put_task_struct() decrements a usage counter and calls
__put_task_struct() if the counter reaches zero.

__put_task_struct() indirectly acquires a spinlock, which is a sleeping
lock under PREEMPT_RT. Therefore, we can't call put_task_struct() in an
atomic context in RT kernels.

This patch series introduces put_task_struct_atomic_safe(), which defers
the call to __put_task_struct() to a process context when compiled with
PREEMPT_RT.

It also fixes known problematic call sites.

Changelog:
==========

v2:
 * Add the put_task_struct_atomic_safe() function that is responsible for
   handling the conditions to call put_task_struct().
 * Replace put_task_struct() by put_task_struct_atomic_safe() in known
   atomic call sites.

Wander Lairson Costa (4):
  sched/task: Add the put_task_struct_atomic_safe function
  sched/deadline: fix inactive_task_timer splat
  sched/rt: use put_task_struct_atomic_safe() to avoid potential splat
  sched/core: use put_task_struct_atomic_safe() to avoid potential splat

 include/linux/sched/task.h | 21 +++++++++++++++++++++
 kernel/fork.c              |  8 ++++++++
 kernel/sched/core.c        |  2 +-
 kernel/sched/deadline.c    |  2 +-
 kernel/sched/rt.c          |  4 ++--
 5 files changed, 33 insertions(+), 4 deletions(-)

-- 
2.39.0

On 20/01/23 12:02, Wander Lairson Costa wrote:
> put_task_struct() decrements a usage counter and calls
> __put_task_struct() if the counter reaches zero.
>
> __put_task_struct() indirectly acquires a spinlock, which is a sleeping
> lock under PREEMPT_RT. Therefore, we can't call put_task_struct() in an
> atomic context in RT kernels.
>
> This patch series introduces put_task_struct_atomic_safe(), which defers
> the call to __put_task_struct() to a process context when compiled with
> PREEMPT_RT.
>
> It also fixes known problematic call sites.
>

Browsing around put_task_struct() callsites gives me the impression there
are more problematic call sites lurking around, which makes me wonder:
should we make the PREEMPT_RT put_task_struct() *always* be done via
call_rcu()?

The task's stack is actually always freed that way in put_task_stack(), cf.

  e540bf3162e8 ("fork: Only cache the VMAP stack in finish_task_switch()")

> Changelog:
> ==========
>
> v2:
>  * Add the put_task_struct_atomic_safe() function that is responsible for
>    handling the conditions to call put_task_struct().
>  * Replace put_task_struct() by put_task_struct_atomic_safe() in known
>    atomic call sites.
>
> Wander Lairson Costa (4):
>   sched/task: Add the put_task_struct_atomic_safe function
>   sched/deadline: fix inactive_task_timer splat
>   sched/rt: use put_task_struct_atomic_safe() to avoid potential splat
>   sched/core: use put_task_struct_atomic_safe() to avoid potential splat
>
>  include/linux/sched/task.h | 21 +++++++++++++++++++++
>  kernel/fork.c              |  8 ++++++++
>  kernel/sched/core.c        |  2 +-
>  kernel/sched/deadline.c    |  2 +-
>  kernel/sched/rt.c          |  4 ++--
>  5 files changed, 33 insertions(+), 4 deletions(-)
>
> --
> 2.39.0

On Fri, Jan 20, 2023 at 2:45 PM Valentin Schneider <vschneid@redhat.com> wrote:
>
> On 20/01/23 at 12:02, Wander Lairson Costa wrote:
> > put_task_struct() decrements a usage counter and calls
> > __put_task_struct() if the counter reaches zero.
> >
> > __put_task_struct() indirectly acquires a spinlock, which is a sleeping
> > lock under PREEMPT_RT. Therefore, we can't call put_task_struct() in an
> > atomic context in RT kernels.
> >
> > This patch series introduces put_task_struct_atomic_safe(), which defers
> > the call to __put_task_struct() to a process context when compiled with
> > PREEMPT_RT.
> >
> > It also fixes known problematic call sites.
> >
>
> Browsing around put_task_struct() callsites gives me the impression there
> are more problematic call sites lurking around, which makes me wonder:
> should we make the PREEMPT_RT put_task_struct() *always* be done via
> call_rcu()?
>

I thought about going on this route, but I was concerned about the
performance side effects this approach could bring. Another idea I had
was to check at runtime if we are in a preemptible context. Again,
this would have a (minor?) performance penalty.

> The task's stack is actually always freed that way in put_task_stack(), cf.
>
>   e540bf3162e8 ("fork: Only cache the VMAP stack in finish_task_switch()")
>
> > Changelog:
> > ==========
> >
> > v2:
> >  * Add the put_task_struct_atomic_safe() function that is responsible for
> >    handling the conditions to call put_task_struct().
> >  * Replace put_task_struct() by put_task_struct_atomic_safe() in known
> >    atomic call sites.
> >
> > Wander Lairson Costa (4):
> >   sched/task: Add the put_task_struct_atomic_safe function
> >   sched/deadline: fix inactive_task_timer splat
> >   sched/rt: use put_task_struct_atomic_safe() to avoid potential splat
> >   sched/core: use put_task_struct_atomic_safe() to avoid potential splat
> >
> >  include/linux/sched/task.h | 21 +++++++++++++++++++++
> >  kernel/fork.c              |  8 ++++++++
> >  kernel/sched/core.c        |  2 +-
> >  kernel/sched/deadline.c    |  2 +-
> >  kernel/sched/rt.c          |  4 ++--
> >  5 files changed, 33 insertions(+), 4 deletions(-)
> >
> > --
> > 2.39.0
>