When running an image that cannot be mounted (e.g.: key missing intentionally
for development purposes), there's a retry loop that takes some time
and slows development down. Add an env var to disable it.
varlink_server_listen_auto() is supposed to be the one-stop solution for
turning simple command line tools into IPC services. They aren't easy to
test/debug however, since you have to invoke them through a service
manager.
Let's make this easier: if the SYSTEMD_VARLINK_LISTEN env var is set,
let's listen on the socket specified therein. This makes things easier
to gdb: just run the service from the cmdline.
Both building and booting a directory image is much faster than
building or booting a disk image so let's default to a directory
image.
In CI, we stick to a disk image to make sure that keeps working as
well.
The only extra dependency this introduces is virtiofsd which is
packaged in all distributions except Debian stable. For users
hacking on systemd on Debian stable, a disk image can be built by
writing the following to mkosi.local.conf:
```
[Output]
Format=disk
```
To make things symmetric to the $SYSTEMD_SSH logic that the varlink
transport supports, let's also honour such a variable in sd-bus when
picking ssh transport.
This uses openssh 9.4's -W support for AF_UNIX. Unfortunately older versions
don't work with this, and I couldn#t figure a way that would work for
older versions too, would not be racy and where we'd still could keep
track of the forked off ssh process.
Unfortunately, on older versions -W will just hang (because it tries to
resolve the AF_UNIX path as regular host name), which sucks, but hopefully this
issue will go away sooner or later on its own, as distributions update.
Fedora is still stuck at 9.3 at the time of posting this (even on
Fedora), even though 9.4, 9.5, 9.6 have all already been released by
now.
Example:
varlinkctl call -j ssh:root@somehost:/run/systemd/io.systemd.Credentials io.systemd.Credentials.Encrypt '{"text":"foobar"}'
Otherwise, udev workers cannot detect slow programs invoked by
IMPORT{program}=, PROGRAM=, or RUN=, and whole worker process may be
killed.
Fixes#30436.
Co-authored-by: sushmbha <sushmita.bhattacharya@oracle.com>
Same as $KERNEL_INSTALL_BYPASS, but for hwdb. This will speed up
cross architecture image builds in mkosi as I can disable package
managers from running the costly hwdb update stuff in qemu user
mode and run it myself with a native systemd-hwdb with --root=.
Now that mkosi-kernel is a thing, this logic in systemd is just mostly
bitrotting since I just use mkosi-kernel these days. If I ever need to
hack on systemd and the kernel in tandem, I'll just add support for
building systemd to mkosi-kernel instead, so let's drop the support for
building a custom kernel in systemd's mkosi configuration.
This code doesn't link when gcc+lld is used:
$ LDFLAGS=-fuse-ld=lld meson setup build-lld && ninja -C build-lld udevadm
...
ld.lld: error: src/shared/libsystemd-shared-255.a(libsystemd-shared-255.a.p/cryptsetup-util.c.o):
symbol crypt_token_external_path@@ has undefined version
collect2: error: ld returned 1 exit status
As a work-around, restrict it to developer mode.
Closes https://github.com/systemd/systemd/issues/30218.
- Use mkosi.images/ instead of mkosi.presets/
- Use the .chroot suffix to run scripts in the image
- Use BuildSources= match for the kernel build
- Move 10-systemd.conf to mkosi.conf and rely on mkosi.local.conf
for local configuration
don't let the devices to be announced just as model "Linux". Let's instead
propagate the underlying block device's model. Also do something
reasonably smart for the serial and firmware version fields.
Introduce a new env variable $SYSTEMD_NSPAWN_CHECK_OS_RELEASE, that can
be used to disable the os-release check for bootable OS trees. Useful
when trying to boot a container with empty /etc/ and bind-mounted /usr/.
Resolves: #29185
I tried to get something similar upstream:
https://gitlab.com/cryptsetup/cryptsetup/-/issues/846
But no luck, it was suggested I use ELF interposition instead. Hence,
let's do so (but not via ugly LD_PRELOAD, but simply by overriding the
relevant symbol natively in our own code).
This makes debugging tokens a ton easier.
Automatically softreboot if the nextroot has been set up with an OS
tree, or automatically kexec if a kernel has been loaded with kexec
--load.
Add SYSTEMCTL_SKIP_AUTO_KEXEC and SYSTEMCTL_SKIP_AUTO_SOFT_REBOOT to
skip the automated switchover.
Instead of using ExtraTrees=, let's use the new RuntimeTrees= option
to mount the full repository into the VM/container. Let's also store
the sources under /usr/src/systemd and update the gdbinit file and
vscode HACKING guide section to match the new location.
Currently we have a 100ms delay which allows for people to enter/show
the boot menu even when timeout is set to zero.
In a handful of cases, that may not be needed - both in terms of access
policy, as well as latency.
For example: the option to provide the boot menu may be hidden behind an
"expert only" UX in the OS, to avoid end users from accidentally
entering it.
In addition, the current 100ms input polling may cause unexpected
additional delays in the boot. Some example numbers from my SteamDeck:
- boot counting/rename/flush doubles 300us -> 600us
- seed/hash setup doubles 900us -> 1800us
- kernel/image load gets ~40% slower 107ms -> 167ms
It's not entirely clear why the UEFI calls gets slower, nevertheless the
information in itself proves useful.
This commit introduces a new option "menu-disabled", which omits the
100ms delay. The option is documented throughout the manual pages as
well as the Boot Loader Specification.
v2:
- use STR_IN_SET
v3:
- drop erroneous whitespace
v4:
- add a new LoaderFeature bit,
- don't change ABI keep TIMEOUT_* tokens the same
- move new token in the 64bit range, update API and storage for it
- change inc/dec behaviour to TIMEOUT_MIN : TIMEOUT_MENU_FORCE
- user cannot opt-in from sd-boot itself, add assert_not_reached()
v5:
- s/Menu disablement control/Menu can be disabled/
- rewrap comments to 109
- use SYNTHETIC_ERRNO(EOPNOTSUPP)
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
To be on the safe side, explicitly mention that apart from the numerical
entries we can allow string ones.
Implementation-wise, bootctl will use internal numerical values that
match sd-boot's ABI. The latter also accepts the string options.
Going forward we'd like to avoid adding more internal magic and be more
explicit.
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Currently we spawn services by forking a child process, doing a bunch
of work, and then exec'ing the service executable.
There are some advantages to this approach:
- quick: we immediately have access to all the enourmous amount of
state simply by virtue of sharing the memory with the parent
- easy to refactor and add features
- part of the same binary, will never be out of sync
There are however significant drawbacks:
- doing work after fork and before exec is against glibc's supported
case for several APIs we call
- copy-on-write trap: anytime any memory is touched in either parent
or child, a copy of that page will be triggered
- memory footprint of the child process will be memory footprint of
PID1, but using the cgroup memory limits of the unit
The last issue is especially problematic on resource constrained
systems where hard memory caps are enforced and swap is not allowed.
As soon as PID1 is under load, with no page out due to no swap, and a
service with a low MemoryMax= tries to start, hilarity ensues.
Add a new systemd-executor binary, that is able to receive all the
required state via memfd, deserialize it, prepare the appropriate
data structures and call exec_child.
Use posix_spawn which uses CLONE_VM + CLONE_VFORK, to ensure there is
no copy-on-write (same address space will be used, and parent process
will be frozen, until exec).
The sd-executor binary is pinned by FD on startup, so that we can
guarantee there will be no incompatibilities during upgrades.
