Re: [PATCH] init/main: Clear boot task idle flag

[Peter Zijlstra]

On Wed, Sep 13, 2023 at 10:51:25AM -0400, Liam R. Howlett wrote:
> * Peter Zijlstra <peterz@xxxxxxxxxxxxx> [230913 09:53]:
> > On Tue, Sep 12, 2023 at 08:56:47PM -0400, Liam R. Howlett wrote:
> > 
> > > diff --git a/init/main.c b/init/main.c
> > > index ad920fac325c..f74772acf612 100644
> > > --- a/init/main.c
> > > +++ b/init/main.c
> > > @@ -696,7 +696,7 @@ noinline void __ref __noreturn rest_init(void)
> > >  	 */
> > >  	rcu_read_lock();
> > >  	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
> > > -	tsk->flags |= PF_NO_SETAFFINITY;
> > > +	tsk->flags |= PF_NO_SETAFFINITY | PF_IDLE;
> > >  	set_cpus_allowed_ptr(tsk, cpumask_of(smp_processor_id()));
> > >  	rcu_read_unlock();
> > >  
> > 
> > Hmm, isn't that pid-1 you're setting PF_IDLE on?
> 
> Yes, thanks.  I think that is what Geert is hitting with my patch.
> 
> debug __might_resched() in kernel/sched/core.c is failing to return in
> that first (complex) if statement.  His report says pid 1 so this is
> likely the issue.
> 
> > 
> > The task becoming idle is 'current' at this point, see the
> > cpu_startup_entry() call below.
> > 
> > Would not something like so be the right thing?
> > 
> > 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 2299a5cfbfb9..802551e0009b 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -9269,7 +9269,7 @@ void __init init_idle(struct task_struct *idle, int cpu)
> >  	 * PF_KTHREAD should already be set at this point; regardless, make it
> >  	 * look like a proper per-CPU kthread.
> >  	 */
> > -	idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
> > +	idle->flags |= PF_KTHREAD | PF_NO_SETAFFINITY;
> 
> I am concerned this will alter more than just the current task, which
> would mean more modifications later.  There is a comment about it being
> called 'more than once' and 'per cpu' so I am hesitant to change the
> function itself.
> 
> Although I am unsure of the call path.. fork_idle() -> init_idle() I
> guess?

There's only 2 ways to get into do_idle(), through cpu_startup_entry()
and play_idle_precise(). The latter already frobs PF_IDLE since it is
the forced idle path, this then leaves cpu_startup_entry() which is the
regular idle path.

All idle threads will end up calling into it, the boot CPU through the
rest_init() and the SMP cpus through arch SMP bringup.

IOW, this ensures all idle loops will have PF_IDLE set but not the
pre-idle loop setup code these threads run.