... (rhbz#2177722)
Oomd was killing a login session (user-*.slice/session-*.scope).
Quoting https://bugzilla.redhat.com/show_bug.cgi?id=2177722#c21:
> In F37 and prior the config was killing based on swap and pressure
> on user-*.slice/user@.service. In 7665e1796f
> it was changed to pressure only on system.slice and all slices under
> user.slice. The relevant point here is that this change now includes
> user-*.slice/session-*.scope which is the critical session bits
> you're seeing killed here.
>
> That session scope should be omitted. The config that I intended
> with the initial PR was for all slices under
> user.slice/user-*.slice/user@.service to be monitored, not for all
> slices under user.slice.
With the file removed:
$ oomctl | rg Path | sort
Path: /system.slice
Path: /user.slice/user-1000.slice/user@1000.service/app.slice
Path: /user.slice/user-1000.slice/user@1000.service/session.slice
See https://github.com/systemd/systemd/pull/26641.
This will allow upstream pull request (and the main branch after the pull
request has been merged) to be built with the new code. This doesn't do
anything for official rpm builds until the new code is part of the sources.
[skip changelog]
... (rhbz#2164594)
The socket exists and is enabled in the initrd. After switch-root, the system
goes into an infinite loop trying to stop the socket while incoming audit
messages trigger start jobs for the socket. This is a bug in the transaction
logic, that'll need to be fixed separately.
We need to preset the socket after the upgrade so that it remains enabled
by default. This should fix the boot issue, though it's not a complete fix,
because we actually want to allow people to disable the socket.
On initial install, the socket is covered by preset-all and gets enabled.
gcc has a new warning which caught a bug of int/enum mismatches.
And we would crash on some architectures when built with -D_FORTIFY_SOURCE=3
because of our malloc_usable_size() use.
This should resolve the build failure in F38 mass build.
- Fixes a few different issues (systemd-timesyncd connectivity problems, broken
emoji output on the console, crashes in pid1 unit dependency logic)
- CVE-2022-4415: systemd: coredump not respecting fs.suid_dumpable kernel
setting
As requested in https://github.com/rhinstaller/anaconda/pull/4368#discussion_r1043839809,
so that it's easier to depend on the appropriate package. Once we have the
signed version built, this provides might be dropped. But let's add it at least
for now so that there's a stable name to depend on.
While at it, let's drop ? from %{_isa}. Systemd is always archful.
This file changes rarely, but it does every one in a while. And since we have an
independent copy, we forget to adjust it. We have had already two bugs because
of this. I submitted a PR upstream to include pam_namespace (because that makes
sense for all distros), so the diff between upstream and us now is just the
inclusion of system-auth (which is not upstreamable).
Effectively, the only difference right now is that 'pam_keyinit force revoke'
is included. It was added upstream with the comment:
We want that systemd --user gets its own keyring as usual, even if the
barebones PAM snippet we ship upstream is used. If we don't do this we get
the basic keyring systemd --system sets up for us.
4047e4fb7b got things very wrong.
The trick with "[ $1 -eq 1 ]" doesn't work for transaction triggers
because the argument is not provided by rpm. We need to use a state
file to propagate the information from %post to %posttrans.
... (for details see https://raw.githubusercontent.com/systemd/systemd/v252-rc1/NEWS)
systemd-pcrphase and systemd-measure and initrd-* units are moved to systemd-udev.
systemd-udev should be part of the initrd, and those tools don't make much sense
in systems without hardware (i.e. containers). (systemd-measure could possibly be
useful, but we can always move it back if there's a good reason.)
- Remove swap policy. Default amount of swap (8GB?) is a lot lower than
what we use internally with the swap policy. Which frequently leads to
GNOME getting killed
(e.g. https://bugzilla.redhat.com/show_bug.cgi?id=1941170, and other
BZs not linked here). Internally we use 0.5x-1x size of physical memory
for swap via swapfiles (this will be documented in systemd upstream).
In simple cases of using more memory than is available (but without
memory pressure), the Kernel OOM killer can handle killing the
offending process.
- Expand the memory pressure policy to system.slice, user-.slice, and
all user owned slices. Support for ManagedOOM*= on user services was
added in https://github.com/systemd/systemd/pull/20690 which allows
us to be more fine grained on the pressure monitoring at the user
level. In addition to the system.slice and user-.slice PSI monitoring
this should result in a better systemd-oomd experience for desktop
systems.
Instead, add systemd-pam to pungi-fedora's multilib whitelist:
https://pagure.io/pungi-fedora/pull-request/1113
This should help with flatpak runtime packaging so that we can avoid
having to ship systemd-pam in the flatpak container.
It turns out that with the Obsoletes, dnf will just install the normal
systemd package if systemd-standalone-* is requested. The commit message
for b36512ad8f which added this says I tested
with local package builds (where it works), but not when going through the
full repo with all packages.
I'm adding the Provides instead, so that it's possible to request on or
the other more easily.
I asked on fedora-devel@, and the lone reply was from Matthew Miller
who tried it once when it was introduced and hasn't used it since.
Dropping this removes the last dependency on libgcrypt and libgpg-error
in libsystemd, significantly reducing our installation footprint.
Right now libmicrohttpd is still linked to libgcrypt, so
libsystemd-journal-remote subpackage will pull libgcrypt in.
When -Dversion-tag was initially added in edaa157918,
I used "v" without any comment. But upstream does not use "v", so we have
versions which don't compare directly:
$ build/systemctl --version|head -n1
systemd 251 (251-66-g7e46a5c+)
$ systemctl --version|head -n1
systemd 251 (v251-1.fc37)
And in 3c4f9413a7, when -Dshared-lib-tag= was
introduced, %{version} was replaced by %{version_no_tilde}, again without any
specific comment. For the shared-lib-tag, it makes sense to use _no_tilde,
because it's enough to have non-conflicting file names, and we don't compare
the tags. I guess I wanted both uses to be consistent. But if we substitute
the tilde, we can't do proper comparisons.
I noticed the following issue: with sd-boot installed from git and a
package, upgrades wouldn't work:
Comparing versions: "systemd-boot v251-1.fc37" < "systemd-boot 251-rc1-390-g3603f15
Skipping "/boot/efi/EFI/systemd/systemd-bootx64.efi", since newer boot loader version in place already.
The two changes should make those comparisons work properly in most
cases.
I tested this with 'sudo dnf --installroot=…', with both
systemd+system-udev installed in one transaction, and in two separate
transactions. Users are created as expected in both cases.
$ rpm -qlv systemd |grep -v 'root root'
-rw-rw-r-- 1 root utmp 0 Jan 22 03:38 /run/utmp
-rw-rw---- 1 root utmp 0 Jan 22 03:38 /var/log/btmp
-rw-rw-r-- 1 root utmp 0 Jan 22 03:38 /var/log/lastlog
-rw-rw-r-- 1 root utmp 0 Jan 22 03:38 /var/log/wtmp
drwxr-sr-x 2 root systemd- 0 Jan 22 03:38 /var/log/journal
During installation rpm would log an error that systemd-journal group
is unknown. We create all our users by calling sysusers in the %post
scriptlet, but that is too late. To avoid the warning we could either
add a %pre scriptlet, but that'd require adding a dependency on
shadow-utils for groupadd, since we can't use our own tools before we
are installed. Let's instead create the directory owned by root.root,
and change the group afterwards. The group ownership is for file
ownership, and in the worst case (we don't assign the group or set
mode +s), unprivileged users will not be able to read the logs.
We also use 'utmp' group, but that is provided by setup.rpm and is not
an issue.
https://bugzilla.redhat.com/show_bug.cgi?id=2018913#c24
For https://fedoraproject.org/wiki/Changes/RenameNobodyUser a scriptlet
was introduced with prevents nss-systemd from synthesizing entries for nobody.
Let's remove the scriptlet: very few people upgrade from such old systems,
and even if they do, having a duplicate entry for nobody is annoying
but hardly a big problem.
(The other side of this, support in nss-systemd remains in place.)
This allows deps on the tools used in the scriptlet to be dropped from -libs.
While at it, also drop noop ldconfig scriptlets.
