| Message ID | 20241108122909.763663-1-patryk.wlazlyn@linux.intel.com |
|---|---|
| Headers | show |
| Series | SRF: Fix offline CPU preventing pc6 entry | expand |
On Tue, Nov 12 2024 at 12:28, Rafael J. Wysocki wrote: > On Tue, Nov 12, 2024 at 12:18 PM Patryk Wlazlyn > <patryk.wlazlyn@linux.intel.com> wrote: >> I don't think so. The old algorithm gives fairly good heuristic for computing >> the mwait hint for the deepest cstate. Even though it's not guaranteed to work, >> it does work on most of the platforms that don't early return. I think we should >> leave it, but prefer idle_driver. > > IOW, as a fallback mechanism, it is as good as it gets. > > As the primary source of information though, not quite. So we have at least 5 places in the kernel which evaluate CPUID leaf 0x5 in different ways. Can we please have _ONE_ function which evaluates the leaf correctly once and provides the required information for all places ready to use? Thanks, tglx
> That's not really working: > > 1) Regular kexec offlines them again. But then, they execute hlt loop in the reboot vector right? I think that's fine. We just don't want to be woken up from the mwait when the RIP points to garbage. > 2) Kexec in panic can't do any of that. Does it make sense to change the kexec, so that every CPU is forced into the reboot vector, which I believe happens for online CPUs? We don't have to online them all the way. Maybe minimal bringup just to be able to send them the IPI? As a side note, It's not that important for this patchset. I think I can make it work without simplifying the existing hack, but crossed my mind at some point that maybe we could do that.
Applied suggestions from Dave and Rafael. Because we now call a smp function, we have to have ifdefs for CONFIG_SMP or rely on mwait_play_dead_with_hint to return immediatelly with an error on non-smp builds. I decided to do the later, but maybe we should return -ENODEV (or other error constant) instead of 1. I am open for suggestions. Removing the existing "kexec hack" by bringing all offlined CPUs back online before proceeding with the kexec would make it even simpler, but I am not sure we can do that. It looks kind of obvious to me, but for some reason the hack exist. Changes since v2: The whole approach changed, but there main things are below. * Split mwait_play_dead (old code) into two parts: 1. Computing mwait hint based on cpuid leaf 0x5 2. Entering while(1) mwait loop with "kexec hack" handling * Prefer cpuidle_play_dead over mwait_play_dead by reordering calls in native_play_dead. * Add implementation for enter_dead() handler in intel_idle that calls executes old mwait_play_dead code, but with the mwait hint from the cpuidle state table. Patryk Wlazlyn (3): x86/smp: Allow calling mwait_play_dead with arbitrary hint x86/smp native_play_dead: Prefer cpuidle_play_dead() over mwait_play_dead() intel_idle: Provide enter_dead() handler for SRF arch/x86/include/asm/smp.h | 6 ++++++ arch/x86/kernel/smpboot.c | 29 ++++++++++++++++++++--------- drivers/idle/intel_idle.c | 16 ++++++++++++++++ 3 files changed, 42 insertions(+), 9 deletions(-)