Suggested in
https://github.com/systemd/systemd/pull/35892#discussion_r2180322856.
This is a tiny amount of code and does not warrant having a separate file
and spawning a separate instance of the compiler during the build.
Note: it took me a while to confirm that the contents of that table and
function don't end up in libsystemd.so. The issue is that they _are_ present in
it, unless LTO is used. We actually use link_whole[libbasic_static] for
libsystemd, so we end up with all that code there. LTO is needed to clean
that up.
Currently, if we use a cgroup namespace together with DelegateSubgroup=,
the subgroup becomes the root of the cgroup namespace because we move the
service process to the subgroup before we unshare the cgroup namespace, and
the current cgroup becomes the root of the cgroup namespace when we unshare
the cgroup namespace.
Let's fix the problem by not moving the service process to the subgroup until
we've unshared the cgroup namespace. Note that this doesn't break the primary use
case of CLONE_INTO_CGROUP since we still use it to immediately clone into the service
main cgroup, just not anymore into the subgroup, but this shouldn't matter in practice.
Additionally, we need special handling for control processes, as those *do*
need to get spawned into the subcgroup immediately if delegation is configured to
avoid violating the cgroupsv2 "no inner processes" rule.
Effectively, this leaves us with the following logic:
- In exec_spawn(), spawn into subgroup if we're spawning a control process
that needs to be spawned into a subgroup immediately. Otherwise, spawn into
main service cgroup.
- In exec_invoke(), move into subgroup early if we don't need a cgroup namespace.
Otherwise, move into subgroup after we've unshared the cgroup namespace.
`ExtensionImages=` and `ExtensionDirectories=` now let you specify
vpick-named extensions; however, since they just get set up once when
the service is started, you can't see newer versions without restarting
the service entirely. Here, also reload confext extensions when you
reload a service. This allows you to deploy a new version of some
configuration and have it picked up at reload time without interruption
to your workload.
Right now, we would only reload confext extensions and leave the sysext
ones behind, since it didn't seem prudent to swap out what is likely
program code at reload. This is made possible by only going for the
`SYSTEMD_CONFEXT_HIERARCHIES` overlays (which only contains `/etc`).
Implementation wise, this uses the new kernel API and two collaborating
child processes under the host & child namespaces in order to gather the
right FDs needed:
- (1) In child, set up the extension images and directories in a slave
mountns, and obtain their FDs.
- (2) Fork into a grandchild under target process namespace, and do a
"fake" unmount to obtain the FD of the underlying target folder
say /etc).
- (3) In the child again, set up new overlay under host NS rights.
We do not want to do I/O heavy jobs inline in PID1 blocking the state
machine, so add separate async states to handle this case.
Co-authored-by: Luca Boccassi <luca.boccassi@gmail.com>
If DefaultDependencies=no, /var/ may not be mounted yet when the service
is being started. Previously, In such case, if the service has
PrivateTmp=disconnected, the service manager created /var/tmp/ on the
root filesystem and mounted the disconnected private tmpfs there. That
poluted the root filesystem and disturbed gpt-auto-generator on next
boot, as /var/ would not be empty anymore. See issue #37258.
This changes PrivateTmp=disconnected as the following:
- If DefaultDependencies=no and RootDirectory=/RootImage= are not set,
then a private tmpfs is mounted _only_ on /tmp/, and set $TMPDIR=/tmp
environment variable to suggest the service to use /tmp/.
- If DefaultDependencies=yes and RootDirectory=/RootImage= are not set,
then implies RequiresMountsFor=/var/, though that is typically
redundant, but anyway. Hence, we can safely mount /var/tmp/.
- Otherwise, i.e. when one of RootDirectory=/RootImage= is set, behaves
as the same as the previous, as the private root filesystem for the
service is explicitly prepared by the service manager, and we can
safely mount a private tmpfs on /var/tmp/ without any extra
dependencies.
Fixes#37258.
Co-authored-by: Mike Yuan <me@yhndnzj.com>
We already have LOG_CONTEXT_PUSH_EXEC() which with two additions
does exactly the same as the custom logging macros, so let's get rid
of the custom logging macros and use LOG_CONTEXT_PUSH_EXEC() instead.
Currently there are various circular dependencies between headers
in core/. Let's get rid of these by making judicious use of forward
declarations and moving includes into implementation files instead of
having them in header files.
Getting rid of circular header includes simplifies the code and makes
various clang based tooling such as iwyu work much better on our code.
The most important change is getting rid of the manager.h include in
unit.h which is possible thanks to the previous commits. We also move
the OOMPolicy and StatusType enums to unit.h to remove the need for
other unit headers to include manager.h to get access to these enums.
This introduce LOG_ITEM() macro that checks arbitrary formats in
log_struct().
Then, drop _printf_ attribute from log_struct_internal(), as it does not
help so much, and compiler checked only the first format string.
Hopefully, this silences false-positive warnings by Coverity.
This delegates one or more namespaces to the service. Concretely,
this setting influences in which order we unshare namespaces. Delegated
namespaces are unshared *after* the user namespace is unshared. Other
namespaces are unshared *before* the user namespace is unshared.
Fixes#35369
(Note: we also change TEST-13-NSPAWN.machined.sh minimally here, because
it checks for byte precise output of a pty allocated for a service
invocation - which it's not going to get if it claims that the pty is an
all-powerful one. After all this PR ensures that we'll generate the new
OSC sequence on non-dumb terminals associated with services. Hence, set
TERM=dumb explicitly to ensure no ANSI sequences are generated, ever.
Which is a nice test btw that TERM=dumb really does its thing here.)
So far we conditioned the logic that issues ansi sequences for resetting
the TTY based on whether something is a pty is not (under the assumption
we need no reset on ptys, since they are shortlived).
This is simply wrong though. The pty that a container getty is invoked
on is generally long-lived: as long as the container is up, and it will
be reused between getty instances/sessions all the time. In such a case
we really should reset properly.
Let's instead make the logic dependent on whether TERM is set to
anything other than "dumb". The previous commit made sure we always set
TERM in a sensible way in systemd-run, hence this
*explicit* logic sounds like a much better choice now, as it mea
To be able to run systemd in a Type=notify transient unit, the notify
socket can't be bind mounted to /run/systemd/notify as systemd in the
transient unit wants to use that as its own notify socket which conflicts
with systemd on the host.
Instead, for sandboxed units, let's bind mount the notify socket to
/run/host/notify as documented in the container interface. Since we don't
guarantee a stable location for the notify socket and insist users use
$NOTIFY_SOCKET to get its path, this is safe to do.
Migrating ProtectHostname to enum will set the stage for adding more
properties like ProtectHostname=private in future commits.
In addition, we add PrivateHostnameEx property to dbus API which uses
string instead of boolean.
This new setting allows unsharing the pid namespace in a unit. Because
you have to fork to get a process into a pid namespace, we fork in
systemd-executor to get into the new pid namespace. The parent then
sends the pid of the child process back to the manager and exits while
the child process continues on with the rest of exec_invoke() and then
executes the actual payload.
Communicating the child pid is done via a new pidref socket pair that is
set up on manager startup.
We unshare the PID namespace right before the mount namespace so we
mount procfs correctly. Note PrivatePIDs=yes always implies MountAPIVFS=yes
to mount procfs.
When running unprivileged in a user session, user namespace is set up first
to allow for PID namespace to be unshared. However, when running in
privileged mode, we unshare the user namespace last to ensure the user
namespace does not own the PID namespace and cannot break out of the sandbox.
Note we disallow Type=forking services from using PrivatePIDs=yes since the
init proess inside the PID namespace must not exit for other processes in
the namespace to exist.
Note Daan De Meyer did the original work for this commit with Ryan Wilson
addressing follow-ups.
Co-authored-by: Daan De Meyer <daan.j.demeyer@gmail.com>
When an exec directory is shared between services, this allows one of the
service to be the producer of files, and the other the consumer, without
letting the consumer modify the shared files.
This will be especially useful in conjunction with id-mapped exec directories
so that fully sandboxed services can share directories in one direction, safely.
Let's make ConfigurationDirectory= a bit less "special-casey", by hiding
the fact that it's the only per-service dir we do not do chown()ing for
inside of a new EXEC_DIRECTORY_TYPE_SHALL_CHOWN() helper.
This commit adds two settings private and strict to
the ProtectControlGroups= property. Private will unshare the cgroup
namespace and mount a read-write private cgroup2 filesystem at /sys/fs/cgroup.
Strict does the same except the mount is read-only. Since the unit is
running in a cgroup namespace, the new root of /sys/fs/cgroup is the unit's
own cgroup.
We also add a new dbus property ProtectControlGroupsEx which accepts strings
instead of boolean. This will allow users to use private/strict via dbus
and systemd-run in addition to service files.
Note private and strict fall back to no and yes respectively if the kernel
doesn't support cgroup2 or system is not using unified hierarchy.
Fixes: #34634
This commit refactors ProtectControlGroups= from using a boolean
in the dbus/execute backend to using an enum. There is no functional
change but this will allow adding new non-boolean values (e.g. strict,
private) a la PrivateHome.
This adds the ExtraFileDescriptor property to StartTransient dbus API
with format "a(hs)" - array of (file descriptor, name) pairs. The FD
will be passed to the unit via sd_notify like Socket and OpenFile.
systemctl show also shows ExtraFileDescriptorName for these transient
units. We only show the name passed to dbus as the FD numbers will
change once passed over the unix socket and are duplicated, so its
confusing to display the numbers.
We do not add this functionality for systemd-run or general systemd
service units as it is not useful for general systemd services.
Arguably, it could be useful for systemd-run in bash scripts but we
prefer to be cautious and not expose the API yet.
Fixes: #34396
One of the major pait points of managing fleets of headless nodes is
that when something fails at startup, unless debug level was already
enabled (which usually isn't, as it's a firehose), one needs to manually
enable it and pray the issue can be reproduced, which often is really
hard and time consuming, just to get extra info. Usually the extra log
messages are enough to triage an issue.
This new option makes it so that when a service fails and is restarted
due to Restart=, log level for that unit is set to debug, so that all
setup code in pid1 and sd-executor logs at debug level, and also a new
DEBUG_INVOCATION=1 env var is passed to the service itself, so that it
knows it should start with a higher log level. Once the unit succeeds
or reaches the rate limit the original level is restored.
This allows for "per-instance" credentials for units. The use case
is best explained with an example. Currently all our getty units
have the following stanzas in their unit file:
"""
ImportCredential=agetty.*
ImportCredential=login.*
"""
This means that setting agetty.autologin=root as a system credential
will make every instance of our all our getty units autologin as the
root user. This prevents us from doing autologin on /dev/hvc0 while
still requiring manual login on all other ttys.
To solve the issue, we introduce support for renaming credentials with
ImportCredential=. This will allow us to add the following to e.g.
serial-getty@.service:
"""
ImportCredential=tty.serial.%I.agetty.*:agetty.
ImportCredential=tty.serial.%I.login.*:login.
"""
which for serial-getty@hvc0.service will make the service manager read
all credentials of the form "tty.serial.hvc0.agetty.xxx" and pass them
to the service in the form "agetty.xxx" (same goes for login). We can
apply the same to each of the getty units to allow setting agetty and
login credentials for individual ttys instead of globally.
And while we are at it, merge exec_context_determine_tty_size() +
exec_context_apply_tty_size().
Let's simplify things, and merge the two funcs, since the latter just
does one more call.
At the same time, let's make sure we actually allow passing separate
input/output fds.
DynamicUser= enables PrivateTmp= implicitly to avoid files owned by reusable uids
leaking into the host. Change it to instead create a fully private tmpfs instance
instead, which also ensures the same result, since it has less impactful semantics
with respect to PrivateTmp=yes, which links the mount namespace to the host's /tmp
instead. If a user specifies PrivateTmp manually, let the existing behaviour
unchanged to ensure backward compatibility is not broken.
Otherwise there's a chance the handover timestamp is less than the
start timestamp depending on the whims of the scheduler.
For non-forking services, we also propagate the start timestamp into
Service->main_exec_status to make sure it starts earlier than the
handoff timestamp.
The expression for calculation of the _do_log values in the log_exec_*
macros need to be the same as the unit_log_level_test() function, used
to calculate _do_log in unit.h. The only difference between execute.h
and unit.h is the lack of the Unit structure.
Fixes: b646fc324a ("core: ensure execute/spawn functions can work without Unit object")
Fixes: 210ca71cb5 ("core/execute: clean up log_exec_full_errno and friends")
Signed-off-by: Łukasz Stelmach <l.stelmach@samsung.com>
Also: rename Handover → Handoff. I think it makes it clearer that this
is not really about handing over any resources, but that the executor is
out off the game from that point on.
This changes the executor to systematically send handoff timestamps to
the service manager if a socket for that is supplied. This drops the
code that did this via Type=exec messages, and reverts that part to the
old behaviour before 93cb78aee2.
Benefits of this approach:
1. We can collect the handoff for any command we fork off, regardless
if it's ExecStart= something else, regardless whether it's Type=exec,
Type=simple or some any other service type, regardless of the unit
type.
2. We collect both CLOCK_REALTIME and CLOCK_MONOTONIC, as we do for the
other process timestamps.
3. It's entirely backwards compatible, as this doesn't change the
protocol between service manager and executor, but just extends it.
Enable the exec_fd logic for Type=notify* services too, and change it
to send a timestamp instead of a '1' byte. Record the timestamp in a
new ExecMainHandoverTimestamp property so that users can track accurately
when control is handed over from systemd to the service payload, so
that latency and startup performance can be trivially and accurately
tracked and attributed.
fresh otherwise
Currently, exec_setup_credential() always rewrite all credentials
upon exec_invoke(), i.e. invocation of each ExecCommand, and within
a single tmpfs instance. This is problematic though:
* When writing each tmp cred file, we essentially double the size
of the credential. Therefore, if one cred is bigger than half
of CREDENTIALS_TOTAL_SIZE_MAX, confusing ENOSPC occurs (see also
https://github.com/systemd/systemd/pull/24734#issuecomment-1925440546)
* Credential is a unit-wide thing and thus should not change
during the whole lifetime of main process. However, if e.g.
a on-disk credential or SetCredential= in unit file
changes between ExecStart= and ExecStartPost=,
the credentials are overwritten when the latter gets to run,
and the already-running main process is suddenly seeing
completely different creds.
So, let's try to reuse final cred dir if the main process has started
and the tmpfs has been populated, so that the creds used is stable
across all ExecStart= and ExecStartPost=-s. We still want to retain
the ability of updating creds through ExecStartPre= though, therefore
we forcibly use a fresh cred dir for those. 'Fresh' means to actually
unmount the old tmpfs first, so the first problem goes away, too.
Zbigniew Jędrzejewski-Szmek
Andres Beltran
Yu Watanabe
Daan De Meyer
Jelle van der Waa
maia x.andLuca Boccassi
Lennart Poettering
Ryan Wilson
Luca Boccassi
Łukasz Stelmach