[PATCH v2] KVM: PPC: Defer vtime accounting 'til after IRQ handling

[Nicholas Piggin]

Excerpts from Laurent Vivier's message of October 20, 2021 4:29 pm:
> On 15/10/2021 04:23, Nicholas Piggin wrote:
>> Excerpts from Laurent Vivier's message of October 13, 2021 7:30 pm:
>>> On 13/10/2021 01:18, Michael Ellerman wrote:
>>>> Laurent Vivier <lvivier at redhat.com> writes:
>>>>> Commit 112665286d08 moved guest_exit() in the interrupt protected
>>>>> area to avoid wrong context warning (or worse), but the tick counter
>>>>> cannot be updated and the guest time is accounted to the system time.
>>>>>
>>>>> To fix the problem port to POWER the x86 fix
>>>>> 160457140187 ("Defer vtime accounting 'til after IRQ handling"):
>>>>>
>>>>> "Defer the call to account guest time until after servicing any IRQ(s)
>>>>>    that happened in the guest or immediately after VM-Exit.  Tick-based
>>>>>    accounting of vCPU time relies on PF_VCPU being set when the tick IRQ
>>>>>    handler runs, and IRQs are blocked throughout the main sequence of
>>>>>    vcpu_enter_guest(), including the call into vendor code to actually
>>>>>    enter and exit the guest."
>>>>>
>>>>> Fixes: 112665286d08 ("KVM: PPC: Book3S HV: Context tracking exit guest context before enabling irqs")
>>>>> Cc: npiggin at gmail.com
>>>>> Cc: <stable at vger.kernel.org> # 5.12
>>>>> Signed-off-by: Laurent Vivier <lvivier at redhat.com>
>>>>> ---
>>>>>
>>>>> Notes:
>>>>>       v2: remove reference to commit 61bd0f66ff92
>>>>>           cc stable 5.12
>>>>>           add the same comment in the code as for x86
>>>>>
>>>>>    arch/powerpc/kvm/book3s_hv.c | 24 ++++++++++++++++++++----
>>>>>    1 file changed, 20 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
>>>>> index 2acb1c96cfaf..a694d1a8f6ce 100644
>>>>> --- a/arch/powerpc/kvm/book3s_hv.c
>>>>> +++ b/arch/powerpc/kvm/book3s_hv.c
>>>> ...
>>>>> @@ -4506,13 +4514,21 @@ int kvmhv_run_single_vcpu(struct kvm_vcpu *vcpu, u64 time_limit,
>>>>>    
>>>>>    	srcu_read_unlock(&kvm->srcu, srcu_idx);
>>>>>    
>>>>> +	context_tracking_guest_exit();
>>>>> +
>>>>>    	set_irq_happened(trap);
>>>>>    
>>>>>    	kvmppc_set_host_core(pcpu);
>>>>>    
>>>>> -	guest_exit_irqoff();
>>>>> -
>>>>>    	local_irq_enable();
>>>>> +	/*
>>>>> +	 * Wait until after servicing IRQs to account guest time so that any
>>>>> +	 * ticks that occurred while running the guest are properly accounted
>>>>> +	 * to the guest.  Waiting until IRQs are enabled degrades the accuracy
>>>>> +	 * of accounting via context tracking, but the loss of accuracy is
>>>>> +	 * acceptable for all known use cases.
>>>>> +	 */
>>>>> +	vtime_account_guest_exit();
>>>>
>>>> This pops a warning for me, running guest(s) on Power8:
>>>>    
>>>>     [  270.745303][T16661] ------------[ cut here ]------------
>>>>     [  270.745374][T16661] WARNING: CPU: 72 PID: 16661 at arch/powerpc/kernel/time.c:311 vtime_account_kernel+0xe0/0xf0
>>>
>>> Thank you, I missed that...
>>>
>>> My patch is wrong, I have to add vtime_account_guest_exit() before the local_irq_enable().
>> 
>> I thought so because if we take an interrupt after exiting the guest that
>> should be accounted to kernel not guest.
>> 
>>>
>>> arch/powerpc/kernel/time.c
>>>
>>>    305 static unsigned long vtime_delta(struct cpu_accounting_data *acct,
>>>    306                                  unsigned long *stime_scaled,
>>>    307                                  unsigned long *steal_time)
>>>    308 {
>>>    309         unsigned long now, stime;
>>>    310
>>>    311         WARN_ON_ONCE(!irqs_disabled());
>>> ...
>>>
>>> But I don't understand how ticks can be accounted now if irqs are still disabled.
>>>
>>> Not sure it is as simple as expected...
>> 
>> I don't know all the timer stuff too well. The
>> !CONFIG_VIRT_CPU_ACCOUNTING case is relying on PF_VCPU to be set when
>> the host timer interrupt runs irqtime_account_process_tick runs so it
>> can accumulate that tick to the guest?
>> 
>> That probably makes sense then, but it seems like we need that in a
>> different place. Timer interrupts are not guaranteed to be the first one
>> to occur when interrupts are enabled.
>> 
>> Maybe a new tick_account_guest_exit() and move PF_VCPU clearing to that
>> for tick based accounting. Call it after local_irq_enable and call the
>> vtime accounting before it. Would that work?
> 
> Hi Nick,
> 
> I think I will not have the time to have a look to fix that?
> 
> Could you try?

Hey Laurent, sure I can take a look at it.

Thanks,
Nick