[v4] sched/rt: move back to RT_GROUP_SCHED and rename it child

  The member back in struct sched_rt_entity only related to RT_GROUP_SCHED,
it should not place out of RT_GROUP_SCHED, move back to RT_GROUP_SCHED
and rename it child.

Init child when parent isn't NULL in init_tg_rt_entry().

Introduce for_each_sched_rt_entity_reverse() to iterate rt_se from
top to down.

Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202308031034.4369b15b-oliver.sang@intel.com
---
 V3 -> V4: Missed rt_se = root in dequeue_rt_stack().
 V2 -> V3: Keep parent is NULL in init_tg_rt_entry().
 V1 -> V2: Add WARN_ON_ONCE in init_tg_rt_entry().
---
 include/linux/sched.h |  2 +-
 kernel/sched/rt.c     | 25 ++++++++++++++++---------
 2 files changed, 17 insertions(+), 10 deletions(-)
  

* Yajun Deng <yajun.deng@linux.dev> wrote:

> The member back in struct sched_rt_entity only related to RT_GROUP_SCHED,
> it should not place out of RT_GROUP_SCHED, move back to RT_GROUP_SCHED
> and rename it child.
> 
> Init child when parent isn't NULL in init_tg_rt_entry().
> 
> Introduce for_each_sched_rt_entity_reverse() to iterate rt_se from
> top to down.
> 
> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
> Reported-by: kernel test robot <oliver.sang@intel.com>
> Closes: https://lore.kernel.org/oe-lkp/202308031034.4369b15b-oliver.sang@intel.com

Yeah, so I agree with these changes, but could you please split up this 
patch into 3 separate patches:

   sched/rt: Move sched_rt_entity::back to under the CONFIG_RT_GROUP_SCHED block
   sched/rt: Rename sched_rt_entity::back to sched_rt_entity::child
   sched/rt: Introduce for_each_sched_rt_entity_reverse() & use it

Thanks,

	Ingo
  

On Thu, Aug 03, 2023 at 01:03:17PM +0800, Yajun Deng wrote:

> @@ -564,6 +566,9 @@ static inline struct task_group *next_task_group(struct task_group *tg)
>  #define for_each_sched_rt_entity(rt_se) \
>  	for (; rt_se; rt_se = rt_se->parent)
>  
> +#define for_each_sched_rt_entity_reverse(rt_se) \
> +	for (; rt_se; rt_se = rt_se->child)
> +
>  static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
>  {
>  	return rt_se->my_q;
> @@ -669,6 +674,9 @@ typedef struct rt_rq *rt_rq_iter_t;
>  #define for_each_sched_rt_entity(rt_se) \
>  	for (; rt_se; rt_se = NULL)
>  
> +#define for_each_sched_rt_entity_reverse(rt_se) \
> +	for_each_sched_rt_entity(rt_se)
> +
>  static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
>  {
>  	return NULL;
> @@ -1481,22 +1489,21 @@ static void __dequeue_rt_entity(struct sched_rt_entity *rt_se, unsigned int flag
>   */
>  static void dequeue_rt_stack(struct sched_rt_entity *rt_se, unsigned int flags)
>  {
> -	struct sched_rt_entity *back = NULL;
> +	struct sched_rt_entity *root;
>  	unsigned int rt_nr_running;
>  
> -	for_each_sched_rt_entity(rt_se) {
> -		rt_se->back = back;
> -		back = rt_se;
> -	}
> +	for_each_sched_rt_entity(rt_se)
> +		root = rt_se;
>  
> -	rt_nr_running = rt_rq_of_se(back)->rt_nr_running;
> +	rt_nr_running = rt_rq_of_se(root)->rt_nr_running;
>  
> -	for (rt_se = back; rt_se; rt_se = rt_se->back) {
> +	rt_se = root;
> +	for_each_sched_rt_entity_reverse(rt_se) {
>  		if (on_rt_rq(rt_se))
>  			__dequeue_rt_entity(rt_se, flags);
>  	}
>  
> -	dequeue_top_rt_rq(rt_rq_of_se(back), rt_nr_running);
> +	dequeue_top_rt_rq(rt_rq_of_se(root), rt_nr_running);
>  }

Urgh, please don't do this. The whole thing is super fragile. Creating
these abstractions makes it appear like
for_each_sched_rt_entity_reverse() is somehow a sane thing to do. Aside
from the name being insanely long, the whole thing makes no sense
what-so-ever unless you first went and did the normal iteration and set
the back pointers.

It's called back for a reason, it walks back the path it first walked.
Normals and unambiguous iteration is up the tree, towards the root,
walking back down is not.
  

On 2023/10/3 17:51, Ingo Molnar wrote:
> * Yajun Deng <yajun.deng@linux.dev> wrote:
>
>> The member back in struct sched_rt_entity only related to RT_GROUP_SCHED,
>> it should not place out of RT_GROUP_SCHED, move back to RT_GROUP_SCHED
>> and rename it child.
>>
>> Init child when parent isn't NULL in init_tg_rt_entry().
>>
>> Introduce for_each_sched_rt_entity_reverse() to iterate rt_se from
>> top to down.
>>
>> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
>> Reported-by: kernel test robot <oliver.sang@intel.com>
>> Closes: https://lore.kernel.org/oe-lkp/202308031034.4369b15b-oliver.sang@intel.com
> Yeah, so I agree with these changes, but could you please split up this
> patch into 3 separate patches:
>
>     sched/rt: Move sched_rt_entity::back to under the CONFIG_RT_GROUP_SCHED block
>     sched/rt: Rename sched_rt_entity::back to sched_rt_entity::child
>     sched/rt: Introduce for_each_sched_rt_entity_reverse() & use it

Okay, This patch makes sense, it saves a few bytes. But Peter seems to 
have a different opinion.

@Peter, I think I split up this patch into 2 separate patches:

          sched/rt: Introduce for_each_sched_rt_entity_back() & use it

         sched/rt: Move sched_rt_entity::back to under the 
CONFIG_RT_GROUP_SCHED block

Comments, please...

>
> Thanks,
>
> 	Ingo
  

On Tue, Oct 03, 2023 at 10:55:50PM +0800, Yajun Deng wrote:

> @Peter, I think I split up this patch into 2 separate patches:
> 
>          sched/rt: Introduce for_each_sched_rt_entity_back() & use it

Why ?? Having that macro makes no sense what so ever. You can't use it,
unless you set up the back pointers first. It is not a self contained
piece of functionality.

>         sched/rt: Move sched_rt_entity::back to under the
> CONFIG_RT_GROUP_SCHED block

This one sure.
  

On 2023/10/4 00:37, Peter Zijlstra wrote:
> On Tue, Oct 03, 2023 at 10:55:50PM +0800, Yajun Deng wrote:
>
>> @Peter, I think I split up this patch into 2 separate patches:
>>
>>           sched/rt: Introduce for_each_sched_rt_entity_back() & use it
> Why ?? Having that macro makes no sense what so ever. You can't use it,
> unless you set up the back pointers first. It is not a self contained
> piece of functionality.


The 2nd patch rely on this patch. It will be used in dequeue_rt_stack. 
We need to add a macro first,

because the 'back' will be not exist after the 2nd patch when disable 
CONFIG_RT_GROUP_SCHED.

diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 88fc98601413..0b6b4a715d6e 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -520,6 +520,9 @@ static inline struct task_group 
*next_task_group(struct task_group *tg)
  #define for_each_sched_rt_entity(rt_se) \
      for (; rt_se; rt_se = rt_se->parent)

+#define for_each_sched_rt_back(rt_se) \
+    for (; rt_se; rt_se = rt_se->back)
+
  static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
  {
      return rt_se->my_q;
@@ -625,6 +628,9 @@ typedef struct rt_rq *rt_rq_iter_t;
  #define for_each_sched_rt_entity(rt_se) \
      for (; rt_se; rt_se = NULL)

+#define for_each_sched_rt_entity_back(rt_se) \
+    for_each_sched_rt_entity(rt_se)
+
  static inline struct rt_rq *group_rt_rq(struct sched_rt_entity *rt_se)
  {
      return NULL;
@@ -1445,7 +1451,8 @@ static void dequeue_rt_stack(struct 
sched_rt_entity *rt_se, unsigned int flags)

      rt_nr_running = rt_rq_of_se(back)->rt_nr_running;

-    for (rt_se = back; rt_se; rt_se = rt_se->back) {
+    rt_se = back;
+    for_each_sched_rt_entity_back(rt_se) {
          if (on_rt_rq(rt_se))
              __dequeue_rt_entity(rt_se, flags);
      }


>
>>          sched/rt: Move sched_rt_entity::back to under the
>> CONFIG_RT_GROUP_SCHED block
> This one sure.