Ideally we would like to hide all other service's credentials for all
services. That would imply for us to enable mount namespacing for all
services, which is something we cannot do, both due to compatibility
with the status quo ante, and because a number of services legitimately
should be able to install mounts in the host hierarchy.
Hence we do the second best thing, we hide the credentials automatically
for all services that opt into mount namespacing otherwise. This is
quite different from other mount sandboxing options: usually you have to
explicitly opt into each. However, given that the credentials logic is a
brand new concept we invented right here and now, and particularly
security sensitive it's OK to reverse this, and by default hide
credentials whenever we can (i.e. whenever mount namespacing is
otherwise opt-ed in to).
Long story short: if you want to hide other service's credentials, the
most basic options is to just turn on PrivateMounts= and there you go,
they should all be gone.
Kernel 5.8 gained a hidepid= implementation that is truly per procfs,
which allows us to mount a distinct once into every unit, with
individual hidepid= settings. Let's expose this via two new settings:
ProtectProc= (wrapping hidpid=) and ProcSubset= (wrapping subset=).
Replaces: #11670
it's not entirely clear what shall be passed via parameter and what via
struct, but these two definitely fit well with the other protect_xyz
fields, hence let's move them over.
We probably should move a lot more more fields into the structure
actuall (most? all even?).
Similarly to "setup" vs. "set up", "fallback" is a noun, and "fall back"
is the verb. (This is pretty clear when we construct a sentence in the
present continous: "we are falling back" not "we are fallbacking").
Follow the same model established for RootImage and RootImageOptions,
and allow to either append a single list of options or tuples of
partition_number:options.
The concept is flawed, and mostly useless. Let's finally remove it.
It has been deprecated since 90a2ec10f2 (6
years ago) and we started to warn since
55dadc5c57 (1.5 years ago).
Let's get rid of it altogether.
Previously, we'd create them from user-runtime-dir@.service. That has
one benefit: since this service runs privileged, we can create the full
set of device nodes. It has one major drawback though: it security-wise
problematic to create files/directories in directories as privileged
user in directories owned by unprivileged users, since they can use
symlinks to redirect what we want to do. As a general rule we hence
avoid this logic: only unpriv code should populate unpriv directories.
Hence, let's move this code to an appropriate place in the service
manager. This means we lose the inaccessible block device node, but
since there's already a fallback in place, this shouldn't be too bad.
Let's make /run/host the sole place we pass stuff from host to container
in and place the "inaccessible" nodes in /run/host too.
In contrast to the previous two commits this is a minor compat break, but
not a relevant one I think. Previously the container manager would place
these nodes in /run/systemd/inaccessible/ and that's where PID 1 in the
container would try to add them too when missing. Container manager and
PID 1 in the container would thus manage the same dir together.
With this change the container manager now passes an immutable directory
to the container and leaves /run/systemd entirely untouched, and managed
exclusively by PID 1 inside the container, which is nice to have clear
separation on who manages what.
In order to make sure systemd then usses the /run/host/inaccesible/
nodes this commit changes PID 1 to look for that dir and if it exists
will symlink it to /run/systemd/inaccessible.
Now, this will work fine if new nspawn and new pid 1 in the container
work together. as then the symlink is created and the difference between
the two dirs won't matter.
For the case where an old nspawn invokes a new PID 1: in this case
things work as they always worked: the dir is managed together.
For the case where different container manager invokes a new PID 1: in
this case the nodes aren't typically passed in, and PID 1 in the
container will try to create them and will likely fail partially (though
gracefully) when trying to create char/block device nodes. THis is fine
though as there are fallbacks in place for that case.
For the case where a new nspawn invokes an old PID1: this is were the
(minor) incompatibily happens: in this case new nspawn will place the
nodes in the /run/host/inaccessible/ subdir, but the PID 1 in the
container won't look for them there. Since the nodes are also not
pre-created in /run/systed/inaccessible/ PID 1 will try to create them
there as if a different container manager sets them up. This is of
course not sexy, but is not a total loss, since as mentioned fallbacks
are in place anyway. Hence I think it's OK to accept this minor
incompatibility.
arg_system == true and getpid() == 1 hold under the very same condition
this early in the main() function (this only changes later when we start
parsing command lines, where arg_system = true is set if users invoke us
in test mode even when getpid() != 1.
Hence, let's simplify things, and merge a couple of if branches and not
pretend they were orthogonal.
Some masks shouldn't be needed externally, so keep their functions in
the module (others would fit there too but they're used in tests) to
think twice if something would depend on them.
Drop unused function cg_attach_many_everywhere.
Use cgroup_realized instead of cgroup_path when we actually ask for
realized.
This should not cause any functional changes.
The usage in unit_get_own_mask is redundant, we only need apply
disable_mask at the end befor application, i.e. calculating enable or
target mask.
(IOW, we allow all configurations, but disabling affects effective
controls.)
Modify tests accordingly and add testing of enable mask.
This is intended as cleanup, with no effect but changing unit_dump
output.
The unit_add_siblings_to_cgroup_realize_queue does more than mere
siblings queueing, hence define a family of a unit as (immediate)
children of the unit and immediate children of all ancestors.
Working with this abstraction simplifies the queuing calls and it
shouldn't change the functionality.
Merge members mask invalidation into
unit_add_siblings_to_cgroup_realize_queue, this way unit_realize_cgroup
needn't be called with members mask invalidation.
We have to retain the members mask invalidation in unit_load -- although
active units would have cgroups (re)realized (unit_load queues for
realization), the realization would happen with potentially stale mask.
unit_free(u) realizes direct parent and invalidates members mask of all
ancestors. This isn't sufficient in v1 controller hierarchies since
siblings of the freed unit may have existed only because of the removed
unit.
We cannot be lazy about the siblings because if parent(u) is also
removed, it'd migrate and rmdir cgroups for siblings(u). However,
realized masks of siblings(u) won't reflect this change.
This was a non-issue earlier, because we weren't removing cgroup
directories properly (effectively matching the stale realized mask),
removal failed because of tasks left by missing migration (see previous
commit).
Therefore, ensure realization of all units necessary to clean up after
the free'd unit.
Fixes: #14149
When we are about to derealize a controller on v1 cgroup, we first
attempt to delete the controller cgroup and migrate afterwards. This
doesn't work in practice because populated cgroup cannot be deleted.
Furthermore, we leave out slices from migration completely, so
(un)setting a control value on them won't realize their controller
cgroup.
Rework actual realization, unit_create_cgroup() becomes
unit_update_cgroup() and make sure that controller hierarchies are
reduced when given controller cgroup ceased to be needed.
Note that with this we introduce slight deviation between v1 and v2 code
-- when a descendant unit turns off a delegated controller, we attempt
to disable it in ancestor slices. On v2 this may fail (kernel enforced,
because of child cgroups using the controller), on v1 we'll migrate
whole subtree and trim the subhierachy. (Previously, we wouldn't take
away delegated controller, however, derealization was broken anyway.)
Fixes: #14149
When available, enable memory_recursiveprot. Realistically it always
makes sense to delegate MemoryLow= and MemoryMin= to all children of a
slice/unit.
The kernel option is not enabled by default as it might cause
regressions in some setups. However, it is the better default in
general, and it results in a more flexible and obvious behaviour.
The alternative to using this option would be for user's to also set
DefaultMemoryLow= on slices when assigning MemoryLow=. However, this
makes the effect of MemoryLow= on some children less obvious, as it
could result in a lower protection rather than increasing it.
From the kernel documentation:
memory_recursiveprot
Recursively apply memory.min and memory.low protection to
entire subtrees, without requiring explicit downward
propagation into leaf cgroups. This allows protecting entire
subtrees from one another, while retaining free competition
within those subtrees. This should have been the default
behavior but is a mount-option to avoid regressing setups
relying on the original semantics (e.g. specifying bogusly
high 'bypass' protection values at higher tree levels).
This was added in kernel commit 8a931f801340c (mm: memcontrol:
recursive memory.low protection), which became available in 5.7 and was
subsequently fixed in kernel 5.7.7 (mm: memcontrol: handle div0 crash
race condition in memory.low).
Follows the same pattern and features as RootImage, but allows an
arbitrary mount point under / to be specified by the user, and
multiple values - like BindPaths.
Original implementation by @topimiettinen at:
https://github.com/systemd/systemd/pull/14451
Reworked to use dissect's logic instead of bare libmount() calls
and other review comments.
Thanks Topi for the initial work to come up with and implement
this useful feature.
This code was changed in this pull request:
https://github.com/systemd/systemd/pull/16571
After some discussion and more investigation, we better understand
what's going on. So, update the comment, so things are more clear
to future readers.
From a report in https://bugzilla.redhat.com/show_bug.cgi?id=1861463:
usb-gadget.target: Failed to load configuration: No such file or directory
usb-gadget.target: Failed to load configuration: No such file or directory
usb-gadget.target: Trying to enqueue job usb-gadget.target/start/fail
usb-gadget.target: Failed to load configuration: No such file or directory
Assertion '!bus_error_is_dirty(e)' failed at src/libsystemd/sd-bus/bus-error.c:239, function bus_error_setfv(). Ignoring.
sys-devices-platform-soc-2100000.bus-2184000.usb-ci_hdrc.0-udc-ci_hdrc.0.device: Failed to enqueue SYSTEMD_WANTS= job, ignoring: Unit usb-gadget.target not found.
I *think* this is the place where the reuse occurs: we call
bus_unit_validate_load_state(unit, e) twice in a row.
The explicit limit is dropped, which means that we return to the kernel default
of 50% of RAM. See 362a55fc14 for a discussion why that is not as much as it
seems. It turns out various applications need more space in /dev/shm and we
would break them by imposing a low limit.
While at it, rename the define and use a single macro for various tmpfs mounts.
We don't really care what the purpose of the given tmpfs is, so it seems
reasonable to use a single macro.
This effectively reverts part of 7d85383edb. Fixes#16617.
Allows to specify mount options for RootImage.
In case of multi-partition images, the partition number can be prefixed
followed by colon. Eg:
RootImageOptions=1:ro,dev 2:nosuid nodev
In absence of a partition number, 0 is assumed.
When we're disabling controller on a direct child of root cgroup, we
forgot to add root slice into cgroup realization queue, which prevented
proper disabling of the controller (on unified hierarchy).
The mechanism relying on "bounce from bottom and propagate up" in
unit_create_cgroup doesn't work on unified hierarchy (leaves needn't be
enabled). Drop it as we rely on the ancestors to be queued -- that's now
intentional but was artifact of combining the two patches:
cb5e3bc37d ("cgroup: Don't explicitly check for member in UNIT_BEFORE") v240~78
65f6b6bdcb ("core: fix re-realization of cgroup siblings") v245-rc1~153^2
Fixes: #14917
The current implementation is LIFO, which is a) confusing b) prevents
some ordered operations on the cgroup tree (e.g. removing children
before parents).
Fix it quickly. Current list implementation turns this from O(1) to O(n)
operation. Rework the lists later.
Previously, we assumed that success meant we definitely got a valid
pointer. There is at least one edge case where this is not true (i.e.,
we can get both a 0 return value, and *also* a NULL pointer):
4246bb550d/libselinux/src/procattr.c (L175)
When this case occurrs, if we don't check the pointer we SIGSEGV in
early initialization.
Fixes#16401.
c80a9a33d0 introduced the .can_fail field,
but didn't set it on .targets. Targets can fail through dependencies.
This leaves .slice and .device units as the types that cannot fail.
$ systemctl cat bad.service bad.target bad-fallback.service
[Service]
Type=oneshot
ExecStart=false
[Unit]
OnFailure=bad-fallback.service
[Service]
Type=oneshot
ExecStart=echo Fixing everythign!
$ sudo systemctl start bad.target
systemd[1]: Starting bad.service...
systemd[1]: bad.service: Main process exited, code=exited, status=1/FAILURE
systemd[1]: bad.service: Failed with result 'exit-code'.
systemd[1]: Failed to start bad.service.
systemd[1]: Dependency failed for bad.target.
systemd[1]: bad.target: Job bad.target/start failed with result 'dependency'.
systemd[1]: bad.target: Triggering OnFailure= dependencies.
systemd[1]: Starting bad-fallback.service...
echo[46901]: Fixing everythign!
systemd[1]: bad-fallback.service: Succeeded.
systemd[1]: Finished bad-fallback.service.
