Re: [PATCH 1/1] locking/qspinlock: Make the 1st spinner only spin on locked_pending bits

[Waiman Long <longman@redhat.com>]

On 5/8/23 23:30, Zhuo, Qiuxu wrote:
>> From: Waiman Long <longman@redhat.com>
>> Sent: Tuesday, May 9, 2023 10:58 AM
>> To: Zhuo, Qiuxu <qiuxu.zhuo@intel.com>; Peter Zijlstra
>> <peterz@infradead.org>; Ingo Molnar <mingo@redhat.com>; Will Deacon
>> <will@kernel.org>
>> Cc: Boqun Feng <boqun.feng@gmail.com>; linux-kernel@vger.kernel.org
>> Subject: Re: [PATCH 1/1] locking/qspinlock: Make the 1st spinner only spin on
>> locked_pending bits
>>
>>
>> On 5/8/23 22:45, Zhuo, Qiuxu wrote:
>>> Hi Waiman,
>>>
>>> Thanks for your review of this patch.
>>> Please see the comments below.
>>>
>>>> From: Waiman Long <longman@redhat.com>
>>>> Sent: Monday, May 8, 2023 11:31 PM
>>>> To: Zhuo, Qiuxu <qiuxu.zhuo@intel.com>; Peter Zijlstra
>>>> <peterz@infradead.org>; Ingo Molnar <mingo@redhat.com>; Will Deacon
>>>> <will@kernel.org>
>>>> Cc: Boqun Feng <boqun.feng@gmail.com>; linux-kernel@vger.kernel.org
>>>> Subject: Re: [PATCH 1/1] locking/qspinlock: Make the 1st spinner only
>>>> spin on locked_pending bits
>>>>
>>>>
>>>> On 5/8/23 04:15, Qiuxu Zhuo wrote:
>>>>> The 1st spinner (header of the MCS queue) spins on the whole
>>>>> qspinlock variable to check whether the lock is released. For a
>>>>> contended qspinlock, this spinning is a hotspot as each CPU queued
>>>>> in the MCS queue performs the spinning when it becomes the 1st
>>>>> spinner (header of
>>>> the MCS queue).
>>>>> The granularity among SMT h/w threads in the same core could be
>> "byte"
>>>>> which the Load-Store Unit (LSU) inside the core handles. Making the
>>>>> 1st spinner only spin on locked_pending bits (not the whole
>>>>> qspinlock) can avoid the false dependency between the tail field and
>>>>> the locked_pending field. So this micro-optimization helps the h/w
>>>>> thread (the 1st spinner) stay in a low power state and prevents it
>>>>> from being woken up by other h/w threads in the same core when they
>>>>> perform
>>>>> xchg_tail() to update the tail field. Please see a similar
>>>>> discussion in the link
>>>> [1].
>>>>> [1]
>>>>> https://lore.kernel.org/r/20230105021952.3090070-1-
>> guoren@kernel.org
>>>>> Signed-off-by: Qiuxu Zhuo <qiuxu.zhuo@intel.com>
>>>>> ---
>>>>>     kernel/locking/qspinlock.c | 13 +++++++++++++
>>>>>     1 file changed, 13 insertions(+)
>>>>>
>>>>> diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
>>>>> index efebbf19f887..e7b990b28610 100644
>>>>> --- a/kernel/locking/qspinlock.c
>>>>> +++ b/kernel/locking/qspinlock.c
>>>>> @@ -513,7 +513,20 @@ void __lockfunc
>>>> queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
>>>>>     	if ((val = pv_wait_head_or_lock(lock, node)))
>>>>>     		goto locked;
>>>>>
>>>>> +#if _Q_PENDING_BITS == 8
>>>>> +	/*
>>>>> +	 * Spinning on the 2-byte locked_pending instead of the 4-byte
>>>> qspinlock
>>>>> +	 * variable can avoid the false dependency between the tail field and
>>>>> +	 * the locked_pending field. This helps the h/w thread (the 1st
>>>> spinner)
>>>>> +	 * stay in a low power state and prevents it from being woken up
>>>>> +by
>>>> other
>>>>> +	 * h/w threads in the same core when they perform xchg_tail() to
>>>> update
>>>>> +	 * the tail field only.
>>>>> +	 */
>>>>> +	smp_cond_load_acquire(&lock->locked_pending, !VAL);
>>>>> +	val = atomic_read_acquire(&lock->val); #else
>>>>>     	val = atomic_cond_read_acquire(&lock->val, !(VAL &
>>>>> _Q_LOCKED_PENDING_MASK));
>>>>> +#endif
>>>>>
>>>>>     locked:
>>>>>     	/*
>>>> What hardware can benefit from this change? Do you have any micro-
>>>> benchmark that can show the performance benefit?
>>> i)  I don't have the hardware to measure the data.
>>>       But I run a benchmark [1] for the contended spinlock on an Intel
>> Sapphire Rapids
>>>       server (192 h/w threads, 2sockets) that showed that the 1st spinner
>> spinning was
>>>       a hotspot (contributed ~55% cache bouncing traffic measured by the
>> perf C2C.
>>>        I don't analyze the cache bouncing here, but just say the spinning is a
>> hotspot).
>> I believe the amount of cacheline bouncing will be the same whether you
>> read 32 or 16 bits from the lock word. At least this is my understanding of the
>> x86 arch. Please correct me if my assumption is incorrect.
> You're right.
> The amount of cache line bouncing was nearly the same either spinning 32 or 16 bits
> (according to my measured perf C2C data on an x86 server).
>   
>>> ii) The similar micro-optimization discussion [2] (looked like it was accepted
>> by you 😉) that
>>>       avoiding the false dependency (between the tail field and the
>> locked_pending field)
>>>       should help some arches (e.g., some ARM64???) the h/w thread
>> (spinner) stay in a
>>>       low-power state without the disruption by its sibling h/w threads in the
>> same core.
>>
>> That is true. However, this patch causes one more read from the lock
>> cacheline which isn't necessary for arches that won't benefit from it.
>> So I am less incline to accept it unless there is evidence of the
>> benefit it can bring.
> This patch removes a bitwise AND operation on the VAL value.
> Does it compensate for the one more read from the cache line?

Register to register operation in the case of bitwise AND doesn't cost 
much, it is the potential reading from a hot cacheline that cause most 
of the delay. Besides there is also an additional acquire barrier. 
Again, if there is no concrete proof of a benefit, there is no point to 
make the code more complicated.

Cheers,
Longman