Messages in this thread | | | Subject | Re: [PATCH] sched/fair: Make cfs_rq::decay_counter non-atomic | From | Kirill Tkhai <> | Date | Wed, 17 Dec 2014 12:09:33 +0300 |
| |
В Вт, 16/12/2014 в 10:00 -0800, bsegall@google.com пишет: > Kirill Tkhai <ktkhai@parallels.com> writes: > > > We update decay_counter in update_cfs_rq_blocked_load() > > only. This function is always called with rq lock locked, > > so we can kill atomic actions. > > > > Signed-off-by: Kirill Tkhai <ktkhai@parallels.com> > > --- > > kernel/sched/fair.c | 11 ++++++----- > > kernel/sched/sched.h | 2 +- > > 2 files changed, 7 insertions(+), 6 deletions(-) > > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > > index 5f3b5a7..af990c4 100644 > > --- a/kernel/sched/fair.c > > +++ b/kernel/sched/fair.c > > @@ -2570,7 +2570,7 @@ static __always_inline int __update_entity_runnable_avg(u64 now, > > static inline u64 __synchronize_entity_decay(struct sched_entity *se) > > { > > struct cfs_rq *cfs_rq = cfs_rq_of(se); > > - u64 decays = atomic64_read(&cfs_rq->decay_counter); > > + u64 decays = ACCESS_ONCE(cfs_rq->decay_counter); > > This is called without rq lock held from migrate_task_rq_fair. (We could > technically change the atomic_add to atomic64_set(atomic64_read() + x), > but I don't know that that is a win) Now, we could do a > min_vruntime-style two-copy thing if this atomic usage is a hot spot on > 32-bit, we just didn't bother initially.
Oh, I forgot that 64-bit read is not everywhere atomic... Thanks, Ben.
Peter, please, ignore this patch.
| |