sched, x86: Avoid unnecessary overflow in sched_clock
authorSalman Qazi <sqazi@google.com>
Tue, 15 Nov 2011 22:12:06 +0000 (14:12 -0800)
committerIngo Molnar <mingo@elte.hu>
Wed, 16 Nov 2011 18:51:25 +0000 (19:51 +0100)
(Added the missing signed-off-by line)

In hundreds of days, the __cycles_2_ns calculation in sched_clock
has an overflow.  cyc * per_cpu(cyc2ns, cpu) exceeds 64 bits, causing
the final value to become zero.  We can solve this without losing
any precision.

We can decompose TSC into quotient and remainder of division by the
scale factor, and then use this to convert TSC into nanoseconds.

Signed-off-by: Salman Qazi <sqazi@google.com>
Acked-by: John Stultz <johnstul@us.ibm.com>
Reviewed-by: Paul Turner <pjt@google.com>
Cc: stable@kernel.org
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20111115221121.7262.88871.stgit@dungbeetle.mtv.corp.google.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/include/asm/timer.h

index fa7b9176b76cb33820034403fd8f4a50dc49709c..431793e5d4846f23bf5947f933fdaf9f5ae1987a 100644 (file)
@@ -32,6 +32,22 @@ extern int no_timer_check;
  *  (mathieu.desnoyers@polymtl.ca)
  *
  *                     -johnstul@us.ibm.com "math is hard, lets go shopping!"
+ *
+ * In:
+ *
+ * ns = cycles * cyc2ns_scale / SC
+ *
+ * Although we may still have enough bits to store the value of ns,
+ * in some cases, we may not have enough bits to store cycles * cyc2ns_scale,
+ * leading to an incorrect result.
+ *
+ * To avoid this, we can decompose 'cycles' into quotient and remainder
+ * of division by SC.  Then,
+ *
+ * ns = (quot * SC + rem) * cyc2ns_scale / SC
+ *    = quot * cyc2ns_scale + (rem * cyc2ns_scale) / SC
+ *
+ *                     - sqazi@google.com
  */
 
 DECLARE_PER_CPU(unsigned long, cyc2ns);
@@ -41,9 +57,14 @@ DECLARE_PER_CPU(unsigned long long, cyc2ns_offset);
 
 static inline unsigned long long __cycles_2_ns(unsigned long long cyc)
 {
+       unsigned long long quot;
+       unsigned long long rem;
        int cpu = smp_processor_id();
        unsigned long long ns = per_cpu(cyc2ns_offset, cpu);
-       ns += cyc * per_cpu(cyc2ns, cpu) >> CYC2NS_SCALE_FACTOR;
+       quot = (cyc >> CYC2NS_SCALE_FACTOR);
+       rem = cyc & ((1ULL << CYC2NS_SCALE_FACTOR) - 1);
+       ns += quot * per_cpu(cyc2ns, cpu) +
+               ((rem * per_cpu(cyc2ns, cpu)) >> CYC2NS_SCALE_FACTOR);
        return ns;
 }