[oom] Fix OOMScoreAdjust crashes from other user services.

[locusf]

No description provided.

[faenil]

lgtm, even though we still don't know why this is needed....@stskeeps?

[hedayat]

Hmmm... it seems that an unprivileged process (systemd --user) can't adjust this value for its children to something lower than its own OOMScoreAdjust value (0 by default). So, this is needed so that OOMScoreAdjust value for user services can be below zero. With this commit, user services can have OOMScoreAdjust value between -750 and 1000. If values lower than -750 is needed, you should decrees this value for user@.service too.

[faenil]

great thanks! where did you find the info?

Edit: I found "The value of /proc//oom_score_adj may be reduced no lower than the last
value set by a CAP_SYS_RESOURCE process. To reduce the value any lower
requires CAP_SYS_RESOURCE." from https://www.kernel.org/doc/Documentation/filesystems/proc.txt

[hedayat]

Yes, the info was from initial guess, 'man proc' (almost same as proc.txt), and some tests. However, when I was thinking about answering you, I found that we have a problem which should be fixed:
"a new process inherits its parent's oom_adj setting." (found in man proc)

This means that since we run 'systemd --user' with OOMScoreAdjust=-750, all user services have this score by default. So, while some services got -250 score to make them less likely to be killed in OOM situations (at least this is what I guess, right?), now they are actually the first user services which might be killed because all other user services are running with -750 score adjust value!

Therefore, I think we should add "OOMScoreAdjust=0" for all other user services.

[faenil]

mm well spotted, I guess this makes sense