[Patch] Fix sys_time to handle intra-tick correction

[Patch] Fix sys_time to handle intra-tick correction

[Vinay Sridhar]

Hi All,

This fix changes sys_time to use do_gettimeofday instead of get_seconds.
Running the stime01 test from LTP triggers this error. Calling sys_stime
and then calling sys_time causes this. "do_settimeofday" subtracts the
nsec offset from the nsec value(0 in this case) passed to it.
Subsequently, "set_normalized_timespec" modifies sec and nsec
accordingly. This compensation is handled in the do_gettimeofday path.
However, sys_time does not handle this case and reports an incorrect
seconds value.


signed-off by : Vinay Sridhar <vinay@linux.vnet.ibm.com>

--- time.c.orig	2008-12-04 06:30:12.000000000 -0600
+++ time.c	2008-12-04 06:31:25.000000000 -0600
@@ -61,7 +61,12 @@ EXPORT_SYMBOL(sys_tz);
  */
 asmlinkage long sys_time(time_t __user * tloc)
 {
-	time_t i = get_seconds();
+	struct timeval t;	
+	time_t i;
+
+	do_gettimeofday(&t);
+
+	i = t.tv_sec;
 
 	if (tloc) {
 		if (put_user(i,tloc))

Re: [Patch] Fix sys_time to handle intra-tick correction

[john stultz]

On Tue, 2008-12-09 at 17:28 +0530, Vinay Sridhar wrote:
> Hi All,
> 
> This fix changes sys_time to use do_gettimeofday instead of get_seconds.
> Running the stime01 test from LTP triggers this error. Calling sys_stime
> and then calling sys_time causes this. "do_settimeofday" subtracts the
> nsec offset from the nsec value(0 in this case) passed to it.
> Subsequently, "set_normalized_timespec" modifies sec and nsec
> accordingly. This compensation is handled in the do_gettimeofday path.
> However, sys_time does not handle this case and reports an incorrect
> seconds value.
> 
> 
> signed-off by : Vinay Sridhar <vinay@linux.vnet.ibm.com>

This does fix the symmetry by using the same interfaces on both sides,
so it seems reasonable to me. However, since do_gettimeofday reads the
clocksource hardware, this may have a slight performance impact to
time() users.

I know privately I said this patch looked fine earlier, so sorry for the
mixed messages, but I really think we should be able to use
get_seconds() without this problem.

Looking more carefully at the code in do_settimeofday(), we should
accumulate the left over cycles, then we set the current time to the
specified time and update the xtime_cache.

So I'm not sure if we're really fixing the right issue here. Further,
running on the x86_64 system, I don't see this issue crop up at all (for
either 64 or 32 bit binaries). I think you mentioned privately that this
was seen on i386, so maybe we need to dig just a bit deeper before going
with this fix.

And sorry again for running hot then cold on this patch. I just didn't
look deeply enough at the issue the first time around. :(

thanks
-john




> --- time.c.orig	2008-12-04 06:30:12.000000000 -0600
> +++ time.c	2008-12-04 06:31:25.000000000 -0600
> @@ -61,7 +61,12 @@ EXPORT_SYMBOL(sys_tz);
>   */
>  asmlinkage long sys_time(time_t __user * tloc)
>  {
> -	time_t i = get_seconds();
> +	struct timeval t;	
> +	time_t i;
> +
> +	do_gettimeofday(&t);
> +
> +	i = t.tv_sec;
> 
>  	if (tloc) {
>  		if (put_user(i,tloc))
> 
> 

Re: [Patch] Fix sys_time to handle intra-tick correction

[john stultz]

On Tue, 2008-12-09 at 16:24 -0800, john stultz wrote:
> On Tue, 2008-12-09 at 17:28 +0530, Vinay Sridhar wrote:
> > Hi All,
> > 
> > This fix changes sys_time to use do_gettimeofday instead of get_seconds.
> > Running the stime01 test from LTP triggers this error. Calling sys_stime
> > and then calling sys_time causes this. "do_settimeofday" subtracts the
> > nsec offset from the nsec value(0 in this case) passed to it.
> > Subsequently, "set_normalized_timespec" modifies sec and nsec
> > accordingly. This compensation is handled in the do_gettimeofday path.
> > However, sys_time does not handle this case and reports an incorrect
> > seconds value.
> > 
> > 
> > signed-off by : Vinay Sridhar <vinay@linux.vnet.ibm.com>
> 
> This does fix the symmetry by using the same interfaces on both sides,
> so it seems reasonable to me. However, since do_gettimeofday reads the
> clocksource hardware, this may have a slight performance impact to
> time() users.
> 
> I know privately I said this patch looked fine earlier, so sorry for the
> mixed messages, but I really think we should be able to use
> get_seconds() without this problem.
> 
> Looking more carefully at the code in do_settimeofday(), we should
> accumulate the left over cycles, then we set the current time to the
> specified time and update the xtime_cache.
> 
> So I'm not sure if we're really fixing the right issue here. Further,
> running on the x86_64 system, I don't see this issue crop up at all (for
> either 64 or 32 bit binaries). I think you mentioned privately that this
> was seen on i386, so maybe we need to dig just a bit deeper before going
> with this fix.
> 
> And sorry again for running hot then cold on this patch. I just didn't
> look deeply enough at the issue the first time around. :(

Hey Vinay,
	So I spent some time checking various kernels and it seems this issue
is already resolved upstream by the clocksource_forward_now() change:
9a055117d3d9cb562f83f8d4cd88772761f4cab0

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=9a055117d3d9cb562f83f8d4cd88772761f4cab0

thanks
-john