Re: [PATCH v15 23/26] sched: early boot clock

[Peter Zijlstra]

On Thu, Jul 19, 2018 at 04:55:42PM -0400, Pavel Tatashin wrote:
> diff --git a/kernel/sched/clock.c b/kernel/sched/clock.c
> index 0e9dbb2d9aea..422cd63f8f17 100644
> --- a/kernel/sched/clock.c
> +++ b/kernel/sched/clock.c
> @@ -202,7 +202,25 @@ static void __sched_clock_gtod_offset(void)
>  
>  void __init sched_clock_init(void)
>  {
> +	unsigned long flags;
> +
> +	/*
> +	 * Set __gtod_offset such that once we mark sched_clock_running,
> +	 * sched_clock_tick() continues where sched_clock() left off.
> +	 *
> +	 * Even if TSC is buggered, we're still UP at this point so it
> +	 * can't really be out of sync.
> +	 */
> +	local_irq_save(flags);
> +	__sched_clock_gtod_offset();
> +	local_irq_restore(flags);
> +
>  	sched_clock_running = 1;
> +
> +	/* Now that sched_clock_running is set adjust scd */
> +	local_irq_save(flags);
> +	sched_clock_tick();
> +	local_irq_restore(flags);
>  }

Sorry, that's still wrong. Because the moment you enable
sched_clock_running we need to have everything set-up for it to run.

The above looks double weird because you could've just done that =1
under the same IRQ-disable section and it would've mostly been OK
(except for NMIs). But the reason it's weird like that is because you're
going to change it into a static key later on.

The below cures things.

---
Subject: sched/clock: Close a hole in sched_clock_init()

All data required for the 'unstable' sched_clock must be set-up _before_
enabling it -- setting sched_clock_running. This includes the
__gtod_offset but also a recent scd stamp.

Make the gtod-offset update also set the csd stamp -- it requires the
same two clock reads _anyway_. This doesn't hurt in the
sched_clock_tick_stable() case and ensures sched_clock_init() gets
everything set-up before use.

Also switch to unconditional IRQ-disable/enable because the static key
stuff already requires this is not ran with IRQs disabled.

Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 kernel/sched/clock.c | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/kernel/sched/clock.c b/kernel/sched/clock.c
index c5c47ad3f386..811a39aca1ce 100644
--- a/kernel/sched/clock.c
+++ b/kernel/sched/clock.c
@@ -197,13 +197,14 @@ void clear_sched_clock_stable(void)
 
 static void __sched_clock_gtod_offset(void)
 {
-	__gtod_offset = (sched_clock() + __sched_clock_offset) - ktime_get_ns();
+	struct sched_clock_data *scd = this_scd();
+
+	__scd_stamp(scd);
+	__gtod_offset = (scd->tick_raw + __sched_clock_offset) - scd->tick_gtod;
 }
 
 void __init sched_clock_init(void)
 {
-	unsigned long flags;
-
 	/*
 	 * Set __gtod_offset such that once we mark sched_clock_running,
 	 * sched_clock_tick() continues where sched_clock() left off.
@@ -211,16 +212,11 @@ void __init sched_clock_init(void)
 	 * Even if TSC is buggered, we're still UP at this point so it
 	 * can't really be out of sync.
 	 */
-	local_irq_save(flags);
+	local_irq_disable();
 	__sched_clock_gtod_offset();
-	local_irq_restore(flags);
+	local_irq_enable();
 
 	static_branch_inc(&sched_clock_running);
-
-	/* Now that sched_clock_running is set adjust scd */
-	local_irq_save(flags);
-	sched_clock_tick();
-	local_irq_restore(flags);
 }
 /*
  * We run this as late_initcall() such that it runs after all built-in drivers,