Re: [PATCH v9 4/5] qrwlock: Use smp_store_release() in write_unlock()

[Waiman Long <waiman.long@hp.com>]

On 01/20/2014 10:18 AM, Peter Zijlstra wrote:
> On Tue, Jan 14, 2014 at 11:44:06PM -0500, Waiman Long wrote:
>> This patch modifies the queue_write_unlock() function to use the new
>> smp_store_release() function (currently in tip). It also removes the
>> temporary implementation of smp_load_acquire() and smp_store_release()
>> function in qrwlock.c.
>>
>> This patch will use atomic subtraction instead if the writer field is
>> not atomic.
>>
>> Signed-off-by: Waiman Long<Waiman.Long@hp.com>
>> ---
>>   include/asm-generic/qrwlock.h |   10 ++++++----
>>   kernel/locking/qrwlock.c      |   34 ----------------------------------
>>   2 files changed, 6 insertions(+), 38 deletions(-)
>>
>> diff --git a/include/asm-generic/qrwlock.h b/include/asm-generic/qrwlock.h
>> index 5abb6ca..68f488b 100644
>> --- a/include/asm-generic/qrwlock.h
>> +++ b/include/asm-generic/qrwlock.h
>> @@ -181,11 +181,13 @@ static inline void queue_read_unlock(struct qrwlock *lock)
>>   static inline void queue_write_unlock(struct qrwlock *lock)
>>   {
>>   	/*
>> -	 * Make sure that none of the critical section will be leaked out.
>> +	 * If the writer field is atomic, it can be cleared directly.
>> +	 * Otherwise, an atomic subtraction will be used to clear it.
>>   	 */
>> -	smp_mb__before_clear_bit();
>> -	ACCESS_ONCE(lock->cnts.writer) = 0;
>> -	smp_mb__after_clear_bit();
>> +	if (__native_word(lock->cnts.writer))
>> +		smp_store_release(&lock->cnts.writer, 0);
>> +	else
>> +		atomic_sub(_QW_LOCKED,&lock->cnts.rwa);
>>   }
> If we're a writer, read-count must be zero. The only way for that not to
> be zero is a concurrent read-(try)lock. If you move all the
> read-(try)locks over to cmpxchg() you can avoid this afaict:

That is not true. A reader may transiently set the reader count to a 
non-zero value in the fast path. Also, a reader in interrupt context 
will force a non-zero reader count to take the read lock as soon as the 
writer is done.

>
> static inline void queue_read_trylock(struct qrwlock *lock)
> {
> 	union qrwcnts cnts
>
> 	cnts = ACCESS_ONCE(lock->cnts);
> 	if (!cnts.writer) {
> 		if (cmpxchg(&lock->cnts.rwc, cnts.rwc, cnts.rwc + _QR_BIAS) == cnts.rwc)
> 			return 1;
> 	}
>
> 	return 0;
> }
>
> static inline void queue_read_lock(struct qrwlock *lock)
> {
> 	if (!queue_read_trylock(lock))
> 		queue_read_lock_slowpath(); // XXX do not assume extra _QR_BIAS
> }
>
> At which point you have the guarantee that read-count == 0, and you can
> write:
>
> static inline void queue_write_unlock(struct qrwlock *lock)
> {
> 	smp_store_release(&lock->cnts.rwc, 0);
> }
>
> No?
>

The current code is optimized for the reader-heavy case. So I used xadd 
for incrementing reader count to reduce the chance of retry due to 
concurrent reader count updates. The downside is the need to back out if 
a writer is here.

I can change the logic to use only cmpxchg for readers, but I don't see 
a compelling reason to do so.

-Longman