Chinaunix

标题: 请教CFS调度器问题(续) [打印本页]

作者: donghaitad 时间: 2013-11-01 21:25
标题: 请教CFS调度器问题(续)
本帖最后由 donghaitad 于 2013-11-02 20:32 编辑

本人想通过修改CFS调度器问题来阻止某一个进程的运行。

内核版本：2.6.24
通过与大家的讨论和调试，目前修改代码如下：
其基本想法是：修改从就绪队列中选择下一个进程。

static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
{
struct sched_entity *se = NULL;

      struct task_struct *tp;

if (first_fair(cfs_rq)) {
se = __pick_next_entity(cfs_rq);
+             tp = task_of(se);
+             if(tp->pid == target_pid)//如果目标进程被选中
+             {
+                   __dequeue_entity(cfs_rq, se);//将该进程移出就绪队列
+                   if (first_fair(cfs_rq)) {//如果就绪队列不为空
+                         se = __pick_next_entity(cfs_rq);;//重新选择进程
+                         set_next_entity(cfs_rq, se);
+                   }
+                   else{//否则选择为空
+                         prinkt("No process is selected!\n");
+                         se = NULL;
+                   }
+             }
+             else
+          set_next_entity(cfs_rq, se);
-             set_next_entity(cfs_rq, se);
}

return se;
}

static struct task_struct *pick_next_task_fair(struct rq *rq)
{
struct cfs_rq *cfs_rq = &rq->cfs;
struct sched_entity *se;

if (unlikely(!cfs_rq->nr_running))
return NULL;

do {
se = pick_next_entity(cfs_rq);
+ if(se==NULL)
+ return NULL;
cfs_rq = group_cfs_rq(se);
} while (cfs_rq);

return task_of(se);
}

这段代码正常运行一段时间后("No process is selected!能够打印出来)，系统就卡住了（但鼠标好像还可以动）。
请大家指教！！！

不好意思重新编辑了一下。

Thanks a lot!

作者: firkraag 时间: 2013-11-02 08:14
本帖最后由 firkraag 于 2013-11-02 08:15 编辑

Pointer tp is not assigned.

What does this mean?
se = set_next_entity(cfs_rq, se);//重新选择进程

prinkt("No process is selected!\n");

作者: donghaitad 时间: 2013-11-02 11:59
回复 2# firkraag

不好意思，我重新编辑了一下。

tp = task_of(se);

+                   if (first_fair(cfs_rq)) {//如果就绪队列不为空
+                         se = __pick_next_entity(cfs_rq);;//重新选择进程
+                   }
+                   else{//否则选择为空
+                         prinkt("No process is selected!\n");
+                         se = NULL;
+                   }

作者: firkraag 时间: 2013-11-02 12:35
Should there be:
"set_next_entity(cfs_rq, se);"
after:
se = __pick_next_entity(cfs_rq);;//重新选择进程 ?

作者: donghaitad 时间: 2013-11-02 20:34
回复 4# firkraag

Thank you very much for your comments, I have added "set_next_entity(cfs_rq, se);"
after:
"se = __pick_next_entity(cfs_rq);"

but it still has the problem.

作者: donghaitad 时间: 2013-11-02 21:18
问题好像已经解决了：把 __dequeue_entity(cfs_rq, se)改成__pick_next_entity(cfs_rq)以后

好像就可以正常运行了。

作者: firkraag 时间: 2013-11-03 10:23
本帖最后由 firkraag 于 2013-11-03 16:45 编辑

Have the 'No process is selected!' ever been printed since the last modify?

Although '__dequeue_entity(cfs_rq, se)' just affects the rbtree and not modifys the data related with the rq and cfs_rq(like nr_running), it seems not a big problem.

Is there any process which can receive the input from user except the process you want to prevent from running?

作者: donghaitad 时间: 2013-11-03 17:31
本帖最后由 donghaitad 于 2013-11-03 19:20 编辑

'No process is selected!' is printed.

"Is there any process which can receive the input from user except the process you want to prevent from running?"

Yes.

但在单CPU环境下还有问题，好像产生死锁了。

作者: firkraag 时间: 2013-11-04 09:50
本帖最后由 firkraag 于 2013-11-04 10:07 编辑

回复 8# donghaitad

'No process is selected!' should not been printed if you replaced '__dequeue_entity(cfs_rq, se)' with '__pick_next_entity(cfs_rq)'.

'__pick_next_entity(cfs_rq)' just returns the se of cfs_rq->rb_leftmost node of cfs_rq and does not modify the rbtree.

Has the problem 'deadlock of printk' been really sloved?

How can you get the 'target_pid'?

作者: firkraag 时间: 2013-11-04 12:30
Maybe you can remove the printk and retry.

作者: donghaitad 时间: 2013-11-04 21:00
本帖最后由 donghaitad 于 2013-11-04 21:06 编辑

我去掉了printk, 但在单CPU模式下还是有问题。

系统卡住不动了，好像死锁了。

试了很多次，多CPU环境下好像没问题。

作者: firkraag 时间: 2013-11-08 11:08
Maybe you can try this:

Before "pick_next_entity" declare the dequeue_task_fair:

static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep);

Modify the pick_next_entity like this:

static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
{
struct sched_entity *se = NULL;
struct task_struct *tp;

if (first_fair(cfs_rq)) {
se = __pick_next_entity(cfs_rq);
if(!(se->my_q)){
tp = task_of(se);
if((tp->pid == target_pid)){
dequeue_task_fair(rq_of(cfs_rq),tp,1);
rq_of(cfs_rq)->nr_running--;
update_load_sub(&rq_of(cfs_rq)->load,se->load.weight);
if (first_fair(cfs_rq)){
se = __pick_next_entity(cfs_rq);
tp=task_of(se);
set_next_entity(cfs_rq, se);
}
else{
se = NULL;
}
}
else{
set_next_entity(cfs_rq, se);
}
}
else{
set_next_entity(cfs_rq, se);

}
}

return se;
}

Modify pick_next_task_fair as you've done.

作者: donghaitad 时间: 2013-11-17 19:07
果然在单CPU环境下也没有问题了，firkraag能帮忙解释一下：
rq_of(cfs_rq)->nr_running--;
update_load_sub(&rq_of(cfs_rq)->load,se->load.weight);
为什么是必须的？

作者: firkraag 时间: 2013-11-18 09:06
They are not necessary for your requirement.

作者: donghaitad 时间: 2013-11-20 21:15
再请教firkraag一个问题：

se->my_q代表调度实体下的就绪队列吧？

Thanks so much!

作者: firkraag 时间: 2013-11-21 09:02
Yes, it is.

欢迎光临 Chinaunix (http://bbs.chinaunix.net/)