Merge pull request #22992 from poettering/loop-dissect-tweaks

loop-util/image dissect fixes

TODO

@@ -169,12 +169,6 @@ Features:
 
 * bootctl: show whether UEFI audit mode is available
 
-* dissect: rework how we access partitions: instead of letting the kernel probe
-  partition tables asynchronously, just pass the stuff we parsed in userspace
-  to the kernel via BLKPG_ADD_PARTITION. Benefit: we don't have to wait for
-  kernel/netlink/udev, but can run this synchronously without chance of losing
-  events or similar.
-
 * sd-event: optionally, if per-event source rate limit is hit, downgrade
   priority, but leave enabled, and once ratelimit window is over, upgrade
   priority again. That way we can combat event source starvation without

src/core/namespace.c

@@ -2055,6 +2055,12 @@ int setup_namespace(
                 if (r < 0)
                         return log_debug_errno(r, "Failed to create loop device for root image: %m");
 
+                /* Make sure udevd won't issue BLKRRPART (which might flush out the loaded partition table)
+                 * while we are still trying to mount things */
+                r = loop_device_flock(loop_device, LOCK_SH);
+                if (r < 0)
+                        return log_debug_errno(r, "Failed to lock loopback device with LOCK_SH: %m");
+
                 r = dissect_image(
                                 loop_device->fd,
                                 &verity,
@@ -2403,6 +2409,14 @@ int setup_namespace(
                         goto finish;
                 }
 
+                /* Now release the block device lock, so that udevd is free to call BLKRRPART on the device
+                 * if it likes. */
+                r = loop_device_flock(loop_device, LOCK_UN);
+                if (r < 0) {
+                        log_debug_errno(r, "Failed to release lock on loopback block device: %m");
+                        goto finish;
+                }
+
                 if (decrypted_image) {
                         r = decrypted_image_relinquish(decrypted_image);
                         if (r < 0) {

src/dissect/dissect.c

@@ -639,6 +639,10 @@ static int action_mount(DissectedImage *m, LoopDevice *d) {
         if (r < 0)
                 return r;
 
+        r = loop_device_flock(d, LOCK_UN);
+        if (r < 0)
+                return log_error_errno(r, "Failed to unlock loopback block device: %m");
+
         if (di) {
                 r = decrypted_image_relinquish(di);
                 if (r < 0)
@@ -687,6 +691,10 @@ static int action_copy(DissectedImage *m, LoopDevice *d) {
 
         mounted_dir = TAKE_PTR(created_dir);
 
+        r = loop_device_flock(d, LOCK_UN);
+        if (r < 0)
+                return log_error_errno(r, "Failed to unlock loopback block device: %m");
+
         if (di) {
                 r = decrypted_image_relinquish(di);
                 if (r < 0)
@@ -845,6 +853,12 @@ 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);
 
+        /* Make sure udevd doesn't issue BLKRRPART underneath us thus making devices disappear in the middle,
+         * that we assume already are there. */
+        r = loop_device_flock(d, LOCK_SH);
+        if (r < 0)
+                return log_error_errno(r, "Failed to lock loopback device: %m");
+
         r = dissect_image_and_warn(
                         d->fd,
                         arg_image,

src/gpt-auto-generator/gpt-auto-generator.c

@@ -1,6 +1,7 @@
 /* SPDX-License-Identifier: LGPL-2.1-or-later */
 
 #include <stdlib.h>
+#include <sys/file.h>
 #include <unistd.h>
 
 #include "sd-device.h"
@@ -696,6 +697,12 @@ static int enumerate_partitions(dev_t devnum) {
         if (r <= 0)
                 return r;
 
+        /* Let's take a LOCK_SH lock on the block device, in case udevd is already running. If we don't take
+         * the lock, udevd might end up issuing BLKRRPART in the middle, and we don't want that, since that
+         * might remove all partitions while we are operating on them. */
+        if (flock(fd, LOCK_SH) < 0)
+                return log_error_errno(errno, "Failed to lock root block device: %m");
+
         r = dissect_image(
                         fd,
                         NULL, NULL,
@@ -703,7 +710,6 @@ static int enumerate_partitions(dev_t devnum) {
                         UINT64_MAX,
                         USEC_INFINITY,
                         DISSECT_IMAGE_GPT_ONLY|
-                        DISSECT_IMAGE_NO_UDEV|
                         DISSECT_IMAGE_USR_NO_ROOT,
                         &m);
         if (r == -ENOPKG) {

src/nspawn/nspawn.c

@@ -5737,6 +5737,13 @@ static int run(int argc, char *argv[]) {
                         goto finish;
                 }
 
+                /* Take a LOCK_SH lock on the device, so that udevd doesn't issue BLKRRPART in our back */
+                r = loop_device_flock(loop, LOCK_SH);
+                if (r < 0) {
+                        log_error_errno(r, "Failed to take lock on loopback block device: %m");
+                        goto finish;
+                }
+
                 r = dissect_image_and_warn(
                                 loop->fd,
                                 arg_image,

src/partition/repart.c

@@ -12,6 +12,7 @@
 #include <sys/ioctl.h>
 #include <sys/stat.h>
 
+#include "sd-device.h"
 #include "sd-id128.h"
 
 #include "alloc-util.h"
@@ -3643,14 +3644,13 @@ static int resolve_copy_blocks_auto_candidate(
                 sd_id128_t *ret_uuid) {
 
         _cleanup_(blkid_free_probep) blkid_probe b = NULL;
-        _cleanup_free_ char *p = NULL;
+        _cleanup_(sd_device_unrefp) sd_device *dev = NULL;
         _cleanup_close_ int fd = -1;
-        const char *pttype, *t;
+        const char *pttype, *t, *p;
         sd_id128_t pt_parsed, u;
         blkid_partition pp;
         dev_t whole_devno;
         blkid_partlist pl;
-        struct stat st;
         int r;
 
         /* Checks if the specified partition has the specified GPT type UUID, and is located on the specified
@@ -3673,21 +3673,19 @@ static int resolve_copy_blocks_auto_candidate(
                                 major(partition_devno), minor(partition_devno),
                                 major(restrict_devno), minor(restrict_devno));
 
-        r = device_path_make_major_minor(S_IFBLK, whole_devno, &p);
+        r = sd_device_new_from_devnum(&dev, 'b', whole_devno);
         if (r < 0)
-                return log_error_errno(r, "Failed to convert block device to device node path: %m");
+                return log_error_errno(r, "Failed to create sd-device for block device %u:%u: %m",
+                                       major(whole_devno), minor(whole_devno));
 
-        fd = open(p, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
+        r = sd_device_get_devname(dev, &p);
+        if (r < 0)
+                return log_error_errno(r, "Failed to get name of block device %u:%u: %m",
+                                       major(whole_devno), minor(whole_devno));
+
+        fd = sd_device_open(dev, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
         if (fd < 0)
-                return log_error_errno(r, "Failed to open '%s': %m", p);
-
-        if (fstat(fd, &st) < 0)
-                return log_error_errno(r, "Failed to stat '%s': %m", p);
-
-        if (!S_ISBLK(st.st_mode) || st.st_rdev != whole_devno)
-                return log_error_errno(
-                                SYNTHETIC_ERRNO(EPERM),
-                                "Opened and determined block device don't match, refusing.");
+                return log_error_errno(fd, "Failed to open block device %s: %m", p);
 
         b = blkid_new_probe();
         if (!b)
@@ -3805,7 +3803,7 @@ static int resolve_copy_blocks_auto(
                 sd_id128_t type_uuid,
                 const char *root,
                 dev_t restrict_devno,
-                char **ret_path,
+                dev_t *ret_devno,
                 sd_id128_t *ret_uuid) {
 
         const char *try1 = NULL, *try2 = NULL;
@@ -3815,8 +3813,6 @@ static int resolve_copy_blocks_auto(
         dev_t devno, found = 0;
         int r;
 
-        assert(ret_path);
-
         /* Enforce some security restrictions: CopyBlocks=auto should not be an avenue to get outside of the
          * --root=/--image= confinement. Specifically, refuse CopyBlocks= in combination with --root= at all,
          * and restrict block device references in the --image= case to loopback block device we set up.
@@ -3926,9 +3922,8 @@ static int resolve_copy_blocks_auto(
                 return log_error_errno(SYNTHETIC_ERRNO(ENXIO),
                                        "Unable to automatically discover suitable partition to copy blocks from.");
 
-        r = device_path_make_major_minor(S_IFBLK, found, ret_path);
-        if (r < 0)
-                return log_error_errno(r, "Failed to convert dev_t to device node path: %m");
+        if (ret_devno)
+                *ret_devno = found;
 
         if (ret_uuid)
                 *ret_uuid = found_uuid;
@@ -3972,32 +3967,43 @@ static int context_open_copy_block_paths(
                                                        "Copying from block device node is not permitted in --image=/--root= mode, refusing.");
 
                 } else if (p->copy_blocks_auto) {
+                        _cleanup_(sd_device_unrefp) sd_device *dev = NULL;
+                        const char *devname;
+                        dev_t devno;
 
-                        r = resolve_copy_blocks_auto(p->type_uuid, root, restrict_devno, &opened, &uuid);
+                        r = resolve_copy_blocks_auto(p->type_uuid, root, restrict_devno, &devno, &uuid);
                         if (r < 0)
                                 return r;
 
-                        source_fd = open(opened, O_RDONLY|O_CLOEXEC|O_NOCTTY);
+                        r = sd_device_new_from_devnum(&dev, 'b', devno);
+                        if (r < 0)
+                                return log_error_errno(r, "Failed to create sd-device object for device %u:%u: %m", major(devno), minor(devno));
+
+                        r = sd_device_get_devname(dev, &devname);
+                        if (r < 0)
+                                return log_error_errno(r, "Failed to get device name of %u:%u: %m", major(devno), minor(devno));
+
+                        opened = strdup(devname);
+                        if (!opened)
+                                return log_oom();
+
+                        source_fd = sd_device_open(dev, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
                         if (source_fd < 0)
-                                return log_error_errno(errno, "Failed to open automatically determined source block copy device '%s': %m", opened);
+                                return log_error_errno(source_fd, "Failed to open automatically determined source block copy device '%s': %m", opened);
 
                         if (fstat(source_fd, &st) < 0)
                                 return log_error_errno(errno, "Failed to stat block copy file '%s': %m", opened);
-
-                        /* If we found it automatically, it must be a block device, let's enforce that */
-                        if (!S_ISBLK(st.st_mode))
-                                return log_error_errno(SYNTHETIC_ERRNO(EBADF),
-                                                       "Automatically detected source block copy device '%s' is not a block device, refusing: %m", opened);
-                }  else
+                } else
                         continue;
 
                 if (S_ISDIR(st.st_mode)) {
-                        _cleanup_free_ char *bdev = NULL;
+                        _cleanup_(sd_device_unrefp) sd_device *dev = NULL;
+                        const char *bdev;
 
                         /* If the file is a directory, automatically find the backing block device */
 
                         if (major(st.st_dev) != 0)
-                                r = device_path_make_major_minor(S_IFBLK, st.st_dev, &bdev);
+                                r = sd_device_new_from_devnum(&dev, 'b', st.st_dev);
                         else {
                                 dev_t devt;
 
@@ -4009,22 +4015,23 @@ static int context_open_copy_block_paths(
                                 if (r < 0)
                                         return log_error_errno(r, "Unable to determine backing block device of '%s': %m", opened);
 
-                                r = device_path_make_major_minor(S_IFBLK, devt, &bdev);
+                                r = sd_device_new_from_devnum(&dev, 'b', devt);
                         }
                         if (r < 0)
-                                return log_error_errno(r, "Failed to determine block device path for block device backing '%s': %m", opened);
+                                return log_error_errno(r, "Failed to create sd-device object for block device backing '%s': %m", opened);
+
+                        r = sd_device_get_devpath(dev, &bdev);
+                        if (r < 0)
+                                return log_error_errno(r, "Failed to get device name for block device backing '%s': %m", opened);
 
                         safe_close(source_fd);
 
-                        source_fd = open(bdev, O_RDONLY|O_CLOEXEC|O_NOCTTY);
+                        source_fd = sd_device_open(dev, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
                         if (source_fd < 0)
-                                return log_error_errno(errno, "Failed to open block device '%s': %m", bdev);
+                                return log_error_errno(source_fd, "Failed to open block device '%s': %m", bdev);
 
                         if (fstat(source_fd, &st) < 0)
                                 return log_error_errno(errno, "Failed to stat block device '%s': %m", bdev);
-
-                        if (!S_ISBLK(st.st_mode))
-                                return log_error_errno(SYNTHETIC_ERRNO(ENOTBLK), "Block device '%s' is not actually a block device, refusing.", bdev);
                 }
 
                 if (S_ISREG(st.st_mode))

src/portable/portable.c

@@ -359,6 +359,10 @@ static int portable_extract_by_path(
                 /* We now have a loopback block device, let's fork off a child in its own mount namespace, mount it
                  * there, and extract the metadata we need. The metadata is sent from the child back to us. */
 
+                r = loop_device_flock(d, LOCK_SH);
+                if (r < 0)
+                        return log_debug_errno(r, "Failed to acquire lock on loopback block device: %m");
+
                 BLOCK_SIGNALS(SIGCHLD);
 
                 r = mkdtemp_malloc("/tmp/inspect-XXXXXX", &tmpdir);

src/shared/discover-image.c

@@ -1196,6 +1196,12 @@ int image_read_metadata(Image *i) {
                 if (r < 0)
                         return r;
 
+                /* Make sure udevd doesn't issue BLKRRPART in the background which might make our partitions
+                 * disappear temporarily. */
+                r = loop_device_flock(d, LOCK_SH);
+                if (r < 0)
+                        return r;
+
                 r = dissect_image(
                                 d->fd,
                                 NULL, NULL,

src/shared/dissect-image.c

@@ -4,6 +4,7 @@
 #include <valgrind/memcheck.h>
 #endif
 
+#include <linux/blkpg.h>
 #include <linux/dm-ioctl.h>
 #include <linux/loop.h>
 #include <sys/mount.h>
@@ -125,389 +126,6 @@ not_found:
 }
 
 #if HAVE_BLKID
-static int enumerator_for_parent(sd_device *d, sd_device_enumerator **ret) {
-        _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL;
-        int r;
-
-        assert(d);
-        assert(ret);
-
-        r = sd_device_enumerator_new(&e);
-        if (r < 0)
-                return r;
-
-        r = sd_device_enumerator_add_match_subsystem(e, "block", true);
-        if (r < 0)
-                return r;
-
-        r = sd_device_enumerator_add_match_parent(e, d);
-        if (r < 0)
-                return r;
-
-        r = sd_device_enumerator_add_match_sysattr(e, "partition", NULL, true);
-        if (r < 0)
-                return r;
-
-        *ret = TAKE_PTR(e);
-        return 0;
-}
-
-static int device_is_partition(
-                sd_device *d,
-                sd_device *expected_parent,
-                blkid_partition pp) {
-
-        const char *v, *parent_syspath, *expected_parent_syspath;
-        blkid_loff_t bsize, bstart;
-        uint64_t size, start;
-        int partno, bpartno, r;
-        sd_device *parent;
-
-        assert(d);
-        assert(expected_parent);
-        assert(pp);
-
-        r = sd_device_get_subsystem(d, &v);
-        if (r < 0)
-                return r;
-        if (!streq(v, "block"))
-                return false;
-
-        if (sd_device_get_devtype(d, &v) < 0 || !streq(v, "partition"))
-                return false;
-
-        r = sd_device_get_parent(d, &parent);
-        if (r < 0)
-                return false; /* Doesn't have a parent? No relevant to us */
-
-        r = sd_device_get_syspath(parent, &parent_syspath); /* Check parent of device of this action */
-        if (r < 0)
-                return r;
-
-        r = sd_device_get_syspath(expected_parent, &expected_parent_syspath); /* Check parent of device we are looking for */
-        if (r < 0)
-                return r;
-
-        if (!path_equal(parent_syspath, expected_parent_syspath))
-                return false; /* Has a different parent than what we need, not interesting to us */
-
-        /* On kernel uevents we may find the partition number in the PARTN= field. Let's use that preferably,
-         * since it's cheaper and more importantly: the sysfs attribute "partition" appears to become
-         * available late, hence let's use the property instead, which is available at the moment we see the
-         * uevent. */
-        r = sd_device_get_property_value(d, "PARTN", &v);
-        if (r == -ENOENT)
-                r = sd_device_get_sysattr_value(d, "partition", &v);
-        if (r < 0)
-                return r;
-
-        r = safe_atoi(v, &partno);
-        if (r < 0)
-                return r;
-
-        errno = 0;
-        bpartno = blkid_partition_get_partno(pp);
-        if (bpartno < 0)
-                return errno_or_else(EIO);
-
-        if (partno != bpartno)
-                return false;
-
-        r = sd_device_get_sysattr_value(d, "start", &v);
-        if (r < 0)
-                return r;
-        r = safe_atou64(v, &start);
-        if (r < 0)
-                return r;
-
-        errno = 0;
-        bstart = blkid_partition_get_start(pp);
-        if (bstart < 0)
-                return errno_or_else(EIO);
-
-        if (start != (uint64_t) bstart)
-                return false;
-
-        r = sd_device_get_sysattr_value(d, "size", &v);
-        if (r < 0)
-                return r;
-        r = safe_atou64(v, &size);
-        if (r < 0)
-                return r;
-
-        errno = 0;
-        bsize = blkid_partition_get_size(pp);
-        if (bsize < 0)
-                return errno_or_else(EIO);
-
-        if (size != (uint64_t) bsize)
-                return false;
-
-        return true;
-}
-
-static int find_partition(
-                sd_device *parent,
-                blkid_partition pp,
-                usec_t timestamp_not_before,
-                DissectImageFlags flags,
-                sd_device **ret) {
-
-        _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL;
-        sd_device *q;
-        int r;
-
-        assert(parent);
-        assert(pp);
-        assert(ret);
-
-        r = enumerator_for_parent(parent, &e);
-        if (r < 0)
-                return r;
-
-        FOREACH_DEVICE(e, q) {
-                uint64_t usec;
-
-                if (!FLAGS_SET(flags, DISSECT_IMAGE_NO_UDEV)) {
-                        r = sd_device_get_usec_initialized(q, &usec);
-                        if (r == -EBUSY) /* Not initialized yet */
-                                continue;
-                        if (r < 0)
-                                return r;
-
-                        if (timestamp_not_before != USEC_INFINITY &&
-                            usec < timestamp_not_before) /* udev database entry older than our attachment? Then it's not ours */
-                                continue;
-                }
-
-                r = device_is_partition(q, parent, pp);
-                if (r < 0)
-                        return r;
-                if (r > 0) {
-                        *ret = sd_device_ref(q);
-                        return 0;
-                }
-        }
-
-        return -ENXIO;
-}
-
-struct wait_data {
-        sd_device *parent_device;
-        blkid_partition blkidp;
-        sd_device *found;
-        uint64_t diskseq;
-        uint64_t uevent_seqnum_not_before;
-        usec_t timestamp_not_before;
-        DissectImageFlags flags;
-};
-
-static inline void wait_data_done(struct wait_data *d) {
-        sd_device_unref(d->found);
-}
-
-static int device_monitor_handler(sd_device_monitor *monitor, sd_device *device, void *userdata) {
-        struct wait_data *w = userdata;
-        int r;
-
-        assert(w);
-
-        if (device_for_action(device, SD_DEVICE_REMOVE))
-                return 0;
-
-        if (w->diskseq != 0) {
-                uint64_t diskseq;
-
-                /* If w->diskseq is non-zero, then we must have a disk seqnum */
-                r = sd_device_get_diskseq(device, &diskseq);
-                if (r < 0) {
-                        log_debug_errno(r, "Dropping event because it has no diskseq, but waiting for %" PRIu64, w->diskseq);
-                        return 0;
-                }
-                if (diskseq < w->diskseq) {
-                        log_debug("Dropping event because diskseq too old (%" PRIu64 " < %" PRIu64 ")",
-                                  diskseq, w->diskseq);
-                        return 0;
-                }
-                if (diskseq > w->diskseq) {
-                        r = -EBUSY;
-                        goto finish; /* Newer than what we were expecting, so we missed it, stop waiting */
-                }
-        } else if (w->uevent_seqnum_not_before != UINT64_MAX) {
-                uint64_t seqnum;
-
-                r = sd_device_get_seqnum(device, &seqnum);
-                if (r < 0)
-                        goto finish;
-
-                if (seqnum <= w->uevent_seqnum_not_before) { /* From an older use of this loop device */
-                        log_debug("Dropping event because seqnum too old (%" PRIu64 " <= %" PRIu64 ")",
-                                  seqnum, w->uevent_seqnum_not_before);
-                        return 0;
-                }
-        }
-
-        r = device_is_partition(device, w->parent_device, w->blkidp);
-        if (r < 0)
-                goto finish;
-        if (r == 0) /* Not the one we need */
-                return 0;
-
-        /* It's the one we need! Yay! */
-        assert(!w->found);
-        w->found = sd_device_ref(device);
-        r = 0;
-
-finish:
-        return sd_event_exit(sd_device_monitor_get_event(monitor), r);
-}
-
-static int timeout_handler(sd_event_source *s, uint64_t usec, void *userdata) {
-        struct wait_data *w = userdata;
-        int r;
-
-        assert(w);
-
-        /* Why partition not appeared within the timeout? We may lost some uevent, as some properties
-         * were not ready when we received uevent... Not sure, but anyway, let's try to find the
-         * partition again before give up. */
-
-        r = find_partition(w->parent_device, w->blkidp, w->timestamp_not_before, w->flags, &w->found);
-        if (r == -ENXIO)
-                return log_debug_errno(SYNTHETIC_ERRNO(ETIMEDOUT),
-                                       "Partition still not appeared after timeout reached.");
-        if (r < 0)
-                return log_debug_errno(r, "Failed to find partition: %m");
-
-        log_debug("Partition appeared after timeout reached.");
-        return sd_event_exit(sd_event_source_get_event(s), 0);
-}
-
-static int retry_handler(sd_event_source *s, uint64_t usec, void *userdata) {
-        struct wait_data *w = userdata;
-        int r;
-
-        assert(w);
-
-        r = find_partition(w->parent_device, w->blkidp, w->timestamp_not_before, w->flags, &w->found);
-        if (r != -ENXIO) {
-                if (r < 0)
-                        return log_debug_errno(r, "Failed to find partition: %m");
-
-                log_debug("Partition found by a periodic search.");
-                return sd_event_exit(sd_event_source_get_event(s), 0);
-        }
-
-        r = sd_event_source_set_time_relative(s, 500 * USEC_PER_MSEC);
-        if (r < 0)
-                return r;
-
-        return sd_event_source_set_enabled(s, SD_EVENT_ONESHOT);
-}
-
-static int wait_for_partition_device(
-                sd_device *parent,
-                blkid_partition pp,
-                usec_t deadline,
-                uint64_t diskseq,
-                uint64_t uevent_seqnum_not_before,
-                usec_t timestamp_not_before,
-                DissectImageFlags flags,
-                sd_device **ret) {
-
-        _cleanup_(sd_event_source_unrefp) sd_event_source *timeout_source = NULL, *retry_source = NULL;
-        _cleanup_(sd_device_monitor_unrefp) sd_device_monitor *monitor = NULL;
-        _cleanup_(sd_event_unrefp) sd_event *event = NULL;
-        int r;
-
-        assert(parent);
-        assert(pp);
-        assert(ret);
-
-        r = find_partition(parent, pp, timestamp_not_before, flags, ret);
-        if (r != -ENXIO)
-                return r;
-
-        r = sd_event_new(&event);
-        if (r < 0)
-                return r;
-
-        r = sd_device_monitor_new(&monitor);
-        if (r < 0)
-                return r;
-
-        r = sd_device_monitor_filter_add_match_subsystem_devtype(monitor, "block", "partition");
-        if (r < 0)
-                return r;
-
-        r = sd_device_monitor_filter_add_match_parent(monitor, parent, true);
-        if (r < 0)
-                return r;
-
-        r = sd_device_monitor_filter_add_match_sysattr(monitor, "partition", NULL, true);
-        if (r < 0)
-                return r;
-
-        r = sd_device_monitor_attach_event(monitor, event);
-        if (r < 0)
-                return r;
-
-        _cleanup_(wait_data_done) struct wait_data w = {
-                .parent_device = parent,
-                .blkidp = pp,
-                .diskseq = diskseq,
-                .uevent_seqnum_not_before = uevent_seqnum_not_before,
-                .timestamp_not_before = timestamp_not_before,
-                .flags = flags,
-        };
-
-        r = sd_device_monitor_start(monitor, device_monitor_handler, &w);
-        if (r < 0)
-                return r;
-
-        /* Check again, the partition might have appeared in the meantime */
-        r = find_partition(parent, pp, timestamp_not_before, flags, ret);
-        if (r != -ENXIO)
-                return r;
-
-        if (deadline != USEC_INFINITY) {
-                r = sd_event_add_time(
-                                event, &timeout_source,
-                                CLOCK_MONOTONIC, deadline, 0,
-                                timeout_handler, &w);
-                if (r < 0)
-                        return r;
-
-                r = sd_event_source_set_exit_on_failure(timeout_source, true);
-                if (r < 0)
-                        return r;
-        }
-
-        /* If we don't have a disk sequence number then we cannot do exact matching,
-         * and we cannot know if we missed it or if it has not been sent yet, so set
-         * up additional retries to increase the chances of receiving the event. */
-        if (diskseq == 0) {
-                r = sd_event_add_time_relative(
-                                event, &retry_source,
-                                CLOCK_MONOTONIC, 500 * USEC_PER_MSEC, 0,
-                                retry_handler, &w);
-                if (r < 0)
-                        return r;
-
-                r = sd_event_source_set_exit_on_failure(retry_source, true);
-                if (r < 0)
-                        return r;
-        }
-
-        r = sd_event_loop(event);
-        if (r < 0)
-                return r;
-
-        assert(w.found);
-        *ret = TAKE_PTR(w.found);
-        return 0;
-}
-
 static void check_partition_flags(
                 const char *node,
                 unsigned long long pflags,
@@ -530,77 +148,8 @@ static void check_partition_flags(
                 log_debug("Unexpected partition flag %llu set on %s!", bit, node);
         }
 }
-
-static int device_wait_for_initialization_harder(
-                sd_device *device,
-                const char *subsystem,
-                usec_t deadline,
-                sd_device **ret) {
-
-        usec_t start, left, retrigger_timeout;
-        int r;
-
-        start = now(CLOCK_MONOTONIC);
-        left = usec_sub_unsigned(deadline, start);
-
-        if (DEBUG_LOGGING) {
-                const char *sn = NULL;
-
-                (void) sd_device_get_sysname(device, &sn);
-                log_device_debug(device,
-                                 "Will wait up to %s for '%s' to initialize…", FORMAT_TIMESPAN(left, 0), strna(sn));
-        }
-
-        if (left != USEC_INFINITY)
-                retrigger_timeout = CLAMP(left / 4, 1 * USEC_PER_SEC, 5 * USEC_PER_SEC); /* A fourth of the total timeout, but let's clamp to 1s…5s range */
-        else
-                retrigger_timeout = 2 * USEC_PER_SEC;
-
-        for (;;) {
-                usec_t local_deadline, n;
-                bool last_try;
-
-                n = now(CLOCK_MONOTONIC);
-                assert(n >= start);
-
-                /* Find next deadline, when we'll retrigger */
-                local_deadline = start +
-                        DIV_ROUND_UP(n - start, retrigger_timeout) * retrigger_timeout;
-
-                if (deadline != USEC_INFINITY && deadline <= local_deadline) {
-                        local_deadline = deadline;
-                        last_try = true;
-                } else
-                        last_try = false;
-
-                r = device_wait_for_initialization(device, subsystem, local_deadline, ret);
-                if (r >= 0 && DEBUG_LOGGING) {
-                        const char *sn = NULL;
-
-                        (void) sd_device_get_sysname(device, &sn);
-                        log_device_debug(device,
-                                         "Successfully waited for device '%s' to initialize for %s.",
-                                         strna(sn),
-                                         FORMAT_TIMESPAN(usec_sub_unsigned(now(CLOCK_MONOTONIC), start), 0));
-
-                }
-                if (r != -ETIMEDOUT || last_try)
-                        return r;
-
-                if (DEBUG_LOGGING)
-                        log_device_debug(device,
-                                         "Device didn't initialize within %s, assuming lost event. Retriggering device.",
-                                         FORMAT_TIMESPAN(usec_sub_unsigned(now(CLOCK_MONOTONIC), start), 0));
-
-                r = sd_device_trigger(device, SD_DEVICE_CHANGE);
-                if (r < 0)
-                        return r;
-        }
-}
 #endif
 
-#define DEVICE_TIMEOUT_USEC (45 * USEC_PER_SEC)
-
 static void dissected_partition_done(DissectedPartition *p) {
         assert(p);
 
@@ -617,6 +166,62 @@ static void dissected_partition_done(DissectedPartition *p) {
         };
 }
 
+#if HAVE_BLKID
+static int ioctl_partition_add(
+                int fd,
+                const char *name,
+                int nr,
+                uint64_t start,
+                uint64_t size) {
+
+        assert(fd >= 0);
+        assert(name);
+        assert(nr > 0);
+
+        struct blkpg_partition bp = {
+                .pno = nr,
+                .start = start,
+                .length = size,
+        };
+
+        struct blkpg_ioctl_arg ba = {
+                .op = BLKPG_ADD_PARTITION,
+                .data = &bp,
+                .datalen = sizeof(bp),
+        };
+
+        if (strlen(name) >= sizeof(bp.devname))
+                return -EINVAL;
+
+        strcpy(bp.devname, name);
+
+        return RET_NERRNO(ioctl(fd, BLKPG, &ba));
+}
+
+static int make_partition_devname(
+                const char *whole_devname,
+                int nr,
+                char **ret) {
+
+        bool need_p;
+
+        assert(whole_devname);
+        assert(nr > 0);
+
+        /* Given a whole block device node name (e.g. /dev/sda or /dev/loop7) generate a partition device
+         * name (e.g. /dev/sda7 or /dev/loop7p5). The rule the kernel uses is simple: if whole block device
+         * node name ends in a digit, then suffix a 'p', followed by the partition number. Otherwise, just
+         * suffix the partition number without any 'p'. */
+
+        if (isempty(whole_devname)) /* Make sure there *is* a last char */
+                return -EINVAL;
+
+        need_p = strchr(DIGITS, whole_devname[strlen(whole_devname)-1]); /* Last char a digit? */
+
+        return asprintf(ret, "%s%s%i", whole_devname, need_p ? "p" : "", nr);
+}
+#endif
+
 int dissect_image(
                 int fd,
                 const VeritySettings *verity,
@@ -638,11 +243,10 @@ int dissect_image(
         _cleanup_(blkid_free_probep) blkid_probe b = NULL;
         _cleanup_free_ char *generic_node = NULL;
         sd_id128_t generic_uuid = SD_ID128_NULL;
-        const char *pttype = NULL, *sysname = NULL;
+        const char *pttype = NULL, *sysname = NULL, *devname = NULL;
         blkid_partlist pl;
         int r, generic_nr = -1, n_partitions;
         struct stat st;
-        usec_t deadline;
 
         assert(fd >= 0);
         assert(ret);
@@ -698,23 +302,6 @@ int dissect_image(
         if (r < 0)
                 return r;
 
-        if (!FLAGS_SET(flags, DISSECT_IMAGE_NO_UDEV)) {
-                _cleanup_(sd_device_unrefp) sd_device *initialized = NULL;
-
-                /* If udev support is enabled, then let's wait for the device to be initialized before we doing anything. */
-
-                r = device_wait_for_initialization_harder(
-                                d,
-                                "block",
-                                usec_add(now(CLOCK_MONOTONIC), DEVICE_TIMEOUT_USEC),
-                                &initialized);
-                if (r < 0)
-                        return r;
-
-                sd_device_unref(d);
-                d = TAKE_PTR(initialized);
-        }
-
         b = blkid_new_probe();
         if (!b)
                 return -ENOMEM;
@@ -770,6 +357,9 @@ int dissect_image(
                 if (r < 0)
                         return r;
         }
+        r = sd_device_get_devname(d, &devname);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to get device devname: %m");
 
         if (!image_name_is_valid(m->image_name)) {
                 log_debug("Image name %s is not valid, ignoring", strempty(m->image_name));
@@ -785,8 +375,8 @@ int dissect_image(
 
                 (void) blkid_probe_lookup_value(b, "USAGE", &usage, NULL);
                 if (STRPTR_IN_SET(usage, "filesystem", "crypto")) {
-                        const char *fstype = NULL, *options = NULL, *devname = NULL;
                         _cleanup_free_ char *t = NULL, *n = NULL, *o = NULL;
+                        const char *fstype = NULL, *options = NULL;
 
                         /* OK, we have found a file system, that's our root partition then. */
                         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
@@ -797,10 +387,6 @@ int dissect_image(
                                         return -ENOMEM;
                         }
 
-                        r = sd_device_get_devname(d, &devname);
-                        if (r < 0)
-                                return r;
-
                         n = strdup(devname);
                         if (!n)
                                 return -ENOMEM;
@@ -873,13 +459,11 @@ int dissect_image(
         if (n_partitions < 0)
                 return errno_or_else(EIO);
 
-        deadline = usec_add(now(CLOCK_MONOTONIC), DEVICE_TIMEOUT_USEC);
         for (int i = 0; i < n_partitions; i++) {
-                _cleanup_(sd_device_unrefp) sd_device *q = NULL;
+                _cleanup_free_ char *node = NULL;
                 unsigned long long pflags;
                 blkid_loff_t start, size;
                 blkid_partition pp;
-                const char *node;
                 int nr;
 
                 errno = 0;
@@ -887,14 +471,6 @@ int dissect_image(
                 if (!pp)
                         return errno_or_else(EIO);
 
-                r = wait_for_partition_device(d, pp, deadline, diskseq, uevent_seqnum_not_before, timestamp_not_before, flags, &q);
-                if (r < 0)
-                        return r;
-
-                r = sd_device_get_devname(q, &node);
-                if (r < 0)
-                        return r;
-
                 pflags = blkid_partition_get_flags(pp);
 
                 errno = 0;
@@ -916,6 +492,31 @@ int dissect_image(
 
                 assert((uint64_t) size < UINT64_MAX/512);
 
+                r = make_partition_devname(devname, nr, &node);
+                if (r < 0)
+                        return r;
+
+                /* So here's the thing: after the main ("whole") block device popped up it might take a while
+                 * before the kernel fully probed the partition table. Waiting for that to finish is icky in
+                 * userspace. So here's what we do instead. We issue the BLKPG_ADD_PARTITION ioctl to add the
+                 * partition ourselves, racing against the kernel. Good thing is: if this call fails with
+                 * EBUSY then the kernel was quicker than us, and that's totally OK, the outcome is good for
+                 * us: the device node will exist. If OTOH our call was successful we won the race. Which is
+                 * also good as the outcome is the same: the partition block device exists, and we can use
+                 * it.
+                 *
+                 * Kernel returns EBUSY if there's already a partition by that number or an overlapping
+                 * partition already existent. */
+
+                r = ioctl_partition_add(fd, node, nr, (uint64_t) start * 512, (uint64_t) size * 512);
+                if (r < 0) {
+                        if (r != -EBUSY)
+                                return log_debug_errno(r, "BLKPG_ADD_PARTITION failed: %m");
+
+                        log_debug_errno(r, "Kernel was quicker than us in adding partition %i.", nr);
+                } else
+                        log_debug("We were quicker than kernel in adding partition %i.", nr);
+
                 if (is_gpt) {
                         PartitionDesignator designator = _PARTITION_DESIGNATOR_INVALID;
                         Architecture architecture = _ARCHITECTURE_INVALID;
@@ -1447,7 +1048,7 @@ int dissect_image(
             (flags & DISSECT_IMAGE_GENERIC_ROOT) &&
             (!verity || !verity->root_hash || verity->designator != PARTITION_USR)) {
 
-                /* OK, we found nothing usable, then check if there's a single generic one distro, and use
+                /* OK, we found nothing usable, then check if there's a single generic partition, and use
                  * that. If the root hash was set however, then we won't fall back to a generic node, because
                  * the root hash decides. */
 
@@ -3379,6 +2980,11 @@ int mount_image_privately_interactively(
         if (r < 0)
                 return log_error_errno(r, "Failed to set up loopback device for %s: %m", image);
 
+        /* Make sure udevd doesn't issue BLKRRPART behind our backs */
+        r = loop_device_flock(d, LOCK_SH);
+        if (r < 0)
+                return r;
+
         r = dissect_image_and_warn(d->fd, image, &verity, NULL, d->diskseq, d->uevent_seqnum_not_before, d->timestamp_not_before, flags, &dissected_image);
         if (r < 0)
                 return r;
@@ -3405,6 +3011,10 @@ int mount_image_privately_interactively(
         if (r < 0)
                 return r;
 
+        r = loop_device_flock(d, LOCK_UN);
+        if (r < 0)
+                return r;
+
         if (decrypted_image) {
                 r = decrypted_image_relinquish(decrypted_image);
                 if (r < 0)
@@ -3485,6 +3095,10 @@ int verity_dissect_and_mount(
         if (r < 0)
                 return log_debug_errno(r, "Failed to create loop device for image: %m");
 
+        r = loop_device_flock(loop_device, LOCK_SH);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to lock loop device: %m");
+
         r = dissect_image(
                         loop_device->fd,
                         &verity,
@@ -3532,6 +3146,10 @@ int verity_dissect_and_mount(
         if (r < 0)
                 return log_debug_errno(r, "Failed to mount image: %m");
 
+        r = loop_device_flock(loop_device, LOCK_UN);
+        if (r < 0)
+                return log_debug_errno(r, "Failed to unlock loopback device: %m");
+
         /* If we got os-release values from the caller, then we need to match them with the image's
          * extension-release.d/ content. Return -EINVAL if there's any mismatch.
          * First, check the distro ID. If that matches, then check the new SYSEXT_LEVEL value if

src/shared/dissect-image.h

@@ -188,19 +188,18 @@ typedef enum DissectImageFlags {
         DISSECT_IMAGE_MOUNT_NON_ROOT_ONLY = 1 << 7,  /* Mount only the non-root and non-/usr partitions */
         DISSECT_IMAGE_VALIDATE_OS         = 1 << 8,  /* Refuse mounting images that aren't identifiable as OS images */
         DISSECT_IMAGE_VALIDATE_OS_EXT     = 1 << 9,  /* Refuse mounting images that aren't identifiable as OS extension images */
-        DISSECT_IMAGE_NO_UDEV             = 1 << 10, /* Don't wait for udev initializing things */
-        DISSECT_IMAGE_RELAX_VAR_CHECK     = 1 << 11, /* Don't insist that the UUID of /var is hashed from /etc/machine-id */
-        DISSECT_IMAGE_FSCK                = 1 << 12, /* File system check the partition before mounting (no effect when combined with DISSECT_IMAGE_READ_ONLY) */
-        DISSECT_IMAGE_NO_PARTITION_TABLE  = 1 << 13, /* Only recognize single file system images */
-        DISSECT_IMAGE_VERITY_SHARE        = 1 << 14, /* When activating a verity device, reuse existing one if already open */
-        DISSECT_IMAGE_MKDIR               = 1 << 15, /* Make top-level directory to mount right before mounting, if missing */
-        DISSECT_IMAGE_USR_NO_ROOT         = 1 << 16, /* If no root fs is in the image, but /usr is, then allow this (so that we can mount the rootfs as tmpfs or so */
-        DISSECT_IMAGE_REQUIRE_ROOT        = 1 << 17, /* Don't accept disks without root partition (or at least /usr partition if DISSECT_IMAGE_USR_NO_ROOT is set) */
-        DISSECT_IMAGE_MOUNT_READ_ONLY     = 1 << 18, /* Make mounts read-only */
+        DISSECT_IMAGE_RELAX_VAR_CHECK     = 1 << 10, /* Don't insist that the UUID of /var is hashed from /etc/machine-id */
+        DISSECT_IMAGE_FSCK                = 1 << 11, /* File system check the partition before mounting (no effect when combined with DISSECT_IMAGE_READ_ONLY) */
+        DISSECT_IMAGE_NO_PARTITION_TABLE  = 1 << 12, /* Only recognize single file system images */
+        DISSECT_IMAGE_VERITY_SHARE        = 1 << 13, /* When activating a verity device, reuse existing one if already open */
+        DISSECT_IMAGE_MKDIR               = 1 << 14, /* Make top-level directory to mount right before mounting, if missing */
+        DISSECT_IMAGE_USR_NO_ROOT         = 1 << 15, /* If no root fs is in the image, but /usr is, then allow this (so that we can mount the rootfs as tmpfs or so */
+        DISSECT_IMAGE_REQUIRE_ROOT        = 1 << 16, /* Don't accept disks without root partition (or at least /usr partition if DISSECT_IMAGE_USR_NO_ROOT is set) */
+        DISSECT_IMAGE_MOUNT_READ_ONLY     = 1 << 17, /* Make mounts read-only */
         DISSECT_IMAGE_READ_ONLY           = DISSECT_IMAGE_DEVICE_READ_ONLY |
                                             DISSECT_IMAGE_MOUNT_READ_ONLY,
-        DISSECT_IMAGE_GROWFS              = 1 << 19, /* Grow file systems in partitions marked for that to the size of the partitions after mount */
-        DISSECT_IMAGE_MOUNT_IDMAPPED      = 1 << 20, /* Mount mounts with kernel 5.12-style userns ID mapping, if file system type doesn't support uid=/gid= */
+        DISSECT_IMAGE_GROWFS              = 1 << 18, /* Grow file systems in partitions marked for that to the size of the partitions after mount */
+        DISSECT_IMAGE_MOUNT_IDMAPPED      = 1 << 19, /* Mount mounts with kernel 5.12-style userns ID mapping, if file system type doesn't support uid=/gid= */
 } DissectImageFlags;
 
 struct DissectedImage {

src/sysext/sysext.c

@@ -533,6 +533,10 @@ static int merge_subprocess(Hashmap *images, const char *workspace) {
                         if (r < 0)
                                 return log_error_errno(r, "Failed to set up loopback device for %s: %m", img->path);
 
+                        r = loop_device_flock(d, LOCK_SH);
+                        if (r < 0)
+                                return log_error_errno(r, "Failed to lock loopback device: %m");
+
                         r = dissect_image_and_warn(
                                         d->fd,
                                         img->path,

src/test/test-loop-block.c

@@ -55,6 +55,9 @@ static void* thread_func(void *ptr) {
 
                 log_notice("Acquired loop device %s, will mount on %s", loop->node, mounted);
 
+                /* Let's make sure udev doesn't call BLKRRPART in the background, while we try to mount the device. */
+                assert_se(loop_device_flock(loop, LOCK_SH) >= 0);
+
                 r = dissect_image(loop->fd, NULL, NULL, loop->diskseq, loop->uevent_seqnum_not_before, loop->timestamp_not_before, DISSECT_IMAGE_READ_ONLY, &dissected);
                 if (r < 0)
                         log_error_errno(r, "Failed dissect loopback device %s: %m", loop->node);
@@ -85,6 +88,10 @@ static void* thread_func(void *ptr) {
                 log_notice_errno(r, "Mounted %s → %s: %m", loop->node, mounted);
                 assert_se(r >= 0);
 
+                /* Now the block device is mounted, we don't need no manual lock anymore, the devices are now
+                 * pinned by the mounts. */
+                assert_se(loop_device_flock(loop, LOCK_UN) >= 0);
+
                 log_notice("Unmounting %s", mounted);
                 mounted = umount_and_rmdir_and_free(mounted);
 
@@ -158,12 +165,6 @@ static int run(int argc, char *argv[]) {
                 return EXIT_TEST_SKIP;
         }
 
-        if (strstr_ptr(ci_environment(), "autopkgtest") || strstr_ptr(ci_environment(), "github-actions")) {
-                // FIXME: we should reenable this one day
-                log_tests_skipped("Skipping test on Ubuntu autopkgtest CI/GH Actions, test too slow and installed udev too flakey.");
-                return EXIT_TEST_SKIP;
-        }
-
         /* This is a test for the loopback block device setup code and it's use by the image dissection
          * logic: since the kernel APIs are hard use and prone to races, let's test this in a heavy duty
          * test: we open a bunch of threads and repeatedly allocate and deallocate loopback block devices in
@@ -221,6 +222,11 @@ static int run(int argc, char *argv[]) {
         pthread_t threads[arg_n_threads];
         sd_id128_t id;
 
+        /* Take an explicit lock while we format the file systems, in accordance with
+         * https://systemd.io/BLOCK_DEVICE_LOCKING/. We don't want udev to interfere and probe while we write
+         * or even issue BLKRRPART or similar while we are working on this. */
+        assert_se(loop_device_flock(loop, LOCK_EX) >= 0);
+
         assert_se(dissect_image(loop->fd, NULL, NULL, loop->diskseq, loop->uevent_seqnum_not_before, loop->timestamp_not_before, 0, &dissected) >= 0);
 
         assert_se(dissected->partitions[PARTITION_ESP].found);
@@ -249,9 +255,21 @@ static int run(int argc, char *argv[]) {
 
         assert_se(mkdtemp_malloc(NULL, &mounted) >= 0);
 
+        /* We are particularly correct here, and now downgrade LOCK → LOCK_SH. That's because we are done
+         * with formatting the file systems, so we don't need the exclusive lock anymore. From now on a
+         * shared one is fine. This way udev can now probe the device if it wants, but still won't call
+         * BLKRRPART on it, and that's good, because that would destroy our partition table while we are at
+         * it. */
+        assert_se(loop_device_flock(loop, LOCK_SH) >= 0);
+
         /* This first (writable) mount will initialize the mount point dirs, so that the subsequent read-only ones can work */
         assert_se(dissected_image_mount(dissected, mounted, UID_INVALID, UID_INVALID, 0) >= 0);
 
+        /* Now we mounted everything, the partitions are pinned. Now it's fine to release the lock
+         * fully. This means udev could now issue BLKRRPART again, but that's OK given this will fail because
+         * we now mounted the device. */
+        assert_se(loop_device_flock(loop, LOCK_UN) >= 0);
+
         assert_se(umount_recursive(mounted, 0) >= 0);
         loop = loop_device_unref(loop);