On Tue, 2025-01-28 at 15:39 +0100, Peter Zijlstra wrote:
On Tue, Jan 28, 2025 at 12:46:07PM +0100, Mike Galbraith wrote:
Seems 6.13 is gripe free thanks to it containing 4423af84b297.
I stumbled upon a reproducer for my x86_64 desktop box: all I need do is fire up a kvm guest in an enterprise configured host. That inspires libvirt goop to engage group scheduling, splat follows instantly.
Back 4423af84b297 out of 6.13, it starts griping, add it to a 6.12 tree containing 6d71a9c61604, it stops doing so.
Ooh, does something like the below (+- reverse-renames as applicable to .12) also help?
Yup, seems to, 6.13 sans 4423af84b297 stopped griping.
Hi Peter, Are there any plans to backport this patch and 6d71a9c61604 into LTS versions as well? I'm seeing the exact same issue on 6.6 as well, where the bug in reweight results in incorrect vruntime calculation, leading to the deadline always being in future, rather than decreasing on each scheduler tick. This causes a task to run for a longer amount of time, starving other tasks on the runqueue.
The issue disappears when testing with 4423af84b297 and your patch to avoid tripping on WARN in place_entity(). I'm happy to push the backported patches into 6.12 and 6.6, if that's alright.