Re: [Qemu-block] [PATCH 6/6] coroutine-lock: make CoRwlock thread-safe and fair

[Paolo Bonzini]

On 15/02/2017 10:23, Fam Zheng wrote:
> On Mon, 02/13 19:12, Paolo Bonzini wrote:
>> This adds a CoMutex around the existing CoQueue.  Because the write-side
> 
> s/CoQueue/CoRwlock/

No, I meant that CoRwlock has a CoQueue, and after this patch it is
wrapped by a CoMutex too.


>> @@ -375,16 +384,20 @@ void qemu_co_rwlock_unlock(CoRwlock *lock)
>>              qemu_co_queue_next(&lock->queue);
>>          }
>>      }
>> -    self->locks_held--;
>> +    qemu_co_mutex_unlock(&lock->mutex);
>>  }
>>  
>>  void qemu_co_rwlock_wrlock(CoRwlock *lock)
>>  {
>> -    Coroutine *self = qemu_coroutine_self();
>> -
>> -    while (lock->writer || lock->reader) {
>> -        qemu_co_queue_wait(&lock->queue, NULL);
>> +    qemu_co_mutex_lock(&lock->mutex);
>> +    lock->pending_writer++;
>> +    while (lock->reader) {
>> +        qemu_co_queue_wait(&lock->queue, &lock->mutex);
>>      }
>> -    lock->writer = true;
>> -    self->locks_held++;
>> +    lock->pending_writer--;
>> +
>> +    /* The rest of the write-side critical section is run with
>> +     * the mutex taken, so that lock->reader remains zero.
>> +     * There is no need to update self->locks_held.
>> +     */
> 
> But is it still better to update self->locks_held anyway for the
> 'assert(!co->locks_held)' in qemu_coroutine_enter? Or is the same thing 
> checked
> elsewhere?

self->locks_held is already incremented by the qemu_co_mutex_lock call
at the beginning of qemu_co_rwlock_wrlock.  It is then decremented in
qemu_co_rwlock_unlock.

For the read side, rdlock _unlocks_ lock->mutex before returning, so
self->locks_held must be incremented by rdlock and decremented by unlock.

Paolo