Re: [RFC PATCH v1 2/4] KVM: arm64: GICv4.1: Try to save hw pending state in save_pending_tables

[Shenming Lu <lushenming@huawei.com>]

On 2020/11/24 16:26, Marc Zyngier wrote:
> On 2020-11-24 07:40, Shenming Lu wrote:
>> On 2020/11/23 17:18, Marc Zyngier wrote:
>>> On 2020-11-23 06:54, Shenming Lu wrote:
>>>> After pausing all vCPUs and devices capable of interrupting, in order
>>>         ^^^^^^^^^^^^^^^^^
>>> See my comment below about this.
>>>
>>>> to save the information of all interrupts, besides flushing the pending
>>>> states in kvm’s vgic, we also try to flush the states of VLPIs in the
>>>> virtual pending tables into guest RAM, but we need to have GICv4.1 and
>>>> safely unmap the vPEs first.
>>>>
>>>> Signed-off-by: Shenming Lu <lushenming@huawei.com>
>>>> ---
>>>>  arch/arm64/kvm/vgic/vgic-v3.c | 62 +++++++++++++++++++++++++++++++----
>>>>  1 file changed, 56 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/kvm/vgic/vgic-v3.c b/arch/arm64/kvm/vgic/vgic-v3.c
>>>> index 9cdf39a94a63..e1b3aa4b2b12 100644
>>>> --- a/arch/arm64/kvm/vgic/vgic-v3.c
>>>> +++ b/arch/arm64/kvm/vgic/vgic-v3.c
>>>> @@ -1,6 +1,8 @@
>>>>  // SPDX-License-Identifier: GPL-2.0-only
>>>>
>>>>  #include <linux/irqchip/arm-gic-v3.h>
>>>> +#include <linux/irq.h>
>>>> +#include <linux/irqdomain.h>
>>>>  #include <linux/kvm.h>
>>>>  #include <linux/kvm_host.h>
>>>>  #include <kvm/arm_vgic.h>
>>>> @@ -356,6 +358,39 @@ int vgic_v3_lpi_sync_pending_status(struct kvm
>>>> *kvm, struct vgic_irq *irq)
>>>>      return 0;
>>>>  }
>>>>
>>>> +/*
>>>> + * With GICv4.1, we can get the VLPI's pending state after unmapping
>>>> + * the vPE. The deactivation of the doorbell interrupt will trigger
>>>> + * the unmapping of the associated vPE.
>>>> + */
>>>> +static void get_vlpi_state_pre(struct vgic_dist *dist)
>>>> +{
>>>> +    struct irq_desc *desc;
>>>> +    int i;
>>>> +
>>>> +    if (!kvm_vgic_global_state.has_gicv4_1)
>>>> +        return;
>>>> +
>>>> +    for (i = 0; i < dist->its_vm.nr_vpes; i++) {
>>>> +        desc = irq_to_desc(dist->its_vm.vpes[i]->irq);
>>>> +        irq_domain_deactivate_irq(irq_desc_get_irq_data(desc));
>>>> +    }
>>>> +}
>>>> +
>>>> +static void get_vlpi_state_post(struct vgic_dist *dist)
>>>
>>> nit: the naming feels a bit... odd. Pre/post what?
>>
>> My understanding is that the unmapping is a preparation for get_vlpi_state...
>> Maybe just call it unmap/map_all_vpes?
> 
> Yes, much better.
> 
> [...]
> 
>>>> +        if (irq->hw) {
>>>> +            WARN_RATELIMIT(irq_get_irqchip_state(irq->host_irq,
>>>> +                        IRQCHIP_STATE_PENDING, &is_pending),
>>>> +                       "IRQ %d", irq->host_irq);
>>>
>>> Isn't this going to warn like mad on a GICv4.0 system where this, by definition,
>>> will generate an error?
>>
>> As we have returned an error in save_its_tables if hw && !has_gicv4_1, we don't
>> have to warn this here?
> 
> Are you referring to the check in vgic_its_save_itt() that occurs in patch 4?
> Fair enough, though I think the use of irq_get_irqchip_state() isn't quite
> what we want, as per my comments on patch #1.
> 
>>>
>>>> +        }
>>>> +
>>>> +        if (stored == is_pending)
>>>>              continue;
>>>>
>>>> -        if (irq->pending_latch)
>>>> +        if (is_pending)
>>>>              val |= 1 << bit_nr;
>>>>          else
>>>>              val &= ~(1 << bit_nr);
>>>>
>>>>          ret = kvm_write_guest_lock(kvm, ptr, &val, 1);
>>>>          if (ret)
>>>> -            return ret;
>>>> +            goto out;
>>>>      }
>>>> -    return 0;
>>>> +
>>>> +out:
>>>> +    get_vlpi_state_post(dist);
>>>
>>> This bit worries me: you have unmapped the VPEs, so any interrupt that has been
>>> generated during that phase is now forever lost (the GIC doesn't have ownership
>>> of the pending tables).
>>
>> In my opinion, during this phase, the devices capable of interrupting
>> should have  already been paused (prevent from sending interrupts),
>> such as VFIO migration protocol has already realized it.
> 
> Is that a hard guarantee? Pausing devices *may* be possible for a limited
> set of endpoints, but I'm not sure that is universally possible to restart
> them and expect a consistent state (you have just dropped a bunch of network
> packets on the floor...).

No, as far as I know, if the VFIO device does not support pause, the migration would
fail early... And the specific action is decided by the vendor driver.
In fact, the VFIO migration is still in an experimental phase... I will pay attention
to the follow-up development.

> 
>>> Do you really expect the VM to be restartable from that point? I don't see how
>>> this is possible.
>>>
>>
>> If the migration has encountered an error, the src VM might be
>> restarted, so we have to map the vPEs back.
> 
> As I said above, I doubt it is universally possible to do so, but
> after all, this probably isn't worse that restarting on the target...
> 
>         M.