On Sat, Jun 06, 2020 at 10:29:42AM -0700, Paul E. McKenney wrote:
> On Fri, Jun 05, 2020 at 05:51:26PM -0700, Paul E. McKenney wrote:
> > On Fri, Jun 05, 2020 at 11:41:59AM -0700, Paul E. McKenney wrote:
> > > On Fri, Jun 05, 2020 at 07:16:07AM -0700, Paul E. McKenney wrote:
> > >
> > > And in case it is helpful, here is the output of "git bisect view",
> > > which lists rather more commits than "git bisect run" claims, but there
> > > are only a few scheduler commits below. I don't see anything that
> > > I can prove can cause this problem, but there are some that are at
> > > least related to this code path.
> > >
> > > Is there anything that is actually relevant?
> >
> > And the run with the WARN and tracing did hit two failures, and the
> > corresponding console logs are in the attached tarball. Both of them
> > triggered a warning as well as the failure.
>
> And the current state of the bisection, for whatever it is worth.
The bisection finished, finally!
90b5363acd47 ("sched: Clean up scheduler_ipi()")
I don't see anything immediately obvious, but then again, I do not
claim to understand this code. If you have additional diagnostics,
please let me know.
Thanx, Paul
------------------------------------------------------------------------
commit 90b5363acd4739769c3f38c1aff16171bd133e8c
Author: Peter Zijlstra (Intel) <pet...@infradead.org>
Date: Fri Mar 27 11:44:56 2020 +0100
sched: Clean up scheduler_ipi()
The scheduler IPI has grown weird and wonderful over the years, time
for spring cleaning.
Move all the non-trivial stuff out of it and into a regular smp function
call IPI. This then reduces the schedule_ipi() to most of it's former NOP
glory and ensures to keep the interrupt vector lean and mean.
Aside of that avoiding the full irq_enter() in the x86 IPI implementation
is incorrect as scheduler_ipi() can be instrumented. To work around that
scheduler_ipi() had an irq_enter/exit() hack when heavy work was
pending. This is gone now.
Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org>
Signed-off-by: Thomas Gleixner <t...@linutronix.de>
Reviewed-by: Alexandre Chartre <alexandre.char...@oracle.com>
Link: https://lkml.kernel.org/r/20200505134058.361859...@linutronix.de
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index b58efb1..cd2070d 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -219,6 +219,13 @@ void update_rq_clock(struct rq *rq)
update_rq_clock_task(rq, delta);
}
+static inline void
+rq_csd_init(struct rq *rq, call_single_data_t *csd, smp_call_func_t func)
+{
+ csd->flags = 0;
+ csd->func = func;
+ csd->info = rq;
+}
#ifdef CONFIG_SCHED_HRTICK
/*
@@ -314,16 +321,14 @@ void hrtick_start(struct rq *rq, u64 delay)
hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
HRTIMER_MODE_REL_PINNED_HARD);
}
+
#endif /* CONFIG_SMP */
static void hrtick_rq_init(struct rq *rq)
{
#ifdef CONFIG_SMP
- rq->hrtick_csd.flags = 0;
- rq->hrtick_csd.func = __hrtick_start;
- rq->hrtick_csd.info = rq;
+ rq_csd_init(rq, &rq->hrtick_csd, __hrtick_start);
#endif
-
hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
rq->hrtick_timer.function = hrtick;
}
@@ -650,6 +655,16 @@ static inline bool got_nohz_idle_kick(void)
return false;
}
+static void nohz_csd_func(void *info)
+{
+ struct rq *rq = info;
+
+ if (got_nohz_idle_kick()) {
+ rq->idle_balance = 1;
+ raise_softirq_irqoff(SCHED_SOFTIRQ);
+ }
+}
+
#else /* CONFIG_NO_HZ_COMMON */
static inline bool got_nohz_idle_kick(void)
@@ -2292,6 +2307,11 @@ void sched_ttwu_pending(void)
rq_unlock_irqrestore(rq, &rf);
}
+static void wake_csd_func(void *info)
+{
+ sched_ttwu_pending();
+}
+
void scheduler_ipi(void)
{
/*
@@ -2300,34 +2320,6 @@ void scheduler_ipi(void)
* this IPI.
*/
preempt_fold_need_resched();
-
- if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
- return;
-
- /*
- * Not all reschedule IPI handlers call irq_enter/irq_exit, since
- * traditionally all their work was done from the interrupt return
- * path. Now that we actually do some work, we need to make sure
- * we do call them.
- *
- * Some archs already do call them, luckily irq_enter/exit nest
- * properly.
- *
- * Arguably we should visit all archs and update all handlers,
- * however a fair share of IPIs are still resched only so this would
- * somewhat pessimize the simple resched case.
- */
- irq_enter();
- sched_ttwu_pending();
-
- /*
- * Check if someone kicked us for doing the nohz idle load balance.
- */
- if (unlikely(got_nohz_idle_kick())) {
- this_rq()->idle_balance = 1;
- raise_softirq_irqoff(SCHED_SOFTIRQ);
- }
- irq_exit();
}
static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
@@ -2336,9 +2328,9 @@ static void ttwu_queue_remote(struct task_struct *p, int
cpu, int wake_flags)
p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
- if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
+ if (llist_add(&p->wake_entry, &rq->wake_list)) {
if (!set_nr_if_polling(rq->idle))
- smp_send_reschedule(cpu);
+ smp_call_function_single_async(cpu, &rq->wake_csd);
else
trace_sched_wake_idle_without_ipi(cpu);
}
@@ -6693,12 +6685,16 @@ void __init sched_init(void)
rq->avg_idle = 2*sysctl_sched_migration_cost;
rq->max_idle_balance_cost = sysctl_sched_migration_cost;
+ rq_csd_init(rq, &rq->wake_csd, wake_csd_func);
+
INIT_LIST_HEAD(&rq->cfs_tasks);
rq_attach_root(rq, &def_root_domain);
#ifdef CONFIG_NO_HZ_COMMON
rq->last_blocked_load_update_tick = jiffies;
atomic_set(&rq->nohz_flags, 0);
+
+ rq_csd_init(rq, &rq->nohz_csd, nohz_csd_func);
#endif
#endif /* CONFIG_SMP */
hrtick_rq_init(rq);
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 46b7bd4..6b7f147 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -10000,12 +10000,11 @@ static void kick_ilb(unsigned int flags)
return;
/*
- * Use smp_send_reschedule() instead of resched_cpu().
- * This way we generate a sched IPI on the target CPU which
+ * This way we generate an IPI on the target CPU which
* is idle. And the softirq performing nohz idle load balance
* will be run before returning from the IPI.
*/
- smp_send_reschedule(ilb_cpu);
+ smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
}
/*
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 978c6fa..21416b3 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -889,9 +889,10 @@ struct rq {
#ifdef CONFIG_SMP
unsigned long last_blocked_load_update_tick;
unsigned int has_blocked_load;
+ call_single_data_t nohz_csd;
#endif /* CONFIG_SMP */
unsigned int nohz_tick_stopped;
- atomic_t nohz_flags;
+ atomic_t nohz_flags;
#endif /* CONFIG_NO_HZ_COMMON */
unsigned long nr_load_updates;
@@ -978,7 +979,7 @@ struct rq {
/* This is used to determine avg_idle's max value */
u64 max_idle_balance_cost;
-#endif
+#endif /* CONFIG_SMP */
#ifdef CONFIG_IRQ_TIME_ACCOUNTING
u64 prev_irq_time;
@@ -1020,6 +1021,7 @@ struct rq {
#endif
#ifdef CONFIG_SMP
+ call_single_data_t wake_csd;
struct llist_head wake_list;
#endif
