Re: [PATCH] riscv: KVM: Remove unnecessary vcpu kick

["Radim Krčmář" <rkrcmar@ventanamicro.com>]

2025-02-19T09:54:26+08:00, BillXiang <xiangwencheng@lanxincomputing.com>:
> Thank you Andrew Jones, forgive my errors in the last email.
> I'm wondering whether it's necessary to kick the virtual hart
> after writing to the vsfile of IMSIC.
> From my understanding, writing to the vsfile should directly
> forward the interrupt as MSI to the virtual hart. This means that
> an additional kick should not be necessary, as it would cause the
> vCPU to exit unnecessarily and potentially degrade performance.

Andrew proposed to avoid the exit overhead, but do a wakeup if the VCPU
is "sleeping".  I talked with Andrew and thought so as well, but now I
agree with you that we shouldn't have anything extra here.

Direct MSIs from IOMMU or other harts won't perform anything afterwards,
so what you want to do correct and KVM has to properly handle the memory
write alone.

> I've tested this behavior in QEMU, and it seems to work perfectly
> fine without the extra kick.

If the rest of KVM behaves correctly is a different question.
A mistake might result in a very rare race condition, so it's better to
do verification rather than generic testing.

For example, is `vsfile_cpu >= 0` the right condition for using direct
interrupts?

I don't see KVM setting vsfile_cpu to -1 before descheduling after
emulating WFI, which could cause a bug as a MSI would never cause a wake
up.  It might still look like it works, because something else could be
waking the VCPU up and then the VCPU would notice this MSI as well.

Please note that I didn't actualy verify the KVM code, so it can be
correct, I just used this to give you an example of what can go wrong
without being able to see it in testing.

I would like to know if KVM needs fixing before this change is accepted.
(It could make bad things worse.)

> Would appreciate any insights or confirmation on this!

Your patch is not acceptable because of its commit message, though.
Please look again at the document that Andrew posted and always reply
the previous thread if you do not send a new patch version.

The commit message should be on point.
Please avoid extraneous information that won't help anyone reading the
commit.  Greeting and commentary can go below the "---" line.
(And possibly above a "---8<---" line, although that is not official and
 may cause issues with some maintainers.)

Thanks.