[PATCH] sched_clock: fix jiffie fallback clock

[PATCH] sched_clock: fix jiffie fallback clock

[Peter Zijlstra]

David pointed out that the default sched_clock() fallback is broken in that it
wraps too soon. Fix this by using the 64 bit jiffie value so that we're large 
enough to overflow properly.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
CC: David Howells <dhowells@redhat.com>
---
 arch/x86/kernel/tsc.c |    6 ++----
 kernel/sched_clock.c  |    2 +-
 2 files changed, 3 insertions(+), 5 deletions(-)

Index: linux-2.6/arch/x86/kernel/tsc.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/tsc.c	2008-09-15 18:41:26.000000000 +0200
+++ linux-2.6/arch/x86/kernel/tsc.c	2008-09-15 18:41:33.000000000 +0200
@@ -46,10 +46,8 @@ u64 native_sched_clock(void)
 	 *   very important for it to be as fast as the platform
 	 *   can achive it. )
 	 */
-	if (unlikely(tsc_disabled)) {
-		/* No locking but a rare wrong value is not a big deal: */
-		return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
-	}
+	if (unlikely(tsc_disabled))
+		return (get_jiffies_64() - INITIAL_JIFFIES) * (NSEC_PER_SEC/HZ);
 
 	/* read the Time Stamp Counter: */
 	rdtscll(this_offset);
Index: linux-2.6/kernel/sched_clock.c
===================================================================
--- linux-2.6.orig/kernel/sched_clock.c	2008-09-15 18:41:26.000000000 +0200
+++ linux-2.6/kernel/sched_clock.c	2008-09-15 18:41:33.000000000 +0200
@@ -38,7 +38,7 @@
  */
 unsigned long long __attribute__((weak)) sched_clock(void)
 {
-	return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
+	return (get_jiffies_64() - INITIAL_JIFFIES) * (NSEC_PER_SEC/HZ);
 }
 
 static __read_mostly int sched_clock_running;

Re: [PATCH] sched_clock: fix jiffie fallback clock

[Ingo Molnar]

* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> David pointed out that the default sched_clock() fallback is broken in 
> that it wraps too soon. Fix this by using the 64 bit jiffie value so 
> that we're large enough to overflow properly.
> 
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> CC: David Howells <dhowells@redhat.com>

applied to tip/sched/urgent, thanks Peter!

	Ingo

Re: [PATCH] sched_clock: fix jiffie fallback clock

[Ingo Molnar]

* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> David pointed out that the default sched_clock() fallback is broken in 
> that it wraps too soon. Fix this by using the 64 bit jiffie value so 
> that we're large enough to overflow properly.
> 
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> CC: David Howells <dhowells@redhat.com>

FYI, -tip testing found bootup hard-lockups today on 32-bit 
testsystems.

I've bisected it down to this change. After some head-scratching and an 
hour of debugging, it turns out that this aspect:

> @@ -46,10 +46,8 @@ u64 native_sched_clock(void)
>        *   very important for it to be as fast as the platform
>        *   can achive it. )
>        */
> -     if (unlikely(tsc_disabled)) {
> -             /* No locking but a rare wrong value is not a big deal: */
> -             return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
> -     }
> +     if (unlikely(tsc_disabled))
> +             return (get_jiffies_64() - INITIAL_JIFFIES) * (NSEC_PER_SEC/HZ);

is the buggy one: it calls get_jiffies_64() which does:

                seq = read_seqbegin(&xtime_lock);

that's not a very wise thing to do within xtime-locked sections - and we 
do call cpu_clock()/sched_clock() in a number of xtime-locked sections.

So i've zapped this commit for the time being, this needs to be solved 
differently.

	Ingo