lkml.org 
[lkml]   [2008]   [Feb]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [RFC, PATCH 1/2] sched: change the fairness model of the CFS group scheduler
http://lkml.org/lkml/2008/2/28/313 has a lot of new bugs.

> +++ linux-2.6.25-rc2/kernel/sched_fair.c

> + struct sched_entity *se = tg->se[cpu], *top_se;

> +
> + for_each_sched_entity(se)
> + top_se = se;
> +

It does unnecesary code execution.
Why don't do once "top_se = se;" instead for each?

> +static inline int depth_se(struct sched_entity *se)
> +{
> + int depth = 0;
> +
> + for_each_sched_entity(se)
> + depth++;
> +
> + return depth;
> +}
> +

It does unnecesary code execution.
Why don't do "depth = the count of sched entities" instead of ++?

> + /* First walk up until both entities are at same depth */
> + se_depth = depth_se(se);
> + pse_depth = depth_se(pse);
> +
> + while (se_depth > pse_depth) {
> + se_depth--;
> + se = parent_entity(se);
> + }
> +
> + while (pse_depth > se_depth) {
> + pse_depth--;
> + pse = parent_entity(pse);
> + }

It does unnecesary code execution.
Exiting the 1st while asserts "se_depth == pse_depth",
the 2nd while never is entered.

> __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)

> - p = rb_entry(curr, struct task_struct, se.run_node);
> - cfs_rq->rb_load_balance_curr = rb_next(curr);
> + /* Skip over entities that are not tasks */
> + do {
> + se = rb_entry(curr, struct sched_entity, run_node);
> + curr = rb_next(curr);
> + } while (curr && !entity_is_task(se));

It's wrong, it says that curr is next to the first found entity that is task.

I think that the correct code could be
curr = rb_next(curr);
while (curr) {
se = rb_entry(curr, struct sched_entity, run_node);
if (entity_is_task(se)) break;
curr = rb_next(curr);
} /* se is next to curr and first entity that is task skipping
non-task entities */

> + cfs_rq->rb_load_balance_curr = curr;
> +
> + if (entity_is_task(se))
> + p = task_of(se);
>
> return p;

You forgot if (unlikely(curr == NULL)) { ... }

> @@ -1210,21 +1267,28 @@ load_balance_fair(struct rq *this_rq, in

> unsigned long maxload, task_load, group_weight;

> - group_weight = se->load.weight;
> + if (!task_load)
> + continue;
> +
> + group_weight = group_cpu_load(tg, busiest->cpu);

> - * 'group_weight' is contributed by tasks of total weight
> + * 'group_weight' is contributed by entities of total weight

> * maxload = (remload / group_weight) * task_load;
> */
> maxload = (rem_load_move * task_load) / group_weight;

The unsigned long division IS SLOW!!! Use multiplication and shift stupid!!!
Be careful with zerodiv exception of "/ group_weight"!!!

> - /*
> - * load_moved holds the task load that was moved. The
> - * effective (group) weight moved would be:
> - * load_moved_eff = load_moved/task_load * group_weight;
> - */
> - load_moved = (group_weight * load_moved) / task_load;
> -
> /* Adjust shares on both cpus to reflect load_moved */
> - group_weight -= load_moved;
> - set_se_shares(se, group_weight);
> + if (likely(se)) {
> + unsigned long load_moved_eff;
> + unsigned long se_shares;
>
> - se = busy_cfs_rq->tg->se[this_cpu];
> - if (!thisload)
> - group_weight = load_moved;
> - else
> - group_weight = se->load.weight + load_moved;
> - set_se_shares(se, group_weight);
> + /*
> + * load_moved holds the task load that was moved. The
> + * effective (group) weight moved would be:
> + * load_moved_eff = load_moved/task_load *
> + * group_weight;
> + */
> + load_moved_eff = (se->load.weight *
> + load_moved) / task_load;
> +
> + set_se_shares(se, se->load.weight - load_moved_eff);
> + if (!thisload)
> + se_shares = load_moved_eff;
> + else
> + se_shares = this_se->load.weight +
> + load_moved_eff;
> + set_se_shares(this_se, se_shares);
> + }

The unsigned long division IS SLOW!!! Use multiplication and shift stupid!!!
Be careful with zerodiv exception of "/ task_load"!!!

J.C.Pizarro


\
 
 \ /
  Last update: 2008-02-29 01:47    [W:0.026 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site