[PATCH] KVM: arm64: Add capability for unconditional WFx passthrough

[Colton Lewis]

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

> Hi Colton,

> On Mon, Jan 29, 2024 at 09:39:17PM +0000, Colton Lewis wrote:
>> Add KVM_CAP_ARM_WFX_PASSTHROUGH capability to always allow WFE/WFI
>> instructions to run without trapping. Current behavior is to only
>> allow this if the vcpu is the only task running. This commit keeps the
>> old behavior when the capability is not set.

>> This allows userspace to set deterministic behavior and increase
>> efficiency for platforms with direct interrupt injection support.

> Marc and I actually had an offlist conversation (shame on us!) about
> this very topic since there are users asking for the _opposite_ of this
> patch (unconditionally trap) [*].

> I had originally wanted something like this, but Marc made the very good
> point that (1) the behavior of WFx traps is in no way user-visible and
> (2) it is entirely an IMP DEF behavior. The architecture only requires
> the traps be effective if the instruction does not complete in finite
> time.

> We need to think of an interface that doesn't depend on
> implementation-specific behavior, such as a control based on runqueue
> depth.

Here's the first thing I came up with after returning to this problem:

We have an ioctl to get/set a threshold value,
wfx_traps_runqueue_depth. If the depth is less than or equal to the
threshold, disable WFx traps. If greater than, enable WFx traps.

Current behavior occurs with a setting of 1. Always trap occurs with a
setting of 0. Never trap occurs with any large enough number.

Of course, having an integer may be more flexible than needed. I can't
imagine a practical use for a number between 1 and UINT_MAX, in which
case it would be better as an enum for different behaviors than a
integer threshold.

What do you think?

Also, do you mean runqueue depth for the current CPU or globally?