diff mbox series

[v6,2/3] sched/task: Add the put_task_struct_atomic_safe() function

Message ID 20230414125532.14958-3-wander@redhat.com
State New
Headers
Series Introduce put_task_struct_atomic_sleep() |

Commit Message

Wander Lairson Costa April 14, 2023, 12:55 p.m. UTC 
  Due to the possibility of indirectly acquiring sleeping locks, it is
unsafe to call put_task_struct() in atomic contexts when the kernel is
compiled with PREEMPT_RT.

To mitigate this issue, this commit introduces
put_task_struct_atomic_safe(), which schedules __put_task_struct()
through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
be a more natural approach, we cannot allocate dynamic memory from
atomic context in PREEMPT_RT, making the code more complex.

This implementation ensures safe execution in atomic contexts and
avoids any potential issues that may arise from using the non-atomic
version.

Signed-off-by: Wander Lairson Costa <wander@redhat.com>
Reported-by: Hu Chunyu <chuhu@redhat.com>
Cc: Paul McKenney <paulmck@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
---
 include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
 kernel/fork.c              |  8 ++++++++
 2 files changed, 39 insertions(+)

Comments

Waiman Long April 17, 2023, 6:51 p.m. UTC | #1 
On 4/14/23 08:55, Wander Lairson Costa wrote:
> Due to the possibility of indirectly acquiring sleeping locks, it is
> unsafe to call put_task_struct() in atomic contexts when the kernel is
> compiled with PREEMPT_RT.
>
> To mitigate this issue, this commit introduces
> put_task_struct_atomic_safe(), which schedules __put_task_struct()
> through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> be a more natural approach, we cannot allocate dynamic memory from
> atomic context in PREEMPT_RT, making the code more complex.
>
> This implementation ensures safe execution in atomic contexts and
> avoids any potential issues that may arise from using the non-atomic
> version.
>
> Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> Reported-by: Hu Chunyu <chuhu@redhat.com>
> Cc: Paul McKenney <paulmck@kernel.org>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> ---
>   include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
>   kernel/fork.c              |  8 ++++++++
>   2 files changed, 39 insertions(+)
>
> diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> index b597b97b1f8f..5c13b83d7008 100644
> --- a/include/linux/sched/task.h
> +++ b/include/linux/sched/task.h
> @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
>   
>   void put_task_struct_rcu_user(struct task_struct *task);
>   
> +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> +
> +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> +{
> +	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> +		/*
> +		 * Decrement the refcount explicitly to avoid unnecessarily
> +		 * calling call_rcu.
> +		 */
> +		if (refcount_dec_and_test(&task->usage))
> +			/*
> +			 * under PREEMPT_RT, we can't call put_task_struct
> +			 * in atomic context because it will indirectly
> +			 * acquire sleeping locks.
> +			 * call_rcu() will schedule delayed_put_task_struct_rcu()
delayed_put_task_struct_rcu()?
> +			 * to be called in process context.
> +			 *
> +			 * __put_task_struct() is called called when
"called called"?
> +			 * refcount_dec_and_test(&t->usage) succeeds.
> +			 *
> +			 * This means that it can't "conflict" with
> +			 * put_task_struct_rcu_user() which abuses ->rcu the same
> +			 * way; rcu_users has a reference so task->usage can't be
> +			 * zero after rcu_users 1 -> 0 transition.

Note that put_task_struct_rcu_user() isn't the only user of task->rcu. 
delayed_free_task() in kernel/fork.c also uses it, though it is only 
called in the error case. Still you may need to take a look to make sure 
that there is no conflict.

Cheers,
Longman
Wander Lairson Costa April 18, 2023, 2:18 p.m. UTC | #2 
On Mon, Apr 17, 2023 at 02:51:45PM -0400, Waiman Long wrote:
> 
> On 4/14/23 08:55, Wander Lairson Costa wrote:
> > Due to the possibility of indirectly acquiring sleeping locks, it is
> > unsafe to call put_task_struct() in atomic contexts when the kernel is
> > compiled with PREEMPT_RT.
> > 
> > To mitigate this issue, this commit introduces
> > put_task_struct_atomic_safe(), which schedules __put_task_struct()
> > through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> > be a more natural approach, we cannot allocate dynamic memory from
> > atomic context in PREEMPT_RT, making the code more complex.
> > 
> > This implementation ensures safe execution in atomic contexts and
> > avoids any potential issues that may arise from using the non-atomic
> > version.
> > 
> > Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> > Reported-by: Hu Chunyu <chuhu@redhat.com>
> > Cc: Paul McKenney <paulmck@kernel.org>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > ---
> >   include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
> >   kernel/fork.c              |  8 ++++++++
> >   2 files changed, 39 insertions(+)
> > 
> > diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> > index b597b97b1f8f..5c13b83d7008 100644
> > --- a/include/linux/sched/task.h
> > +++ b/include/linux/sched/task.h
> > @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
> >   void put_task_struct_rcu_user(struct task_struct *task);
> > +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> > +
> > +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> > +{
> > +	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> > +		/*
> > +		 * Decrement the refcount explicitly to avoid unnecessarily
> > +		 * calling call_rcu.
> > +		 */
> > +		if (refcount_dec_and_test(&task->usage))
> > +			/*
> > +			 * under PREEMPT_RT, we can't call put_task_struct
> > +			 * in atomic context because it will indirectly
> > +			 * acquire sleeping locks.
> > +			 * call_rcu() will schedule delayed_put_task_struct_rcu()
> delayed_put_task_struct_rcu()?
> > +			 * to be called in process context.
> > +			 *
> > +			 * __put_task_struct() is called called when
> "called called"?
> > +			 * refcount_dec_and_test(&t->usage) succeeds.
> > +			 *
> > +			 * This means that it can't "conflict" with
> > +			 * put_task_struct_rcu_user() which abuses ->rcu the same
> > +			 * way; rcu_users has a reference so task->usage can't be
> > +			 * zero after rcu_users 1 -> 0 transition.
> 
> Note that put_task_struct_rcu_user() isn't the only user of task->rcu.
> delayed_free_task() in kernel/fork.c also uses it, though it is only called
> in the error case. Still you may need to take a look to make sure that there
> is no conflict.
> 

delayed_free_task() is called when a process fails to start. Therefore, AFAICT,
there is no way it can conflict with put_task_struct().

> Cheers,
> Longman
>
Waiman Long April 18, 2023, 2:26 p.m. UTC | #3 
On 4/18/23 10:18, Wander Lairson Costa wrote:
> On Mon, Apr 17, 2023 at 02:51:45PM -0400, Waiman Long wrote:
>> On 4/14/23 08:55, Wander Lairson Costa wrote:
>>> Due to the possibility of indirectly acquiring sleeping locks, it is
>>> unsafe to call put_task_struct() in atomic contexts when the kernel is
>>> compiled with PREEMPT_RT.
>>>
>>> To mitigate this issue, this commit introduces
>>> put_task_struct_atomic_safe(), which schedules __put_task_struct()
>>> through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
>>> be a more natural approach, we cannot allocate dynamic memory from
>>> atomic context in PREEMPT_RT, making the code more complex.
>>>
>>> This implementation ensures safe execution in atomic contexts and
>>> avoids any potential issues that may arise from using the non-atomic
>>> version.
>>>
>>> Signed-off-by: Wander Lairson Costa <wander@redhat.com>
>>> Reported-by: Hu Chunyu <chuhu@redhat.com>
>>> Cc: Paul McKenney <paulmck@kernel.org>
>>> Cc: Thomas Gleixner <tglx@linutronix.de>
>>> ---
>>>    include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
>>>    kernel/fork.c              |  8 ++++++++
>>>    2 files changed, 39 insertions(+)
>>>
>>> diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
>>> index b597b97b1f8f..5c13b83d7008 100644
>>> --- a/include/linux/sched/task.h
>>> +++ b/include/linux/sched/task.h
>>> @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
>>>    void put_task_struct_rcu_user(struct task_struct *task);
>>> +extern void __delayed_put_task_struct(struct rcu_head *rhp);
>>> +
>>> +static inline void put_task_struct_atomic_safe(struct task_struct *task)
>>> +{
>>> +	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
>>> +		/*
>>> +		 * Decrement the refcount explicitly to avoid unnecessarily
>>> +		 * calling call_rcu.
>>> +		 */
>>> +		if (refcount_dec_and_test(&task->usage))
>>> +			/*
>>> +			 * under PREEMPT_RT, we can't call put_task_struct
>>> +			 * in atomic context because it will indirectly
>>> +			 * acquire sleeping locks.
>>> +			 * call_rcu() will schedule delayed_put_task_struct_rcu()
>> delayed_put_task_struct_rcu()?
>>> +			 * to be called in process context.
>>> +			 *
>>> +			 * __put_task_struct() is called called when
>> "called called"?
>>> +			 * refcount_dec_and_test(&t->usage) succeeds.
>>> +			 *
>>> +			 * This means that it can't "conflict" with
>>> +			 * put_task_struct_rcu_user() which abuses ->rcu the same
>>> +			 * way; rcu_users has a reference so task->usage can't be
>>> +			 * zero after rcu_users 1 -> 0 transition.
>> Note that put_task_struct_rcu_user() isn't the only user of task->rcu.
>> delayed_free_task() in kernel/fork.c also uses it, though it is only called
>> in the error case. Still you may need to take a look to make sure that there
>> is no conflict.
>>
> delayed_free_task() is called when a process fails to start. Therefore, AFAICT,
> there is no way it can conflict with put_task_struct().

I think so too, but for completeness, you should document somewhere that 
it is a possible conflicting user.

Cheers,
Longman
Paul E. McKenney April 24, 2023, 6:09 p.m. UTC | #4 
On Fri, Apr 14, 2023 at 09:55:28AM -0300, Wander Lairson Costa wrote:
> Due to the possibility of indirectly acquiring sleeping locks, it is
> unsafe to call put_task_struct() in atomic contexts when the kernel is
> compiled with PREEMPT_RT.
> 
> To mitigate this issue, this commit introduces
> put_task_struct_atomic_safe(), which schedules __put_task_struct()
> through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> be a more natural approach, we cannot allocate dynamic memory from
> atomic context in PREEMPT_RT, making the code more complex.
> 
> This implementation ensures safe execution in atomic contexts and
> avoids any potential issues that may arise from using the non-atomic
> version.
> 
> Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> Reported-by: Hu Chunyu <chuhu@redhat.com>
> Cc: Paul McKenney <paulmck@kernel.org>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> ---
>  include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
>  kernel/fork.c              |  8 ++++++++
>  2 files changed, 39 insertions(+)
> 
> diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> index b597b97b1f8f..5c13b83d7008 100644
> --- a/include/linux/sched/task.h
> +++ b/include/linux/sched/task.h
> @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
>  
>  void put_task_struct_rcu_user(struct task_struct *task);
>  
> +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> +
> +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> +{
> +	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> +		/*
> +		 * Decrement the refcount explicitly to avoid unnecessarily
> +		 * calling call_rcu.
> +		 */
> +		if (refcount_dec_and_test(&task->usage))
> +			/*
> +			 * under PREEMPT_RT, we can't call put_task_struct
> +			 * in atomic context because it will indirectly
> +			 * acquire sleeping locks.
> +			 * call_rcu() will schedule delayed_put_task_struct_rcu()
> +			 * to be called in process context.
> +			 *
> +			 * __put_task_struct() is called called when
> +			 * refcount_dec_and_test(&t->usage) succeeds.
> +			 *
> +			 * This means that it can't "conflict" with
> +			 * put_task_struct_rcu_user() which abuses ->rcu the same
> +			 * way; rcu_users has a reference so task->usage can't be
> +			 * zero after rcu_users 1 -> 0 transition.
> +			 */
> +			call_rcu(&task->rcu, __delayed_put_task_struct);

This will invoke __delayed_put_task_struct() with softirqs disabled.
Or do softirq-disabled contexts count as non-atomic in PREEMPT_RT?

							Thanx, Paul

> +	} else {
> +		put_task_struct(task);
> +	}
> +}
> +
>  /* Free all architecture-specific resources held by a thread. */
>  void release_thread(struct task_struct *dead_task);
>  
> diff --git a/kernel/fork.c b/kernel/fork.c
> index 0c92f224c68c..9884794fe4b8 100644
> --- a/kernel/fork.c
> +++ b/kernel/fork.c
> @@ -854,6 +854,14 @@ void __put_task_struct(struct task_struct *tsk)
>  }
>  EXPORT_SYMBOL_GPL(__put_task_struct);
>  
> +void __delayed_put_task_struct(struct rcu_head *rhp)
> +{
> +	struct task_struct *task = container_of(rhp, struct task_struct, rcu);
> +
> +	__put_task_struct(task);
> +}
> +EXPORT_SYMBOL_GPL(__delayed_put_task_struct);
> +
>  void __init __weak arch_task_cache_init(void) { }
>  
>  /*
> -- 
> 2.39.2
>
Wander Lairson Costa April 24, 2023, 6:43 p.m. UTC | #5 
On Mon, Apr 24, 2023 at 3:09 PM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> On Fri, Apr 14, 2023 at 09:55:28AM -0300, Wander Lairson Costa wrote:
> > Due to the possibility of indirectly acquiring sleeping locks, it is
> > unsafe to call put_task_struct() in atomic contexts when the kernel is
> > compiled with PREEMPT_RT.
> >
> > To mitigate this issue, this commit introduces
> > put_task_struct_atomic_safe(), which schedules __put_task_struct()
> > through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> > be a more natural approach, we cannot allocate dynamic memory from
> > atomic context in PREEMPT_RT, making the code more complex.
> >
> > This implementation ensures safe execution in atomic contexts and
> > avoids any potential issues that may arise from using the non-atomic
> > version.
> >
> > Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> > Reported-by: Hu Chunyu <chuhu@redhat.com>
> > Cc: Paul McKenney <paulmck@kernel.org>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > ---
> >  include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
> >  kernel/fork.c              |  8 ++++++++
> >  2 files changed, 39 insertions(+)
> >
> > diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> > index b597b97b1f8f..5c13b83d7008 100644
> > --- a/include/linux/sched/task.h
> > +++ b/include/linux/sched/task.h
> > @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
> >
> >  void put_task_struct_rcu_user(struct task_struct *task);
> >
> > +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> > +
> > +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> > +{
> > +     if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> > +             /*
> > +              * Decrement the refcount explicitly to avoid unnecessarily
> > +              * calling call_rcu.
> > +              */
> > +             if (refcount_dec_and_test(&task->usage))
> > +                     /*
> > +                      * under PREEMPT_RT, we can't call put_task_struct
> > +                      * in atomic context because it will indirectly
> > +                      * acquire sleeping locks.
> > +                      * call_rcu() will schedule delayed_put_task_struct_rcu()
> > +                      * to be called in process context.
> > +                      *
> > +                      * __put_task_struct() is called called when
> > +                      * refcount_dec_and_test(&t->usage) succeeds.
> > +                      *
> > +                      * This means that it can't "conflict" with
> > +                      * put_task_struct_rcu_user() which abuses ->rcu the same
> > +                      * way; rcu_users has a reference so task->usage can't be
> > +                      * zero after rcu_users 1 -> 0 transition.
> > +                      */
> > +                     call_rcu(&task->rcu, __delayed_put_task_struct);
>
> This will invoke __delayed_put_task_struct() with softirqs disabled.
> Or do softirq-disabled contexts count as non-atomic in PREEMPT_RT?
>

softirqs execute in thread context in PREEMPT_RT. We are good here.

>                                                         Thanx, Paul
>
> > +     } else {
> > +             put_task_struct(task);
> > +     }
> > +}
> > +
> >  /* Free all architecture-specific resources held by a thread. */
> >  void release_thread(struct task_struct *dead_task);
> >
> > diff --git a/kernel/fork.c b/kernel/fork.c
> > index 0c92f224c68c..9884794fe4b8 100644
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -854,6 +854,14 @@ void __put_task_struct(struct task_struct *tsk)
> >  }
> >  EXPORT_SYMBOL_GPL(__put_task_struct);
> >
> > +void __delayed_put_task_struct(struct rcu_head *rhp)
> > +{
> > +     struct task_struct *task = container_of(rhp, struct task_struct, rcu);
> > +
> > +     __put_task_struct(task);
> > +}
> > +EXPORT_SYMBOL_GPL(__delayed_put_task_struct);
> > +
> >  void __init __weak arch_task_cache_init(void) { }
> >
> >  /*
> > --
> > 2.39.2
> >
>
Paul E. McKenney April 24, 2023, 6:52 p.m. UTC | #6 
On Mon, Apr 24, 2023 at 03:43:09PM -0300, Wander Lairson Costa wrote:
> On Mon, Apr 24, 2023 at 3:09 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > On Fri, Apr 14, 2023 at 09:55:28AM -0300, Wander Lairson Costa wrote:
> > > Due to the possibility of indirectly acquiring sleeping locks, it is
> > > unsafe to call put_task_struct() in atomic contexts when the kernel is
> > > compiled with PREEMPT_RT.
> > >
> > > To mitigate this issue, this commit introduces
> > > put_task_struct_atomic_safe(), which schedules __put_task_struct()
> > > through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> > > be a more natural approach, we cannot allocate dynamic memory from
> > > atomic context in PREEMPT_RT, making the code more complex.
> > >
> > > This implementation ensures safe execution in atomic contexts and
> > > avoids any potential issues that may arise from using the non-atomic
> > > version.
> > >
> > > Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> > > Reported-by: Hu Chunyu <chuhu@redhat.com>
> > > Cc: Paul McKenney <paulmck@kernel.org>
> > > Cc: Thomas Gleixner <tglx@linutronix.de>
> > > ---
> > >  include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
> > >  kernel/fork.c              |  8 ++++++++
> > >  2 files changed, 39 insertions(+)
> > >
> > > diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> > > index b597b97b1f8f..5c13b83d7008 100644
> > > --- a/include/linux/sched/task.h
> > > +++ b/include/linux/sched/task.h
> > > @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
> > >
> > >  void put_task_struct_rcu_user(struct task_struct *task);
> > >
> > > +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> > > +
> > > +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> > > +{
> > > +     if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> > > +             /*
> > > +              * Decrement the refcount explicitly to avoid unnecessarily
> > > +              * calling call_rcu.
> > > +              */
> > > +             if (refcount_dec_and_test(&task->usage))
> > > +                     /*
> > > +                      * under PREEMPT_RT, we can't call put_task_struct
> > > +                      * in atomic context because it will indirectly
> > > +                      * acquire sleeping locks.
> > > +                      * call_rcu() will schedule delayed_put_task_struct_rcu()
> > > +                      * to be called in process context.
> > > +                      *
> > > +                      * __put_task_struct() is called called when
> > > +                      * refcount_dec_and_test(&t->usage) succeeds.
> > > +                      *
> > > +                      * This means that it can't "conflict" with
> > > +                      * put_task_struct_rcu_user() which abuses ->rcu the same
> > > +                      * way; rcu_users has a reference so task->usage can't be
> > > +                      * zero after rcu_users 1 -> 0 transition.
> > > +                      */
> > > +                     call_rcu(&task->rcu, __delayed_put_task_struct);
> >
> > This will invoke __delayed_put_task_struct() with softirqs disabled.
> > Or do softirq-disabled contexts count as non-atomic in PREEMPT_RT?
> 
> softirqs execute in thread context in PREEMPT_RT. We are good here.

So the sleeping lock is a spinlock rather than (say) a mutex?

                                                        Thanx, Paul

> > > +     } else {
> > > +             put_task_struct(task);
> > > +     }
> > > +}
> > > +
> > >  /* Free all architecture-specific resources held by a thread. */
> > >  void release_thread(struct task_struct *dead_task);
> > >
> > > diff --git a/kernel/fork.c b/kernel/fork.c
> > > index 0c92f224c68c..9884794fe4b8 100644
> > > --- a/kernel/fork.c
> > > +++ b/kernel/fork.c
> > > @@ -854,6 +854,14 @@ void __put_task_struct(struct task_struct *tsk)
> > >  }
> > >  EXPORT_SYMBOL_GPL(__put_task_struct);
> > >
> > > +void __delayed_put_task_struct(struct rcu_head *rhp)
> > > +{
> > > +     struct task_struct *task = container_of(rhp, struct task_struct, rcu);
> > > +
> > > +     __put_task_struct(task);
> > > +}
> > > +EXPORT_SYMBOL_GPL(__delayed_put_task_struct);
> > > +
> > >  void __init __weak arch_task_cache_init(void) { }
> > >
> > >  /*
> > > --
> > > 2.39.2
> > >
> >
>
Wander Lairson Costa April 24, 2023, 8:34 p.m. UTC | #7 
On Mon, Apr 24, 2023 at 3:52 PM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> On Mon, Apr 24, 2023 at 03:43:09PM -0300, Wander Lairson Costa wrote:
> > On Mon, Apr 24, 2023 at 3:09 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > >
> > > On Fri, Apr 14, 2023 at 09:55:28AM -0300, Wander Lairson Costa wrote:
> > > > Due to the possibility of indirectly acquiring sleeping locks, it is
> > > > unsafe to call put_task_struct() in atomic contexts when the kernel is
> > > > compiled with PREEMPT_RT.
> > > >
> > > > To mitigate this issue, this commit introduces
> > > > put_task_struct_atomic_safe(), which schedules __put_task_struct()
> > > > through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> > > > be a more natural approach, we cannot allocate dynamic memory from
> > > > atomic context in PREEMPT_RT, making the code more complex.
> > > >
> > > > This implementation ensures safe execution in atomic contexts and
> > > > avoids any potential issues that may arise from using the non-atomic
> > > > version.
> > > >
> > > > Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> > > > Reported-by: Hu Chunyu <chuhu@redhat.com>
> > > > Cc: Paul McKenney <paulmck@kernel.org>
> > > > Cc: Thomas Gleixner <tglx@linutronix.de>
> > > > ---
> > > >  include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
> > > >  kernel/fork.c              |  8 ++++++++
> > > >  2 files changed, 39 insertions(+)
> > > >
> > > > diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> > > > index b597b97b1f8f..5c13b83d7008 100644
> > > > --- a/include/linux/sched/task.h
> > > > +++ b/include/linux/sched/task.h
> > > > @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
> > > >
> > > >  void put_task_struct_rcu_user(struct task_struct *task);
> > > >
> > > > +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> > > > +
> > > > +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> > > > +{
> > > > +     if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> > > > +             /*
> > > > +              * Decrement the refcount explicitly to avoid unnecessarily
> > > > +              * calling call_rcu.
> > > > +              */
> > > > +             if (refcount_dec_and_test(&task->usage))
> > > > +                     /*
> > > > +                      * under PREEMPT_RT, we can't call put_task_struct
> > > > +                      * in atomic context because it will indirectly
> > > > +                      * acquire sleeping locks.
> > > > +                      * call_rcu() will schedule delayed_put_task_struct_rcu()
> > > > +                      * to be called in process context.
> > > > +                      *
> > > > +                      * __put_task_struct() is called called when
> > > > +                      * refcount_dec_and_test(&t->usage) succeeds.
> > > > +                      *
> > > > +                      * This means that it can't "conflict" with
> > > > +                      * put_task_struct_rcu_user() which abuses ->rcu the same
> > > > +                      * way; rcu_users has a reference so task->usage can't be
> > > > +                      * zero after rcu_users 1 -> 0 transition.
> > > > +                      */
> > > > +                     call_rcu(&task->rcu, __delayed_put_task_struct);
> > >
> > > This will invoke __delayed_put_task_struct() with softirqs disabled.
> > > Or do softirq-disabled contexts count as non-atomic in PREEMPT_RT?
> >
> > softirqs execute in thread context in PREEMPT_RT. We are good here.
>
> So the sleeping lock is a spinlock rather than (say) a mutex?
>

Yes, under PREEMPT_RT, spinlocks are implemented in terms of rtmutex.

>                                                         Thanx, Paul
>
> > > > +     } else {
> > > > +             put_task_struct(task);
> > > > +     }
> > > > +}
> > > > +
> > > >  /* Free all architecture-specific resources held by a thread. */
> > > >  void release_thread(struct task_struct *dead_task);
> > > >
> > > > diff --git a/kernel/fork.c b/kernel/fork.c
> > > > index 0c92f224c68c..9884794fe4b8 100644
> > > > --- a/kernel/fork.c
> > > > +++ b/kernel/fork.c
> > > > @@ -854,6 +854,14 @@ void __put_task_struct(struct task_struct *tsk)
> > > >  }
> > > >  EXPORT_SYMBOL_GPL(__put_task_struct);
> > > >
> > > > +void __delayed_put_task_struct(struct rcu_head *rhp)
> > > > +{
> > > > +     struct task_struct *task = container_of(rhp, struct task_struct, rcu);
> > > > +
> > > > +     __put_task_struct(task);
> > > > +}
> > > > +EXPORT_SYMBOL_GPL(__delayed_put_task_struct);
> > > > +
> > > >  void __init __weak arch_task_cache_init(void) { }
> > > >
> > > >  /*
> > > > --
> > > > 2.39.2
> > > >
> > >
> >
>
Steven Rostedt April 24, 2023, 8:34 p.m. UTC | #8 
On Mon, 24 Apr 2023 11:52:52 -0700
"Paul E. McKenney" <paulmck@kernel.org> wrote:

> > softirqs execute in thread context in PREEMPT_RT. We are good here.  
> 
> So the sleeping lock is a spinlock rather than (say) a mutex?

local_bh_disable() on RT basically turns into:

	local_lock(&softirq_ctrl.lock);
	rcu_read_lock();

Which grabs a per CPU mutex that is taken by softirqs, and also calls
rcu_read_lock(). This allows bottom halves to still run as threads but
maintain the same synchronization as they do on mainline.

-- Steve
Paul E. McKenney April 24, 2023, 9:54 p.m. UTC | #9 
On Mon, Apr 24, 2023 at 05:34:29PM -0300, Wander Lairson Costa wrote:
> On Mon, Apr 24, 2023 at 3:52 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > On Mon, Apr 24, 2023 at 03:43:09PM -0300, Wander Lairson Costa wrote:
> > > On Mon, Apr 24, 2023 at 3:09 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > > > On Fri, Apr 14, 2023 at 09:55:28AM -0300, Wander Lairson Costa wrote:
> > > > > Due to the possibility of indirectly acquiring sleeping locks, it is
> > > > > unsafe to call put_task_struct() in atomic contexts when the kernel is
> > > > > compiled with PREEMPT_RT.
> > > > >
> > > > > To mitigate this issue, this commit introduces
> > > > > put_task_struct_atomic_safe(), which schedules __put_task_struct()
> > > > > through call_rcu() when PREEMPT_RT is enabled. While a workqueue would
> > > > > be a more natural approach, we cannot allocate dynamic memory from
> > > > > atomic context in PREEMPT_RT, making the code more complex.
> > > > >
> > > > > This implementation ensures safe execution in atomic contexts and
> > > > > avoids any potential issues that may arise from using the non-atomic
> > > > > version.
> > > > >
> > > > > Signed-off-by: Wander Lairson Costa <wander@redhat.com>
> > > > > Reported-by: Hu Chunyu <chuhu@redhat.com>
> > > > > Cc: Paul McKenney <paulmck@kernel.org>
> > > > > Cc: Thomas Gleixner <tglx@linutronix.de>
> > > > > ---
> > > > >  include/linux/sched/task.h | 31 +++++++++++++++++++++++++++++++
> > > > >  kernel/fork.c              |  8 ++++++++
> > > > >  2 files changed, 39 insertions(+)
> > > > >
> > > > > diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
> > > > > index b597b97b1f8f..5c13b83d7008 100644
> > > > > --- a/include/linux/sched/task.h
> > > > > +++ b/include/linux/sched/task.h
> > > > > @@ -141,6 +141,37 @@ static inline void put_task_struct_many(struct task_struct *t, int nr)
> > > > >
> > > > >  void put_task_struct_rcu_user(struct task_struct *task);
> > > > >
> > > > > +extern void __delayed_put_task_struct(struct rcu_head *rhp);
> > > > > +
> > > > > +static inline void put_task_struct_atomic_safe(struct task_struct *task)
> > > > > +{
> > > > > +     if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
> > > > > +             /*
> > > > > +              * Decrement the refcount explicitly to avoid unnecessarily
> > > > > +              * calling call_rcu.
> > > > > +              */
> > > > > +             if (refcount_dec_and_test(&task->usage))
> > > > > +                     /*
> > > > > +                      * under PREEMPT_RT, we can't call put_task_struct
> > > > > +                      * in atomic context because it will indirectly
> > > > > +                      * acquire sleeping locks.
> > > > > +                      * call_rcu() will schedule delayed_put_task_struct_rcu()
> > > > > +                      * to be called in process context.
> > > > > +                      *
> > > > > +                      * __put_task_struct() is called called when
> > > > > +                      * refcount_dec_and_test(&t->usage) succeeds.
> > > > > +                      *
> > > > > +                      * This means that it can't "conflict" with
> > > > > +                      * put_task_struct_rcu_user() which abuses ->rcu the same
> > > > > +                      * way; rcu_users has a reference so task->usage can't be
> > > > > +                      * zero after rcu_users 1 -> 0 transition.
> > > > > +                      */
> > > > > +                     call_rcu(&task->rcu, __delayed_put_task_struct);
> > > >
> > > > This will invoke __delayed_put_task_struct() with softirqs disabled.
> > > > Or do softirq-disabled contexts count as non-atomic in PREEMPT_RT?
> > >
> > > softirqs execute in thread context in PREEMPT_RT. We are good here.
> >
> > So the sleeping lock is a spinlock rather than (say) a mutex?
> 
> Yes, under PREEMPT_RT, spinlocks are implemented in terms of rtmutex.

Reviewed-by: Paul E. McKenney <paulmck@kernel.org>

                                                        Thanx, Paul

> > > > > +     } else {
> > > > > +             put_task_struct(task);
> > > > > +     }
> > > > > +}
> > > > > +
> > > > >  /* Free all architecture-specific resources held by a thread. */
> > > > >  void release_thread(struct task_struct *dead_task);
> > > > >
> > > > > diff --git a/kernel/fork.c b/kernel/fork.c
> > > > > index 0c92f224c68c..9884794fe4b8 100644
> > > > > --- a/kernel/fork.c
> > > > > +++ b/kernel/fork.c
> > > > > @@ -854,6 +854,14 @@ void __put_task_struct(struct task_struct *tsk)
> > > > >  }
> > > > >  EXPORT_SYMBOL_GPL(__put_task_struct);
> > > > >
> > > > > +void __delayed_put_task_struct(struct rcu_head *rhp)
> > > > > +{
> > > > > +     struct task_struct *task = container_of(rhp, struct task_struct, rcu);
> > > > > +
> > > > > +     __put_task_struct(task);
> > > > > +}
> > > > > +EXPORT_SYMBOL_GPL(__delayed_put_task_struct);
> > > > > +
> > > > >  void __init __weak arch_task_cache_init(void) { }
> > > > >
> > > > >  /*
> > > > > --
> > > > > 2.39.2
> > > > >
> > > >
> > >
> >
>
diff mbox series

Patch

diff --git a/include/linux/sched/task.h b/include/linux/sched/task.h
index b597b97b1f8f..5c13b83d7008 100644
--- a/include/linux/sched/task.h
+++ b/include/linux/sched/task.h
@@ -141,6 +141,37 @@  static inline void put_task_struct_many(struct task_struct *t, int nr)
 
 void put_task_struct_rcu_user(struct task_struct *task);
 
+extern void __delayed_put_task_struct(struct rcu_head *rhp);
+
+static inline void put_task_struct_atomic_safe(struct task_struct *task)
+{
+	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
+		/*
+		 * Decrement the refcount explicitly to avoid unnecessarily
+		 * calling call_rcu.
+		 */
+		if (refcount_dec_and_test(&task->usage))
+			/*
+			 * under PREEMPT_RT, we can't call put_task_struct
+			 * in atomic context because it will indirectly
+			 * acquire sleeping locks.
+			 * call_rcu() will schedule delayed_put_task_struct_rcu()
+			 * to be called in process context.
+			 *
+			 * __put_task_struct() is called called when
+			 * refcount_dec_and_test(&t->usage) succeeds.
+			 *
+			 * This means that it can't "conflict" with
+			 * put_task_struct_rcu_user() which abuses ->rcu the same
+			 * way; rcu_users has a reference so task->usage can't be
+			 * zero after rcu_users 1 -> 0 transition.
+			 */
+			call_rcu(&task->rcu, __delayed_put_task_struct);
+	} else {
+		put_task_struct(task);
+	}
+}
+
 /* Free all architecture-specific resources held by a thread. */
 void release_thread(struct task_struct *dead_task);
 
diff --git a/kernel/fork.c b/kernel/fork.c
index 0c92f224c68c..9884794fe4b8 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -854,6 +854,14 @@  void __put_task_struct(struct task_struct *tsk)
 }
 EXPORT_SYMBOL_GPL(__put_task_struct);
 
+void __delayed_put_task_struct(struct rcu_head *rhp)
+{
+	struct task_struct *task = container_of(rhp, struct task_struct, rcu);
+
+	__put_task_struct(task);
+}
+EXPORT_SYMBOL_GPL(__delayed_put_task_struct);
+
 void __init __weak arch_task_cache_init(void) { }
 
 /*