Shrikanth Hegde [off-list ref] writes:

On 11/9/24 22:24, Shrikanth Hegde wrote:

quoted

On 11/9/24 00:36, Ankur Arora wrote:

quoted

Shrikanth Hegde [off-list ref] writes:

quoted

Define preempt lazy bit for Powerpc. Use bit 9 which is free and within
16 bit range of NEED_RESCHED, so compiler can issue single andi.

Since Powerpc doesn't use the generic entry/exit, add lazy check at exit
to user. CONFIG_PREEMPTION is defined for lazy/full/rt so use it for
return to kernel.

Ran a few benchmarks and db workload on Power10. Performance is close to
preempt=none/voluntary. It is possible that some patterns would
differ in lazy[2]. More details of preempt lazy is here [1]

Since Powerpc system can have large core count and large memory,
preempt lazy is going to be helpful in avoiding soft lockup issues.

[1]: https://lore.kernel.org/ lkml/20241007074609.447006177@infradead.org/
[2]: https://lore.kernel.org/all/1a973dda-c79e-4d95-935b-
e4b93eb077b8@linux.ibm.com/

Signed-off-by: Shrikanth Hegde <redacted>

Looks good. Reviewed-by: [off-list ref]

Thank you Ankur for taking a look and rwb tag.

quoted

However, I just checked and powerpc does not have
CONFIG_KVM_XFER_TO_GUEST_WORK. Do you need this additional patch
for handling the lazy bit at KVM guest entry?

It doesn't use the generic kvm entry/exit either AFAIK. I need to understand
more of this kvm maze. There are quite a lot of combinations.

The generic kvm entry/exit is gated by CONFIG_KVM_XFER_TO_GUEST_WORK.

quoted

will take a look. Thanks for the pointers.

quoted

diff --git a/arch/powerpc/kvm/powerpc.c b/arch/powerpc/kvm/powerpc.c
index f14329989e9a..7bdf7015bb65 100644
--- a/arch/powerpc/kvm/powerpc.c
+++ b/arch/powerpc/kvm/powerpc.c

@@ -84,7 +84,8 @@ int kvmppc_prepare_to_enter(struct kvm_vcpu *vcpu)
         hard_irq_disable();

         while (true) {
-               if (need_resched()) {
+               unsigned long tf = read_thread_flags();
+               if (tf & (_TIF_NEED_RESCHED | _TIF_NEED_RESCHED_LAZY)) {
                         local_irq_enable();
                         cond_resched();
                         hard_irq_disable();

This is not going help since, with LAZY, cond_resched is nop. So it doesn't call
to schedule. Same is true with preempt=full. I need to figure out if kvm stuff
was tested for preempt=full.

Instead of cond_resched this needs to use schedule instead. Need to test it out
and also see other places for kvm.

Oh yeah. Missed that it was calling cond_resched().

So, i need to spend more time on this and figure out, will send the patches
after that.

quoted

quoted

Ankur

--
ankur