Merge pull request #26693 from poettering/udev-loop-links

udev: add /dev/loop/by-inode/… + /dev/loop/by-ref/… loopback block device symlinks

TODO

@@ -132,6 +132,12 @@ Features:
 * set MS_NOSYMFOLLOW for ESP and XBOOTLDR mounts both in gpt-generator and in
   dissect.c
 
+* rework loopback support in fstab: when "loop" option is used, then
+  instantiate a new systemd-loop@.service for the source path, set the
+  lo_file_name field for it to something recognizable derived from the fstab
+  line, and then generate a mount unit for it using a udev generated symlink
+  based on lo_file_name.
+
 * remove tomoyo support, it's obsolete and unmaintained apparently
 
 * journald: add varlink service that allows subscribing to certain log events,
@@ -277,9 +283,6 @@ Features:
 * systemd-sysext: for sysext DDIs picked up via EFI stub, set much stricter
   image policy by default
 
-* systemd-dissect: maybe add "--attach" and "--detach" verbs which
-  synchronously attach a DDI to a loopback device but not actually mount them.
-
 * pam_systemd_home: add module parameter to control whether to only accept
   only password or only pcks11/fido2 auth, and then use this to hook nicely
   into two of the three PAM stacks gdm provides.

man/systemd-dissect.xml

@@ -32,6 +32,12 @@
     <cmdsynopsis>
       <command>systemd-dissect <arg choice="opt" rep="repeat">OPTIONS</arg> <option>--umount</option> <arg choice="plain"><replaceable>PATH</replaceable></arg></command>
     </cmdsynopsis>
+    <cmdsynopsis>
+      <command>systemd-dissect <arg choice="opt" rep="repeat">OPTIONS</arg> <option>--attach</option> <arg choice="plain"><replaceable>IMAGE</replaceable></arg></command>
+    </cmdsynopsis>
+    <cmdsynopsis>
+      <command>systemd-dissect <arg choice="opt" rep="repeat">OPTIONS</arg> <option>--detach</option> <arg choice="plain"><replaceable>PATH</replaceable></arg></command>
+    </cmdsynopsis>
     <cmdsynopsis>
       <command>systemd-dissect <arg choice="opt" rep="repeat">OPTIONS</arg> <option>--list</option> <arg choice="plain"><replaceable>IMAGE</replaceable></arg></command>
     </cmdsynopsis>
@@ -173,6 +179,25 @@
         <listitem><para>This is a shortcut for <option>--umount --rmdir</option>.</para></listitem>
       </varlistentry>
 
+      <varlistentry>
+        <term><option>--attach</option></term>
+
+        <listitem><para>Attach the specified disk image to an automatically allocated loopback block device,
+        and print the path to the loopback block device to standard output. This is similar to an invocation
+        of <command>losetup --find --show</command>, but will validate the image as DDI before attaching, and
+        derive the correct sector size to use automatically. Moreover, it ensures the per-partition block
+        devices are created before returning. Takes a path to a disk image file.</para></listitem>
+      </varlistentry>
+
+      <varlistentry>
+        <term><option>--detach</option></term>
+
+        <listitem><para>Detach the specified disk image from a loopback block device. This undoes the effect
+        of <option>--attach</option> above. This expects either a path to a loopback block device as an
+        argument, or the path to the backing image file. In the latter case it will automatically determine
+        the right device to detach.</para></listitem>
+      </varlistentry>
+
       <varlistentry>
         <term><option>--list</option></term>
         <term><option>-l</option></term>
@@ -361,6 +386,25 @@
         </para></listitem>
       </varlistentry>
 
+      <varlistentry>
+        <term><option>--loop-ref=</option></term>
+
+        <listitem><para>Configures the "reference" string the kernel shall report as backing file for the
+        loopback block device. While this is supposed to be a path or filename referencing the backing file,
+        this is not enforced and the kernel accepts arbitrary free-form strings, chosen by the user. Accepts
+        arbitrary strings up to a length of 63 characters. This sets the kernel's
+        <literal>.lo_file_name</literal> field for the block device. Note this is distinct from the
+        <filename>/sys/class/block/loopX/loop/backing_file</filename> attribute file that always reports a
+        path referring to the actual backing file. The latter is subject to mount namespace translation, the
+        former is not.</para>
+
+        <para>This setting is particularly useful in combination with the <option>--attach</option> command,
+        as it allows later referencing the allocated loop device via <filename>/dev/loop/by-ref/…</filename>
+        symlinks. Example: first, set up the loopback device via <command>systemd-dissect attach
+        --loop-ref=quux foo.raw</command>, and then reference it in a command via the specified filename:
+        <command>cfdisk /dev/loop/by-ref/quux</command>.</para></listitem>
+      </varlistentry>
+
       <xi:include href="standard-options.xml" xpointer="no-pager" />
       <xi:include href="standard-options.xml" xpointer="no-legend" />
       <xi:include href="standard-options.xml" xpointer="json" />

rules.d/60-persistent-storage.rules.in

@@ -141,4 +141,15 @@ ENV{ID_PART_ENTRY_SCHEME}=="gpt", ENV{ID_PART_ENTRY_NAME}=="?*", SYMLINK+="disk/
 ENV{DISKSEQ}=="?*", ENV{DEVTYPE}!="partition", ENV{ID_IGNORE_DISKSEQ}!="1", SYMLINK+="disk/by-diskseq/$env{DISKSEQ}"
 ENV{DISKSEQ}=="?*", ENV{DEVTYPE}=="partition", ENV{ID_IGNORE_DISKSEQ}!="1", SYMLINK+="disk/by-diskseq/$env{DISKSEQ}-part%n"
 
+# Create symlinks that allow referencing loopback devices by their backing file's inode number
+ENV{DEVTYPE}!="partition", ENV{ID_LOOP_BACKING_DEVICE}!="", ENV{ID_LOOP_BACKING_INODE}!="", SYMLINK+="loop/by-inode/$env{ID_LOOP_BACKING_DEVICE}-$env{ID_LOOP_BACKING_INODE}"
+ENV{DEVTYPE}=="partition", ENV{ID_LOOP_BACKING_DEVICE}!="", ENV{ID_LOOP_BACKING_INODE}!="", SYMLINK+="loop/by-inode/$env{ID_LOOP_BACKING_DEVICE}-$env{ID_LOOP_BACKING_INODE}-part%n"
+
+# Similar, but uses the .lo_file_name field of the loopback device (note that
+# this is basically just a free-form string passed from userspace to the kernel
+# when the device is created, it is not necessarily a file system path like the
+# "loop/backing_file" sysfs attribute, which is always an absolute path)
+ENV{DEVTYPE}!="partition", ENV{ID_LOOP_BACKING_FILENAME_ENC}!="", SYMLINK+="loop/by-ref/$env{ID_LOOP_BACKING_FILENAME_ENC}"
+ENV{DEVTYPE}=="partition", ENV{ID_LOOP_BACKING_FILENAME_ENC}!="", SYMLINK+="loop/by-ref/$env{ID_LOOP_BACKING_FILENAME_ENC}-part%n"
+
 LABEL="persistent_storage_end"

src/dissect/dissect.c

@@ -52,6 +52,8 @@ static enum {
         ACTION_DISSECT,
         ACTION_MOUNT,
         ACTION_UMOUNT,
+        ACTION_ATTACH,
+        ACTION_DETACH,
         ACTION_LIST,
         ACTION_MTREE,
         ACTION_WITH,
@@ -79,9 +81,11 @@ static bool arg_legend = true;
 static bool arg_rmdir = false;
 static bool arg_in_memory = false;
 static char **arg_argv = NULL;
+static char *arg_loop_ref = NULL;
 
 STATIC_DESTRUCTOR_REGISTER(arg_verity_settings, verity_settings_done);
 STATIC_DESTRUCTOR_REGISTER(arg_argv, strv_freep);
+STATIC_DESTRUCTOR_REGISTER(arg_loop_ref, freep);
 
 static int help(void) {
         _cleanup_free_ char *link = NULL;
@@ -94,6 +98,8 @@ static int help(void) {
         printf("%1$s [OPTIONS...] IMAGE\n"
                "%1$s [OPTIONS...] --mount IMAGE PATH\n"
                "%1$s [OPTIONS...] --umount PATH\n"
+               "%1$s [OPTIONS...] --attach IMAGE\n"
+               "%1$s [OPTIONS...] --detach PATH\n"
                "%1$s [OPTIONS...] --list IMAGE\n"
                "%1$s [OPTIONS...] --mtree IMAGE\n"
                "%1$s [OPTIONS...] --with IMAGE [COMMAND…]\n"
@@ -119,6 +125,7 @@ static int help(void) {
                "                          not embedded in IMAGE\n"
                "     --json=pretty|short|off\n"
                "                          Generate JSON output\n"
+               "     --loop-ref=NAME      Set reference string for loopback device\n"
                "\n%3$sCommands:%4$s\n"
                "  -h --help               Show this help\n"
                "     --version            Show package version\n"
@@ -126,6 +133,8 @@ static int help(void) {
                "  -M                      Shortcut for --mount --mkdir\n"
                "  -u --umount             Unmount the image from the specified directory\n"
                "  -U                      Shortcut for --umount --rmdir\n"
+               "     --attach             Attach the disk image to a loopback block device\n"
+               "     --detach             Detach a loopback block device gain\n"
                "  -l --list               List all the files and directories of the specified\n"
                "                          OS image\n"
                "     --mtree              Show BSD mtree manifest of OS image\n"
@@ -206,6 +215,9 @@ static int parse_argv(int argc, char *argv[]) {
                 ARG_JSON,
                 ARG_MTREE,
                 ARG_DISCOVER,
+                ARG_ATTACH,
+                ARG_DETACH,
+                ARG_LOOP_REF,
         };
 
         static const struct option options[] = {
@@ -215,6 +227,8 @@ static int parse_argv(int argc, char *argv[]) {
                 { "no-legend",     no_argument,       NULL, ARG_NO_LEGEND     },
                 { "mount",         no_argument,       NULL, 'm'               },
                 { "umount",        no_argument,       NULL, 'u'               },
+                { "attach",        no_argument,       NULL, ARG_ATTACH        },
+                { "detach",        no_argument,       NULL, ARG_DETACH        },
                 { "with",          no_argument,       NULL, ARG_WITH          },
                 { "read-only",     no_argument,       NULL, 'r'               },
                 { "discard",       required_argument, NULL, ARG_DISCARD       },
@@ -232,6 +246,7 @@ static int parse_argv(int argc, char *argv[]) {
                 { "copy-to",       no_argument,       NULL, 'a'               },
                 { "json",          required_argument, NULL, ARG_JSON          },
                 { "discover",      no_argument,       NULL, ARG_DISCOVER      },
+                { "loop-ref",      required_argument, NULL, ARG_LOOP_REF      },
                 {}
         };
 
@@ -291,6 +306,14 @@ static int parse_argv(int argc, char *argv[]) {
                         arg_rmdir = true;
                         break;
 
+                case ARG_ATTACH:
+                        arg_action = ACTION_ATTACH;
+                        break;
+
+                case ARG_DETACH:
+                        arg_action = ACTION_DETACH;
+                        break;
+
                 case 'l':
                         arg_action = ACTION_LIST;
                         arg_flags |= DISSECT_IMAGE_READ_ONLY;
@@ -417,6 +440,20 @@ static int parse_argv(int argc, char *argv[]) {
                         arg_action = ACTION_DISCOVER;
                         break;
 
+                case ARG_LOOP_REF:
+                        if (isempty(optarg)) {
+                                arg_loop_ref = mfree(arg_loop_ref);
+                                break;
+                        }
+
+                        if (strlen(optarg) >= sizeof_field(struct loop_info64, lo_file_name))
+                                return log_error_errno(SYNTHETIC_ERRNO(EINVAL), "Loop device ref string '%s' is too long.", optarg);
+
+                        r = free_and_strdup_warn(&arg_loop_ref, optarg);
+                        if (r < 0)
+                                return r;
+                        break;
+
                 case '?':
                         return -EINVAL;
 
@@ -454,6 +491,22 @@ static int parse_argv(int argc, char *argv[]) {
                 arg_path = argv[optind];
                 break;
 
+        case ACTION_ATTACH:
+                if (optind + 1 != argc)
+                        return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
+                                               "Expected an image file path as only argument.");
+
+                arg_image = argv[optind];
+                break;
+
+        case ACTION_DETACH:
+                if (optind + 1 != argc)
+                        return log_error_errno(SYNTHETIC_ERRNO(EINVAL),
+                                               "Expected an image file path or loopback device as only argument.");
+
+                arg_image = argv[optind];
+                break;
+
         case ACTION_LIST:
         case ACTION_MTREE:
                 if (optind + 1 != argc)
@@ -1486,6 +1539,113 @@ static int action_discover(void) {
         return table_print_with_pager(t, arg_json_format_flags, arg_pager_flags, arg_legend);
 }
 
+static int action_attach(DissectedImage *m, LoopDevice *d) {
+        int r;
+
+        assert(m);
+        assert(d);
+
+        r = loop_device_set_autoclear(d, false);
+        if (r < 0)
+                return log_error_errno(r, "Failed to disable auto-clear logic on loopback device: %m");
+
+        r = dissected_image_relinquish(m);
+        if (r < 0)
+                return log_error_errno(r, "Failed to relinquish DM and loopback block devices: %m");
+
+        puts(d->node);
+        return 0;
+}
+
+static int action_detach(const char *path) {
+        _cleanup_(loop_device_unrefp) LoopDevice *loop = NULL;
+        _cleanup_close_ int fd = -EBADF;
+        struct stat st;
+        int r;
+
+        fd = open(path, O_PATH|O_CLOEXEC);
+        if (fd < 0)
+                return log_error_errno(errno, "Failed to open '%s': %m", path);
+
+        if (fstat(fd, &st) < 0)
+                return log_error_errno(errno, "Failed to stat '%s': %m", path);
+
+        if (S_ISBLK(st.st_mode)) {
+                r = loop_device_open_from_fd(fd, O_RDONLY, LOCK_EX, &loop);
+                if (r < 0)
+                        return log_error_errno(r, "Failed to open '%s' as loopback block device: %m", path);
+
+        } else if (S_ISREG(st.st_mode)) {
+                _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL;
+                sd_device *d;
+
+                /* If a regular file is specified, search for a loopback block device that is backed by it */
+
+                r = sd_device_enumerator_new(&e);
+                if (r < 0)
+                        return log_error_errno(r, "Failed to allocate enumerator: %m");
+
+                r = sd_device_enumerator_add_match_subsystem(e, "block", true);
+                if (r < 0)
+                        return log_error_errno(r, "Failed to match block devices: %m");
+
+                r = sd_device_enumerator_add_match_sysname(e, "loop*");
+                if (r < 0)
+                        return log_error_errno(r, "Failed to match loopback block devices: %m");
+
+                (void) sd_device_enumerator_allow_uninitialized(e);
+
+                FOREACH_DEVICE(e, d) {
+                        _cleanup_(loop_device_unrefp) LoopDevice *entry_loop = NULL;
+                        const char *name, *devtype;
+
+                        r = sd_device_get_sysname(d, &name);
+                        if (r < 0) {
+                                log_warning_errno(r, "Failed to get enumerated device's sysname, skipping: %m");
+                                continue;
+                        }
+
+                        r = sd_device_get_devtype(d, &devtype);
+                        if (r < 0) {
+                                log_warning_errno(r, "Failed to get devtype of '%s', skipping: %m", name);
+                                continue;
+                        }
+
+                        if (!streq(devtype, "disk")) /* Filter out partition block devices */
+                                continue;
+
+                        r = loop_device_open(d, O_RDONLY, LOCK_SH, &entry_loop);
+                        if (r < 0) {
+                                log_warning_errno(r, "Failed to open loopback block device '%s', skipping: %m", name);
+                                continue;
+                        }
+
+                        if (entry_loop->backing_devno == st.st_dev && entry_loop->backing_inode == st.st_ino) {
+                                /* Found it! The kernel allows attaching a single file to multiple loopback
+                                 * devices. Let's destruct them in reverse order, i.e. find the last matching
+                                 * loopback device here, rather than the first. */
+
+                                loop_device_unref(loop);
+                                loop = TAKE_PTR(entry_loop);
+                        }
+                }
+
+                if (!loop)
+                        return log_error_errno(SYNTHETIC_ERRNO(ENXIO), "No loopback block device backed by '%s' found.", path);
+
+                r = loop_device_flock(loop, LOCK_EX);
+                if (r < 0)
+                        return log_error_errno(r, "Failed to upgrade device lock: %m");
+        }
+
+        r = loop_device_set_autoclear(loop, true);
+        if (r < 0)
+                log_warning_errno(r, "Failed to enable autoclear logic on '%s', ignoring: %m", loop->node);
+
+        loop_device_unrelinquish(loop);
+        return 0;
+}
+
 static int run(int argc, char *argv[]) {
         _cleanup_(dissected_image_unrefp) DissectedImage *m = NULL;
         _cleanup_(loop_device_unrefp) LoopDevice *d = NULL;
@@ -1503,6 +1663,8 @@ static int run(int argc, char *argv[]) {
 
         if (arg_action == ACTION_UMOUNT)
                 return action_umount(arg_path);
+        if (arg_action == ACTION_DETACH)
+                return action_detach(arg_image);
         if (arg_action == ACTION_DISCOVER)
                 return action_discover();
 
@@ -1528,6 +1690,12 @@ static int run(int argc, char *argv[]) {
         if (r < 0)
                 return log_error_errno(r, "Failed to set up loopback device for %s: %m", arg_image);
 
+        if (arg_loop_ref) {
+                r = loop_device_set_filename(d, arg_loop_ref);
+                if (r < 0)
+                        log_warning_errno(r, "Failed to set loop reference string to '%s', ignoring: %m", arg_loop_ref);
+        }
+
         r = dissect_loop_device_and_warn(
                         d,
                         &arg_verity_settings,
@@ -1537,6 +1705,9 @@ static int run(int argc, char *argv[]) {
         if (r < 0)
                 return r;
 
+        if (arg_action == ACTION_ATTACH)
+                return action_attach(m, d);
+
         r = dissected_image_load_verity_sig_partition(
                         m,
                         d->fd,
@@ -1547,29 +1718,23 @@ static int run(int argc, char *argv[]) {
         switch (arg_action) {
 
         case ACTION_DISSECT:
-                r = action_dissect(m, d);
-                break;
+                return action_dissect(m, d);
 
         case ACTION_MOUNT:
-                r = action_mount(m, d);
-                break;
+                return action_mount(m, d);
 
         case ACTION_LIST:
         case ACTION_MTREE:
         case ACTION_COPY_FROM:
         case ACTION_COPY_TO:
-                r = action_list_or_mtree_or_copy(m, d);
-                break;
+                return action_list_or_mtree_or_copy(m, d);
 
         case ACTION_WITH:
-                r = action_with(m, d);
-                break;
+                return action_with(m, d);
 
         default:
                 assert_not_reached();
         }
-
-        return r;
 }
 
 DEFINE_MAIN_FUNCTION(run);

src/shared/loop-util.c

@@ -595,6 +595,8 @@ static int loop_device_make_internal(
         }
 
         d->backing_file = TAKE_PTR(backing_file);
+        d->backing_inode = st.st_ino;
+        d->backing_devno = st.st_dev;
 
         log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64,
                   d->node,
@@ -841,11 +843,12 @@ int loop_device_open(
 
         _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF;
         _cleanup_free_ char *node = NULL, *backing_file = NULL;
+        dev_t devnum, backing_devno = 0;
         struct loop_info64 info;
+        ino_t backing_inode = 0;
         uint64_t diskseq = 0;
         LoopDevice *d;
         const char *s;
-        dev_t devnum;
         int r, nr = -1;
 
         assert(dev);
@@ -878,6 +881,9 @@ int loop_device_open(
                         if (!backing_file)
                                 return -ENOMEM;
                 }
+
+                backing_devno = info.lo_device;
+                backing_inode = info.lo_inode;
         }
 
         r = fd_get_diskseq(fd, &diskseq);
@@ -913,6 +919,8 @@ int loop_device_open(
                 .node = TAKE_PTR(node),
                 .dev = sd_device_ref(dev),
                 .backing_file = TAKE_PTR(backing_file),
+                .backing_inode = backing_inode,
+                .backing_devno = backing_devno,
                 .relinquished = true, /* It's not ours, don't try to destroy it when this object is freed */
                 .devno = devnum,
                 .diskseq = diskseq,
@@ -1103,3 +1111,57 @@ int loop_device_sync(LoopDevice *d) {
 
         return RET_NERRNO(fsync(d->fd));
 }
+
+int loop_device_set_autoclear(LoopDevice *d, bool autoclear) {
+        struct loop_info64 info;
+
+        assert(d);
+
+        if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0)
+                return -errno;
+
+        if (autoclear == FLAGS_SET(info.lo_flags, LO_FLAGS_AUTOCLEAR))
+                return 0;
+
+        SET_FLAG(info.lo_flags, LO_FLAGS_AUTOCLEAR, autoclear);
+
+        if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0)
+                return -errno;
+
+        return 1;
+}
+
+int loop_device_set_filename(LoopDevice *d, const char *name) {
+        struct loop_info64 info;
+
+        assert(d);
+
+        /* Sets the .lo_file_name of the loopback device. This is supposed to contain the path to the file
+         * backing the block device, but is actually just a free-form string you can pass to the kernel. Most
+         * tools that actually care for the backing file path use the sysfs attribute file loop/backing_file
+         * which is a kernel generated string, subject to file system namespaces and such.
+         *
+         * .lo_file_name is useful since userspace can select it freely when creating a loopback block
+         * device, and we can use it for /dev/loop/by-ref/ symlinks, and similar, so that apps can recognize
+         * their own loopback files. */
+
+        if (name && strlen(name) >= sizeof(info.lo_file_name))
+                return -ENOBUFS;
+
+        if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0)
+                return -errno;
+
+        if (strneq((char*) info.lo_file_name, strempty(name), sizeof(info.lo_file_name)))
+                return 0;
+
+        if (name) {
+                strncpy((char*) info.lo_file_name, name, sizeof(info.lo_file_name)-1);
+                info.lo_file_name[sizeof(info.lo_file_name)-1] = 0;
+        } else
+                memzero(info.lo_file_name, sizeof(info.lo_file_name));
+
+        if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0)
+                return -errno;
+
+        return 1;
+}

src/shared/loop-util.h

@@ -14,12 +14,14 @@ struct LoopDevice {
         unsigned n_ref;
         int fd;
         int lock_fd;
-        int nr;   /* The loopback device index (i.e. 4 for /dev/loop4); if this object encapsulates a non-loopback block device, set to -1 */
-        dev_t devno;
+        int nr;         /* The loopback device index (i.e. 4 for /dev/loop4); if this object encapsulates a non-loopback block device, set to -1 */
+        dev_t devno;    /* The loopback device's own dev_t */
         char *node;
         sd_device *dev;
         char *backing_file;
         bool relinquished;
+        dev_t backing_devno; /* The backing file's dev_t */
+        ino_t backing_inode; /* The backing file's ino_t */
         uint64_t diskseq; /* Block device sequence number, monothonically incremented by the kernel on create/attach, or 0 if we don't know */
         uint64_t uevent_seqnum_not_before; /* uevent sequm right before we attached the loopback device, or UINT64_MAX if we don't know */
         usec_t timestamp_not_before; /* CLOCK_MONOTONIC timestamp taken immediately before attaching the loopback device, or USEC_INFINITY if we don't know */
@@ -47,3 +49,6 @@ int loop_device_refresh_size(LoopDevice *d, uint64_t offset, uint64_t size);
 
 int loop_device_flock(LoopDevice *d, int operation);
 int loop_device_sync(LoopDevice *d);
+
+int loop_device_set_autoclear(LoopDevice *d, bool autoclear);
+int loop_device_set_filename(LoopDevice *d, const char *name);

src/udev/udev-builtin-blkid.c

@@ -5,11 +5,17 @@
  * Copyright © 2011 Karel Zak <kzak@redhat.com>
  */
 
+#if HAVE_VALGRIND_MEMCHECK_H
+#include <valgrind/memcheck.h>
+#endif
+
 #include <errno.h>
 #include <fcntl.h>
 #include <getopt.h>
+#include <linux/loop.h>
 #include <stdio.h>
 #include <stdlib.h>
+#include <sys/ioctl.h>
 #include <sys/stat.h>
 
 #include "sd-id128.h"
@@ -17,6 +23,7 @@
 #include "alloc-util.h"
 #include "blkid-util.h"
 #include "device-util.h"
+#include "devnum-util.h"
 #include "efi-loader.h"
 #include "errno-util.h"
 #include "fd-util.h"
@@ -234,11 +241,86 @@ static int probe_superblocks(blkid_probe pr) {
         return blkid_do_safeprobe(pr);
 }
 
+static int read_loopback_backing_inode(
+                sd_device *dev,
+                int fd,
+                dev_t *ret_devno,
+                ino_t *ret_inode,
+                char **ret_fname) {
+
+        _cleanup_free_ char *fn = NULL;
+        struct loop_info64 info;
+        const char *name;
+        int r;
+
+        assert(dev);
+        assert(fd >= 0);
+        assert(ret_devno);
+        assert(ret_inode);
+        assert(ret_fname);
+
+        /* Retrieves various fields of the current loopback device backing file, so that we can ultimately
+         * use it to create stable symlinks to loopback block devices, based on what they are backed by. We
+         * pick up inode/device as well as file name field. Note that we pick up the "lo_file_name" field
+         * here, which is an arbitrary free-form string provided by userspace. We do not return the sysfs
+         * attribute loop/backing_file here, because that is directly accessible from udev rules anyway. And
+         * sometimes, depending on context, it's a good thing to return the string userspace can freely pick
+         * over the string automatically generated by the kernel. */
+
+        r = sd_device_get_sysname(dev, &name);
+        if (r < 0)
+                return r;
+
+        if (!startswith(name, "loop"))
+                goto notloop;
+
+        if (ioctl(fd, LOOP_GET_STATUS64, &info) < 0) {
+                if (ERRNO_IS_NOT_SUPPORTED(errno))
+                        goto notloop;
+
+                return -errno;
+        }
+
+#if HAVE_VALGRIND_MEMCHECK_H
+        VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
+#endif
+
+        if (isempty((char*) info.lo_file_name) ||
+            strnlen((char*) info.lo_file_name, sizeof(info.lo_file_name)-1) == sizeof(info.lo_file_name)-1)
+                /* Don't pick up file name if it is unset or possibly truncated. (Note: we can't really know
+                 * the file file name is truncated if it uses sizeof(info.lo_file_name)-1 as length; it could
+                 * also just mean the string is just that long and wasn't truncated — but the fact is simply
+                 * that we cannot know in that case if it was truncated or not, hence we assume the worst and
+                 * suppress — at least for now. For shorter strings we know for sure it wasn't truncated,
+                 * hence that's always safe.) */
+                fn = NULL;
+        else {
+                fn = memdup_suffix0(info.lo_file_name, sizeof(info.lo_file_name));
+                if (!fn)
+                        return -ENOMEM;
+        }
+
+        *ret_inode = info.lo_inode;
+        *ret_devno = info.lo_device;
+        *ret_fname = TAKE_PTR(fn);
+        return 1;
+
+
+notloop:
+        *ret_devno = 0;
+        *ret_inode = 0;
+        *ret_fname = NULL;
+        return 0;
+}
+
 static int builtin_blkid(sd_device *dev, sd_netlink **rtnl, int argc, char *argv[], bool test) {
         const char *devnode, *root_partition = NULL, *data, *name;
         _cleanup_(blkid_free_probep) blkid_probe pr = NULL;
+        _cleanup_free_ char *backing_fname = NULL;
         bool noraid = false, is_gpt = false;
         _cleanup_close_ int fd = -EBADF;
+        ino_t backing_inode = 0;
+        dev_t backing_devno = 0;
         int64_t offset = 0;
         int r;
 
@@ -352,6 +434,32 @@ static int builtin_blkid(sd_device *dev, sd_netlink **rtnl, int argc, char *argv
         if (is_gpt)
                 find_gpt_root(dev, pr, test);
 
+        r = read_loopback_backing_inode(
+                        dev,
+                        fd,
+                        &backing_devno,
+                        &backing_inode,
+                        &backing_fname);
+        if (r < 0)
+                log_device_debug_errno(dev, r, "Failed to read loopback backing inode, ignoring: %m");
+        else if (r > 0) {
+                udev_builtin_add_propertyf(dev, test, "ID_LOOP_BACKING_DEVICE", DEVNUM_FORMAT_STR, DEVNUM_FORMAT_VAL(backing_devno));
+                udev_builtin_add_propertyf(dev, test, "ID_LOOP_BACKING_INODE", "%" PRIu64, (uint64_t) backing_inode);
+
+                if (backing_fname) {
+                        /* In the worst case blkid_encode_string() will blow up to 4x the string
+                         * length. Hence size the buffer to 4x of the longest string
+                         * read_loopback_backing_inode() might return */
+                        char encoded[sizeof_field(struct loop_info64, lo_file_name) * 4 + 1];
+
+                        assert(strlen(backing_fname) < ELEMENTSOF(encoded) / 4);
+                        blkid_encode_string(backing_fname, encoded, ELEMENTSOF(encoded));
+
+                        udev_builtin_add_property(dev, test, "ID_LOOP_BACKING_FILENAME", backing_fname);
+                        udev_builtin_add_property(dev, test, "ID_LOOP_BACKING_FILENAME_ENC", encoded);
+                }
+        }
+
         return 0;
 }
 

src/udev/udev-builtin.c

@@ -134,3 +134,21 @@ int udev_builtin_add_property(sd_device *dev, bool test, const char *key, const
 
         return 0;
 }
+
+int udev_builtin_add_propertyf(sd_device *dev, bool test, const char *key, const char *valf, ...) {
+        _cleanup_free_ char *val = NULL;
+        va_list ap;
+        int r;
+
+        assert(dev);
+        assert(key);
+        assert(valf);
+
+        va_start(ap, valf);
+        r = vasprintf(&val, valf, ap);
+        va_end(ap);
+        if (r < 0)
+                return log_oom_debug();
+
+        return udev_builtin_add_property(dev, test, key, val);
+}

src/udev/udev-builtin.h

@@ -6,6 +6,8 @@
 #include "sd-device.h"
 #include "sd-netlink.h"
 
+#include "macro.h"
+
 typedef enum UdevBuiltinCommand {
 #if HAVE_BLKID
         UDEV_BUILTIN_BLKID,
@@ -78,5 +80,6 @@ int udev_builtin_run(sd_device *dev, sd_netlink **rtnl, UdevBuiltinCommand cmd,
 void udev_builtin_list(void);
 bool udev_builtin_should_reload(void);
 int udev_builtin_add_property(sd_device *dev, bool test, const char *key, const char *val);
+int udev_builtin_add_propertyf(sd_device *dev, bool test, const char *key, const char *valf, ...) _printf_(4, 5);
 int udev_builtin_hwdb_lookup(sd_device *dev, const char *prefix, const char *modalias,
                              const char *filter, bool test);

test/units/testsuite-50.sh

@@ -430,6 +430,28 @@ mount -t ddi "${image}.gpt" "$T" -o ro,X-mount.mkdir,discard
 umount -R "$T"
 rmdir "$T"
 
+LOOP="$(systemd-dissect --attach --loop-ref=waldo "${image}.raw")"
+
+# Wait until the symlinks we want to test are established
+udevadm trigger -w "$LOOP"
+
+# Check if the /dev/loop/* symlinks really reference the right device
+test /dev/loop/by-ref/waldo -ef "$LOOP"
+
+if [ "$(stat -c '%Hd:%Ld' "${image}.raw")" != '?d:?d' ] ; then
+   # Old stat didn't know the %Hd and %Ld specifiers and turned them into ?d
+   # instead. Let's simply skip the test on such old systems.
+   test "$(stat -c '/dev/loop/by-inode/%Hd:%Ld-%i' "${image}.raw")" -ef "$LOOP"
+fi
+
+# Detach by loopback device
+systemd-dissect --detach "$LOOP"
+
+# Detach by backing inode
+systemd-dissect --attach --loop-ref=waldo "${image}.raw"
+systemd-dissect --detach "${image}.raw"
+(! systemd-dissect --detach "${image}.raw")
+
 echo OK >/testok
 
 exit 0