[PATCH 2/2] sched: Fix "divide error: 0000" in find_busiest_group

[PATCH 2/2] sched: Fix "divide error: 0000" in find_busiest_group

[Terry Loftin]

Add a check to sched_avg_update() to detect and reset age_stamp if the
clock value has wrapped.  Because __cycles_2_ns() includes an offset
to account for start up time, the clock may not wrap to zero, so use
the current clock value instead.

Signed-off-by: Terry Loftin <terry.loftin@hp.com>
Signed-off-by: Bob Montgomery <bob.montgomery@hp.com>
---
diff --git a/kernel/sched.c b/kernel/sched.c
index 18d38e4..b39cae1 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -1256,6 +1256,8 @@ static void sched_avg_update(struct rq *rq)
 {
 	s64 period = sched_avg_period();

+	if (unlikely(rq->age_stamp > rq->clock))
+		rq->age_stamp = rq->clock;
 	while ((s64)(rq->clock - rq->age_stamp) > period) {
 		/*
 		 * Inline assembly required to prevent the compiler

Re: [PATCH 2/2] sched: Fix "divide error: 0000" in find_busiest_group

[Peter Zijlstra]

On Tue, 2011-07-19 at 14:58 -0600, Terry Loftin wrote:
> Add a check to sched_avg_update() to detect and reset age_stamp if the
> clock value has wrapped.  Because __cycles_2_ns() includes an offset
> to account for start up time, the clock may not wrap to zero, so use
> the current clock value instead.

So you're running on a platform (unspecified) where we use a raw
sched_clock() that is buggy. Again, you're fixing symptoms not causes. 

> Signed-off-by: Terry Loftin <terry.loftin@hp.com>
> Signed-off-by: Bob Montgomery <bob.montgomery@hp.com>
> ---
> diff --git a/kernel/sched.c b/kernel/sched.c
> index 18d38e4..b39cae1 100644
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -1256,6 +1256,8 @@ static void sched_avg_update(struct rq *rq)
>  {
>  	s64 period = sched_avg_period();
> 
> +	if (unlikely(rq->age_stamp > rq->clock))
> +		rq->age_stamp = rq->clock;
>  	while ((s64)(rq->clock - rq->age_stamp) > period) {
>  		/*
>  		 * Inline assembly required to prevent the compiler

Re: [PATCH 2/2] sched: Fix "divide error: 0000" in find_busiest_group

[Terry Loftin]

On 07/19/2011 03:18 PM, Peter Zijlstra wrote:
> On Tue, 2011-07-19 at 14:58 -0600, Terry Loftin wrote:
>> Add a check to sched_avg_update() to detect and reset age_stamp if the
>> clock value has wrapped.  Because __cycles_2_ns() includes an offset
>> to account for start up time, the clock may not wrap to zero, so use
>> the current clock value instead.
> 
> So you're running on a platform (unspecified) where we use a raw
> sched_clock() that is buggy. Again, you're fixing symptoms not causes. 
> 
This x86_64.  This is the actual cause, unless the rq->clock
value should never roll, in which case, the clock roll is the
actual cause and you can disregard these patches.

-T
>> Signed-off-by: Terry Loftin <terry.loftin@hp.com>
>> Signed-off-by: Bob Montgomery <bob.montgomery@hp.com>
>> ---
>> diff --git a/kernel/sched.c b/kernel/sched.c
>> index 18d38e4..b39cae1 100644
>> --- a/kernel/sched.c
>> +++ b/kernel/sched.c
>> @@ -1256,6 +1256,8 @@ static void sched_avg_update(struct rq *rq)
>>  {
>>  	s64 period = sched_avg_period();
>>
>> +	if (unlikely(rq->age_stamp > rq->clock))
>> +		rq->age_stamp = rq->clock;
>>  	while ((s64)(rq->clock - rq->age_stamp) > period) {
>>  		/*
>>  		 * Inline assembly required to prevent the compiler
> 
> 
> 

Re: [PATCH 2/2] sched: Fix "divide error: 0000" in find_busiest_group

[Peter Zijlstra]

On Tue, 2011-07-19 at 16:21 -0600, Terry Loftin wrote:
> > So you're running on a platform (unspecified) where we use a raw
> > sched_clock() that is buggy. Again, you're fixing symptoms not causes. 
> > 
> This x86_64.  This is the actual cause, unless the rq->clock
> value should never roll, in which case, the clock roll is the
> actual cause and you can disregard these patches.

Its supposed to roll over on the full 64bit, and I think x86_64 only
suffers this if you have sched_clock_stable set to 1.

So I think the correct fix is disabling that logic for now. John Stultz
was working on some patches to fix __cycles_2_ns().

Something like the below perhaps.

---
 arch/x86/kernel/cpu/intel.c |    2 --
 1 files changed, 0 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/cpu/intel.c b/arch/x86/kernel/cpu/intel.c
index 1edf5ba..dba0482 100644
--- a/arch/x86/kernel/cpu/intel.c
+++ b/arch/x86/kernel/cpu/intel.c
@@ -91,8 +91,6 @@ static void __cpuinit early_init_intel(struct cpuinfo_x86 *c)
 	if (c->x86_power & (1 << 8)) {
 		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
 		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
-		if (!check_tsc_unstable())
-			sched_clock_stable = 1;
 	}
 
 	/*

Re: [PATCH 2/2] sched: Fix "divide error: 0000" in find_busiest_group

[Terry Loftin]

On 07/19/2011 04:33 PM, Peter Zijlstra wrote:
> On Tue, 2011-07-19 at 16:21 -0600, Terry Loftin wrote:
>>> So you're running on a platform (unspecified) where we use a raw
>>> sched_clock() that is buggy. Again, you're fixing symptoms not causes. 
>>>
>> This x86_64.  This is the actual cause, unless the rq->clock
>> value should never roll, in which case, the clock roll is the
>> actual cause and you can disregard these patches.
> 
> Its supposed to roll over on the full 64bit, and I think x86_64 only
> suffers this if you have sched_clock_stable set to 1.
> 
> So I think the correct fix is disabling that logic for now. John Stultz
> was working on some patches to fix __cycles_2_ns().
> 
> Something like the below perhaps.
OK - this would also avoid the problem I came across.  Thank you
for your time.

-T
> 
> ---
>  arch/x86/kernel/cpu/intel.c |    2 --
>  1 files changed, 0 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/kernel/cpu/intel.c b/arch/x86/kernel/cpu/intel.c
> index 1edf5ba..dba0482 100644
> --- a/arch/x86/kernel/cpu/intel.c
> +++ b/arch/x86/kernel/cpu/intel.c
> @@ -91,8 +91,6 @@ static void __cpuinit early_init_intel(struct cpuinfo_x86 *c)
>  	if (c->x86_power & (1 << 8)) {
>  		set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
>  		set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
> -		if (!check_tsc_unstable())
> -			sched_clock_stable = 1;
>  	}
>  
>  	/*
> 
>