On 01/23/2019 03:40 PM, Peter Zijlstra wrote:
> On Wed, Jan 23, 2019 at 03:11:19PM -0500, Waiman Long wrote:
>> On 01/23/2019 04:34 AM, Will Deacon wrote:
>>> On Tue, Jan 22, 2019 at 10:49:08PM -0500, Waiman Long wrote:
>>>> @@ -412,6 +412,21 @@ void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
>>>> idx = node->count++;
>>>> tail = encode_tail(smp_processor_id(), idx);
>>>> + if (unlikely(idx >= MAX_NODES)) {
>>>> + while (!queued_spin_trylock(lock))
>>>> + cpu_relax();
>>>> + goto release;
>>>> + }
>> So the additional code checks the idx value and branch to the end of the
>> function when the condition is true. There isn't too much overhead here.
> So something horrible we could do (and I'm not at all advocating we do
> this), is invert node->count. That is, start at 3 and decrement and
> detect sign flips.
>
> That avoids the additional compare. It would require we change the
> structure layout though, otherwise we keep hitting that second line by
> default, which would suck.
The cost of the additional compare will not be noticeable if the branch
prediction logic is working properly. Inverting the loop logic, however,
will be a much bigger change and it may not guarantee it will be faster
anyway. So I don't think we should down go this route :-)
Cheers,
Longman
