[PATCH v4 14/20] KVM: arm/arm64: Avoid timer save/restore in vcpu entry/exit

[Marc Zyngier]

On Fri, Oct 20 2017 at  1:49:33 pm BST, Christoffer Dall <christoffer.dall at linaro.org> wrote:
> From: Christoffer Dall <cdall at linaro.org>
>
> We don't need to save and restore the hardware timer state and examine
> if it generates interrupts on on every entry/exit to the guest.  The
> timer hardware is perfectly capable of telling us when it has expired
> by signaling interrupts.
>
> When taking a vtimer interrupt in the host, we don't want to mess with
> the timer configuration, we just want to forward the physical interrupt
> to the guest as a virtual interrupt.  We can use the split priority drop
> and deactivate feature of the GIC to do this, which leaves an EOI'ed
> interrupt active on the physical distributor, making sure we don't keep
> taking timer interrupts which would prevent the guest from running.  We
> can then forward the physical interrupt to the VM using the HW bit in
> the LR of the GIC, like we do already, which lets the guest directly
> deactivate both the physical and virtual timer simultaneously, allowing
> the timer hardware to exit the VM and generate a new physical interrupt
> when the timer output is again asserted later on.
>
> We do need to capture this state when migrating VCPUs between physical
> CPUs, however, which we use the vcpu put/load functions for, which are
> called through preempt notifiers whenever the thread is scheduled away
> from the CPU or called directly if we return from the ioctl to
> userspace.
>
> One caveat is that we have to save and restore the timer state in both
> kvm_timer_vcpu_[put/load] and kvm_timer_[schedule/unschedule], because
> we can have the following flows:
>
>   1. kvm_vcpu_block
>   2. kvm_timer_schedule
>   3. schedule
>   4. kvm_timer_vcpu_put (preempt notifier)
>   5. schedule (vcpu thread gets scheduled back)
>   6. kvm_timer_vcpu_load (preempt notifier)
>   7. kvm_timer_unschedule
>
> And a version where we don't actually call schedule:
>
>   1. kvm_vcpu_block
>   2. kvm_timer_schedule
>   7. kvm_timer_unschedule
>
> Since kvm_timer_[schedule/unschedule] may not be followed by put/load,
> but put/load also may be called independently, we call the timer
> save/restore functions from both paths.  Since they rely on the loaded
> flag to never save/restore when unnecessary, this doesn't cause any
> harm, and we ensure that all invokations of either set of functions work
> as intended.
>
> An added benefit beyond not having to read and write the timer sysregs
> on every entry and exit is that we no longer have to actively write the
> active state to the physical distributor, because we configured the
> irq for the vtimer to only get a priority drop when handling the
> interrupt in the GIC driver (we called irq_set_vcpu_affinity()), and
> the interrupt stays active after firing on the host.
>
> Signed-off-by: Christoffer Dall <cdall at linaro.org>

That was a pretty interesting read! :-)

Reviewed-by: Marc Zyngier <marc.zyngier at arm.com>

	M.
-- 
Jazz is not dead. It just smells funny.