sched/fair: Make sure to try to detach at least one movable task
authorVincent Guittot <vincent.guittot@linaro.org>
Thu, 25 Aug 2022 12:27:23 +0000 (14:27 +0200)
committerPeter Zijlstra <peterz@infradead.org>
Thu, 15 Sep 2022 14:13:51 +0000 (16:13 +0200)
During load balance, we try at most env->loop_max time to move a task.
But it can happen that the loop_max LRU tasks (ie tail of
the cfs_tasks list) can't be moved to dst_cpu because of affinity.
In this case, loop in the list until we found at least one.

The maximum of detached tasks remained the same as before.

Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20220825122726.20819-2-vincent.guittot@linaro.org
kernel/sched/fair.c

index 4e5b171b11718cfd69500801cba6c8b56ebe4a9b..dae3bfae1bc0e958a54ec816b3f645f21bde6659 100644 (file)
@@ -8049,8 +8049,12 @@ static int detach_tasks(struct lb_env *env)
                p = list_last_entry(tasks, struct task_struct, se.group_node);
 
                env->loop++;
-               /* We've more or less seen every task there is, call it quits */
-               if (env->loop > env->loop_max)
+               /*
+                * We've more or less seen every task there is, call it quits
+                * unless we haven't found any movable task yet.
+                */
+               if (env->loop > env->loop_max &&
+                   !(env->flags & LBF_ALL_PINNED))
                        break;
 
                /* take a breather every nr_migrate tasks */
@@ -10179,7 +10183,9 @@ more_balance:
 
                if (env.flags & LBF_NEED_BREAK) {
                        env.flags &= ~LBF_NEED_BREAK;
-                       goto more_balance;
+                       /* Stop if we tried all running tasks */
+                       if (env.loop < busiest->nr_running)
+                               goto more_balance;
                }
 
                /*