shared: add generic helper tools for installing files/dir trees

This adds a bit of generic helper tools for installing files/dir trees.
"installing" is supposed to mean the final step when preparing a disk
image or directory tree, where the result is renamed to its final name.
It has some bells and whistles, as it is able to replace existing files
sanely, can fsync() things carefully and can mark things read-only in a
nice way.

This is supposed to be generic, unified code that can be used eventually
for any of our tools that prepare disk images/directory trees, including
importd, nspawn's --template= mechanism, the discover-image.c logic,
and more.

src/shared/install-file.c

@@ -0,0 +1,269 @@
+/* SPDX-License-Identifier: LGPL-2.1-or-later */
+
+#include <sys/ioctl.h>
+
+#include "btrfs-util.h"
+#include "chattr-util.h"
+#include "errno-util.h"
+#include "fd-util.h"
+#include "fs-util.h"
+#include "install-file.h"
+#include "missing_syscall.h"
+#include "rm-rf.h"
+
+int fs_make_very_read_only(int fd) {
+        struct stat st;
+        int r;
+
+        assert(fd >= 0);
+
+        /* Tries to make the specified fd "comprehensively" read-only. Primary usecase for this is OS images,
+         * i.e. either loopback files or larger directory hierarchies. Depending on the inode type and
+         * backing file system this means something different:
+         *
+         * 1. If the fd refers to a btrfs subvolume we'll mark it read-only as a whole
+         * 2. If the fd refers to any other directory we'll set the FS_IMMUTABLE_FL flag on it
+         * 3. If the fd refers to a regular file we'll drop the w bits.
+         * 4. If the fd refers to a block device, use BLKROSET to set read-only state
+         *
+         * You might wonder why not drop the x bits for directories. That's because we want to guarantee that
+         * everything "inside" the image remains largely the way it is, in case you mount it. And since the
+         * mode of the root dir of the image is pretty visible we don't want to modify it. btrfs subvol flags
+         * and the FS_IMMUTABLE_FL otoh are much less visible. Changing the mode of regular files should be
+         * OK though, since after all this is supposed to be used for disk images, i.e. the fs in the disk
+         * image doesn't make the mode of the loopback file it is stored in visible. */
+
+        if (fstat(fd, &st) < 0)
+                return -errno;
+
+        switch (st.st_mode & S_IFMT) {
+
+        case S_IFDIR:
+                if (btrfs_might_be_subvol(&st)) {
+                        r = btrfs_subvol_set_read_only_fd(fd, true);
+                        if (r >= 0)
+                                return 0;
+
+                        if (!ERRNO_IS_NOT_SUPPORTED(r) && r != -EINVAL)
+                                return r;
+                }
+
+                r = chattr_fd(fd, FS_IMMUTABLE_FL, FS_IMMUTABLE_FL, NULL);
+                if (r < 0)
+                        return r;
+
+                break;
+
+        case S_IFREG:
+                if ((st.st_mode & 0222) != 0)
+                        if (fchmod(fd, st.st_mode & 07555) < 0)
+                                return -errno;
+
+                break;
+
+        case S_IFBLK: {
+                int ro = 1;
+
+                if (ioctl(fd, BLKROSET, &ro) < 0)
+                        return -errno;
+
+                break;
+        }
+
+        default:
+                return -EBADFD;
+        }
+
+        return 0;
+}
+
+static int unlinkat_maybe_dir(int dirfd, const char *pathname) {
+
+        /* Invokes unlinkat() for regular files first, and if this fails with EISDIR tries again with
+         * AT_REMOVEDIR */
+
+        if (unlinkat(dirfd, pathname, 0) < 0) {
+                if (errno != EISDIR)
+                        return -errno;
+
+                if (unlinkat(dirfd, pathname, AT_REMOVEDIR) < 0)
+                        return -errno;
+        }
+
+        return 0;
+}
+
+int install_file(int source_atfd, const char *source_name,
+                 int target_atfd, const char *target_name,
+                 InstallFileFlags flags) {
+
+        _cleanup_close_ int rofd = -1;
+        int r;
+
+        /* Moves a file or directory tree into place, with some bells and whistles:
+         *
+         * 1. Optionally syncs before/after to ensure file installation can be used as barrier
+         * 2. Optionally marks the file/directory read-only using fs_make_very_read_only()
+         * 3. Optionally operates in replacing or in non-replacing mode.
+         * 4. If it replaces will remove the old tree if needed.
+         */
+
+        assert(source_atfd >= 0 || source_atfd == AT_FDCWD);
+        assert(source_name);
+        assert(target_atfd >= 0 || target_atfd == AT_FDCWD);
+
+        /* If target_name is specified as NULL no renaming takes place. Instead it is assumed the file is
+         * already in place, and only the syncing/read-only marking shall be applied. Note that with
+         * target_name=NULL and flags=0 this call is a NOP */
+
+        if ((flags & (INSTALL_FSYNC|INSTALL_FSYNC_FULL|INSTALL_SYNCFS|INSTALL_READ_ONLY)) != 0) {
+                _cleanup_close_ int pfd = -1;
+                struct stat st;
+
+                /* Open an O_PATH fd for the source if we need to sync things or mark things read only. */
+
+                pfd = openat(source_atfd, source_name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
+                if (pfd < 0)
+                        return -errno;
+
+                if (fstat(pfd, &st) < 0)
+                        return -errno;
+
+                switch (st.st_mode & S_IFMT) {
+
+                case S_IFREG: {
+                        _cleanup_close_ int regfd = -1;
+
+                        regfd = fd_reopen(pfd, O_RDONLY|O_CLOEXEC);
+                        if (regfd < 0)
+                                return regfd;
+
+                        if ((flags & (INSTALL_FSYNC_FULL|INSTALL_SYNCFS)) != 0) {
+                                /* If this is just a regular file (as oppose to a fully populated directory)
+                                 * let's downgrade INSTALL_SYNCFS to INSTALL_FSYNC_FULL, after all this is
+                                 * going to be a single inode we install */
+                                r = fsync_full(regfd);
+                                if (r < 0)
+                                        return r;
+                        } else if (flags & INSTALL_FSYNC) {
+                                if (fsync(regfd) < 0)
+                                        return -errno;
+                        }
+
+                        if (flags & INSTALL_READ_ONLY)
+                                rofd = TAKE_FD(regfd);
+
+                        break;
+                }
+
+                case S_IFDIR: {
+                        _cleanup_close_ int dfd = -1;
+
+                        dfd = fd_reopen(pfd, O_RDONLY|O_DIRECTORY|O_CLOEXEC);
+                        if (dfd < 0)
+                                return dfd;
+
+                        if (flags & INSTALL_SYNCFS) {
+                                if (syncfs(dfd) < 0)
+                                        return -errno;
+                        } else if (flags & INSTALL_FSYNC_FULL) {
+                                r = fsync_full(dfd);
+                                if (r < 0)
+                                        return r;
+                        } else if (flags & INSTALL_FSYNC) {
+                                if (fsync(dfd) < 0)
+                                        return -errno;
+                        }
+
+                        if (flags & INSTALL_READ_ONLY)
+                                rofd = TAKE_FD(dfd);
+
+                        break;
+                }
+
+                default:
+                        /* Other inodes: char/block device inodes, fifos, symlinks, sockets don't need
+                         * syncing themselves, as they only exist in the directory, and have no contents on
+                         * disk */
+
+                        if (target_name && (flags & (INSTALL_FSYNC_FULL|INSTALL_SYNCFS)) != 0) {
+                                r = fsync_directory_of_file(pfd);
+                                if (r < 0)
+                                        return r;
+                        }
+
+                        break;
+                }
+        }
+
+        if (target_name) {
+                /* Rename the file */
+
+                if (flags & INSTALL_REPLACE) {
+                        /* First, try a simple renamat(), maybe that's enough */
+                        if (renameat(source_atfd, source_name, target_atfd, target_name) < 0) {
+                                _cleanup_close_ int dfd = -1;
+
+                                if (!IN_SET(errno, EEXIST, ENOTDIR, ENOTEMPTY, EISDIR, EBUSY))
+                                        return -errno;
+
+                                /* Hmm, the target apparently existed already. Let's try to use
+                                 * RENAME_EXCHANGE. But let's first open the inode if it's a directory, so
+                                 * that we can later remove its contents if it's a directory. Why do this
+                                 * before the rename()? Mostly because if we have trouble opening the thing
+                                 * we want to know before we start actually modifying the file system. */
+
+                                dfd = openat(target_atfd, target_name, O_RDONLY|O_DIRECTORY|O_CLOEXEC, 0);
+                                if (dfd < 0 && errno != ENOTDIR)
+                                        return -errno;
+
+                                if (renameat2(source_atfd, source_name, target_atfd, target_name, RENAME_EXCHANGE) < 0) {
+
+                                        if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL)
+                                                return -errno;
+
+                                        /* The exchange didn't work, let's remove the target first, and try again */
+
+                                        if (dfd >= 0)
+                                                (void) rm_rf_children(TAKE_FD(dfd), REMOVE_PHYSICAL|REMOVE_SUBVOLUME|REMOVE_CHMOD, NULL);
+
+                                        r = unlinkat_maybe_dir(target_atfd, target_name);
+                                        if (r < 0)
+                                                return log_debug_errno(r, "Failed to remove target directory: %m");
+
+                                        if (renameat(source_atfd, source_name, target_atfd, target_name) < 0)
+                                                return -errno;
+                                } else {
+                                        /* The exchange worked, hence let's remove the source (i.e. the old target) */
+                                        if (dfd >= 0)
+                                                (void) rm_rf_children(TAKE_FD(dfd), REMOVE_PHYSICAL|REMOVE_SUBVOLUME|REMOVE_CHMOD, NULL);
+
+                                        r = unlinkat_maybe_dir(source_atfd, source_name);
+                                        if (r < 0)
+                                                return log_debug_errno(r, "Failed to remove replaced target directory: %m");
+                                }
+                        }
+                } else {
+                        r = rename_noreplace(source_atfd, source_name, target_atfd, target_name);
+                        if (r < 0)
+                                return r;
+                }
+        }
+
+        if (rofd >= 0) {
+                r = fs_make_very_read_only(rofd);
+                if (r < 0)
+                        return r;
+        }
+
+        if ((flags & (INSTALL_FSYNC_FULL|INSTALL_SYNCFS)) != 0) {
+                if (target_name)
+                        r = fsync_parent_at(target_atfd, target_name);
+                else
+                        r = fsync_parent_at(source_atfd, source_name);
+                if (r < 0)
+                        return r;
+        }
+
+        return 0;
+}

src/shared/install-file.h

@@ -0,0 +1,14 @@
+/* SPDX-License-Identifier: LGPL-2.1-or-later */
+#pragma once
+
+int fs_make_very_read_only(int fd);
+
+typedef enum InstallFileFlags {
+        INSTALL_REPLACE    = 1 << 0, /* Replace an existing inode */
+        INSTALL_READ_ONLY  = 1 << 1, /* Call fs_make_very_read_only() to make the inode comprehensively read-only */
+        INSTALL_FSYNC      = 1 << 2, /* fsync() file contents before moving file in */
+        INSTALL_FSYNC_FULL = 1 << 3, /* like INSTALL_FSYNC, but also fsync() parent dir before+after moving file in */
+        INSTALL_SYNCFS     = 1 << 4, /* syncfs() before moving file in, fsync() parent dir after moving file in */
+} InstallFileFlags;
+
+int install_file(int source_atfd, const char *source_name, int target_atfd, const char *target_name, InstallFileFlags flags);

src/shared/meson.build

@@ -152,6 +152,8 @@ shared_sources = files('''
         import-util.c
         import-util.h
         initreq.h
+        install-file.c
+        install-file.h
         install-printf.c
         install-printf.h
         install.c

src/test/meson.build

@@ -215,6 +215,8 @@ tests += [
 
         [['src/test/test-fs-util.c']],
 
+        [['src/test/test-install-file.c']],
+
         [['src/test/test-umask-util.c']],
 
         [['src/test/test-proc-cmdline.c']],

src/test/test-install-file.c

@@ -0,0 +1,72 @@
+/* SPDX-License-Identifier: LGPL-2.1-or-later */
+
+#include "fileio.h"
+#include "install-file.h"
+#include "path-util.h"
+#include "rm-rf.h"
+#include "tests.h"
+#include "tmpfile-util.h"
+#include "umask-util.h"
+
+static void test_install_file(void) {
+        _cleanup_(rm_rf_physical_and_freep) char *p = NULL;
+        _cleanup_free_ char *a = NULL, *b = NULL, *c = NULL;
+        struct stat stat1, stat2;
+
+        log_info("/* %s */", __func__);
+
+        assert_se(mkdtemp_malloc(NULL, &p) >= 0);
+        assert_se(a = path_join(p, "foo"));
+        assert_se(b = path_join(p, "bar"));
+
+        RUN_WITH_UMASK(0077)
+                assert_se(write_string_file(a, "wups", WRITE_STRING_FILE_CREATE) >= 0);
+
+        assert_se(lstat(a, &stat1) >= 0);
+        assert_se(S_ISREG(stat1.st_mode));
+
+        assert_se(install_file(AT_FDCWD, a, AT_FDCWD, b, 0) >= 0);
+        assert_se(install_file(AT_FDCWD, b, AT_FDCWD, a, INSTALL_FSYNC) >= 0);
+
+        assert_se(write_string_file(b, "ttss", WRITE_STRING_FILE_CREATE) >= 0);
+        assert_se(install_file(AT_FDCWD, a, AT_FDCWD, b, INSTALL_FSYNC_FULL) == -EEXIST);
+        assert_se(install_file(AT_FDCWD, a, AT_FDCWD, b, INSTALL_FSYNC_FULL|INSTALL_REPLACE) >= 0);
+
+        assert_se(stat(b, &stat2) >= 0);
+        assert_se(stat1.st_dev == stat2.st_dev);
+        assert_se(stat1.st_ino == stat2.st_ino);
+        assert_se((stat2.st_mode & 0222) != 0); /* writable */
+
+        assert_se(install_file(AT_FDCWD, b, AT_FDCWD, a, INSTALL_FSYNC_FULL|INSTALL_REPLACE|INSTALL_READ_ONLY) >= 0);
+
+        assert_se(stat(a, &stat2) >= 0);
+        assert_se(stat1.st_dev == stat2.st_dev);
+        assert_se(stat1.st_ino == stat2.st_ino);
+        assert_se((stat2.st_mode & 0222) == 0); /* read-only */
+
+        assert_se(mkdir(b, 0755) >= 0);
+        assert_se(c = path_join(b, "dir"));
+        assert_se(mkdir(c, 0755) >= 0);
+        free(c);
+        assert_se(c = path_join(b, "reg"));
+        assert_se(mknod(c, S_IFREG|0755, 0) >= 0);
+        free(c);
+        assert_se(c = path_join(b, "fifo"));
+        assert_se(mknod(c, S_IFIFO|0755, 0) >= 0);
+
+        assert_se(install_file(AT_FDCWD, b, AT_FDCWD, a, INSTALL_FSYNC_FULL) == -EEXIST);
+        assert_se(install_file(AT_FDCWD, b, AT_FDCWD, a, INSTALL_FSYNC_FULL|INSTALL_REPLACE) == 0);
+
+        assert_se(write_string_file(b, "ttss", WRITE_STRING_FILE_CREATE) >= 0);
+
+        assert_se(install_file(AT_FDCWD, b, AT_FDCWD, a, INSTALL_FSYNC_FULL) == -EEXIST);
+        assert_se(install_file(AT_FDCWD, b, AT_FDCWD, a, INSTALL_FSYNC_FULL|INSTALL_REPLACE) == 0);
+}
+
+int main(int argc, char *argv[]) {
+        test_setup_logging(LOG_INFO);
+
+        test_install_file();
+
+        return 0;
+}