Re: [PATCH] steal_account_process_tick() should return jiffies

[Frederic Weisbecker]

On Sat, Mar 05, 2016 at 11:27:01AM +0100, Thomas Gleixner wrote:
> Chris,
> 
> On Fri, 4 Mar 2016, Chris Friesen wrote:
> 
> First of all the subject line should contain a subsystem prefix,
> i.e. "sched/cputime:"
> 
> > The callers of steal_account_process_tick() expect it to return whether
> > the last jiffy was stolen or not.
> >     
> > Currently the return value of steal_account_process_tick() is in units
> > of cputime, which vary between either jiffies or nsecs depending on
> > CONFIG_VIRT_CPU_ACCOUNTING_GEN.
> 
> Sure, but what is the actual problem? The return value is boolean and tells
> whether there was stolen time accounted or not.
>      
> > The fix is to change steal_account_process_tick() to always return
> > jiffies.  If CONFIG_VIRT_CPU_ACCOUNTING_GEN is not enabled then this
> > is a no-op.
> 
> What does that fix? 
>  
> > As far as I can tell this bug has been present since commit dee08a72.
> 
> Which bug?
> 
> > Signed-off-by: Chris Friesen <chris.friesen@xxxxxxxxxxxxx>
> > ---
> > 
> >  kernel/sched/cputime.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
> > index b2ab2ff..e724496 100644
> > --- a/kernel/sched/cputime.c
> > +++ b/kernel/sched/cputime.c
> > @@ -276,7 +276,7 @@ static __always_inline bool steal_account_process_tick(void)
> >  		this_rq()->prev_steal_time += cputime_to_nsecs(steal_ct);
> >  
> >  		account_steal_time(steal_ct);
> > -		return steal_ct;
> > +		return cputime_to_jiffies(steal_ct);
> 
> So if steal time is close to a jiffie, then cputime_to_jiffies will return 0
> and you account a full jiffie to user/system/whatever.
> 
> Without a proper explanation of the problem and the resulting "bug" I really
> cannot figure out why we want that change.

Indeed the changelog should better explain the problem. So I think the issue is that
if the cputime has nsecs granularity and we have a tiny stolen time to account (lets say
a few nanosecs, in fact anything that is below a jiffy), we are not going to account the
tick on user/system.

But the fix doesn't look right to me because we are still accounting the steal time
if it is lower than a jiffy and that steal time will never be substracted to user/system
time if it never reach a jiffy.

Instead the fix should accumulate the steal time and account it only once it's worth
a jiffy and then substract it from system/user time accordingly. Something like that:

diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
index b2ab2ff..d38e25f 100644
--- a/kernel/sched/cputime.c
+++ b/kernel/sched/cputime.c
@@ -262,7 +262,7 @@ static __always_inline bool steal_account_process_tick(void)
 #ifdef CONFIG_PARAVIRT
 	if (static_key_false(&paravirt_steal_enabled)) {
 		u64 steal;
-		cputime_t steal_ct;
+		unsigned long steal_jiffies;
 
 		steal = paravirt_steal_clock(smp_processor_id());
 		steal -= this_rq()->prev_steal_time;
@@ -272,11 +272,11 @@ static __always_inline bool steal_account_process_tick(void)
 		 * based on jiffies). Lets cast the result to cputime
 		 * granularity and account the rest on the next rounds.
 		 */
-		steal_ct = nsecs_to_cputime(steal);
-		this_rq()->prev_steal_time += cputime_to_nsecs(steal_ct);
+		steal_jiffies = nsecs_to_jiffies(steal);
+		this_rq()->prev_steal_time += jiffies_to_nsecs(steal_jiffies);
 
 		account_steal_time(steal_ct);
-		return steal_ct;
+		return steal_jiffies;
 	}
 #endif
 	return false;