Re: [PATCH v7 2/3] sched/task: Add the put_task_struct_atomic_safe() function

[Wander Lairson Costa]

On Thu, May 4, 2023 at 5:16 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>
> Hi Wander,
>
> I certainly missed something ;) plus I am already sleeping. but let me try to
> reply anyway.
>
> On 05/04, Wander Lairson Costa wrote:
> > On Thu, May 4, 2023 at 4:23 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> > >
> > > On 05/04, Wander Lairson Costa wrote:
> > > >
> > > > On Thu, May 4, 2023 at 12:23 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> > > > >
> > > > > Yes, but as Sebastian explained CONFIG_PROVE_RAW_LOCK_NESTING won't like it.
> > > > >
> > > > >         https://lore.kernel.org/all/Y+zFNrCjBn53%2F+Q2@xxxxxxxxxxxxx/
> > > > >
> > > >
> > > > I think that was my confusion in that thread. My understanding is that
> > > > CONFIG_PROVE_RAW_LOCK_NESTING will check lock ordering but not
> > > > context.
> > >
> > > Sorry, I don't understand... perhaps I missed something. But iiuc
> > > the problem is simple.
> > >
> > > So, this code
> > >
> > >         raw_spin_lock(one);
> > >         spin_lock(two);
> > >
> > > is obviously wrong if CONFIG_PREEMPT_RT.
> > >
> > > Without PREEMPT_RT this code is fine because raw_spinlock_t and spinlock_t
> > > are the same thing. Except they have different lockdep annotations if
> > > CONFIG_PROVE_RAW_LOCK_NESTING is true, LD_WAIT_SPIN and LD_WAIT_CONFIG.
> > >
> > > So if CONFIG_PROVE_RAW_LOCK_NESTING is set, lockdep will complain even
> > > on the !PREEMPT_RT kernel, iow it checks the nesting as if the code runs
> > > on with PREEMPT_RT.
> > >
> > > Cough... not sure my explanation can help ;) It looks very confusing when
> > > I read it.
> > >
> >
> > Thanks for the explanation. That's my understanding too. The part I
> > don't get is why this would fail with a call_rcu() inside
> > put_task_struct().
>
> the problem is that call_rcu() won't be called if !IS_ENABLED(PREEMPT_RT),
> ___put_task_struct() will be called.
>
> CONFIG_PROVE_RAW_LOCK_NESTING can't know this can't happen if PREEMPT_RT
> is set.
>
> IOW. To simplify, suppose we have
>
>         // can be called in atomic context, e.g. under
>         // raw_spin_lock() so it is wrong with PREEMPT_RT
>         void __put_task_struct(struct task_struct *tsk)
>         {
>                 spin_lock(some_lock);
>         }
>
> lets "fix" the code above, lets change __put_task_struct,
>
>         void __put_task_struct(struct task_struct *tsk)
>         {
>                 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
>                         return;
>
>                 spin_lock(some_lock);
>         }
>
> Now, if CONFIG_PREEMPT_RT is true then __put_task_struct() is fine
> wrt lock nesting.
>
> But, if CONFIG_PREEMPT_RT is not set, then __put_task_struct() still
> does the same:
>
>         void __put_task_struct(struct task_struct *tsk)
>         {
>                 spin_lock(some_lock);
>         }
>
> and CONFIG_PROVE_RAW_LOCK_NESTING will complain. Because, once again,
> it checks the nesting as if CONFIG_PREEMPT_RT is true, and in this case
> __put_task_struct() if it is called under raw_spin_lock().
>

IIUC, this is a problem with the current implementation, isn't it?
Because the holder of raw_spin_lock is the put_task_struct() caller.

> Oleg.
>