Syndrome:
Two CPUs might do idle balance in the same time.
One CPU does idle balance and pulls some tasks.
However before pick next task, ALL task are pulled back to other CPU.
That results in infinite loop in both CPUs.
=========================================
code flow:
in pick_next_task_fair()
again:
if nr_running == 0
goto idle
pick next task
return
idle:
idle_balance
/* pull some tasks from other CPU,
* However other CPU are also do idle balance,
* and pull back these task */
go to again
=========================================
The result to pull ALL tasks back when the task_h_load
is incorrect and too low.
static unsigned long task_h_load(struct task_struct *p)
{
struct cfs_rq *cfs_rq = task_cfs_rq(p);
update_cfs_rq_h_load(cfs_rq);
return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
cfs_rq->runnable_load_avg + 1);
}
The cfs_rq->h_load is incorrect and might too small.
The original idea of cfs_rq::last_h_load_update will not
update cfs_rq::h_load more than once a jiffies.
When the Two CPUs pull each other in the pick_next_task_fair,
the irq disabled and result in jiffie not update.
(Other CPUs wait for runqueue lock locked by the two CPUs.
So, ALL CPUs are irq disabled.)
Solution:
cfs_rq h_load might not update due to irq disable
use sched_clock instead jiffies
Signed-off-by: YT Chang <[email protected]>
---
kernel/sched/fair.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 83ab35e..231c53f 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -7578,9 +7578,11 @@ static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
{
struct rq *rq = rq_of(cfs_rq);
struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
- unsigned long now = jiffies;
+ u64 now = sched_clock_cpu(cpu_of(rq));
unsigned long load;
+ now = now * HZ >> 30;
+
if (cfs_rq->last_h_load_update == now)
return;
--
1.9.1
On Thu, Nov 21, 2019 at 04:30:09PM +0800, YT Chang wrote:
> Syndrome:
>
> Two CPUs might do idle balance in the same time.
> One CPU does idle balance and pulls some tasks.
> However before pick next task, ALL task are pulled back to other CPU.
> That results in infinite loop in both CPUs.
Can you easily reproduce this?
> =========================================
> code flow:
>
> in pick_next_task_fair()
>
> again:
>
> if nr_running == 0
> goto idle
> pick next task
> return
>
> idle:
> idle_balance
> /* pull some tasks from other CPU,
> * However other CPU are also do idle balance,
> * and pull back these task */
>
> go to again
>
> =========================================
> The result to pull ALL tasks back when the task_h_load
> is incorrect and too low.
Clearly you're not running a PREEMPT kernel, otherwise the break in
detach_tasks() would've saved you, right?
> static unsigned long task_h_load(struct task_struct *p)
> {
> struct cfs_rq *cfs_rq = task_cfs_rq(p);
>
> update_cfs_rq_h_load(cfs_rq);
> return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
> cfs_rq->runnable_load_avg + 1);
> }
>
> The cfs_rq->h_load is incorrect and might too small.
> The original idea of cfs_rq::last_h_load_update will not
> update cfs_rq::h_load more than once a jiffies.
> When the Two CPUs pull each other in the pick_next_task_fair,
> the irq disabled and result in jiffie not update.
> (Other CPUs wait for runqueue lock locked by the two CPUs.
> So, ALL CPUs are irq disabled.)
This cannot be true; because the loop drops rq->lock, so other CPUs
should have an opportunity to acquire the lock and make progress.
> Solution:
> cfs_rq h_load might not update due to irq disable
> use sched_clock instead jiffies
>
> Signed-off-by: YT Chang <[email protected]>
> ---
> kernel/sched/fair.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 83ab35e..231c53f 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -7578,9 +7578,11 @@ static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
> {
> struct rq *rq = rq_of(cfs_rq);
> struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
> - unsigned long now = jiffies;
> + u64 now = sched_clock_cpu(cpu_of(rq));
> unsigned long load;
>
> + now = now * HZ >> 30;
> +
> if (cfs_rq->last_h_load_update == now)
> return;
>
This is disguisting and wrong. That is not the correct relation between
sched_clock() and jiffies.
On Thu, 2019-11-21 at 13:38 +0100, Peter Zijlstra wrote:
> On Thu, Nov 21, 2019 at 04:30:09PM +0800, YT Chang wrote:
> > Syndrome:
> >
> > Two CPUs might do idle balance in the same time.
> > One CPU does idle balance and pulls some tasks.
> > However before pick next task, ALL task are pulled back to other CPU.
> > That results in infinite loop in both CPUs.
>
> Can you easily reproduce this?
No, I can't easily reproduce this.
>
> > =========================================
> > code flow:
> >
> > in pick_next_task_fair()
> >
> > again:
> >
> > if nr_running == 0
> > goto idle
> > pick next task
> > return
> >
> > idle:
> > idle_balance
> > /* pull some tasks from other CPU,
> > * However other CPU are also do idle balance,
> > * and pull back these task */
> >
> > go to again
> >
> > =========================================
> > The result to pull ALL tasks back when the task_h_load
> > is incorrect and too low.
>
> Clearly you're not running a PREEMPT kernel, otherwise the break in
> detach_tasks() would've saved you, right?
>
> > static unsigned long task_h_load(struct task_struct *p)
> > {
> > struct cfs_rq *cfs_rq = task_cfs_rq(p);
> >
> > update_cfs_rq_h_load(cfs_rq);
> > return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
> > cfs_rq->runnable_load_avg + 1);
> > }
> >
> > The cfs_rq->h_load is incorrect and might too small.
> > The original idea of cfs_rq::last_h_load_update will not
> > update cfs_rq::h_load more than once a jiffies.
> > When the Two CPUs pull each other in the pick_next_task_fair,
> > the irq disabled and result in jiffie not update.
> > (Other CPUs wait for runqueue lock locked by the two CPUs.
> > So, ALL CPUs are irq disabled.)
>
> This cannot be true; because the loop drops rq->lock, so other CPUs
> should have an opportunity to acquire the lock and make progress.
I recheck other CPUs situation.
Other CPUs are irq disabled due to wait for lock (Not runqueue lock).
The root cause should be why other CPUs are waiting for the lock
rather than replacing jiffies with sched_clock().
>
> > Solution:
> > cfs_rq h_load might not update due to irq disable
> > use sched_clock instead jiffies
> >
> > Signed-off-by: YT Chang <[email protected]>
> > ---
> > kernel/sched/fair.c | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 83ab35e..231c53f 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -7578,9 +7578,11 @@ static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
> > {
> > struct rq *rq = rq_of(cfs_rq);
> > struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
> > - unsigned long now = jiffies;
> > + u64 now = sched_clock_cpu(cpu_of(rq));
> > unsigned long load;
> >
> > + now = now * HZ >> 30;
> > +
> > if (cfs_rq->last_h_load_update == now)
> > return;
> >
>
> This is disguisting and wrong. That is not the correct relation between
> sched_clock() and jiffies.