Re: [PATCH v2 repost] sched: cputime: avoid multiplication overflow (in common cases)

[Stanislaw Gruszka <sgruszka@redhat.com>]

On Wed, Jan 09, 2013 at 07:33:03PM +0100, Frederic Weisbecker wrote:
> On Mon, Jan 07, 2013 at 12:31:45PM +0100, Stanislaw Gruszka wrote:
> > We scale stime, utime values based on rtime (sum_exec_runtime converted
> > to jiffies). During scaling we multiple rtime * utime, what seems to be
> > fine, since both values are converted to u64, but is not.
> > 
> > Let assume HZ is 1000 - 1ms tick. Process consist of 64 threads, run
> > for 1 day, threads utilize 100% cpu on user space. Machine has 64 cpus.
> > 
> > Process rtime = utime will be 64 * 24 * 60 * 60 * 1000 jiffies, what is
> > 0x149970000. Multiplication rtime * utime result is 0x1a855771100000000,
> > which can not be covered in 64 bits.
> > 
> > Result of overflow is stall of utime values visible in user space
> > (prev_utime in kernel), even if application still consume lot of CPU
> > time.
> > 
> > Probably good fix for the problem, will be using 128 bit variable and
> > proper mul128 and div_u128_u64 primitives. While mul128 is on it's
> > way to kernel, there is no 128 bit division yet. I'm not sure, if we
> > want to add it to kernel. Perhaps we could also change the way how
> > stime and utime are calculated, but I don't know how, so I come with
> > the below solution for the problem.
> > 
> > To avoid overflow patch change value we scale to min(stime, utime). This
> > is more like workaround, but will work for processes, which perform
> > mostly on user space or mostly on kernel space. Unfortunately processes,
> > which perform on kernel and user space equally, and additionally utilize
> > lot of CPU time, still will hit this overflow pretty quickly. However
> > such processes seems to be uncommon.
> > 
> > Signed-off-by: Stanislaw Gruszka <sgruszka@redhat.com>
> 
> I can easily imagine that overflow to happen with user time on intensive
> CPU bound loads, or may be guests.
> 
> But can we easily reach the same for system time? Even on intensive I/O bound
> loads we shouldn't spend that much time in the kernel. Most of it probably goes
> to idle.
> 
> What do you think?

I think you are right :-)
 
> If that assumption is right in most cases, the following patch should solve the
> issue:

I'm fine with this patch, it achives the same effect as my patch, but is simpler.

Thanks
Stanislaw

> diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
> index 293b202..0650dd4 100644
> --- a/kernel/sched/cputime.c
> +++ b/kernel/sched/cputime.c
> @@ -509,11 +509,11 @@ EXPORT_SYMBOL_GPL(vtime_account);
>  # define nsecs_to_cputime(__nsecs)	nsecs_to_jiffies(__nsecs)
>  #endif
>  
> -static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
> +static cputime_t scale_utime(cputime_t stime, cputime_t rtime, cputime_t total)
>  {
>  	u64 temp = (__force u64) rtime;
>  
> -	temp *= (__force u64) utime;
> +	temp *= (__force u64) stime;
>  
>  	if (sizeof(cputime_t) == 4)
>  		temp = div_u64(temp, (__force u32) total);
> @@ -531,10 +531,10 @@ static void cputime_adjust(struct task_cputime *curr,
>  			   struct cputime *prev,
>  			   cputime_t *ut, cputime_t *st)
>  {
> -	cputime_t rtime, utime, total;
> +	cputime_t rtime, stime, total;
>  
> -	utime = curr->utime;
> -	total = utime + curr->stime;
> +	stime = curr->stime;
> +	total = stime + curr->utime;
>  
>  	/*
>  	 * Tick based cputime accounting depend on random scheduling
> @@ -549,17 +549,17 @@ static void cputime_adjust(struct task_cputime *curr,
>  	rtime = nsecs_to_cputime(curr->sum_exec_runtime);
>  
>  	if (total)
> -		utime = scale_utime(utime, rtime, total);
> +		stime = scale_stime(stime, rtime, total);
>  	else
> -		utime = rtime;
> +		stime = rtime;
>  
>  	/*
>  	 * If the tick based count grows faster than the scheduler one,
>  	 * the result of the scaling may go backward.
>  	 * Let's enforce monotonicity.
>  	 */
> -	prev->utime = max(prev->utime, utime);
> -	prev->stime = max(prev->stime, rtime - prev->utime);
> +	prev->stime = max(prev->stime, stime);
> +	prev->utime = max(prev->utime, rtime - prev->stime);
>  
>  	*ut = prev->utime;
>  	*st = prev->stime;