[PATCH] KVM: arm/arm64: Fix arch timers with userspace irqchips

[Christoffer Dall]

On Wed, Jan 31, 2018 at 12:19 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
> On 31/01/18 10:05, Christoffer Dall wrote:
>> On Wed, Jan 31, 2018 at 09:37:31AM +0000, Marc Zyngier wrote:
>>> On 30/01/18 12:46, Christoffer Dall wrote:
>>>> When introducing support for irqchip in userspace we needed a way to
>>>> mask the timer signal to prevent the guest continuously exiting due to a
>>>> screaming timer.
>>>>
>>>> We did this by disabling the corresponding percpu interrupt on the
>>>> host interrupt controller, because we cannot rely on the host system
>>>> having a GIC, and therefore cannot make any assumptions about having an
>>>> active state to hide the timer signal.
>>>>
>>>> Unfortunately, when introducing this feature, it became entirely
>>>> possible that a VCPU which belongs to a VM that has a userspace irqchip
>>>> can disable the vtimer irq on the host on some physical CPU, and then go
>>>> away without ever enabling the vimter irq on that physical CPU again.
>>>
>>> vtimer
>>>
>>>>
>>>> This means that using irqchips in userspace on a system that also
>>>> supports running VMs with an in-kernel GIC can prevent forward progress
>>>> from in-kernel GIC VMs.
>>>>
>>>> Later on, when we started taking virtual timer interrupts in the arch
>>>> timer code, we would also leave this timer state active for userspace
>>>> irqchip VMs, because we leave it up to a VGIC-enabled guest to
>>>> deactivate the hardware IRQ using the HW bit in the LR.
>>>>
>>>> Both issues are solved by only using the enable/disable trick on systems
>>>> that do not have a host GIC which supports the active state, because all
>>>> VMs on such systems must use irqchips in userspace.  Systems that have a
>>>> working GIC with support for an active state use the active state to
>>>> mask the timer signal for both userspace an in-kernel irqchips.
>>
>> this should also be "and"
>>
>>>>
>>>> Cc: Alexander Graf <agraf at suse.de>
>>>> Cc: <stable at vger.kernel.org> # v4.12+
>>>> Fixes: d9e139778376 ("KVM: arm/arm64: Support arch timers with a userspace gic")
>>>> Signed-off-by: Christoffer Dall <christoffer.dall at linaro.org>
>>>> ---
>>>> This conflicts horribly with everything when applied to either
>>>> kvmarm/queue or kvmarm/master.  Therefore, this patch is written for
>>>> (and applies to) v4.15 with kvmarm/queue merged and should therefore
>>>> apply cleanly after v4.16-rc1.  An example with this patch applied can
>>>> be found on kvmarm/temp-for-v4.16-rc2.  I plan on sending this along
>>>> with any other potential fixes post v4.16-rc1.
>>>>
>>>>  virt/kvm/arm/arch_timer.c | 77 ++++++++++++++++++++++++++---------------------
>>>>  1 file changed, 42 insertions(+), 35 deletions(-)
>>>>
>>>> diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c
>>>> index 70268c0bec79..228906ceb722 100644
>>>> --- a/virt/kvm/arm/arch_timer.c
>>>> +++ b/virt/kvm/arm/arch_timer.c
>>>> @@ -35,6 +35,7 @@
>>>>  static struct timecounter *timecounter;
>>>>  static unsigned int host_vtimer_irq;
>>>>  static u32 host_vtimer_irq_flags;
>>>> +static bool has_gic_active_state;
>>>>
>>>>  static const struct kvm_irq_level default_ptimer_irq = {
>>>>     .irq    = 30,
>>>> @@ -69,25 +70,6 @@ static void soft_timer_cancel(struct hrtimer *hrt, struct work_struct *work)
>>>>             cancel_work_sync(work);
>>>>  }
>>>>
>>>> -static void kvm_vtimer_update_mask_user(struct kvm_vcpu *vcpu)
>>>> -{
>>>> -   struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
>>>> -
>>>> -   /*
>>>> -    * When using a userspace irqchip with the architected timers, we must
>>>> -    * prevent continuously exiting from the guest, and therefore mask the
>>>> -    * physical interrupt by disabling it on the host interrupt controller
>>>> -    * when the virtual level is high, such that the guest can make
>>>> -    * forward progress.  Once we detect the output level being
>>>> -    * de-asserted, we unmask the interrupt again so that we exit from the
>>>> -    * guest when the timer fires.
>>>> -    */
>>>> -   if (vtimer->irq.level)
>>>> -           disable_percpu_irq(host_vtimer_irq);
>>>> -   else
>>>> -           enable_percpu_irq(host_vtimer_irq, 0);
>>>> -}
>>>> -
>>>>  static irqreturn_t kvm_arch_timer_handler(int irq, void *dev_id)
>>>>  {
>>>>     struct kvm_vcpu *vcpu = *(struct kvm_vcpu **)dev_id;
>>>> @@ -107,8 +89,8 @@ static irqreturn_t kvm_arch_timer_handler(int irq, void *dev_id)
>>>>             kvm_timer_update_irq(vcpu, true, vtimer);
>>>>
>>>>     if (static_branch_unlikely(&userspace_irqchip_in_use) &&
>>>> -       unlikely(!irqchip_in_kernel(vcpu->kvm)))
>>>> -           kvm_vtimer_update_mask_user(vcpu);
>>>> +       unlikely(!irqchip_in_kernel(vcpu->kvm)) && !has_gic_active_state)
>>>> +           disable_percpu_irq(host_vtimer_irq);
>>>
>>> nit: this is the negated version of the fix you posted earlier
>>> (0e23cb34af26 in kvmarm/queue), plus a new term... How about rewriting
>>> this as:
>>>
>>> static inline bool userspace_irqchip(struct kvm *kvm)
>>> {
>>>      return (static_branch_unlikely(&userspace_irqchip_in_use) &&
>>>              unlikely(!irqchip_in_kernel(kvm)));
>>> }
>>>
>>> and this becomes:
>>>
>>>      if (userspace_irqchip(vcpu->kvm) && !has_gic_active_state)
>>>              [...]
>>>
>>> which I find much more readable.
>>>
>>> Same for kvm_timer_update_irq():
>>>
>>>      if (!userspace_irqchip(vcpu->kvm))
>>>              [...]
>>>
>>
>> yes, good idea, I find it more readable too.
>>
>>>>
>>>>     return IRQ_HANDLED;
>>>>  }
>>>> @@ -460,13 +442,16 @@ static void set_cntvoff(u64 cntvoff)
>>>>     kvm_call_hyp(__kvm_timer_set_cntvoff, low, high);
>>>>  }
>>>>
>>>> -static void kvm_timer_vcpu_load_vgic(struct kvm_vcpu *vcpu)
>>>> +static void kvm_timer_vcpu_load_gic(struct kvm_vcpu *vcpu)
>>>>  {
>>>>     struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
>>>>     bool phys_active;
>>>>     int ret;
>>>>
>>>> -   phys_active = kvm_vgic_map_is_active(vcpu, vtimer->irq.irq);
>>>> +   if (irqchip_in_kernel(vcpu->kvm))
>>>> +           phys_active = kvm_vgic_map_is_active(vcpu, vtimer->irq.irq);
>>>> +   else
>>>> +           phys_active = vtimer->irq.level;
>>>
>>> You could use the above helper here too.
>>>
>>
>> yes
>>
>>>>
>>>>     ret = irq_set_irqchip_state(host_vtimer_irq,
>>>>                                 IRQCHIP_STATE_ACTIVE,
>>>> @@ -474,9 +459,24 @@ static void kvm_timer_vcpu_load_vgic(struct kvm_vcpu *vcpu)
>>>>     WARN_ON(ret);
>>>>  }
>>>>
>>>> -static void kvm_timer_vcpu_load_user(struct kvm_vcpu *vcpu)
>>>> +static void kvm_timer_vcpu_load_nogic(struct kvm_vcpu *vcpu)
>>>>  {
>>>> -   kvm_vtimer_update_mask_user(vcpu);
>>>> +   struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
>>>> +
>>>> +   /*
>>>> +    * When using a userspace irqchip with the architected timers and a
>>>> +    * host interrupt controller that doesn't support an active state, we
>>>> +    * must still we must prevent continuously exiting from the guest, and
>>
>> and here I should s/we must//
>>
>>>> +    * therefore mask the physical interrupt by disabling it on the host
>>>> +    * interrupt controller when the virtual level is high, such that the
>>>> +    * guest can make forward progress.  Once we detect the output level
>>>> +    * being de-asserted, we unmask the interrupt again so that we exit
>>>> +    * from the guest when the timer fires.
>>>> +    */
>>>> +   if (vtimer->irq.level)
>>>> +           disable_percpu_irq(host_vtimer_irq);
>>>> +   else
>>>> +           enable_percpu_irq(host_vtimer_irq, host_vtimer_irq_flags);
>>>>  }
>>>>
>>>>  void kvm_timer_vcpu_load(struct kvm_vcpu *vcpu)
>>>> @@ -487,10 +487,10 @@ void kvm_timer_vcpu_load(struct kvm_vcpu *vcpu)
>>>>     if (unlikely(!timer->enabled))
>>>>             return;
>>>>
>>>> -   if (unlikely(!irqchip_in_kernel(vcpu->kvm)))
>>>> -           kvm_timer_vcpu_load_user(vcpu);
>>>> +   if (has_gic_active_state)
>>>
>>> likely()?
>>>
>>
>> yes, will be fixes if using a static key.
>>
>>>> +           kvm_timer_vcpu_load_gic(vcpu);
>>>>     else
>>>> -           kvm_timer_vcpu_load_vgic(vcpu);
>>>> +           kvm_timer_vcpu_load_nogic(vcpu);
>>>>
>>>>     set_cntvoff(vtimer->cntvoff);
>>>>
>>>> @@ -555,18 +555,23 @@ static void unmask_vtimer_irq_user(struct kvm_vcpu *vcpu)
>>>>  {
>>>>     struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
>>>>
>>>> -   if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
>>>> -           __timer_snapshot_state(vtimer);
>>>> -           if (!kvm_timer_should_fire(vtimer)) {
>>>> -                   kvm_timer_update_irq(vcpu, false, vtimer);
>>>> -                   kvm_vtimer_update_mask_user(vcpu);
>>>> -           }
>>>> +   __timer_snapshot_state(vtimer);
>>>> +   if (!kvm_timer_should_fire(vtimer)) {
>>>> +           kvm_timer_update_irq(vcpu, false, vtimer);
>>>> +           if (!has_gic_active_state)
>>>
>>> unlikely()?
>>>
>>
>> same
>>
>>>> +                   enable_percpu_irq(host_vtimer_irq, host_vtimer_irq_flags);
>>
>> So, reading your reply made me think.  Shouldn't this actually be:
>>
>>               if (static_key_likely(has_gic_active_state))
>>                       enable_percpu_irq(host_vtimer_irq, host_vtimer_irq_flags);
>>               else
>>                       irq_set_irqchip_state(host_vtimer_irq, IRQCHIP_STATE_ACTIVE, false);
>
> Shouldn't it be the other way around? Use the active state when we have
> a GIC and mask if we don't?
>

Yes, it should.  Doh.

>>
>> ... which makes me want to wrap the irqchip call in a small wrapper
>> called from here and from load().  What do you think?
>
> Indeed. Some operation that "does the right thing", no matter what what
> bizarre configuration we're in. It would definitely help understanding
> the flow which has become a bit unwieldy.
>

I tried reworking the load function to first branch off
userspace_irqchip() and then call the sub-function from here, but it
was no more readable, so I've done something in between for v2.  Stay
tuned...

Did I mention that I hate this feature, which keeps breaking, and
which really isn't covered by a simple kvm-unit-test script?

-Christoffer