Re: [PATCH 1/2] KVM: register asynchronous page fault waits with interrupts disabled

[Paolo Bonzini]

Il mar 3 nov 2026, 15:43 Sean Christopherson <seanjc@xxxxxxxxxx> ha scritto:
>
> On Tue, Nov 03, 2026, Paolo Bonzini wrote:
> > KVM cannot use a node on the stack like kvm_async_pf_task_wait_schedule() does
>
> Oh!  Yes, KVM can!  If we fully handle the PV async #PF in vcpu_enter_guest(),
> then the node can be put on the stack, right?  Juggling preemption and guest
> timing is heinous, but on the plus side it consolidates the PV async #PF code,
> so I think it's a net positive?  Especially when factoring in the node stuff.
>
> Fully handle PV async #PFs (when running as a KVM guest) before enabling
> IRQs so that it's impossible for the current task, i.e. the vCPU task, to
> be preempted before invoking kvm_async_pf_task_wait_schedule().  This will
> allow the guest-side async #PF code to eliminate its dummy allocation,
> which is needed only to handle the case where the page ready notification
> is delivered before handling the not-present #PF.
>
> Add dedicated handling for PV async #PFs instead of wrapping/piggybacking
> kvm_x86_ops.handle_exit_irqoff(), as the act of handling an async #PF may
> enable IRQs and schedule out the current task, i.e. needs to be done just
> before KVM itself explicitly enables IRQs, and needs to enable preemption
> before calling into the kernel's PV async #PF handler.
>
> Opportunistically wrap the code with CONFIG_KVM_GUEST to help document the
> context, and drop the CONFIG_KVM_GUEST=n stubs.
>
> Suggested-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>

More like Instigated-by. :)

The main advantage would be that it bypasses the missing lock issue
that Sashiko reported, but I don't know... That;s surely nice, but at
the same time it is not great having to add a new callback, and to
duplicate the guest_time_exit/preempt_enable code into
kvm_handle_pv_async_pf().

Not that the extra field in vcpu->arch is beautiful, I admit that, but
it seems extra unwarranted complexity.

Unless you insist I will send a v2 that takes the lock properly.

Paolo