Re: [PATCH v3 0/3] locking/mutex: Enable optimistic spinning of lock waiter

[Jason Low <jason.low2@hpe.com>]

On Tue, 2016-03-22 at 13:46 -0400, Waiman Long wrote:
> v2->v3:
>  - Remove patch 4 as it is not useful.
>  - Allow need_resched() check for waiter & add more comments about
>    changes to address issues raised by PeterZ.
> 
> v1->v2:
>  - Set task state to running before doing optimistic spinning.
>  - Add 2 more patches to handle possible missed wakeups and wasteful
>    spinning in try_to_wake_up() function.
> 
> This patchset is a variant of PeterZ's "locking/mutex: Avoid spinner
> vs waiter starvation" patch. The major difference is that the
> waiter-spinner won't enter into the OSQ used by the spinners. Instead,
> it will spin directly on the lock in parallel with the queue head
> of the OSQ. So there will be a bit more cacheline contention on the
> lock cacheline, but that shouldn't cause noticeable impact on system
> performance.
> 
> This patchset tries to address 2 issues with Peter's patch:
> 
>  1) Ding Tianhong still find that hanging task could happen in some cases.
>  2) Jason Low found that there was performance regression for some AIM7
>     workloads.
> 
> By making the waiter-spinner to spin directly on the mutex, it will
> increase the chance for the waiter-spinner to get the lock instead
> of waiting in the OSQ for its turn.
> 
> Patch 1 modifies the mutex_optimistic_spin() function to enable it
> to be called by a waiter-spinner that doesn't need to go into the OSQ.
> 
> Patch 2 modifies the mutex locking slowpath to make the waiter call
> mutex_optimistic_spin() to do spinning after being waken up.

Just in case, here was the Waiman's patch that I was referring to for
addressing some of the starvation issues with mutex by allowing the
waiter to do optimistic spinning after being woken up.

> Patch 3 reverses the sequence of setting task state and changing
> mutex count to -1 to prevent the possibility of missed wakeup.
> 
> Patch 4 modifies the wakeup code to abandon the wakeup operation
> while spinning on the on_cpu flag if the task has changed back to a
> non-sleeping state.
> 
> My own test on a 4-socket E7-4820 v3 system showed a regression of
> about 4% in the high_systime workload with Peter's patch which this
> new patch effectively eliminates.
> 
> Testing on an 8-socket Westmere-EX server, however, has performance
> change from -9% to than +140% on the fserver workload of AIM7
> depending on how the system was set up.
> 
> Waiman Long (3):
>   locking/mutex: Add waiter parameter to mutex_optimistic_spin()
>   locking/mutex: Enable optimistic spinning of woken task in wait queue
>   locking/mutex: Avoid missed wakeup of mutex waiter
> 
>  kernel/locking/mutex.c |  126 ++++++++++++++++++++++++++++++++++--------------
>  1 files changed, 90 insertions(+), 36 deletions(-)