On 16 December 2016 at 09:55, Vincent Guittot
[off-list ref] wrote:

On 15 December 2016 at 22:42, Peter Zijlstra [off-list ref] wrote:

quoted

On Thu, Dec 01, 2016 at 05:38:53PM +0100, Vincent Guittot wrote:

quoted

The update of the share of a cfs_rq is done when its load_avg is updated
but before the group_entity's load_avg has been updated for the past time
slot. This generates wrong load_avg accounting which can be significant
when small tasks are involved in the scheduling.

Let take the example of a task TA that is dequeued of its task group TG1.
TA was the only task in TG1 which becomes idle.

We have the sequence:

- dequeue_entity TA->se
    - update_load_avg(TA->se)
    - dequeue_entity_load_avg(TG1->cfs_rq, TA->se)
    - account_entity_dequeue(TG1->cfs_rq, TA->se)
          TG1->cfs_rq->load.weight = 0
    - update_cfs_shares(TG1->cfs_rq)
              TG1->se->load.weight is updated with the new share of
              cfs_rq. TG1->se->load.weight = 0.
- dequeue_entity TG1->se
    - update_load_avg(TG1->se) but its weight is now null so the last time
slot (up to a tick) will be accounted with its new weight (0 in our case)
instead of its real weight. The last time slot is accounted as an idle one
whereas it was a running one.

If the running time of TA is short enough that no tick happens when it
runs, all running time of TG1->se will be accounted as idle time.

Instead, we should update the share of a cfs_rq (in fact the weight of its
group entity) only after having updated the load_avg of the group_entity.

update_cfs_shares() now takes the sched_entity as parameter instead of the
cfs_rq and the weight of the group_entity is updated only once its load_avg
has been synced with current time.

Urgh, brain hurt, also those names don't help; s/TG1/A/ s/TA/a/

So the problem is that in our for_each_sched_entity(se) loop we end up
changing the next se before we get there.


                root
              (cfs_rq)
                  \
                  (se)
                    A
                 (cfs_rq)
                      \
                      (se)
                       a


Starting at a's se, we update_cfs_shares() on A's cfs_rq, which then
updates A's se, which is the next se in our iteration and mucks with
state before we get there.

So you change update_cfs_shares() to go downward while we go upward,
ensuring we only update things that we've finished with.

yes

quoted

Makes sense..

quoted

 kernel/sched/fair.c | 27 ++++++++++++++++-----------
 1 file changed, 16 insertions(+), 11 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 18d9e75..19092fa 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c

@@ -2689,15 +2689,18 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,

 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);

-static void update_cfs_shares(struct cfs_rq *cfs_rq)
+static void update_cfs_shares(struct sched_entity *se)
 {
      struct task_group *tg;
-     struct sched_entity *se;
+     struct cfs_rq *cfs_rq = group_cfs_rq(se);
      long shares;

please keep them ordered by length.

Ok

quoted

quoted

+     if (entity_is_task(se))

can be: !cfs_rq, which is the same and we already done that load.

yes. My goal was to keep it more readable about the meaning of the
test and I was expecting that the compiler would be smart enough to
use the same one load for both cfs_rq = group_cfs_rq(se) and
entity_is_task(se)

I can change with !cfs_rq

quoted

quoted

+             return;
+
      tg = cfs_rq->tg;

This load isn't needed here yet, can be moved down a bit.

Indeed

quoted

quoted

-     se = tg->se[cpu_of(rq_of(cfs_rq))];
-     if (!se || throttled_hierarchy(cfs_rq))
+
+     if (throttled_hierarchy(cfs_rq))
              return;
 #ifndef CONFIG_SMP
      if (likely(se->load.weight == tg->shares))

quoted

@@ -3583,9 +3588,9 @@ enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
              se->vruntime += cfs_rq->min_vruntime;

      update_load_avg(se, UPDATE_TG);
+     update_cfs_shares(se);
      enqueue_entity_load_avg(cfs_rq, se);
      account_entity_enqueue(cfs_rq, se);
-     update_cfs_shares(cfs_rq);

      if (flags & ENQUEUE_WAKEUP)
              place_entity(cfs_rq, se, 0);

So here we need to update_cfs_shares() _before_ enqueue_entity, because
the update_cfs_shares() will affect this se's load, right?

exactly

In fact, the only constraint is that update_cfs_shares() must be done
before account_entity_enqueue(). But there no constraint with
enqueue_entity_load_avg() so it's probably better to put manipulation
of load together and manipulation of weight together:

update_load_avg(se, UPDATE_TG);
enqueue_entity_load_avg(cfs_rq, se);
update_cfs_shares(se);
account_entity_enqueue(cfs_rq, se);

quoted

quoted

@@ -3681,7 +3686,7 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
      /* return excess runtime on last dequeue */
      return_cfs_rq_runtime(cfs_rq);

-     update_cfs_shares(cfs_rq);
+     update_cfs_shares(se);

      /*
       * Now advance min_vruntime if @se was the entity holding it back,

But this one hurts my brain..

It must be done after dequeue_entity_load_avg() such that we subtract
the load as was seen until now.

 update_cfs_shares(A's se) must be done after update_load_avg(A's se,
UPDATE_TG); so the update od A's se ->load-avg will be updated with
the previous load to update load_avg for the previous time slot.

update_cfs_shares(A's se) could be done before or after
dequeue_entity_load_avg(A's se) because the root's cfs_rq is kept sync
during the reweight of A's se. Nevertheless, i see one advantage of
doing that after: reweight_entity will be faster because A's se->on_rq
will have been cleared in the meantime

quoted

Could we please add comments explaining this ordering, because I forever
need to think about this (both enqueue and dequeue).

OK

quoted

quoted

@@ -3864,7 +3869,7 @@ entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
       * Ensure that runnable average is periodically updated.
       */
      update_load_avg(curr, UPDATE_TG);
-     update_cfs_shares(cfs_rq);
+     update_cfs_shares(curr);

 #ifdef CONFIG_SCHED_HRTICK
      /*

@@ -4761,7 +4766,7 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
                      break;

              update_load_avg(se, UPDATE_TG);
-             update_cfs_shares(cfs_rq);
+             update_cfs_shares(se);
      }

      if (!se)

@@ -4820,7 +4825,7 @@ static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
                      break;

              update_load_avg(se, UPDATE_TG);
-             update_cfs_shares(cfs_rq);
+             update_cfs_shares(se);
      }

      if (!se)

This has a distinct pattern to it though; should we think about
something like: UPDATE_SHARES for update_load_avg() or does that confuse
things?

IMHO, keeping update_cfs_shares separated from update_load_avg make it
clear about when we update the shares and enable some optimization
like for  dequeue_entity

quoted

quoted

@@ -9316,7 +9321,7 @@ int sched_group_set_shares(struct task_group *tg, unsigned long shares)
              /* Possible calls to update_curr() need rq clock */
              update_rq_clock(rq);
              for_each_sched_entity(se)
-                     update_cfs_shares(group_cfs_rq(se));
+                     update_cfs_shares(se);

Should we not also catch up with our load before we frob the shares?

yes you're right, an update_load_avg is missing

quoted

quoted

              raw_spin_unlock_irqrestore(&rq->lock, flags);
      }