sched/fair: Fix fixed point arithmetic width for shares and effective load

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Hello,

Please backport
commit ab522e33f91799661aad47bebb691f241a9f6bb8
('sched/fair: Fix fixed point arithmetic width for shares and effective load')
to 4.8.

It was apparently not backported as of 4.8.16, while it fixes a huge
performance regression in our tests, see the graphs between 19320.5 and
19451.5 on

http://starpu.gforge.inria.fr/testing/trunk/benchmarks/tasks_size_overhead_total_lws-200.png

which happened to be using a kernel without this fix.

FTR, here is the patch again.

Samuel

commit ab522e33f91799661aad47bebb691f241a9f6bb8
Author: Dietmar Eggemann <dietmar.eggemann@arm.com>
Date:   Mon Aug 22 15:00:41 2016 +0100

    sched/fair: Fix fixed point arithmetic width for shares and effective load
    
    Since commit:
    
      2159197d6677 ("sched/core: Enable increased load resolution on 64-bit kernels")
    
    we now have two different fixed point units for load:
    
    - 'shares' in calc_cfs_shares() has 20 bit fixed point unit on 64-bit
      kernels. Therefore use scale_load() on MIN_SHARES.
    
    - 'wl' in effective_load() has 10 bit fixed point unit. Therefore use
      scale_load_down() on tg->shares which has 20 bit fixed point unit on
      64-bit kernels.
    
    Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
    Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
    Cc: Linus Torvalds <torvalds@linux-foundation.org>
    Cc: Mike Galbraith <efault@gmx.de>
    Cc: Peter Zijlstra <peterz@infradead.org>
    Cc: Thomas Gleixner <tglx@linutronix.de>
    Link: http://lkml.kernel.org/r/1471874441-24701-1-git-send-email-dietmar.eggemann@arm.com
    Signed-off-by: Ingo Molnar <mingo@kernel.org>

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 8fb4d19..786ef94 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5017,9 +5017,9 @@ static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
 		 * wl = S * s'_i; see (2)
 		 */
 		if (W > 0 && w < W)
-			wl = (w * (long)tg->shares) / W;
+			wl = (w * (long)scale_load_down(tg->shares)) / W;
 		else
-			wl = tg->shares;
+			wl = scale_load_down(tg->shares);
 
 		/*
 		 * Per the above, wl is the new se->load.weight value; since