We're still seeing problems in mkosi CI where we lose notify messages
sent over vsock from virtual machines because the virtual machine shuts
down before the socket buffers are flushed.
Let's try setting SO_LINGER on vsock sockets so that the kernel waits
until the buffers are flushed when we close the file descriptor.
To make sure it works, let's get rid of our previous workaround that
sends EXIT_STATUS earlier to reduce the chance of it not arriving.
This way, if it doesn't work, we'll notice it immediately due to CI
failures.
There's a race condition where the EXIT_STATUS= message we send
just before shutting down the VM doesn't arrive on the host,
presumably because the VM is shut down before the kernel has had a
chance to forward the message to the host.
Since there's no obvious way to wait until the message has been
flushed to the host, let's send the message before we execute the
final sync() instead of after executing the final sync(). In my
testing, this seems to either guarantee the message is sent or
introduces sufficient delay that the kernel always has time to flush
its socket buffers to the host.
We use it for more than just pipe() arrays. For example also for
socketpair(). Hence let's give it a generic name.
Also add EBADF_TRIPLET to mirror this for things like
stdin/stdout/stderr arrays, which we use a bunch of times.
This reverts commit 6b219b74de.
This commit doesn't look right to me. We have to unmount everything
recursively *before* we MS_MOVE because the MS_MOVE will not get rid of
it for us, and we simply cannot access these mounts after the MS_MOVE is
complete anymore.
This is a fundamental difference between MS_MOVE and pivot_root(). The
latter repivots the entire mount table getting rid of anything outside
of the new root. MS_MOVE otoh just mounts a bunch of mount points to the
top, leaving in place whatever might be underneath it.
Thus, if we go through the MS_MOVE codepath we must unmount everything
explicitly before doing so because otherwise the mounts will be pinned
forever, but be entirely invisble to userspace.
Negative flags always raise eyebrows. Let's normalize
SWITCH_ROOT_SKIP_RECURSIVE_RUN to become SWITCH_ROOT_RECURSIVE_RUN, i.e.
make recursive behaviour opt-in, rather than opt-out. We only want it
for the initrd→host transition, and in all other cases we'd prefer to
avoid it.
This allows us to simplify some code. Also, normalize the mount point
table in switch_root() a bit, to be static const, and then just select
between two mount flag sets via SWITCH_ROOT_RECURSIVE_RUN, where the
mount flag without MS_BIND (i.e. zero) just means "skip this entry".
This does not actually change anything in behaviour, it's just
refactoring.
Follow-up for: b12d41a8bb
Whether we put some binaries in /bin or in /usr/bin should not have any
effect on unmounting during shutdown. Even if people split /usr/ off we
should not try to unmount it, no matter what as it is simply where
binaries are placed.
Let's make sure implicitly that the target directory is a mount point,
instead of doing so manually beforehand. This allows us to drop this
step from the transition into the /run/initramfs/ dir at shutdown.
During the initrd→host transition the switch root operations so far
where towards pre-existing mount points, but there are cetrainly
usecases where it might make sense to siwtch into arbitrary
subdirectories, too.
Our shutdown binary that takes over as PID 1 when shutting down puts
great efforts into a sync() that comes with a time-out once sync'ing
process stops. If we'd add another dumb sync() here, we kinda defeat all
it is good for. Hence, let's keep the sync() in for most codepats, but
let's disable it for the final shutdown logic when we transition back
into the exitrd. After all we sync()ed more than enough here, no need to
sync() even more.
Let's replace the current boolean param with a proper flags param. With
a single flag this doesn't appear to make much sense, though it does
already make things more readable I think.
However, once we add a second flag, it starts to make more sense.
Also, while we are at it, condition the "istmp" determinaton with this
flag too, since we only need it when the flag is set.
We previously would use MS_MOVE to move the old procfs, sysfs, /dev/ and
/run to the new place in some places, and MS_BIND in others.
The logic when to use MS_MOVE and when to use MS_BIND was pretty
arbitrary so far: we'd use MS_MOVE during the initrd → host transition
and MS_BIND when transitioning from host into the exitrd during
shutdown.
Traditionally, using MS_MOVE was preferable, because we didn't bother
with unmounting the old mount hierarchy before the switch root, and thus
using MS_MOVE did some clean-up as side-effect (because the old mounts
went away this way). But since we nowadays properly umount all remaining
mount points (since 268d1244e8) when
transitioning it's pointless.
Let's just use MS_BIND always. Let's tweak it though: let's use
MS_BIND|MS_REC for the kernel API VFS, and MS_BIND without MS_REC for
/run/. The latter reflects the fact that the submounts /run/ has usually
are not so much about just accessing kernel APIs but about auxiliary
user resources. Hence let's only move the main mount over for that.
While we are at it, also set up the base filesystem *before* we move the
mounts from the old to the new root, since the base filesystem setup
logic creates various needed inodes for us, which we really should make
use of instead of creating on our own.
This adds a new mechanism for rebooting, a form of "userspace reboot"
hereby dubbed "soft-reboot". It will stop all services as in a usual
shutdown, possibly transition into a new root fs and then issue a fresh
initial transaction. The kernel is not replaced.
File descriptors can be passed over, thus opening the door for leaving
certain resources around between such reboots.
Usecase: this is an extremely quick way to reset userspace fully when
updating image based systems, without going through a full
hardware/firmware/boot loader/kernel/initrd cycle. It minimizes "grayout time"
for OS updates. (In particular when combined with kernel live patching)
If we are not able to detach all MD/DM/loopback devices this is not
necessarily a failure, it's simply because we might be running off them.
Hence let's tone down our language a bit, and just say "Unable to"
rather than "Failed to".
While we are at it, let's also clean this up a bit: unlike DM/MD devices
loopback devices are likely partitioned, hence trace the block device
through the partition layer and LUKS.
So far detach-swap.[ch] were still using the MountPoint structure to
store swap device info in. Since it was only using a single field of it
sharing the whole structure is kinda pointless. Hence, let's decouple
this and only add the field we really need.
So far, we invoked pivot_root() specifying /mnt/ as second argument,
which then unmounted right-after. We'd create /mnt/ if needed. This
sucks, because it means /mnt/ must strictly be pre-created on immutable
images.
Remove this limitation, by using pivot_root() with "." as source and
target, which will result in two stacked mounts afterwards: the new one
underneath, the old one ontop. We can then simply unmount the top one,
and have what we want without needing any extra /mnt/ dir.
Since we don't need /mnt/ anymore we can get rid of the extra
unmount_old_root parameter and simply specify it as NULL if we don't
want the old mount to stick around.
When running in a container, we can propagate the exit status of
pid1 as usual via the process exit status. This is not possible
when running in a VM. Instead, let's send EXIT_STATUS=%i via the
notify socket if one is configured. The user running the VM can then
pick up the exit status from the notify socket after the VM has shut
down.
Otherwise, if getopt() and friends are used before parse_argv(), then
the GNU extensions may be ignored.
This should not change any behavior at least now, as we usually use
getopt_long() only once per invocation. But in the next commit,
getopt_long() will be used for other arrays, hence this change will
become necessary.
Chasing symlinks is a core function that's used in a lot of places
so it deservers a less verbose names so let's rename it to chase()
and chaseat().
We also slightly change the pattern used for the chaseat() helpers
so we get chase_and_openat() and similar.
Although this slightly more verbose it makes it much easier to reason
about. The code that produces the tests heavily benefits from this.
Test lists are also now sorted by test name.
When systemd-shutdown is run, shared propagation is not needed anymore. Making
mounts private allow to move mount points. This is needed to untangle cyclic
mounts.
There is a case that confuses systemd-shutdown: a filesystem has been moved to
a mount point which is part of another filesystem from an image from that
former filesystem. systemd-shutdown cannot unmount any of those two
filesystems. It needs first to move the filesystem containing the image of the
other out of the tree of that image.
Here we move leaf mount points when they are busy so that they do not block
parent mounts. We can only move leafs at each iteration since moving mount
points also move sub mount points which would invalidate we read from
`/proc/self/mountinfo`.
Daan De Meyer
Lennart Poettering
Yu Watanabe
OMOJOLA JOSHUA
Jan Janssen
Valentin David