Use FEAT_WFxT (armv8.7) for spin-waiting

[EgorBo]

Currently we use yield instruction (which is basically an alias of nop) and it has an impact on CPU usage. We should consider switching to "Wait-for-Event with timeout" instruction:

Example (generated by GPT):

        // Armv8.7-A + FEAT_WFxT
        // quick_spin_wait(): brief pause (~50µs) using WFET
        .p2align 4
quick_spin_wait:
        mrs     x9,  CNTVCT_EL0         // now
        mrs     x10, CNTFRQ_EL0         // counter Hz
        mov     x11, #50                // microseconds
        mul     x12, x10, x11           // Hz * usec
        mov     x13, #1000000
        udiv    x12, x12, x13           // ticks for ~50us
        add     x12, x9,  x12           // absolute deadline

        sevl
        wfet    x12                     // clear event reg / quick bail
        wfet    x12                     // sleep until event/timeout

        ret

Hopefully, all of these are allowed in the user space.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[tannergooding]

We should ideally update to use WaitPkg on x64 at the same time, which allows for similar handling and providing an explicit timing and sleep level for the wait.

[vcsjones]

Doesn't modern Linux trap this and basically turn it in to a nop? https://patchwork.kernel.org/project/linux-arm-kernel/patch/20180807093326.5090-1-marc.zyngier@arm.com/

[EgorBo]

Doesn't modern Linux trap this and basically turn it in to a nop? https://patchwork.kernel.org/project/linux-arm-kernel/patch/20180807093326.5090-1-marc.zyngier@arm.com/

WFI is a different instruction (quite old), but Linux may do the same for the new one as well (I don't see any trap for WFET in Linux sources, for WFI I do see it).

[snickolls-arm]

This blog might be of interest for implementing this. It also has a backup option for when FEAT_WFxT is not enabled.

https://community.arm.com/arm-community-blogs/b/architectures-and-processors-blog/posts/multi-threaded-applications-arm

[EgorBo]

Prototype is #130362, but the actual feature requires at least C200 HW to be adopted.

[snickolls-arm]

Prototype is #130362, but the actual feature requires at least C200 HW to be adopted.

It would be worth trying the other implementation suggested in the blog as well, using instruction synchronization barriers (isb) as the delay mechanism. It should be an improvement on yield and is available from Armv8.0-A, so it would be suitable as a baseline implementation to fall back to when FEAT_WFxT is not available.