[PATCH] cpufreq: stats: Add memory barrier to store_reset()

[PATCH] cpufreq: stats: Add memory barrier to store_reset()

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

There is nothing to prevent the CPU or the compiler from reordering
the writes to stats->reset_time and stats->reset_pending in
store_reset(), in which case the readers of stats->reset_time may see
a stale value.  Moreover, on 32-bit arches the write to reset_time
cannot be completed in one go, so the readers of it may see a
partially updated value in that case.

To prevent that from happening, add a write memory barrier between
the writes to stats->reset_time and stats->reset_pending in
store_reset().

Fixes: 40c3bd4cfa6f ("cpufreq: stats: Defer stats update to cpufreq_stats_record_transition()")
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---

I couldn't convince myself that it was OK to leave the code as it was.

linux-next material.

---
 drivers/cpufreq/cpufreq_stats.c |    7 +++++++
 1 file changed, 7 insertions(+)

Index: linux-pm/drivers/cpufreq/cpufreq_stats.c
===================================================================
--- linux-pm.orig/drivers/cpufreq/cpufreq_stats.c
+++ linux-pm/drivers/cpufreq/cpufreq_stats.c
@@ -99,6 +99,13 @@ static ssize_t store_reset(struct cpufre
 	 * avoid races.
 	 */
 	WRITE_ONCE(stats->reset_time, get_jiffies_64());
+	/*
+	 * The memory barrier below is to prevent the readers of reset_time from
+	 * seeing a stale or partially updated value. Note that they both access
+	 * reset_time only if reset_pending is 1, so corresponding read barriers
+	 * are not needed.
+	 */
+	smp_wmb();
 	WRITE_ONCE(stats->reset_pending, 1);
 
 	return count;

Re: [PATCH] cpufreq: stats: Add memory barrier to store_reset()

[Rafael J. Wysocki]

On Tue, Oct 6, 2020 at 1:59 PM Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
>
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>
> There is nothing to prevent the CPU or the compiler from reordering
> the writes to stats->reset_time and stats->reset_pending in
> store_reset(), in which case the readers of stats->reset_time may see
> a stale value.  Moreover, on 32-bit arches the write to reset_time
> cannot be completed in one go, so the readers of it may see a
> partially updated value in that case.
>
> To prevent that from happening, add a write memory barrier between
> the writes to stats->reset_time and stats->reset_pending in
> store_reset().
>
> Fixes: 40c3bd4cfa6f ("cpufreq: stats: Defer stats update to cpufreq_stats_record_transition()")
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
>
> I couldn't convince myself that it was OK to leave the code as it was.
>
> linux-next material.
>
> ---
>  drivers/cpufreq/cpufreq_stats.c |    7 +++++++
>  1 file changed, 7 insertions(+)
>
> Index: linux-pm/drivers/cpufreq/cpufreq_stats.c
> ===================================================================
> --- linux-pm.orig/drivers/cpufreq/cpufreq_stats.c
> +++ linux-pm/drivers/cpufreq/cpufreq_stats.c
> @@ -99,6 +99,13 @@ static ssize_t store_reset(struct cpufre
>          * avoid races.
>          */
>         WRITE_ONCE(stats->reset_time, get_jiffies_64());
> +       /*
> +        * The memory barrier below is to prevent the readers of reset_time from
> +        * seeing a stale or partially updated value. Note that they both access
> +        * reset_time only if reset_pending is 1, so corresponding read barriers
> +        * are not needed.

I'm taking this back after double-checking memory-barriers.txt.

Will send a v2.

> +        */
> +       smp_wmb();
>         WRITE_ONCE(stats->reset_pending, 1);
>
>         return count;
>
>
>