Re: [PATCH v2] locking/pvqspinlock: Relax cmpxchg's to improve performance on some archs

[Pan Xinhui]

å 2017/2/8 14:09, Boqun Feng åé:
> On Wed, Feb 08, 2017 at 12:05:40PM +0800, Boqun Feng wrote:
> > On Wed, Feb 08, 2017 at 11:39:10AM +0800, Xinhui Pan wrote:
> > > 2016-12-26 4:26 GMT+08:00 Waiman Long <longman@xxxxxxxxxx>:
> > >
> > > > A number of cmpxchg calls in qspinlock_paravirt.h were replaced by more
> > > > relaxed versions to improve performance on architectures that use LL/SC.
> > > >
> > > > All the locking related cmpxchg's are replaced with the _acquire
> > > > variants:
> > > >  - pv_queued_spin_steal_lock()
> > > >  - trylock_clear_pending()
> > > >
> > > > The cmpxchg's related to hashing are replaced by either by the _release
> > > > or the _relaxed variants. See the inline comment for details.
> > > >
> > > > Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
> > > >
> > > >  v1->v2:
> > > >   - Add comments in changelog and code for the rationale of the change.
> > > >
> > > > ---
> > > >  kernel/locking/qspinlock_paravirt.h | 50 ++++++++++++++++++++++++------
> > > > -------
> > > >  1 file changed, 33 insertions(+), 17 deletions(-)
> > > >
> > > >
> > > > @@ -323,8 +329,14 @@ static void pv_wait_node(struct mcs_spinlock *node,
> > > > struct mcs_spinlock *prev)
> > > >                  * If pv_kick_node() changed us to vcpu_hashed, retain that
> > > >                  * value so that pv_wait_head_or_lock() knows to not also
> > > > try
> > > >                  * to hash this lock.
> > > > +                *
> > > > +                * The smp_store_mb() and control dependency above will
> > > > ensure
> > > > +                * that state change won't happen before that.
> > > > Synchronizing
> > > > +                * with pv_kick_node() wrt hashing by this waiter or by the
> > > > +                * lock holder is done solely by the state variable. There
> > > > is
> > > > +                * no other ordering requirement.
> > > >                  */
> > > > -               cmpxchg(&pn->state, vcpu_halted, vcpu_running);
> > > > +               cmpxchg_relaxed(&pn->state, vcpu_halted, vcpu_running);
> > > >
> > > >                 /*
> > > >                  * If the locked flag is still not set after wakeup, it is
> > > > a
> > > > @@ -360,9 +372,12 @@ static void pv_kick_node(struct qspinlock *lock,
> > > > struct mcs_spinlock *node)
> > > >          * pv_wait_node(). If OTOH this fails, the vCPU was running and
> > > > will
> > > >          * observe its next->locked value and advance itself.
> > > >          *
> > > > -        * Matches with smp_store_mb() and cmpxchg() in pv_wait_node()
> > > > +        * Matches with smp_store_mb() and cmpxchg_relaxed() in
> > > > pv_wait_node().
> > > > +        * A release barrier is used here to ensure that node->locked is
> > > > +        * always set before changing the state. See comment in
> > > > pv_wait_node().
> > > >          */
> > > > -       if (cmpxchg(&pn->state, vcpu_halted, vcpu_hashed) != vcpu_halted)
> > > > +       if (cmpxchg_release(&pn->state, vcpu_halted, vcpu_hashed)
> > > > +                       != vcpu_halted)
> > > >                 return;
> > > >
> > > > hi, Waiman
> > > We can't use _release here, a full barrier is needed.
> > >
> > > There is pv_kick_node vs pv_wait_head_or_lock
> > >
> > > [w] l->locked = _Q_SLOW_VAL  //reordered here
> > >
> > > if (READ_ONCE(pn->state) == vcpu_hashed) //False.
> > >
> > >                    lp = (struct qspinlock **)1;
> > >
> > > [STORE] pn->state = vcpu_hashed                        lp = pv_hash(lock,
> > > pn);
> > > pv_hash()                                                                if
> > > (xchg(&l->locked, _Q_SLOW_VAL) == 0) // fasle, not unhashed.
> >
> > This analysis is correct, but..
>
> Hmm.. look at this again, I don't think this analysis is meaningful,
> let's say the reordering didn't happen, we still got(similar to your
> case):
>
> 						if (READ_ONCE(pn->state) == vcpu_hashed) // false.
> 						  lp = (struct qspinlock **)1;
>
> cmpxchg(pn->state, vcpu_halted, vcpu_hashed);
> 						  if(!lp) {
> 						    lp = pv_hash(lock, pn);
> WRITE_ONCE(l->locked, _Q_SLOW_VAL);
> pv_hash();
> 						    if (xchg(&l->locked, _Q_SLOW_VAL) == 0) // fasle, not unhashed.
>
> , right?
>
> Actually, I think this or your case could not happen because we have
>
> 	cmpxchg(pn->state, vcpu_halted, vcpu_running);
>
> in pv_wait_node(), which makes us either observe vcpu_hashed or set
> pn->state to vcpu_running before pv_kick_node() trying to do the hash.
yep, there is still a race. We have to fix it.
so I think
we must check old = xchg(&l->locked, _Q_SLOW_VAL)
if (old  == 0)
do something
else if (old  == _Q_SLOW_VAL)
do something else

> I may miss something subtle, but does switching back to cmpxchg() could
> fix the RCU stall you observed?
yes, just fix this cmpxchg and then no RCU stall.

> Regards,
> Boqun
>
> > > Then the same lock has hashed twice but only unhashed once. So at last as
> > > the hash table grows big, we hit RCU stall.
> > >
> > > I hit RCU stall when I run netperf benchmark
> >
> > how will a big hash table hit RCU stall? Do you have the call trace for
> > your RCU stall?
maybe too many time on hashing? I am not sure.

> > Regards,
> > Boqun
> >
> > > thanks
> > > xinhui
> > >
> > >
> > > > --
> > > > 1.8.3.1