Re: [patch] sched, cgroups: fix MIN_SHARES on 64 bit boxen

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Sat, 2011-06-04 at 11:29 +0200, Mike Galbraith wrote:
> c8b28116 claimed to have no user-visible effect, but allows setting cpu.shares
> to < MIN_SHARES, which the user then indeed sees.
> 
> Signed-off-by: Mike Galbraith <efault@gmx.de>
> ---
>  kernel/sched.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> Index: linux-2.6/kernel/sched.c
> ===================================================================
> --- linux-2.6.orig/kernel/sched.c
> +++ linux-2.6/kernel/sched.c
> @@ -292,7 +292,7 @@ static DEFINE_SPINLOCK(task_group_lock);
>   * (The default weight is 1024 - so there's no practical
>   *  limitation from this.)
>   */
> -#define MIN_SHARES	2
> +#define MIN_SHARES	(scale_load(2))
>  #define MAX_SHARES	(1UL << (18 + SCHED_LOAD_RESOLUTION))
>  
>  static int root_task_group_load = ROOT_TASK_GROUP_LOAD;

Hurm, but that destroys most of the gains from that patch, the whole
point was being able to have finer graunlarities, but now
calc_cfs_shares() and effective_load() are clipped the coarse
granularity.

So maybe explicitly change the MIN_SHARES usage in
sched_group_set_shares(). That wants to become a clamp user anyway,
something like:

shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));

That way MAX_SHARES can also loose its SCHED_LOAD_RESOLUTION factor and
bring is back in line with MIN_SHARES.