Skip to contentSkip to conversation

x86, tsc add an initial read offset to __cycles_2_ns() calculations

PATCH In review Prarit Bhargava 3 participants 3 messages Jul 24, 2013 – Jul 26, 2013 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86, tsc add an initial read offset to __cycles_2_ns() calculations

Note that the E5 Sandybridge processor does not have IA32_TSC_ADJUST MSR implemented so attempting to resynch the TSCs is not possible on the problem hardware.

Thanks to John for the suggestion below.

P.

----8<----

The TSC can have non-zero values at boot time on Intel Xeon E5 (family 6, model 45) aka "SandyBridge" processors. This is documented in the Errata for the E5 processors as BT81.

The __cycles_2_ns() calculation is known to overflow if a large value of cycles is passed into the function. This is done by design to improve precision for smaller significant digits in the calculation. Since the E5 processor can pass in a large value, we need to snapshot the TSC's initial value to avoid calculation overflows in the conversions of cycles to nanoseconds.

Tested successfully on various Sandybridge systems as well as a few older and newer systems without any issues.

Also, remove the unused cycles_2_ns() function.

---
 arch/x86/include/asm/timer.h |   15 +++------------
 arch/x86/kernel/tsc.c        |   13 +++++++++++++
 2 files changed, 16 insertions(+), 12 deletions(-)
arch/x86/include/asm/timer.h +3 -12
index 34baa0e..f9d666b 100644
@@ -12,6 +12,8 @@ extern int recalibrate_cpu_khz(void);
 
 extern int no_timer_check;
 
+extern unsigned long long tsc_initial_value;
+
 /* Accelerators for sched_clock()
  * convert from cycles(64bits) => nanoseconds (64bits)
  *  basic equation:
@@ -59,21 +61,10 @@ static inline unsigned long long __cycles_2_ns(unsigned long long cyc)
 {
 	int cpu = smp_processor_id();
 	unsigned long long ns = per_cpu(cyc2ns_offset, cpu);
+	cyc -= tsc_initial_value;
 	ns += mult_frac(cyc, per_cpu(cyc2ns, cpu),
 			(1UL << CYC2NS_SCALE_FACTOR));
 	return ns;
 }
 
-static inline unsigned long long cycles_2_ns(unsigned long long cyc)
-{
-	unsigned long long ns;
-	unsigned long flags;
-
-	local_irq_save(flags);
-	ns = __cycles_2_ns(cyc);
-	local_irq_restore(flags);
-
-	return ns;
-}
-
 #endif /* _ASM_X86_TIMER_H */
arch/x86/kernel/tsc.c +13 -0
index 6ff4924..63ed8cc 100644
@@ -38,6 +38,16 @@ static int __read_mostly tsc_unstable;
 static int __read_mostly tsc_disabled = -1;
 
 int tsc_clocksource_reliable;
+
+/*
+ * TSC can have non-zero values at boot time on Intel Xeon E5 (family 6,
+ * model 45) aka "SandyBridge" processors.  This is documented in the
+ * Errata for the processors as BT81.  As a result, we need to snapshot
+ * the TSC's initial value to avoid calculation overflows in the conversions
+ * of cycles to nanoseconds.
+ */
+unsigned long long tsc_initial_value;
+
 /*
  * Scheduler clock - returns current time in nanosec units.
  */
@@ -979,6 +989,9 @@ void __init tsc_init(void)
 		return;
 	}
 
+	tsc_initial_value = get_cycles();
+	pr_info("TSC: tsc initial value = %lld\n", tsc_initial_value);
+
 	pr_info("Detected %lu.%03lu MHz processor\n",
 		(unsigned long)cpu_khz / 1000,
 		(unsigned long)cpu_khz % 1000);
Signature · 1 line
1.7.9.3
#1 lore raw
Quoted 38 lines · Prarit Bhargava

The TSC can have non-zero values at boot time on Intel Xeon E5 (family 6, model 45) aka "SandyBridge" processors. This is documented in the Errata for the E5 processors as BT81.

The __cycles_2_ns() calculation is known to overflow if a large value of cycles is passed into the function. This is done by design to improve precision for smaller significant digits in the calculation. Since the E5 processor can pass in a large value, we need to snapshot the TSC's initial value to avoid calculation overflows in the conversions of cycles to nanoseconds.

Tested successfully on various Sandybridge systems as well as a few older and newer systems without any issues.

---
 arch/x86/include/asm/timer.h |   15 +++------------
 arch/x86/kernel/tsc.c        |   13 +++++++++++++
 2 files changed, 16 insertions(+), 12 deletions(-)
arch/x86/include/asm/timer.h +2 -0
index 34baa0e..f9d666b 100644
@@ -12,6 +12,8 @@ extern int recalibrate_cpu_khz(void);
 
 extern int no_timer_check;
 
+extern unsigned long long tsc_initial_value;
+
 /* Accelerators for sched_clock()
  * convert from cycles(64bits) => nanoseconds (64bits)
  *  basic equation:

arch/x86/include/asm/timer.h static inline unsigned long long __cycles_2_ns(unsigned long long cyc) · line 68

Quoted 8 lines
arch/x86/include/asm/timer.h +1 -0
@@ -59,21 +61,10 @@ static inline unsigned long long __cycles_2_ns(unsigned long long cyc)
 {
 	int cpu = smp_processor_id();
 	unsigned long long ns = per_cpu(cyc2ns_offset, cpu);
+	cyc -= tsc_initial_value;
 	ns += mult_frac(cyc, per_cpu(cyc2ns, cpu),
 			(1UL << CYC2NS_SCALE_FACTOR));
 	return ns;
 }

Hurm.. but eventually the TSC value _will_ get that large again, right? So shouldn't we fix the actual problem?

maintainer
#2 lore raw
In reply to Prarit Bhargava
Quoted 13 lines · Prarit Bhargava
arch/x86/kernel/tsc.c +9 -0
+/*
+ * TSC can have non-zero values at boot time on Intel Xeon E5 (family 6,
+ * model 45) aka "SandyBridge" processors.  This is documented in the
+ * Errata for the processors as BT81.  As a result, we need to snapshot
+ * the TSC's initial value to avoid calculation overflows in the conversions
+ * of cycles to nanoseconds.
+ */
+unsigned long long tsc_initial_value;
+
 /*
  * Scheduler clock - returns current time in nanosec units.
  */

arch/x86/kernel/tsc.c void __init tsc_init(void) · line 994

Quoted 6 lines
arch/x86/kernel/tsc.c +3 -0
@@ -979,6 +989,9 @@ void __init tsc_init(void)
 		return;
 	}
 
+	tsc_initial_value = get_cycles();
+	pr_info("TSC: tsc initial value = %lld\n", tsc_initial_value);
+

And we probably don't need that info on processors unaffected by the erratum...

Signature · 5 lines
Regards/Gruss,
    Boris.

Sent from a fat crate under my desk. Formatting is fine.
--