Public bug reported:
There is a growing need to design sched_ext schedulers that can monitor
specific PMU hardware events (e.g., branch-prediction misses, cache
misses, CPU cycles, etc.) either to mitigate potential hardware-related
regressions or implement more informed scheduling policies that can
improve performance.
Unfortunately, the 6.14 kernel does not provide access to PMU counters
from BPF, which makes it impossible to run such schedulers on linux-
nvidia-6.14. Exposing perf_event_read_output() to all BPF program types
would enable these schedulers to work on linux-nvidia-6.14 as well.
The required change is a clean cherry pick of a single upstream commit
(bpf: Make perf_event_read_output accessible in all program types.)
which makes a kfunc usable from multiple contexts and it's pretty self
contained in the BPF subsystem, therefore regression potential is
minimal.
** Affects: linux-nvidia (Ubuntu)
Importance: Undecided
Status: New
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2131947
Title:
Make perf counters accessible from sched_ext schedulers
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-nvidia/+bug/2131947/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
