Re: [PATCH] pv-qspinlock: Try to re-hash the lock after spurious_wakeup

[xinhui]

On 2016å05æ28æ 11:41, Waiman Long wrote:
> On 05/27/2016 06:32 AM, xinhui wrote:
> > On 2016å05æ27æ 02:31, Waiman Long wrote:
> > > On 05/25/2016 02:09 AM, Pan Xinhui wrote:
> > > > In pv_wait_head_or_lock, if there is a spurious_wakeup, and it fails to
> > > > get the lock as there is lock stealing, then after a short spin, we need
> > > > hash the lock again and enter pv_wait to yield.
> > > >
> > > > Currently after a spurious_wakeup, as l->locked is not _Q_SLOW_VAL,
> > > > pv_wait might do nothing and return directly, that is not
> > > > paravirt-friendly because pv_wait_head_or_lock will just spin on the
> > > > lock then.
> > > >
> > > > Signed-off-by: Pan Xinhui<xinhui.pan@xxxxxxxxxxxxxxxxxx>
> > > > ---
> > > >   kernel/locking/qspinlock_paravirt.h | 39 +++++++++++++++++++++++++++++--------
> > > >   1 file changed, 31 insertions(+), 8 deletions(-)
> > >
> > > Is this a problem you can easily reproduce on PPC? I have not observed this issue when testing on x86.
> > Hi, Waiman
> >     I notice the spurious_wakeup count is very high when I do benchmark tests and stress tests. So after a simple investigation,
> > I find pv_wait_head_or_lock() just keep loops.
>
> That shouldn't happen in normal case. When testing on x86, I typically get the following stat data for an over-commited guest:
>
> pv_lock_slowpath=9256211
> pv_lock_stealing=36398363
> pv_spurious_wakeup=311
> pv_wait_again=294
> pv_wait_early=3255605
> pv_wait_head=173
> pv_wait_node=3256280
OK, here is the result after run command  perf bench sched messaging -g 512

pv_lock_slowpath=2331407
pv_lock_stealing=192038
pv_spurious_wakeup=236319
pv_wait_again=215668
pv_wait_early=177299
pv_wait_head=9206
pv_wait_node=228781

> The queue head don't call pv_wait that often. There are a bit of spurious wakeup, but it is mostly caused by lock stealing. How long is a cpu_relax() in PPC takes?
946012160 cpu_relax loops with 10 seconds. So if SPIN_THRESHOLD is 1<<15, it costs 0.3ms to spin on the lock. How about x86?

And only 10134976 pv_wait/pv_kick hyper-call loops within 10 seconds. so every hyper-call itself(the so-called latency) costs less than 1us.

> >     Here is my story, in my pv-qspinlcok patchset V1&&v2, pv_wait on ppc ignore the first two parameters of *ptr and val, that makes lock_stealing hit too much.
>
> The pvqspinlock code does depend on pv_wait() doing a final check to see if the lock value change. The code may not work reliably without that.
agree, So pv_wait now do the check of *ptr and val.

> > and when I change SPIN_THRESHOLD to a small value, system is very much unstable because waiter will enter pv_wait quickly and no one will kick waiter's cpu if
> > we enter pv_wait twice thanks to the lock_stealing.
> >     So what I do in my pv-qspinlcok patchset V3 is that add if (*ptr == val) in pv_wait. However as I mentioned above, then spurious_wakeup count is too high, that also means our cpu
> > slice is wasted.
>
> The SPIN_THRESHOLD should be sufficiently big. A small value will cause too many waits and wake-up's which may not be good. Anyway, more testing and tuning may be needed to make the pvqspinlock code work well with PPC.
agree , but I think the SPIN_THRESHOLD (1<<15) for ppc is a little large.

I even come up with an idea that make SPIN_THRESHOLD an extern variable on ppc. But I am busy and I wonder if it's worth doing that.

> Cheers,
> Longman