Re: [PATCH 1/2] rwsem: check the lock before cpmxchg in down_write_trylock and rwsem_do_wake

[Alex Shi <alex.shi@intel.com>]

On 06/22/2013 03:21 PM, Peter Hurley wrote:
> On 06/21/2013 07:51 PM, Tim Chen wrote:
>> Doing cmpxchg will cause cache bouncing when checking
>> sem->count. This could cause scalability issue
>> in a large machine (e.g. a 80 cores box).
>>
>> A pre-read of sem->count can mitigate this.
>>
>> Signed-off-by: Alex Shi <alex.shi@intel.com>
>> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
>> ---
>>   include/asm-generic/rwsem.h |    8 ++++----
>>   lib/rwsem.c                 |   21 +++++++++++++--------
>>   2 files changed, 17 insertions(+), 12 deletions(-)
>>
>> diff --git a/include/asm-generic/rwsem.h b/include/asm-generic/rwsem.h
>> index bb1e2cd..052d973 100644
>> --- a/include/asm-generic/rwsem.h
>> +++ b/include/asm-generic/rwsem.h
>> @@ -70,11 +70,11 @@ static inline void __down_write(struct
>> rw_semaphore *sem)
>>
>>   static inline int __down_write_trylock(struct rw_semaphore *sem)
>>   {
>> -    long tmp;
>> +    if (unlikely(&sem->count != RWSEM_UNLOCKED_VALUE))
>                      ^^^^^^^^^^^
> 
> This is probably not what you want.
> 

this function logical is quite simple. check the sem->count before
cmpxchg is no harm this logical.

So could you like to tell us what should we want?

> 
>> +        return 0;
>>
>> -    tmp = cmpxchg(&sem->count, RWSEM_UNLOCKED_VALUE,
>> -              RWSEM_ACTIVE_WRITE_BIAS);
>> -    return tmp == RWSEM_UNLOCKED_VALUE;
>> +    return cmpxchg(&sem->count, RWSEM_UNLOCKED_VALUE,
>> +        RWSEM_ACTIVE_WRITE_BIAS) == RWSEM_UNLOCKED_VALUE;
>>   }
>>
>>   /*
>> diff --git a/lib/rwsem.c b/lib/rwsem.c
>> index 19c5fa9..2072af5 100644
>> --- a/lib/rwsem.c
>> +++ b/lib/rwsem.c
>> @@ -75,7 +75,7 @@ __rwsem_do_wake(struct rw_semaphore *sem, enum
>> rwsem_wake_type wake_type)
>>                * will block as they will notice the queued writer.
>>                */
>>               wake_up_process(waiter->task);
>> -        goto out;
>> +        return sem;
> 
> Please put these flow control changes in a separate patch.

I had sent the split patches to Tim&Davidlohr. They will send them out
as a single patchset.
> 
> 
>>       }
>>
>>       /* Writers might steal the lock before we grant it to the next
>> reader.
>> @@ -85,15 +85,21 @@ __rwsem_do_wake(struct rw_semaphore *sem, enum
>> rwsem_wake_type wake_type)
>>       adjustment = 0;
>>       if (wake_type != RWSEM_WAKE_READ_OWNED) {
>>           adjustment = RWSEM_ACTIVE_READ_BIAS;
>> - try_reader_grant:
>> -        oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
>> -        if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
>> -            /* A writer stole the lock. Undo our reader grant. */
>> +        while (1) {
>> +            /* A writer stole the lock. */
>> +            if (sem->count < RWSEM_WAITING_BIAS)
>> +                return sem;
> 
> I'm all for structured looping instead of goto labels but this optimization
> is only useful on the 1st iteration. IOW, on the second iteration you
> already
> know that you need to try for reclaiming the lock.
> 

sorry. could you like to say more clear, what's the 1st or 2nd iteration
or others?
> 
>> +
>> +            oldcount = rwsem_atomic_update(adjustment, sem)
>> +                                - adjustment;
>> +            if (likely(oldcount >= RWSEM_WAITING_BIAS))
>> +                break;
>> +
>> +             /* A writer stole the lock.  Undo our reader grant. */
>>               if (rwsem_atomic_update(-adjustment, sem) &
>>                           RWSEM_ACTIVE_MASK)
>> -                goto out;
>> +                return sem;
>>               /* Last active locker left. Retry waking readers. */
>> -            goto try_reader_grant;
>>           }
>>       }
>>
>> @@ -136,7 +142,6 @@ __rwsem_do_wake(struct rw_semaphore *sem, enum
>> rwsem_wake_type wake_type)
>>       sem->wait_list.next = next;
>>       next->prev = &sem->wait_list;
>>
>> - out:
>>       return sem;
>>   }
> 
> 
> Alex and Tim,
> 
> Was there a v1 of this series; ie., is this v2 (or higher)?
> 
> How are you validating lock correctness/behavior with this series?

some benchmark tested against this patch, mainly aim7. plus by eyes, we
didn't change the logical except check the lock value before  do locking
> 
> Regards,
> Peter Hurley
> 


-- 
Thanks
    Alex