On 19/01/2018 13:36, Pavel Dovgalyuk wrote:
>> From: Paolo Bonzini [mailto:[email protected]]
>> Sent: Friday, January 19, 2018 3:26 PM
>> To: Pavel Dovgalyuk; 'Pavel Dovgalyuk'; [email protected]
>> Cc: [email protected]; [email protected]; [email protected];
>> [email protected];
>> [email protected]; [email protected]; [email protected];
>> [email protected];
>> [email protected]; [email protected]
>> Subject: Re: [RFC PATCH v4 13/23] cpus: only take BQL for sleeping threads
>>
>> On 19/01/2018 13:25, Pavel Dovgalyuk wrote:
>>>>> It means, that I'll have to fix all the has_work function to avoid races,
>>>>> because x86_cpu_has_work may have them?
>>>> Why only x86_cpu_has_work?
>>>>
>>>> Even reading cs->interrupt_request outside the mutex is unsafe.
>>> All the vcpu function that access interrupt controller or peripheral state
>>> may be unsafe?
>>> How can it work safely then?
>>
>> They do it inside the big QEMU lock.
>
> Right. Without these patches.
> They are within the replay lock. And BQL is not covering vcpu execution with
> these patches.
> Therefore RR will be ok and regular execution may encounter races?
> It means that I missed something in Alex ideas, because he prepared the
> initial patches.
Yes.
>> But here you're calling cpu_has_work (via all_cpu_threads_idle) outside the
>> lock.
>
> Yes, I see, but what we have to do?
I don't know. But the idiom in these patches,
while(...) {
lock()
cond_wait()
unlock()
}
is unsafe as well, so the issue is more than just cpu_has_work.
Paolo
