ARM: 7643/1: sched: correct update_sched_clock()
Joonsoo Kim [Sat, 9 Feb 2013 04:52:45 +0000 (05:52 +0100)]
commit 7c4e9ced424be4d36df6a3e3825763e97ee97607 upstream.

If we want load epoch_cyc and epoch_ns atomically,
we should update epoch_cyc_copy first of all.
This notify reader that updating is in progress.

If we update epoch_cyc first like as current implementation,
there is subtle error case.
Look at the below example.

<Initial Condition>
cyc = 9
ns = 900
cyc_copy = 9

== CASE 1 ==
<CPU A = reader>           <CPU B = updater>
                           write cyc = 10
read cyc = 10
read ns = 900
                           write ns = 1000
                           write cyc_copy = 10
read cyc_copy = 10

output = (10, 900)

== CASE 2 ==
<CPU A = reader>           <CPU B = updater>
read cyc = 9
                           write cyc = 10
                           write ns = 1000
read ns = 1000
read cyc_copy = 9
                           write cyc_copy = 10
output = (9, 1000)

If atomic read is ensured, output should be (9, 900) or (10, 1000).
But, output in example case are not.

So, change updating sequence in order to correct this problem.

Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

arch/arm/kernel/sched_clock.c

index 27d186a..6bbf936 100644 (file)
@@ -84,11 +84,11 @@ static void notrace update_sched_clock(void)
         * detectable in cyc_to_fixed_sched_clock().
         */
        raw_local_irq_save(flags);
-       cd.epoch_cyc = cyc;
+       cd.epoch_cyc_copy = cyc;
        smp_wmb();
        cd.epoch_ns = ns;
        smp_wmb();
-       cd.epoch_cyc_copy = cyc;
+       cd.epoch_cyc = cyc;
        raw_local_irq_restore(flags);
 }