatively small we'll stop trying after 100 iterations, giving up. */ enum { PHASE_SUGGESTED, /* the first phase, reusing directory ownership UIDs */ PHASE_HASHED, /* the second phase, deriving a UID from the username by hashing */ PHASE_RANDOM, /* the last phase, randomly picking UIDs */ } phase = PHASE_SUGGESTED; static const uint8_t hash_key[] = { 0x37, 0x53, 0x7e, 0x31, 0xcf, 0xce, 0x48, 0xf5, 0x8a, 0xbb, 0x39, 0x57, 0x8d, 0xd9, 0xec, 0x59 }; unsigned n_tries = 100, current_suggested = 0; int r; (void) mkdir("/run/systemd/dynamic-uid", 0755); for (;;) { char lock_path[STRLEN("/run/systemd/dynamic-uid/") + DECIMAL_STR_MAX(uid_t) + 1]; _cleanup_close_ int lock_fd = -1; uid_t candidate; ssize_t l; if (--n_tries <= 0) /* Give up retrying eventually */ return -EBUSY; switch (phase) { case PHASE_SUGGESTED: { struct stat st; if (!suggested_paths || !suggested_paths[current_suggested]) { /* We reached the end of the suggested paths list, let's try by hashing the name */ phase = PHASE_HASHED; continue; } if (stat(suggested_paths[current_suggested++], &st) < 0) continue; /* We can't read the UID of this path, but that doesn't matter, just try the next */ candidate = st.st_uid; break; } case PHASE_HASHED: /* A static user by this name does not exist yet. Let's find a free ID then, and use that. We * start with a UID generated as hash from the user name. */ candidate = UID_CLAMP_INTO_RANGE(siphash24(name, strlen(name), hash_key)); /* If this one fails, we should proceed with random tries */ phase = PHASE_RANDOM; break; case PHASE_RANDOM: /* Pick another random UID, and see if that works for us. */ random_bytes(&candidate, sizeof(candidate)); candidate = UID_CLAMP_INTO_RANGE(candidate); break; default: assert_not_reached("unknown phase"); } /* Make sure whatever we picked here actually is in the right range */ if (!uid_is_dynamic(candidate)) continue; xsprintf(lock_path, "/run/systemd/dynamic-uid/" UID_FMT, candidate); for (;;) { struct stat st; lock_fd = open(lock_path, O_CREAT|O_RDWR|O_NOFOLLOW|O_CLOEXEC|O_NOCTTY, 0600); if (lock_fd < 0) return -errno; r = flock(lock_fd, LOCK_EX|LOCK_NB); /* Try to get a BSD file lock on the UID lock file */ if (r < 0) { if (IN_SET(errno, EBUSY, EAGAIN)) goto next; /* already in use */ return -errno; } if (fstat(lock_fd, &st) < 0) return -errno; if (st.st_nlink > 0) break; /* Oh, bummer, we got the lock, but the file was unlinked between the time we opened it and * got the lock. Close it, and try again. */ lock_fd = safe_close(lock_fd); } /* Some superficial check whether this UID/GID might already be taken by some static user */ if (getpwuid(candidate) || getgrgid((gid_t) candidate) || search_ipc(candidate, (gid_t) candidate) != 0) { (void) unlink(lock_path); continue; } /* Let's store the user name in the lock file, so that we can use it for looking up the username for a UID */ l = pwritev(lock_fd, (struct iovec[2]) { IOVEC_INIT_STRING(name), IOVEC_INIT((char[1]) { '\n' }, 1), }, 2, 0); if (l < 0) { r = -errno; (void) unlink(lock_path); return r; } (void) ftruncate(lock_fd, l); (void) make_uid_symlinks(candidate, name, true); /* also add direct lookup symlinks */ *ret_uid = candidate; return TAKE_FD(lock_fd); next: ; } } static int dynamic_user_pop(DynamicUser *d, uid_t *ret_uid, int *ret_lock_fd) { uid_t uid = UID_INVALID; struct iovec iov = IOVEC_INIT(&uid, sizeof(uid)); int lock_fd; ssize_t k; assert(d); assert(ret_uid); assert(ret_lock_fd); /* Read the UID and lock fd that is stored in the storage AF_UNIX socket. This should be called with the lock * on the socket taken. */ k = receive_one_fd_iov(d->storage_socket[0], &iov, 1, MSG_DONTWAIT, &lock_fd); if (k < 0) return (int) k; *ret_uid = uid; *ret_lock_fd = lock_fd; return 0; } static int dynamic_user_push(DynamicUser *d, uid_t uid, int lock_fd) { struct iovec iov = IOVEC_INIT(&uid, sizeof(uid)); assert(d); /* Store the UID and lock_fd in the storage socket. This should be called with the socket pair lock taken. */ return send_one_fd_iov(d->storage_socket[1], lock_fd, &iov, 1, MSG_DONTWAIT); } static void unlink_uid_lock(int lock_fd, uid_t uid, const char *name) { char lock_path[STRLEN("/run/systemd/dynamic-uid/") + DECIMAL_STR_MAX(uid_t) + 1]; if (lock_fd < 0) return; xsprintf(lock_path, "/run/systemd/dynamic-uid/" UID_FMT, uid); (void) unlink(lock_path); (void) make_uid_symlinks(uid, name, false); /* remove direct lookup symlinks */ } static int lockfp(int fd, int *fd_lock) { if (lockf(fd, F_LOCK, 0) < 0) return -errno; *fd_lock = fd; return 0; } static void unlockfp(int *fd_lock) { if (*fd_lock < 0) return; lockf(*fd_lock, F_ULOCK, 0); *fd_lock = -1; } static int dynamic_user_realize( DynamicUser *d, char **suggested_dirs, uid_t *ret_uid, gid_t *ret_gid, bool is_user) { _cleanup_(unlockfp) int storage_socket0_lock = -1; _cleanup_close_ int uid_lock_fd = -1; _cleanup_close_ int etc_passwd_lock_fd = -1; uid_t num = UID_INVALID; /* a uid if is_user, and a gid otherwise */ gid_t gid = GID_INVALID; /* a gid if is_user, ignored otherwise */ bool flush_cache = false; int r; assert(d); assert(is_user == !!ret_uid); assert(ret_gid); /* Acquire a UID for the user name. This will allocate a UID for the user name if the user doesn't exist * yet. If it already exists its existing UID/GID will be reused. */ r = lockfp(d->storage_socket[0], &storage_socket0_lock); if (r < 0) return r; r = dynamic_user_pop(d, &num, &uid_lock_fd); if (r < 0) { int new_uid_lock_fd; uid_t new_uid; if (r != -EAGAIN) return r; /* OK, nothing stored yet, let's try to find something useful. While we are working on this release the * lock however, so that nobody else blocks on our NSS lookups. */ unlockfp(&storage_socket0_lock); /* Let's see if a proper, static user or group by this name exists. Try to take the lock on * /etc/passwd, if that fails with EROFS then /etc is read-only. In that case it's fine if we don't * take the lock, given that users can't be added there anyway in this case. */ etc_passwd_lock_fd = take_etc_passwd_lock(NULL); if (etc_passwd_lock_fd < 0 && etc_passwd_lock_fd != -EROFS) return etc_passwd_lock_fd; /* First, let's parse this as numeric UID */ r = parse_uid(d->name, &num); if (r < 0) { struct passwd *p; struct group *g; if (is_user) { /* OK, this is not a numeric UID. Let's see if there's a user by this name */ p = getpwnam(d->name); if (p) { num = p->pw_uid; gid = p->pw_gid; } else { /* if the user does not exist but the group with the same name exists, refuse operation */ g = getgrnam(d->name); if (g) return -EILSEQ; } } else { /* Let's see if there's a group by this name */ g = getgrnam(d->name); if (g) num = (uid_t) g->gr_gid; else { /* if the group does not exist but the user with the same name exists, refuse operation */ p = getpwnam(d->name); if (p) return -EILSEQ; } } } if (num == UID_INVALID) { /* No static UID assigned yet, excellent. Let's pick a new dynamic one, and lock it. */ uid_lock_fd = pick_uid(suggested_dirs, d->name, &num); if (uid_lock_fd < 0) return uid_lock_fd; } /* So, we found a working UID/lock combination. Let's see if we actually still need it. */ r = lockfp(d->storage_socket[0], &storage_socket0_lock); if (r < 0) { unlink_uid_lock(uid_lock_fd, num, d->name); return r; } r = dynamic_user_pop(d, &new_uid, &new_uid_lock_fd); if (r < 0) { if (r != -EAGAIN) { /* OK, something bad happened, let's get rid of the bits we acquired. */ unlink_uid_lock(uid_lock_fd, num, d->name); return r; } /* Great! Nothing is stored here, still. Store our newly acquired data. */ flush_cache = true; } else { /* Hmm, so as it appears there's now something stored in the storage socket. Throw away what we * acquired, and use what's stored now. */ unlink_uid_lock(uid_lock_fd, num, d->name); safe_close(uid_lock_fd); num = new_uid; uid_lock_fd = new_uid_lock_fd; } } else if (is_user && !uid_is_dynamic(num)) { struct passwd *p; /* Statically allocated user may have different uid and gid. So, let's obtain the gid. */ errno = 0; p = getpwuid(num); if (!p) return errno > 0 ? -errno : -ESRCH; gid = p->pw_gid; } /* If the UID/GID was already allocated dynamically, push the data we popped out back in. If it was already * allocated statically, push the UID back too, but do not push the lock fd in. If we allocated the UID * dynamically right here, push that in along with the lock fd for it. */ r = dynamic_user_push(d, num, uid_lock_fd); if (r < 0) return r; if (flush_cache) { /* If we allocated a new dynamic UID, refresh nscd, so that it forgets about potentially cached * negative entries. But let's do so after we release the /etc/passwd lock, so that there's no * potential for nscd wanting to lock that for completing the invalidation. */ etc_passwd_lock_fd = safe_close(etc_passwd_lock_fd); (void) nscd_flush_cache(STRV_MAKE("passwd", "group")); } if (is_user) { *ret_uid = num; *ret_gid = gid != GID_INVALID ? gid : num; } else *ret_gid = num; return 0; } int dynamic_user_current(DynamicUser *d, uid_t *ret) { _cleanup_(unlockfp) int storage_socket0_lock = -1; _cleanup_close_ int lock_fd = -1; uid_t uid; int r; assert(d); assert(ret); /* Get the currently assigned UID for the user, if there's any. This simply pops the data from the storage socket, and pushes it back in right-away. */ r = lockfp(d->storage_socket[0], &storage_socket0_lock); if (r < 0) return r; r = dynamic_user_pop(d, &uid, &lock_fd); if (r < 0) return r; r = dynamic_user_push(d, uid, lock_fd); if (r < 0) return r; *ret = uid; return 0; } static DynamicUser* dynamic_user_unref(DynamicUser *d) { if (!d) return NULL; /* Note that this doesn't actually release any resources itself. If a dynamic user should be fully destroyed * and its UID released, use dynamic_user_destroy() instead. NB: the dynamic user table may contain entries * with no references, which is commonly the case right before a daemon reload. */ assert(d->n_ref > 0); d->n_ref--; return NULL; } static int dynamic_user_close(DynamicUser *d) { _cleanup_(unlockfp) int storage_socket0_lock = -1; _cleanup_close_ int lock_fd = -1; uid_t uid; int r; /* Release the user ID, by releasing the lock on it, and emptying the storage socket. After this the user is * unrealized again, much like it was after it the DynamicUser object was first allocated. */ r = lockfp(d->storage_socket[0], &storage_socket0_lock); if (r < 0) return r; r = dynamic_user_pop(d, &uid, &lock_fd); if (r == -EAGAIN) /* User wasn't realized yet, nothing to do. */ return 0; if (r < 0) return r; /* This dynamic user was realized and dynamically allocated. In this case, let's remove the lock file. */ unlink_uid_lock(lock_fd, uid, d->name); (void) nscd_flush_cache(STRV_MAKE("passwd", "group")); return 1; } static DynamicUser* dynamic_user_destroy(DynamicUser *d) { if (!d) return NULL; /* Drop a reference to a DynamicUser object, and destroy the user completely if this was the last * reference. This is called whenever a service is shut down and wants its dynamic UID gone. Note that * dynamic_user_unref() is what is called whenever a service is simply freed, for example during a reload * cycle, where the dynamic users should not be destroyed, but our datastructures should. */ dynamic_user_unref(d); if (d->n_ref > 0) return NULL; (void) dynamic_user_close(d); return dynamic_user_free(d); } int dynamic_user_serialize(Manager *m, FILE *f, FDSet *fds) { DynamicUser *d; Iterator i; assert(m); assert(f); assert(fds); /* Dump the dynamic user database into the manager serialization, to deal with daemon reloads. */ HASHMAP_FOREACH(d, m->dynamic_users, i) { int copy0, copy1; copy0 = fdset_put_dup(fds, d->storage_socket[0]); if (copy0 < 0) return log_error_errno(copy0, "Failed to add dynamic user storage fd to serialization: %m"); copy1 = fdset_put_dup(fds, d->storage_socket[1]); if (copy1 < 0) return log_error_errno(copy1, "Failed to add dynamic user storage fd to serialization: %m"); (void) serialize_item_format(f, "dynamic-user", "%s %i %i", d->name, copy0, copy1); } return 0; } void dynamic_user_deserialize_one(Manager *m, const char *value, FDSet *fds) { _cleanup_free_ char *name = NULL, *s0 = NULL, *s1 = NULL; int r, fd0, fd1; assert(m); assert(value); assert(fds); /* Parse the serialization again, after a daemon reload */ r = extract_many_words(&value, NULL, 0, &name, &s0, &s1, NULL); if (r != 3 || !isempty(value)) { log_debug("Unable to parse dynamic user line."); return; } if (safe_atoi(s0, &fd0) < 0 || !fdset_contains(fds, fd0)) { log_debug("Unable to process dynamic user fd specification."); return; } if (safe_atoi(s1, &fd1) < 0 || !fdset_contains(fds, fd1)) { log_debug("Unable to process dynamic user fd specification."); return; } r = dynamic_user_add(m, name, (int[]) { fd0, fd1 }, NULL); if (r < 0) { log_debug_errno(r, "Failed to add dynamic user: %m"); return; } (void) fdset_remove(fds, fd0); (void) fdset_remove(fds, fd1); } void dynamic_user_vacuum(Manager *m, bool close_user) { DynamicUser *d; Iterator i; assert(m); /* Empty the dynamic user database, optionally cleaning up orphaned dynamic users, i.e. destroy and free users * to which no reference exist. This is called after a daemon reload finished, in order to destroy users which * might not be referenced anymore. */ HASHMAP_FOREACH(d, m->dynamic_users, i) { if (d->n_ref > 0) continue; if (close_user) { log_debug("Removing orphaned dynamic user %s", d->name); (void) dynamic_user_close(d); } dynamic_user_free(d); } } int dynamic_user_lookup_uid(Manager *m, uid_t uid, char **ret) { char lock_path[STRLEN("/run/systemd/dynamic-uid/") + DECIMAL_STR_MAX(uid_t) + 1]; _cleanup_free_ char *user = NULL; uid_t check_uid; int r; assert(m); assert(ret); /* A friendly way to translate a dynamic user's UID into a name. */ if (!uid_is_dynamic(uid)) return -ESRCH; xsprintf(lock_path, "/run/systemd/dynamic-uid/" UID_FMT, uid); r = read_one_line_file(lock_path, &user); if (IN_SET(r, -ENOENT, 0)) return -ESRCH; if (r < 0) return r; /* The lock file might be stale, hence let's verify the data before we return it */ r = dynamic_user_lookup_name(m, user, &check_uid); if (r < 0) return r; if (check_uid != uid) /* lock file doesn't match our own idea */ return -ESRCH; *ret = TAKE_PTR(user); return 0; } int dynamic_user_lookup_name(Manager *m, const char *name, uid_t *ret) { DynamicUser *d; int r; assert(m); assert(name); assert(ret); /* A friendly call for translating a dynamic user's name into its UID */ d = hashmap_get(m->dynamic_users, name); if (!d) return -ESRCH; r = dynamic_user_current(d, ret); if (r == -EAGAIN) /* not realized yet? */ return -ESRCH; return r; } int dynamic_creds_acquire(DynamicCreds *creds, Manager *m, const char *user, const char *group) { bool acquired = false; int r; assert(creds); assert(m); /* A DynamicUser object encapsulates an allocation of both a UID and a GID for a specific name. However, some * services use different user and groups. For cases like that there's DynamicCreds containing a pair of user * and group. This call allocates a pair. */ if (!creds->user && user) { r = dynamic_user_acquire(m, user, &creds->user); if (r < 0) return r; acquired = true; } if (!creds->group) { if (creds->user && (!group || streq_ptr(user, group))) creds->group = dynamic_user_ref(creds->user); else { r = dynamic_user_acquire(m, group, &creds->group); if (r < 0) { if (acquired) creds->user = dynamic_user_unref(creds->user); return r; } } } return 0; } int dynamic_creds_realize(DynamicCreds *creds, char **suggested_paths, uid_t *uid, gid_t *gid) { uid_t u = UID_INVALID; gid_t g = GID_INVALID; int r; assert(creds); assert(uid); assert(gid); /* Realize both the referenced user and group */ if (creds->user) { r = dynamic_user_realize(creds->user, suggested_paths, &u, &g, true); if (r < 0) return r; } if (creds->group && creds->group != creds->user) { r = dynamic_user_realize(creds->group, suggested_paths, NULL, &g, false); if (r < 0) return r; } *uid = u; *gid = g; return 0; } void dynamic_creds_unref(DynamicCreds *creds) { assert(creds); creds->user = dynamic_user_unref(creds->user); creds->group = dynamic_user_unref(creds->group); } void dynamic_creds_destroy(DynamicCreds *creds) { assert(creds); creds->user = dynamic_user_destroy(creds->user); creds->group = dynamic_user_destroy(creds->group); }