[PATCH] KVM: arm64: vgic: Fix soft lockup during VM teardown

[Shanker Donthineni]

Hi Marc,

On 1/19/23 08:01, Marc Zyngier wrote:
> External email: Use caution opening links or attachments
> 
> 
> On Thu, 19 Jan 2023 13:00:49 +0000,
> Shanker Donthineni <sdonthineni at nvidia.com> wrote:
>>
>>
>>
>> On 1/19/23 01:11, Marc Zyngier wrote:
>>> So you can see the VM being torn down while the vgic save sequence is
>>> still in progress?
>>>
>>> If you can actually see that, then this is a much bigger bug than the
>>> simple race you are describing, and we're missing a reference on the
>>> kvm structure. This would be a *MAJOR* bug.
>>>
>> How do we know vGIC save sequence is in progress while VM is being
>> teardown?  I'm launching/terminating ~32 VMs in a loop to reproduce
>> the issue.
> 
> Errr... *you* know when you are issuing the save ioctl, right? You
> also know when you are terminating the VM (closing its fd or killing
> the VMM).
> 
>>
>>> Please post the full traces, not snippets. The absolutely full kernel
>>> log, the configuration, what you run, how you run it, *EVERYTHING*. I
>>> need to be able to reproduce this.
>> Sure, I'll share the complete boot log messages of host kernel next run.
>>
>>>
>>>>
>>>>>>
>>>>>> irqreturn_t handle_irq_event(struct irq_desc *desc)
>>>>>> {
>>>>>>        irqd_set(&desc->irq_data, IRQD_IRQ_INPROGRESS);
>>>>>>        raw_spin_unlock(&desc->lock);
>>>>>>
>>>>>>        ret = handle_irq_event_percpu(desc);
>>>>>>
>>>>>>        raw_spin_lock(&desc->lock);
>>>>>>        irqd_clear(&desc->irq_data, IRQD_IRQ_INPROGRESS);
>>>>>> }
>>>>>
>>>>> How is that relevant to this trace? Do you see this function running
>>>>> concurrently with the teardown? If it matters here, it must be a VPE
>>>>> doorbell, right? But you claim that this is on a GICv4 platform, while
>>>>> this would only affect GICv4.1... Or are you using GICv4.1?
>>>>>
>>>> handle_irq_event() is running concurrently with irq_domain_activate_irq()
>>>> which happens before free_irq() called. Corruption at [78.983544] and
>>>> teardown started at [87.360891].
>>>
>>> But that doesn't match the description you made of concurrent
>>> events. Does it take more than 9 seconds for the vgic state to be
>>> saved to memory?
>>
>> Are there any other possibilities of corrupting IRQD_IRQ_INPROGRESS
>> state bit other than concurrent accesses?
> 
> Forget about this bit. You said that we could see the VM teardown
> happening *at the same time* as the vgic state saving, despite the
> vgic device holding a reference on the kvm structure. If that's the
> case, this bit is the least of our worries. Think of the consequences
> for a second...
> 
> [...]
> 
>> Using the below steps for launching/terminating 32 VMs in loop. The
>> failure is intermittent. The same issue is reproducible with KVMTOOL
>> also.
> 
> kvmtool never issue a KVM_DEV_ARM_VGIC_GRP_CTRL with the
> KVM_DEV_ARM_ITS_SAVE_TABLES argument, so the code path we discussed is
> never used. What is the exact problem you're observing with kvmtool
> as the VMM?
> 
Unfortunately I didn't capture the log messages of kvmtool based VM boot.
Saw once kernel crash. Now I'm trying to reproduce the issue with QEMU tool
and share complete log messages.

>          M.
> 
> --
> Without deviation from the norm, progress is not possible.