LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
* [PATCH] system timer: fix crash in <100Hz system timer
@ 2008-01-05 22:16 David Fries
  2008-01-07 23:36 ` Andrew Morton
  2008-01-09  7:15 ` Andrew Morton
  0 siblings, 2 replies; 5+ messages in thread
From: David Fries @ 2008-01-05 22:16 UTC (permalink / raw)
  To: Thomas Gleixner, Andrew Morton; +Cc: linux-kernel

[-- Attachment #1: Type: text/plain, Size: 2078 bytes --]

The kernel has a divide by zero crash when trying to run the system
timer less than 100Hz.  The problem is x/(HZ/USER_HZ) and related.
Now x*(USER_HZ/HZ) will be used if HZ<USER_HZ.

I'm running the Linux kernel under qemu and went to run a slower
system timer to take less CPU (and battery) on the host.  I found that
the kernel paniced under emulation because of a divide by zero in
three places.  Here is the patch.  The base git was updated today
01-05-2008.  I went for a 20Hz system time by adding config HZ_20 etc
to kernel/Kconfig.hz.  With this patch I verified the system timer by
looking at /proc/interrupts.

Signed-off-by: David Fries <david@fries.net>

-- 
David Fries <david@fries.net>
http://fries.net/~david/ (PGP encryption key available)


diff --git a/include/linux/acct.h b/include/linux/acct.h
index 302eb72..86b848d 100644
--- a/include/linux/acct.h
+++ b/include/linux/acct.h
@@ -173,7 +173,11 @@ typedef struct acct acct_t;
 static inline u32 jiffies_to_AHZ(unsigned long x)
 {
 #if (TICK_NSEC % (NSEC_PER_SEC / AHZ)) == 0
-	return x / (HZ / AHZ);
+	#if HZ < AHZ
+		return x * (AHZ / HZ);
+	#else
+		return x / (HZ / AHZ);
+	#endif
 #else
         u64 tmp = (u64)x * TICK_NSEC;
         do_div(tmp, (NSEC_PER_SEC / AHZ));
diff --git a/kernel/time.c b/kernel/time.c
index 09d3c45..23af26f 100644
--- a/kernel/time.c
+++ b/kernel/time.c
@@ -565,7 +565,11 @@ EXPORT_SYMBOL(jiffies_to_timeval);
 clock_t jiffies_to_clock_t(long x)
 {
 #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
+	#if HZ < USER_HZ
+	return x * (USER_HZ / HZ);
+	#else
 	return x / (HZ / USER_HZ);
+	#endif
 #else
 	u64 tmp = (u64)x * TICK_NSEC;
 	do_div(tmp, (NSEC_PER_SEC / USER_HZ));
@@ -598,7 +602,12 @@ EXPORT_SYMBOL(clock_t_to_jiffies);
 u64 jiffies_64_to_clock_t(u64 x)
 {
 #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
-	do_div(x, HZ / USER_HZ);
+	#if HZ < USER_HZ
+		x *= USER_HZ;
+		do_div(x, HZ);
+	#else
+		do_div(x, HZ / USER_HZ);
+	#endif
 #else
 	/*
 	 * There are better ways that don't overflow early,

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [PATCH] system timer: fix crash in <100Hz system timer
  2008-01-05 22:16 [PATCH] system timer: fix crash in <100Hz system timer David Fries
@ 2008-01-07 23:36 ` Andrew Morton
  2008-01-07 23:51   ` H. Peter Anvin
  2008-01-09  7:15 ` Andrew Morton
  1 sibling, 1 reply; 5+ messages in thread
From: Andrew Morton @ 2008-01-07 23:36 UTC (permalink / raw)
  To: David Fries; +Cc: tglx, linux-kernel, H. Peter Anvin

On Sat, 5 Jan 2008 16:16:55 -0600
David Fries <david@fries.net> wrote:

> The kernel has a divide by zero crash when trying to run the system
> timer less than 100Hz.  The problem is x/(HZ/USER_HZ) and related.
> Now x*(USER_HZ/HZ) will be used if HZ<USER_HZ.
> 
> I'm running the Linux kernel under qemu and went to run a slower
> system timer to take less CPU (and battery) on the host.  I found that
> the kernel paniced under emulation because of a divide by zero in
> three places.  Here is the patch.  The base git was updated today
> 01-05-2008.  I went for a 20Hz system time by adding config HZ_20 etc
> to kernel/Kconfig.hz.  With this patch I verified the system timer by
> looking at /proc/interrupts.
> 
> Signed-off-by: David Fries <david@fries.net>
> 
> -- 
> David Fries <david@fries.net>
> http://fries.net/~david/ (PGP encryption key available)
> 
> 
> diff --git a/include/linux/acct.h b/include/linux/acct.h
> index 302eb72..86b848d 100644
> --- a/include/linux/acct.h
> +++ b/include/linux/acct.h
> @@ -173,7 +173,11 @@ typedef struct acct acct_t;
>  static inline u32 jiffies_to_AHZ(unsigned long x)
>  {
>  #if (TICK_NSEC % (NSEC_PER_SEC / AHZ)) == 0
> -	return x / (HZ / AHZ);
> +	#if HZ < AHZ
> +		return x * (AHZ / HZ);
> +	#else
> +		return x / (HZ / AHZ);
> +	#endif
>  #else
>          u64 tmp = (u64)x * TICK_NSEC;
>          do_div(tmp, (NSEC_PER_SEC / AHZ));
> diff --git a/kernel/time.c b/kernel/time.c
> index 09d3c45..23af26f 100644
> --- a/kernel/time.c
> +++ b/kernel/time.c
> @@ -565,7 +565,11 @@ EXPORT_SYMBOL(jiffies_to_timeval);
>  clock_t jiffies_to_clock_t(long x)
>  {
>  #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
> +	#if HZ < USER_HZ
> +	return x * (USER_HZ / HZ);
> +	#else
>  	return x / (HZ / USER_HZ);
> +	#endif
>  #else
>  	u64 tmp = (u64)x * TICK_NSEC;
>  	do_div(tmp, (NSEC_PER_SEC / USER_HZ));
> @@ -598,7 +602,12 @@ EXPORT_SYMBOL(clock_t_to_jiffies);
>  u64 jiffies_64_to_clock_t(u64 x)
>  {
>  #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
> -	do_div(x, HZ / USER_HZ);
> +	#if HZ < USER_HZ
> +		x *= USER_HZ;
> +		do_div(x, HZ);
> +	#else
> +		do_div(x, HZ / USER_HZ);
> +	#endif
>  #else
>  	/*
>  	 * There are better ways that don't overflow early,

Alas, I get 100% rejects due to conflicting changes from Peter's
avoid-overflows-in-kernel-timec.patch.

Peter, did that patch propagate this failure, or might it have happily
fixed it?


^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [PATCH] system timer: fix crash in <100Hz system timer
  2008-01-07 23:36 ` Andrew Morton
@ 2008-01-07 23:51   ` H. Peter Anvin
  2008-01-08  0:09     ` Andrew Morton
  0 siblings, 1 reply; 5+ messages in thread
From: H. Peter Anvin @ 2008-01-07 23:51 UTC (permalink / raw)
  To: Andrew Morton; +Cc: David Fries, tglx, linux-kernel

Andrew Morton wrote:
>>
>> diff --git a/include/linux/acct.h b/include/linux/acct.h
>> index 302eb72..86b848d 100644
>> --- a/include/linux/acct.h
>> +++ b/include/linux/acct.h
>> @@ -173,7 +173,11 @@ typedef struct acct acct_t;
>>  static inline u32 jiffies_to_AHZ(unsigned long x)
>>  {
>>  #if (TICK_NSEC % (NSEC_PER_SEC / AHZ)) == 0
>> -	return x / (HZ / AHZ);
>> +	#if HZ < AHZ
>> +		return x * (AHZ / HZ);
>> +	#else
>> +		return x / (HZ / AHZ);
>> +	#endif
>>  #else
>>          u64 tmp = (u64)x * TICK_NSEC;
>>          do_div(tmp, (NSEC_PER_SEC / AHZ));
>> diff --git a/kernel/time.c b/kernel/time.c
>> index 09d3c45..23af26f 100644
>> --- a/kernel/time.c
>> +++ b/kernel/time.c
>> @@ -565,7 +565,11 @@ EXPORT_SYMBOL(jiffies_to_timeval);
>>  clock_t jiffies_to_clock_t(long x)
>>  {
>>  #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
>> +	#if HZ < USER_HZ
>> +	return x * (USER_HZ / HZ);
>> +	#else
>>  	return x / (HZ / USER_HZ);
>> +	#endif
>>  #else
>>  	u64 tmp = (u64)x * TICK_NSEC;
>>  	do_div(tmp, (NSEC_PER_SEC / USER_HZ));
>> @@ -598,7 +602,12 @@ EXPORT_SYMBOL(clock_t_to_jiffies);
>>  u64 jiffies_64_to_clock_t(u64 x)
>>  {
>>  #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
>> -	do_div(x, HZ / USER_HZ);
>> +	#if HZ < USER_HZ
>> +		x *= USER_HZ;
>> +		do_div(x, HZ);
>> +	#else
>> +		do_div(x, HZ / USER_HZ);
>> +	#endif
>>  #else
>>  	/*
>>  	 * There are better ways that don't overflow early,
> 
> Alas, I get 100% rejects due to conflicting changes from Peter's
> avoid-overflows-in-kernel-timec.patch.
> 
> Peter, did that patch propagate this failure, or might it have happily
> fixed it?
> 

My patch doesn't touch any of these functions, nor touches any code 
within 70 lines of this patch -- the last line touched is line 478 -- 
and doesn't touch linux/acct.h at all, so how could it cause a conflict?

But no, it doesn't fix this particular problem, even if using a similar 
technique might very well be a better way to do this kind of conversion.

	-hpa

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [PATCH] system timer: fix crash in <100Hz system timer
  2008-01-07 23:51   ` H. Peter Anvin
@ 2008-01-08  0:09     ` Andrew Morton
  0 siblings, 0 replies; 5+ messages in thread
From: Andrew Morton @ 2008-01-08  0:09 UTC (permalink / raw)
  To: H. Peter Anvin; +Cc: david, tglx, linux-kernel

On Mon, 07 Jan 2008 15:51:09 -0800
"H. Peter Anvin" <hpa@zytor.com> wrote:

> >> +		do_div(x, HZ / USER_HZ);
> >> +	#endif
> >>  #else
> >>  	/*
> >>  	 * There are better ways that don't overflow early,
> > 
> > Alas, I get 100% rejects due to conflicting changes from Peter's
> > avoid-overflows-in-kernel-timec.patch.
> > 
> > Peter, did that patch propagate this failure, or might it have happily
> > fixed it?
> > 
> 
> My patch doesn't touch any of these functions, nor touches any code 
> within 70 lines of this patch -- the last line touched is line 478 -- 
> and doesn't touch linux/acct.h at all, so how could it cause a conflict?

oop, sorry, it was the mime-mess-mangles-everything problem and I failed to
spot it.

> But no, it doesn't fix this particular problem, even if using a similar 
> technique might very well be a better way to do this kind of conversion.

OK.

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [PATCH] system timer: fix crash in <100Hz system timer
  2008-01-05 22:16 [PATCH] system timer: fix crash in <100Hz system timer David Fries
  2008-01-07 23:36 ` Andrew Morton
@ 2008-01-09  7:15 ` Andrew Morton
  1 sibling, 0 replies; 5+ messages in thread
From: Andrew Morton @ 2008-01-09  7:15 UTC (permalink / raw)
  To: David Fries; +Cc: Thomas Gleixner, linux-kernel, Ingo Molnar

On Sat, 5 Jan 2008 16:16:55 -0600 David Fries <david@fries.net> wrote:

> --- a/kernel/time.c
> +++ b/kernel/time.c
> @@ -565,7 +565,11 @@ EXPORT_SYMBOL(jiffies_to_timeval);
>  clock_t jiffies_to_clock_t(long x)
>  {
>  #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
> +	#if HZ < USER_HZ
> +	return x * (USER_HZ / HZ);
> +	#else
>  	return x / (HZ / USER_HZ);
> +	#endif
>  #else
>  	u64 tmp = (u64)x * TICK_NSEC;
>  	do_div(tmp, (NSEC_PER_SEC / USER_HZ));
> @@ -598,7 +602,12 @@ EXPORT_SYMBOL(clock_t_to_jiffies);
>  u64 jiffies_64_to_clock_t(u64 x)
>  {
>  #if (TICK_NSEC % (NSEC_PER_SEC / USER_HZ)) == 0
> -	do_div(x, HZ / USER_HZ);
> +	#if HZ < USER_HZ
> +		x *= USER_HZ;
> +		do_div(x, HZ);
> +	#else
> +		do_div(x, HZ / USER_HZ);
> +	#endif
>  #else

Somwhat off-topic:

I guess HZ=USER_HZ is a not-uncommon case, and it's pretty silly calling
do_div(x, 1) all the time.  How about we optimise that case?

Perhaps there are other places...


--- a/kernel/time.c~speed-up-jiffies-conversion-functions-if-hz==user_hz
+++ a/kernel/time.c
@@ -618,8 +618,10 @@ u64 jiffies_64_to_clock_t(u64 x)
 # if HZ < USER_HZ
 	x *= USER_HZ;
 	do_div(x, HZ);
-# else
+# elif HZ > USER_HZ
 	do_div(x, HZ / USER_HZ);
+# else
+	/* Nothing to do */
 # endif
 #else
 	/*
_


^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2008-01-09  7:15 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2008-01-05 22:16 [PATCH] system timer: fix crash in <100Hz system timer David Fries
2008-01-07 23:36 ` Andrew Morton
2008-01-07 23:51   ` H. Peter Anvin
2008-01-08  0:09     ` Andrew Morton
2008-01-09  7:15 ` Andrew Morton

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).