Messages in this thread | | | Date | Thu, 21 Apr 2022 15:24:33 +0800 | Subject | Re: [External] Re: [PATCH] sched/core: Avoid obvious double update_rq_clock warning | From | Hao Jia <> |
| |
On 2022/4/21 Dietmar Eggemann wrote: > On 20/04/2022 10:29, Hao Jia wrote: >> On 4/19/22 6:48 PM, Peter Zijlstra wrote: >>> On Mon, Apr 18, 2022 at 05:09:29PM +0800, Hao Jia wrote: > > [...] > >>> I'm really not sure about this part though. This is a bit of a mess. The >>> balancer doesn't really need the pinning stuff. I realize you did that >>> because we got the clock annotation mixed up with that, but urgh. >>> >>> Basically we want double_rq_lock() / double_lock_balance() to clear >>> RQCF_UPDATED, right? Perhaps do that directly? >>> >>> (maybe with an inline helper and a wee comment?) >>> >>> The only immediate problem with this would appear to be that >>> _double_rq_lock() behaves differently when it returns 0. Not sure that >>> matters. >>> >>> Hmm? >> >> Thanks for your review comments. >> As you have prompted, the WARN_DOUBLE_CLOCK warning is still triggered >> when _double_rq_lock() returns 0. >> Please review the solution below, and based on your review, I will >> submit the v2 patch as soon as possible. >> Thanks. > > > [...] > > Maybe something like this: > > -->8-- > > From: Dietmar Eggemann <dietmar.eggemann@arm.com> > Date: Wed, 20 Apr 2022 11:12:10 +0200 > Subject: [PATCH] sched/core: Clear RQCF_UPDATED in _double_lock_balance() & > double_rq_lock() > > Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com> > --- > kernel/sched/core.c | 6 +++--- > kernel/sched/sched.h | 20 ++++++++++++++++---- > 2 files changed, 19 insertions(+), 7 deletions(-) > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c > index 068c088e9584..f4cfe7eea861 100644 > --- a/kernel/sched/core.c > +++ b/kernel/sched/core.c > @@ -610,10 +610,10 @@ void double_rq_lock(struct rq *rq1, struct rq *rq2) > swap(rq1, rq2); > > raw_spin_rq_lock(rq1); > - if (__rq_lockp(rq1) == __rq_lockp(rq2)) > - return; > + if (__rq_lockp(rq1) != __rq_lockp(rq2)) > + raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING); > > - raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING); > + rq_clock_clear_update(rq1, rq2); > } > #endif > > diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h > index 58263f90c559..3a77b10d7cc4 100644 > --- a/kernel/sched/sched.h > +++ b/kernel/sched/sched.h > @@ -2515,6 +2515,16 @@ static inline bool rq_order_less(struct rq *rq1, struct rq *rq2) > > extern void double_rq_lock(struct rq *rq1, struct rq *rq2); > > +#ifdef CONFIG_SCHED_DEBUG > +static inline void rq_clock_clear_update(struct rq *rq1, struct rq *rq2) > +{ > + rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); > + rq2->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP); > +} > +#else > +static inline void rq_clock_clear_update(struct rq *rq1, struct rq *rq2) {} > +#endif > + > #ifdef CONFIG_PREEMPTION > > /* > @@ -2549,14 +2559,15 @@ static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest) > __acquires(busiest->lock) > __acquires(this_rq->lock) > { > - if (__rq_lockp(this_rq) == __rq_lockp(busiest)) > - return 0; > - > - if (likely(raw_spin_rq_trylock(busiest))) > + if (__rq_lockp(this_rq) == __rq_lockp(busiest) || > + likely(raw_spin_rq_trylock(busiest))) { > + rq_clock_clear_update(this_rq, busiest); > return 0; > + } > > if (rq_order_less(this_rq, busiest)) { > raw_spin_rq_lock_nested(busiest, SINGLE_DEPTH_NESTING); > + rq_clock_clear_update(this_rq, busiest); > return 0; > } > > @@ -2650,6 +2661,7 @@ static inline void double_rq_lock(struct rq *rq1, struct rq *rq2) > BUG_ON(rq1 != rq2); > raw_spin_rq_lock(rq1); > __acquire(rq2->lock); /* Fake it out ;) */ > + rq_clock_clear_update(rq1, rq2);
Thanks for your review. This is very helpful to me. If CONFIG_SMP is not enabled, should we just clear the RQCF_UPDATED of one of rq1 and q2?
like this: rq1->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
Thanks.
> } > > /*
| |