Re: [External] Re: [PATCH] sched/core: Avoid WARN_DOUBLE_CLOCK warning when CONFIG_SCHED_CORE

[Hao Jia <jiahao.os@bytedance.com>]

On 2023/1/16 Peter Zijlstra wrote:
> On Tue, Dec 06, 2022 at 03:05:50PM +0800, Hao Jia wrote:
>> When we need to call update_rq_clock() to update the rq clock of
>> other CPUs on the same core, before that we need to clear RQCF_UPDATED
>> of rq->clock_update_flags to avoid the WARN_DOUBLE_CLOCK warning.
>> Because at this time the rq->clock_update_flags of other CPUs
>> may be RQCF_UPDATED.
> 
> So you've found that the WARN_DOUBLE_CLOCK machinery doesn't work for
> core-sched -- but then instead of fixing that machinery, you put
> band-aids on it :/
> 
> 
Hi, Peter
Sorry for the late reply. I just finished my holiday.

I am trying to adapt WARN_DOUBLE_CLOCK machinery for core-sched.

If sched_core_enabled(), we will get a core wide rq->lock, so we can
safely clear RQCF_UPDATED of rq->clock_update_flags of all CPUs on this
core.
This avoids a WARN_DOUBLE_CLOCK warning when we call update_rq_clock()
to update the rq clock of other cpus on the same core.

We cannot clear rq->clock_update_flags of other cpus on the same core in
rq_pin_lock(). Because in some functions, we will temporarily give up
core wide rq->lock, and then use raw_spin_rq_lock() to obtain core wide
rq->lock, such as newidle_balance() and _double_lock_balance().


Thanks,
Hao

---

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index e838feb6adc5..f279912e30b3 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -435,6 +435,21 @@ sched_core_dequeue(struct rq *rq, struct 
task_struct *p, int flags) { }

  #endif /* CONFIG_SCHED_CORE */

+static inline void sched_core_rq_clock_clear_update(struct rq *rq)
+{
+#ifdef CONFIG_SCHED_DEBUG
+       const struct cpumask *smt_mask;
+       int i;
+       if (rq->core_enabled) {
+               smt_mask = cpu_smt_mask(rq->cpu);
+               for_each_cpu(i, smt_mask) {
+                       if (rq->cpu != i)
+                               cpu_rq(i)->clock_update_flags &= 
(RQCF_REQ_SKIP|RQCF_ACT_SKIP);
+               }
+       }
+#endif
+}
+
  /*
   * Serialization rules:
   *
@@ -546,6 +561,7 @@ void raw_spin_rq_lock_nested(struct rq *rq, int 
subclass)
                 if (likely(lock == __rq_lockp(rq))) {
                         /* preempt_count *MUST* be > 1 */
                         preempt_enable_no_resched();
+                       sched_core_rq_clock_clear_update(rq);
                         return;
                 }
                 raw_spin_unlock(lock);