To clarify, the regression appears to be the same problem that the
rlimit increase is fixing, but the applications failing now are simply
bigger. In general, any application that calls mlockall() with
MCL_FUTURE, but doesn't adjust its rlimit (or change its systemd service
file to adjust LimitMEMLOCK) is very likely destined to crash later in
its life.
I believe the only lp bugs for this regression are bug 1890394 and bug
1902879, which are both fix-committed and verified, so this bug should
be ok to release (again) after those are released. Also I will note both
of those applications (slick-greeter and lightdm-gtk-greeter) were fixed
by commenting out their calls to mlockall.
There is also bug 1902871 and bug 1903199 but I believe those are both
dups of bug 1900394.
Also finally to reflect on cryptsetup's use of mlockall(), since it's
the origin for this bug; cryptsetup is maybe "better" about its use of
mlockall() since it keeps the mlock only for the duration of an
'action':
if (action->required_memlock)
crypt_memory_lock(NULL, 1);
set_int_handler(0);
r = action->handler();
if (action->required_memlock)
crypt_memory_lock(NULL, 0);
however, as this bug shows, that action handler function can still
attempt to allocate enough memory to reach the rlimit and cause
allocation failures. Personally, I think cryptsetup should be fixed
upstream to call setrlimit() to increase its RLIMIT_MEMLOCK to infinity,
at least while the mlock is in effect.
--
You received this bug notification because you are a member of Ubuntu
Touch seeded packages, which is subscribed to systemd in Ubuntu.
https://bugs.launchpad.net/bugs/1830746
Title:
memlock setting in systemd (pid 1) too low for containers (bionic)
Status in systemd package in Ubuntu:
Fix Released
Status in systemd source package in Bionic:
Fix Released
Status in systemd source package in Cosmic:
Won't Fix
Status in systemd source package in Disco:
Won't Fix
Status in systemd source package in Eoan:
Fix Released
Status in systemd source package in Focal:
Fix Released
Bug description:
[Impact]
* Since systemd commit fb3ae275cb ("main: bump RLIMIT_NOFILE for the root
user substantially") [https://github.com/systemd/systemd/commit/fb3ae275cb],
which is present in Bionic, the memlock ulimit value was bumped to 16M. It's an
adjustable limit, but the default (in previous Ubuntu releases/systemd
versions) was really small.
* Although bumping this value was a good thing, 16M is not enough and we can
see failures on mlock'ed allocations on Bionic, like the one hereby reported by
Kees or the recent introduced cryptsetup build failures (due to PPA builder
updates to Bionic) - see https://bugs.launchpad.net/bugs//1891473.
* It's especially harmful in containers to have such "small" limit, so
we are hereby SRUing a more recent bump from upstream systemd, in the
form of commit 91cfdd8d29 ("core: bump mlock ulimit to 64Mb")
[https://github.com/systemd/systemd/commit/91cfdd8d29]. Latest Ubuntu
releases, like Focal and subsequent ones, already include this patch
so effectively we're putting Bionic on-par with newer releases.
* A discussion about this topic (leading to this SRU) is present in
ubuntu-devel ML: https://lists.ubuntu.com/archives/ubuntu-
devel/2020-September/041159.html.
[Test Case]
* The straightforward test is to just look "ulimit -l" and "ulimit -Hl" in a
current Bionic system, and then install an updated version with the hereby
proposed SRU to see such limit bump from 16M to 64M (after a reboot) - a
version containing this fix is available at my PPA as of 2020-09-10 [0] (likely
to be deleted in next month or so).
* A more interesting test is to run a Focal container in a current
Bionic system and try to build the cryptsetup package - it'll fail in
some tests. After updating the host (Bionic) systemd to include the
mlock bump patch, the build succeeds in the Focal container.
[Regression Potential]
* Since it's a simple bump and it makes Bionic behave like Focal, I don't
foresee regressions. One potential issue would be if some users rely on the
lower default limit (16M) and this value is bumped by a package update, but
that could be circumvented by setting a lower limit in limits.conf. The
benefits for such bump are likely much bigger than any "regression" caused for
users relying on such default limit.
[0] https://launchpad.net/~gpiccoli/+archive/ubuntu/test1830746
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/systemd/+bug/1830746/+subscriptions
--
Mailing list: https://launchpad.net/~touch-packages
Post to : touch-packages@lists.launchpad.net
Unsubscribe : https://launchpad.net/~touch-packages
More help : https://help.launchpad.net/ListHelp
