Re: [PATCH] KVM: arm64: nv: do not inject L2-bound IRQs to L1 hypervisor

[Volodymyr Babchuk <Volodymyr_Babchuk@epam.com>]

Hi Oliver,

Oliver Upton <oliver.upton@linux.dev> writes:

> Hi Volodymyr,
>
> On Thu, Oct 02, 2025 at 09:00:11PM +0000, Volodymyr Babchuk wrote:
>> Difference between nested virtualization and "baremetal" case is that
>> real GIC can track lots of active interrupts simultaneously, but vGIC
>> is limited only to 4..16 LRs.
>
> There isn't an architectural limitation here. Nothing prevents a
> virtualized GIC from representing more active IRQs than there are LRs in
> hardware.
>
> ICH_HCR_EL2.LRENPIE allows you to take a trap when an EOI is received
> for an IRQ that exists outside of teh list registers which would allow
> the deactivation of the SW view of that IRQ.

Ah, right. I am missed this feature.

>
> As Marc suggested, the correct thing to do is adjust the sorting of IRQs
> such that pending IRQs fill LRs before those in an active state.

Looks like I missed an email where he suggested that, but yeah, this was
my first ideal as well, I just didn't knew about LRENPIE bit.

[...]

> I'd really like to see a solution similar to Marc's proposal which
> addresses the fundamental problem of active IRQs overflowing the list
> registers.
>

I'm sorry but it appears that I won't be able to put any more effort
into this issue right now. My boss wants me to switch to another pressing
task... So yeah, I won't mind if someone else will continue the work.

Also, I am sorry that I can't share reproducer, but as I said, that
would require all the whopping Android build and believe me, you don't
want this. I suspect that this issue can be reproduced by firing lots of
interrupts simultaneously in kvmtool. And of course in this case there
is nothing Xen-specific, so KVM as L1 hypervisor should be fine too. But
I have no time to verify this.

-- 
WBR, Volodymyr