[RFC PATCH] sched: Do not copy user_cpus_ptr when parent is reset_on_fork

[RFC PATCH] sched: Do not copy user_cpus_ptr when parent is reset_on_fork

[Xuewen Yan]

Now, the task's user_cpus_ptr would dup from parent's user_cpus_ptr.
It is better reset the user_cpus_ptr when parent's reset_on_fork
is set.

Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
---
 kernel/sched/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index f3951e4a55e5..2fbae00cd1dc 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2666,7 +2666,7 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
 	 * do_set_cpus_allowed().
 	 */
 	raw_spin_lock_irqsave(&src->pi_lock, flags);
-	if (src->user_cpus_ptr) {
+	if (src->user_cpus_ptr && !src->sched_reset_on_fork) {
 		swap(dst->user_cpus_ptr, user_mask);
 		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
 	}
-- 
2.25.1

Re: [RFC PATCH] sched: Do not copy user_cpus_ptr when parent is reset_on_fork

[Waiman Long]

On 9/5/24 05:04, Xuewen Yan wrote:
> Now, the task's user_cpus_ptr would dup from parent's user_cpus_ptr.
> It is better reset the user_cpus_ptr when parent's reset_on_fork
> is set.

According to sched(7):

        Each thread has a reset-on-fork scheduling flag.  When this flag
        is set, children created by fork(2) do not inherit privileged
        scheduling policies.

It can be argued what are considered privileged scheduling policies. 
AFAICS, a restricted affinity doesn't seem to be a "privileged" 
scheduling policy. That is my own opinion strictly from the definition 
point of view, I will let others weigh in on that and I am OK to go 
either way.

Cheers,
Longman

>
> Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
> ---
>   kernel/sched/core.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index f3951e4a55e5..2fbae00cd1dc 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2666,7 +2666,7 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
>   	 * do_set_cpus_allowed().
>   	 */
>   	raw_spin_lock_irqsave(&src->pi_lock, flags);
> -	if (src->user_cpus_ptr) {
> +	if (src->user_cpus_ptr && !src->sched_reset_on_fork) {
>   		swap(dst->user_cpus_ptr, user_mask);
>   		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
>   	}

Re: [RFC PATCH] sched: Do not copy user_cpus_ptr when parent is reset_on_fork

[Phil Auld]

On Thu, Sep 05, 2024 at 08:42:33AM -0400 Waiman Long wrote:
> On 9/5/24 05:04, Xuewen Yan wrote:
> > Now, the task's user_cpus_ptr would dup from parent's user_cpus_ptr.
> > It is better reset the user_cpus_ptr when parent's reset_on_fork
> > is set.
> 
> According to sched(7):
> 
>        Each thread has a reset-on-fork scheduling flag.  When this flag
>        is set, children created by fork(2) do not inherit privileged
>        scheduling policies.
> 
> It can be argued what are considered privileged scheduling policies. AFAICS,
> a restricted affinity doesn't seem to be a "privileged" scheduling policy.
> That is my own opinion strictly from the definition point of view, I will
> let others weigh in on that and I am OK to go either way.
>

I think that one could argue that clearing a restricted affinity is
increasing the privilege and not preventing inheritence of same.
i.e. it would be the opposite of what reset-on-fork means.

I'd say NAK to this one if I had that power.

Cheers,
Phil

> Cheers,
> Longman
> 
> > 
> > Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
> > ---
> >   kernel/sched/core.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index f3951e4a55e5..2fbae00cd1dc 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -2666,7 +2666,7 @@ int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
> >   	 * do_set_cpus_allowed().
> >   	 */
> >   	raw_spin_lock_irqsave(&src->pi_lock, flags);
> > -	if (src->user_cpus_ptr) {
> > +	if (src->user_cpus_ptr && !src->sched_reset_on_fork) {
> >   		swap(dst->user_cpus_ptr, user_mask);
> >   		cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
> >   	}
> 
> 

-- 

Re: [RFC PATCH] sched: Do not copy user_cpus_ptr when parent is reset_on_fork

[Waiman Long]

On 9/5/24 09:12, Phil Auld wrote:
> On Thu, Sep 05, 2024 at 08:42:33AM -0400 Waiman Long wrote:
>> On 9/5/24 05:04, Xuewen Yan wrote:
>>> Now, the task's user_cpus_ptr would dup from parent's user_cpus_ptr.
>>> It is better reset the user_cpus_ptr when parent's reset_on_fork
>>> is set.
>> According to sched(7):
>>
>>         Each thread has a reset-on-fork scheduling flag.  When this flag
>>         is set, children created by fork(2) do not inherit privileged
>>         scheduling policies.
>>
>> It can be argued what are considered privileged scheduling policies. AFAICS,
>> a restricted affinity doesn't seem to be a "privileged" scheduling policy.
>> That is my own opinion strictly from the definition point of view, I will
>> let others weigh in on that and I am OK to go either way.
>>
> I think that one could argue that clearing a restricted affinity is
> increasing the privilege and not preventing inheritence of same.
> i.e. it would be the opposite of what reset-on-fork means.
>
> I'd say NAK to this one if I had that power.

Maybe I am not clear enough in my previous mail. My position is the same 
as yours. I think this patch is not necessary. More reasons should be 
provided as to why it is right to not inherited the restricted affinity 
when reset-on-fork flag is reset.

Cheers,
Longman

Re: [RFC PATCH] sched: Do not copy user_cpus_ptr when parent is reset_on_fork

[Phil Auld]

On Thu, Sep 05, 2024 at 10:00:36AM -0400 Waiman Long wrote:
> 
> On 9/5/24 09:12, Phil Auld wrote:
> > On Thu, Sep 05, 2024 at 08:42:33AM -0400 Waiman Long wrote:
> > > On 9/5/24 05:04, Xuewen Yan wrote:
> > > > Now, the task's user_cpus_ptr would dup from parent's user_cpus_ptr.
> > > > It is better reset the user_cpus_ptr when parent's reset_on_fork
> > > > is set.
> > > According to sched(7):
> > > 
> > >         Each thread has a reset-on-fork scheduling flag.  When this flag
> > >         is set, children created by fork(2) do not inherit privileged
> > >         scheduling policies.
> > > 
> > > It can be argued what are considered privileged scheduling policies. AFAICS,
> > > a restricted affinity doesn't seem to be a "privileged" scheduling policy.
> > > That is my own opinion strictly from the definition point of view, I will
> > > let others weigh in on that and I am OK to go either way.
> > > 
> > I think that one could argue that clearing a restricted affinity is
> > increasing the privilege and not preventing inheritence of same.
> > i.e. it would be the opposite of what reset-on-fork means.
> > 
> > I'd say NAK to this one if I had that power.
> 
> Maybe I am not clear enough in my previous mail. My position is the same as
> yours. I think this patch is not necessary. More reasons should be provided
> as to why it is right to not inherited the restricted affinity when
> reset-on-fork flag is reset.
>

No, you were clear. I was just providing my opinion, weighing in :)  

Cheers,
Phil

> Cheers,
> Longman
> 

--