On Wed, 2023-11-08 at 11:17 +0000, Nicolas Saenz Julienne wrote:
> A number of use cases have surfaced where it'd be beneficial to have a
> vCPU stop its execution in user-space, as opposed to having it sleep
> in-kernel. Be it in order to make better use of the pCPU's time while
> the vCPU is halted, or to implement security features like Hyper-V's
> VSM.
>
> A problem with this approach is that user-space has no way of knowing
> whether the vCPU has pending events (interrupts, timers, etc...), so we
> need a new interface to query if they are. poll() turned out to be a
> very good fit.
>
> So enable polling vCPUs. The poll() interface considers a vCPU has a
> pending event if it didn't enter the guest since being kicked by an
> event source (being kicked forces a guest exit). Kicking a vCPU that has
> pollers wakes up the polling threads.
>
> NOTES:
> - There is a race between the 'vcpu->kicked' check in the polling
> thread and the vCPU thread re-entering the guest. This hardly affects
> the use-cases stated above, but needs to be fixed.
>
> - This was tested alongside a WIP Hyper-V Virtual Trust Level
> implementation which makes ample use of the poll() interface.
>
> Signed-off-by: Nicolas Saenz Julienne <[email protected]>
> ---
> arch/x86/kvm/x86.c | 2 ++
> include/linux/kvm_host.h | 2 ++
> virt/kvm/kvm_main.c | 30 ++++++++++++++++++++++++++++++
> 3 files changed, 34 insertions(+)
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 57f9c58e1e32..bf4891bc044e 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -10788,6 +10788,8 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> goto cancel_injection;
> }
>
> + WRITE_ONCE(vcpu->kicked, false);
> +
> if (req_immediate_exit) {
> kvm_make_request(KVM_REQ_EVENT, vcpu);
> static_call(kvm_x86_request_immediate_exit)(vcpu);
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index 687589ce9f63..71e1e8cf8936 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -336,6 +336,7 @@ struct kvm_vcpu {
> #endif
> int mode;
> u64 requests;
> + bool kicked;
> unsigned long guest_debug;
>
> struct mutex mutex;
> @@ -395,6 +396,7 @@ struct kvm_vcpu {
> */
> struct kvm_memory_slot *last_used_slot;
> u64 last_used_slot_gen;
> + wait_queue_head_t wqh;
> };
>
> /*
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index ad9aab898a0c..fde004a0ac46 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -497,12 +497,14 @@ static void kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
> kvm_vcpu_set_dy_eligible(vcpu, false);
> vcpu->preempted = false;
> vcpu->ready = false;
> + vcpu->kicked = false;
> preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
> vcpu->last_used_slot = NULL;
>
> /* Fill the stats id string for the vcpu */
> snprintf(vcpu->stats_id, sizeof(vcpu->stats_id), "kvm-%d/vcpu-%d",
> task_pid_nr(current), id);
> + init_waitqueue_head(&vcpu->wqh);
> }
>
> static void kvm_vcpu_destroy(struct kvm_vcpu *vcpu)
> @@ -3970,6 +3972,10 @@ void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
> if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
> smp_send_reschedule(cpu);
> }
> +
> + if (!cmpxchg(&vcpu->kicked, false, true))
> + wake_up_interruptible(&vcpu->wqh);
> +
> out:
> put_cpu();
> }
> @@ -4174,6 +4180,29 @@ static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
> return 0;
> }
>
> +static __poll_t kvm_vcpu_poll(struct file *file, poll_table *wait)
> +{
> + struct kvm_vcpu *vcpu = file->private_data;
> +
> + poll_wait(file, &vcpu->wqh, wait);
> +
> + /*
> + * Make sure we read vcpu->kicked after adding the vcpu into
> + * the waitqueue list. Otherwise we might have the following race:
> + *
> + * READ_ONCE(vcpu->kicked)
> + * cmpxchg(&vcpu->kicked, false, true))
> + * wake_up_interruptible(&vcpu->wqh)
> + * list_add_tail(wait, &vcpu->wqh)
> + */
> + smp_mb();
> + if (READ_ONCE(vcpu->kicked)) {
> + return EPOLLIN;
> + }
> +
> + return 0;
> +}
> +
> static int kvm_vcpu_release(struct inode *inode, struct file *filp)
> {
> struct kvm_vcpu *vcpu = filp->private_data;
> @@ -4186,6 +4215,7 @@ static const struct file_operations kvm_vcpu_fops = {
> .release = kvm_vcpu_release,
> .unlocked_ioctl = kvm_vcpu_ioctl,
> .mmap = kvm_vcpu_mmap,
> + .poll = kvm_vcpu_poll,
> .llseek = noop_llseek,
> KVM_COMPAT(kvm_vcpu_compat_ioctl),
> };
A few ideas on the design:
I think that we can do this in a simpler way.
I am thinking about the following API:
-> vCPU does vtlcall and KVM exits to the userspace.
-> The userspace sets the vCPU to the new MP runstate (KVM_MP_STATE_HALTED_USERSPACE) which is just like regular halt
but once vCPU is ready to run, the KVM instead exits to userspace.
-> The userspace does another KVM_RUN which blocks till an event comes and exits back to userspace.
-> The userspace can now decide what to do, and it might for example send signal to vCPU thread which runs VTL0,
to kick it out of the guest mode, and resume the VTL1 vCPU.
Best regards,
Maxim Levitsky