Lennart Poettering
Lennart Poettering
> The linux implementation of `kill(-1, signal)` holds `tasklist_lock` for reading for the loop duration. > > https://github.com/torvalds/linux/blob/v5.17/kernel/signal.c#L1599-L1619 interesting...
so, the issue is traversal of the root fs, right? so what about this: open an fd to /proc/ first, and only use that for traversal. That way we should...
I am sorry, but no, let's not make this configurable. "Compatibility" you say, but with what? Programs that hardcode their cgroup path are borked anyway, we don't have to bother....
I am still very much against fiddling with threaded cgroups, as mentioned before. Let people do that down the hierarchy. Again, I am still very sure that if there's more...
Patch looks generally OK, but please, do per-session timer event sources, not a global one with a manual loop to figure out which the next idle session will be.
CI failures appear unrelated
hmm, so i understand this is all supposed to address issues around inotify watch descriptor usage? is that really something that is easy to trigger? wouldn't a nicer way be...