On Sat, May 06, 2017 at 08:08:09PM +0200, Christoffer Dall wrote:
> On Wed, May 03, 2017 at 06:06:29PM +0200, Andrew Jones wrote:
> > VCPU halting/resuming is partially implemented with VCPU requests.
> > When kvm_arm_halt_guest() is called all VCPUs get the EXIT request,
> > telling them to exit guest mode and look at the state of 'pause',
> > which will be true, telling them to sleep. As ARM's VCPU RUN
> > implements the memory barrier pattern described in "Ensuring Requests
> > Are Seen" of Documentation/virtual/kvm/vcpu-requests.rst, there's
> > no way for a VCPU halted by kvm_arm_halt_guest() to miss the pause
> > state change. However, before this patch, a single VCPU halted with
> > kvm_arm_halt_vcpu() did not get a request, opening a tiny race window.
> > This patch adds the request, closing the race window and also allowing
> > us to remove the final check of pause in VCPU RUN, as the final check
> > for requests is sufficient.
> >
> > Signed-off-by: Andrew Jones <drjones@xxxxxxxxxx>
> >
> > ---
> >
> > I have two questions about the halting/resuming.
> >
> > Question 1:
> >
> > Do we even need kvm_arm_halt_vcpu()/kvm_arm_resume_vcpu()? It should
> > only be necessary if one VCPU can activate or inactivate the private
> > IRQs of another VCPU, right? That doesn't seem like something that
> > should be possible, but I'm GIC-illiterate...
>
> True, it shouldn't be possible. I wonder if we were thinking of
> userspace access to the CPU-specific data, but we already ensure that no
> VCPUs are running at that time, so I don't think it should be necessary.
>
> >
> > Question 2:
> >
> > It's not clear to me if we have another problem with halting/resuming
> > or not. If it's possible for VCPU1 and VCPU2 to race in
> > vgic_mmio_write_s/cactive(), then the following scenario could occur,
> > leading to VCPU3 being in guest mode when it should not be. Does the
> > hardware prohibit more than one VCPU entering trap handlers that lead
> > to these functions at the same time? If not, then I guess pause needs
> > to be a counter instead of a boolean.
> >
> > VCPU1 VCPU2 VCPU3
> > ----- ----- -----
> > VCPU3->pause = true;
> > halt(VCPU3);
> > if (pause)
> > sleep();
> > VCPU3->pause = true;
> > halt(VCPU3);
> > VCPU3->pause = false;
> > resume(VCPU3);
> > ...wake up...
> > if (!pause)
> > Enter guest mode. Bad!
> > VCPU3->pause = false;
> > resume(VCPU3);
> >
> > (Yes, the "Bad!" is there to both identify something we don't want
> > occurring and to make fun of Trump's tweeting style.)
>
> I think it's bad, and it might be even worse, because it could lead to a
> CPU looping forever in the host kernel, since there's no guarantee to
> exit from the VM in the other VCPU thread.
>
> But I think simply taking the kvm->lock mutex to serialize the mmio
> active change operations should be sufficient.
>
> If we agree on this I can send a patch with your reported by that fixes
> that issue, which gets rid of kvm_arm_halt_vcpu and requires you to
> modify your first patch to clear the KVM_REQ_VCPU_EXIT flag for each
> vcpu in kvm_arm_halt_guest instead and you can fold the remaining change
> from this patch into a patch that completely gets rid of the pause flag.
Yup, seems reasonable to me to lock the kvm mutex on a stop the guest type
action.
>
> See untested patch draft at the end of this mail.
>
> Thanks,
> -Christoffer
>
> > ---
> > arch/arm/kvm/arm.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c
> > index 47f6c7fdca96..9174ed13135a 100644
> > --- a/arch/arm/kvm/arm.c
> > +++ b/arch/arm/kvm/arm.c
> > @@ -545,6 +545,7 @@ void kvm_arm_halt_guest(struct kvm *kvm)
> > void kvm_arm_halt_vcpu(struct kvm_vcpu *vcpu)
> > {
> > vcpu->arch.pause = true;
> > + kvm_make_request(KVM_REQ_VCPU_EXIT, vcpu);
> > kvm_vcpu_kick(vcpu);
> > }
> >
> > @@ -664,7 +665,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
> >
> > if (ret <= 0 || need_new_vmid_gen(vcpu->kvm) ||
> > kvm_request_pending(vcpu) ||
> > - vcpu->arch.power_off || vcpu->arch.pause) {
> > + vcpu->arch.power_off) {
> > vcpu->mode = OUTSIDE_GUEST_MODE;
> > local_irq_enable();
> > kvm_pmu_sync_hwstate(vcpu);
> > --
> > 2.9.3
> >
>
>
> Untested draft patch:
>
> diff --git a/arch/arm/include/asm/kvm_host.h b/arch/arm/include/asm/kvm_host.h
> index d488b88..b77a3af 100644
> --- a/arch/arm/include/asm/kvm_host.h
> +++ b/arch/arm/include/asm/kvm_host.h
> @@ -234,8 +234,6 @@ struct kvm_vcpu *kvm_arm_get_running_vcpu(void);
> struct kvm_vcpu __percpu **kvm_get_running_vcpus(void);
> void kvm_arm_halt_guest(struct kvm *kvm);
> void kvm_arm_resume_guest(struct kvm *kvm);
> -void kvm_arm_halt_vcpu(struct kvm_vcpu *vcpu);
> -void kvm_arm_resume_vcpu(struct kvm_vcpu *vcpu);
>
> int kvm_arm_copy_coproc_indices(struct kvm_vcpu *vcpu, u64 __user *uindices);
> unsigned long kvm_arm_num_coproc_regs(struct kvm_vcpu *vcpu);
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 578df18..7a38d5a 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -334,8 +334,6 @@ struct kvm_vcpu *kvm_arm_get_running_vcpu(void);
> struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void);
> void kvm_arm_halt_guest(struct kvm *kvm);
> void kvm_arm_resume_guest(struct kvm *kvm);
> -void kvm_arm_halt_vcpu(struct kvm_vcpu *vcpu);
> -void kvm_arm_resume_vcpu(struct kvm_vcpu *vcpu);
>
> u64 __kvm_call_hyp(void *hypfn, ...);
> #define kvm_call_hyp(f, ...) __kvm_call_hyp(kvm_ksym_ref(f), ##__VA_ARGS__)
> diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c
> index 7941699..932788a 100644
> --- a/virt/kvm/arm/arm.c
> +++ b/virt/kvm/arm/arm.c
> @@ -542,27 +542,15 @@ void kvm_arm_halt_guest(struct kvm *kvm)
> kvm_make_all_cpus_request(kvm, KVM_REQ_VCPU_EXIT);
> }
>
> -void kvm_arm_halt_vcpu(struct kvm_vcpu *vcpu)
> -{
> - vcpu->arch.pause = true;
> - kvm_vcpu_kick(vcpu);
> -}
> -
> -void kvm_arm_resume_vcpu(struct kvm_vcpu *vcpu)
> -{
> - struct swait_queue_head *wq = kvm_arch_vcpu_wq(vcpu);
> -
> - vcpu->arch.pause = false;
> - swake_up(wq);
> -}
> -
> void kvm_arm_resume_guest(struct kvm *kvm)
> {
> int i;
> struct kvm_vcpu *vcpu;
>
> - kvm_for_each_vcpu(i, vcpu, kvm)
> - kvm_arm_resume_vcpu(vcpu);
> + kvm_for_each_vcpu(i, vcpu, kvm) {
> + vcpu->arch.pause = false;
> + swake_up(kvm_arch_vcpu_wq(vcpu));
> + }
> }
>
> static void vcpu_sleep(struct kvm_vcpu *vcpu)
> diff --git a/virt/kvm/arm/vgic/vgic-mmio.c b/virt/kvm/arm/vgic/vgic-mmio.c
> index 2a5db13..c143add 100644
> --- a/virt/kvm/arm/vgic/vgic-mmio.c
> +++ b/virt/kvm/arm/vgic/vgic-mmio.c
> @@ -231,23 +231,21 @@ static void vgic_mmio_change_active(struct kvm_vcpu *vcpu, struct vgic_irq *irq,
> * be migrated while we don't hold the IRQ locks and we don't want to be
> * chasing moving targets.
> *
> - * For private interrupts, we only have to make sure the single and only VCPU
> - * that can potentially queue the IRQ is stopped.
> + * For private interrupts we don't have to do anything because userspace
> + * accesses to the VGIC state already require all VCPUs to be stopped, and
> + * only the VCPU itself can modify its private interrupts active state, which
> + * guarantees that the VCPU is not running.
> */
> static void vgic_change_active_prepare(struct kvm_vcpu *vcpu, u32 intid)
> {
> - if (intid < VGIC_NR_PRIVATE_IRQS)
> - kvm_arm_halt_vcpu(vcpu);
> - else
> + if (intid > VGIC_NR_PRIVATE_IRQS)
> kvm_arm_halt_guest(vcpu->kvm);
> }
>
> /* See vgic_change_active_prepare */
> static void vgic_change_active_finish(struct kvm_vcpu *vcpu, u32 intid)
> {
> - if (intid < VGIC_NR_PRIVATE_IRQS)
> - kvm_arm_resume_vcpu(vcpu);
> - else
> + if (intid > VGIC_NR_PRIVATE_IRQS)
> kvm_arm_resume_guest(vcpu->kvm);
> }
>
> @@ -258,6 +256,7 @@ void vgic_mmio_write_cactive(struct kvm_vcpu *vcpu,
> u32 intid = VGIC_ADDR_TO_INTID(addr, 1);
> int i;
>
> + mutex_lock(&vcpu->kvm->lock);
> vgic_change_active_prepare(vcpu, intid);
> for_each_set_bit(i, &val, len * 8) {
> struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, intid + i);
> @@ -265,6 +264,7 @@ void vgic_mmio_write_cactive(struct kvm_vcpu *vcpu,
> vgic_put_irq(vcpu->kvm, irq);
> }
> vgic_change_active_finish(vcpu, intid);
> + mutex_unlock(&vcpu->kvm->lock);
> }
>
> void vgic_mmio_write_sactive(struct kvm_vcpu *vcpu,
> @@ -274,6 +274,7 @@ void vgic_mmio_write_sactive(struct kvm_vcpu *vcpu,
> u32 intid = VGIC_ADDR_TO_INTID(addr, 1);
> int i;
>
> + mutex_lock(&vcpu->kvm->lock);
> vgic_change_active_prepare(vcpu, intid);
> for_each_set_bit(i, &val, len * 8) {
> struct vgic_irq *irq = vgic_get_irq(vcpu->kvm, vcpu, intid + i);
> @@ -281,6 +282,7 @@ void vgic_mmio_write_sactive(struct kvm_vcpu *vcpu,
> vgic_put_irq(vcpu->kvm, irq);
> }
> vgic_change_active_finish(vcpu, intid);
> + mutex_unlock(&vcpu->kvm->lock);
> }
>
> unsigned long vgic_mmio_read_priority(struct kvm_vcpu *vcpu,
Looks good to me. How about adding kvm->lock to the locking order comment
at the top of virt/kvm/arm/vgic/vgic.c too. With that, you can add my R-b
on the posting.
I'll rebase this series on your posting.
Thanks,
drew
