This extends the close_all_fds() logic to overmount /proc with an empty
tmpfs, and/or to block close_range() via seccomp, so that we run the
test case for the function with the fallback paths.
This should make sure that we don't regress in limited environments or
older kernels.
Add special case optimization for a single exception fd. It's a
pretty common case in our codebase, and the optimization is simple
and means we don't need to copy/sort the exception array, so do it.
This new helper is not used yet, but it's useful for apply UID/GID
shifts so that the underlying home dir can use an arbitrary UID (for
example "nobody") and we'll still make it appear as owned by the target
UID.
This operates roughly like this:
1. The relevant underlying UID is mapped to the target UID
2. Everything in the homed UID range except for the target UID is left
unmapped (and thus will appear as "nobody")
3. Everything in the 16bit UID range outside of the homed UID
range/target UID/nobody user is mapped to itself
4. Everything else is left unmapped (in particular everything outside of
the 16 bit range).
Why do it like this?
The 2nd rule done to ensure that any files from homed's managed UID
range that do not match the user's own UID will be shown as "unmapped"
basically. Of course, IRL this should never happen, except if people
managed to manipulate the underlying fs directly.
The 3rd rule is to allow that if devs untar an OS image it more or
less just works as before: 16bit UIDs outside of the homed range will
be mapped onto themselves: you can untar things and tar it back up and
things will just work.
This does what we already do for the LUKS backend: instead of mounting
the source directory directly to the final home dir, we instead bind
mount it to /run/systemd/user-home-mount (where /run/ is unshared and
specific to our own mount namespace), then adjust its mount flags and
then bind mount it in a single atomic operation into the final
destination, fully set up.
This doesn't improve much on its own, but it makes things a tiny bit
more correct: this way MS_NODEV/MS_NOEXEC/MS_NOSUID will already be
applied when the bind mount appears in the host mount namespace, instead
of being adjusted after the fact.
Doing things this way also makes things work more like the LUKS backend,
reducing surprises. Most importantly it's preparation for doing
uidmapping for directory homes, added in a later commit.
For the other backends we synthesize a "binding" section in the json
record of the user that stores meta info how a user record is "bound" to
the local host. It declares storage info and such. Let's do the same for
the directory/subvolume backends.
Let's migrate home_create_directory_or_subvolume() to also use HomeSetup
for storing its runtime objects we'd like to destroy in case of failure.
In the beginning this is just the root_fd, but later on we can add more.
No change in behaviour, just shifting things around.
Previously, valid_until (or preferred_until for preferred lifetime) was
calculated from lifetime. So, when an upstream interface acquire a
dynamic prefix (e.g. through DHCPv6-PD) with long lifetime, then sd-radv
advertise the same lifetime. It may not be desired for some situations.
Let's make sure '@' is never written as entry ID into any EFI variable,
as we want the ability to add new ids like this later on, with them
resulting in a clear error on older implementations.
mkosi automatically builds for the host distro which seems a much
better default to encourage since dnf won't be installed on any host
system that's not Fedora anyway.
Lennart Poettering
Yu Watanabe
Luca Boccassi
Zbigniew Jędrzejewski-Szmek
Daan De Meyer