Re: [Qemu-devel] [RFC PATCH] Use atomic cmpxchg to atomically check the exclusive value in a STREX

[Frederic Konrad <fred.konrad@greensocs.com>]

On 10/06/2015 10:03, Frederic Konrad wrote:
> On 09/06/2015 15:55, Alex Bennée wrote:
>> Alex Bennée <alex.bennee@linaro.org> writes:
>>
>>> fred.konrad@greensocs.com writes:
>>>
>>>> From: KONRAD Frederic <fred.konrad@greensocs.com>
>>>>
>>>> This mechanism replaces the existing load/store exclusive mechanism 
>>>> which seems
>>>> to be broken for multithread.
>>>> It follows the intention of the existing mechanism and stores the 
>>>> target address
>>>> and data values during a load operation and checks that they remain 
>>>> unchanged
>>>> before a store.
>>>>
>>>> In common with the older approach, this provides weaker semantics 
>>>> than required
>>>> in that it could be that a different processor writes the same 
>>>> value as a
>>>> non-exclusive write, however in practise this seems to be irrelevant.
>> <snip>
>>>>   +/* Protect cpu_exclusive_* variable .*/
>>>> +__thread bool cpu_have_exclusive_lock;
>>>> +QemuMutex cpu_exclusive_lock;
>>>> +
>>>> +inline void arm_exclusive_lock(void)
>>>> +{
>>>> +    if (!cpu_have_exclusive_lock) {
>>>> +        qemu_mutex_lock(&cpu_exclusive_lock);
>>>> +        cpu_have_exclusive_lock = true;
>>>> +    }
>>>> +}
>>>> +
>>>> +inline void arm_exclusive_unlock(void)
>>>> +{
>>>> +    if (cpu_have_exclusive_lock) {
>>>> +        cpu_have_exclusive_lock = false;
>>>> +        qemu_mutex_unlock(&cpu_exclusive_lock);
>>>> +    }
>>>> +}
>>> I don't quite follow. If these locks are mean to be protecting 
>>> access to
>>> variables then how do they do that? The lock won't block if another
>>> thread is currently messing with the protected values.
>> Having re-read after coffee I'm still wondering why we need the
>> per-thread bool? All the lock/unlock pairs are for critical sections so
>> don't we just want to serialise on the qemu_mutex_lock(), what do the
>> flags add apart from allowing you to next locks that shouldn't happen?
>>
>>
> You are probably right, this might be a rest of the old approach.
> There were branches so we needed to allow next locks.
>

Hmm actually it seems necessary as we mutex the entire atomic_cmpxchg64.
If tlb_fill trigger an exception we re-lock the mutex.

> Thanks,
> Fred