Re: [PATCH] sched: Avoid division by zero - really

[Eric Dumazet <eric.dumazet@gmail.com>]

Peter Zijlstra a écrit :
> When re-computing the shares for each task group's cpu representation we
> need the ratio of weight on each cpu vs the total weight of the sched
> domain.
> 
> Since load-balancing is loosely (read not) synchronized, the weight of
> individual cpus can change between doing the sum and calculating the
> ratio.
> 
> The previous patch dealt with only one of the race scenarios, this patch
> side steps them all by saving a snapshot of all the individual cpu
> weights, thereby always working on a consistent set.
> 
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> ---
>  kernel/sched.c |   50 +++++++++++++++++++++++++++++---------------------
>  1 files changed, 29 insertions(+), 21 deletions(-)
> 
> diff --git a/kernel/sched.c b/kernel/sched.c
> index 0e76b17..4591054 100644
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -1515,30 +1515,29 @@ static unsigned long cpu_avg_load_per_task(int cpu)
>  
>  #ifdef CONFIG_FAIR_GROUP_SCHED
>  
> +struct update_shares_data {
> +	unsigned long rq_weight[NR_CPUS];
> +};
> +
> +static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);

ouch... thats quite large IMHO, up to 4096*8 = 32768 bytes per cpu...

Now we have nice dynamic per cpu allocations, we could use it here,
and use nr_cpus instead of NR_CPUS as the array size ?