On Tue, Mar 12, 2024, Sean Christopherson wrote:
> Fix a bug in dynamic preemption where the kernel will yield contended
> spinlocks (and rwlocks) even if the selected preemption model is "none" or
> "voluntary". I say "bug" because this divergence from PREEMPT_DYNAMIC=n
> behavior effectively broke existing KVM configurations, e.g. vCPUs would
> get stuck and become unresponsive for multiple seconds if there was heavy
> KSM or NUMA balancing activity in the host.
>
> This isn't super urgent, as 6.8 has a fix in KVM for the over-aggressive
> yielding (commit d02c357e5bfa ("KVM: x86/mmu: Retry fault before acquiring
> mmu_lock if mapping is changing"), but I wouldn't be surprised if the
> behavior is causing other performance issues/regressions that are less
> severe and/or less visible.
Anyone have any thoughts on how to move this forward? I have a hard time
believing no one has opinions on this code :-)
> v2:
> - Rebase onto Linus' tree to deal with the code movement to spinlock.h.
> - Opportunistically document the behavior.
> - Add the PREEMPT_AUTO folks to Cc to get their eyeballs/input.
>
> v1: https://lore.kernel.org/all/20240110214723.695930-1-seanjc@xxxxxxxxxx
>
> Sean Christopherson (2):
> sched/core: Move preempt_model_*() helpers from sched.h to preempt.h
> sched/core: Drop spinlocks on contention iff kernel is preemptible
>
> .../admin-guide/kernel-parameters.txt | 4 +-
> include/linux/preempt.h | 41 +++++++++++++++++++
> include/linux/sched.h | 41 -------------------
> include/linux/spinlock.h | 14 +++----
> 4 files changed, 50 insertions(+), 50 deletions(-)
>
>
> base-commit: b29f377119f68b942369a9366bdcb1fec82b2cda
> --
> 2.44.0.278.ge034bb2e1d-goog
>
