[RFC][PATCH] freezer,sched: Rewrite core freezer logic

[Peter Zijlstra]

On Wed, Jun 02, 2021 at 01:54:53PM +0100, Will Deacon wrote:

> There's also Documentation/power/freezing-of-tasks.rst to update. I'm not

Since it's .rst, the only update I'm willing to do is delete it
outright.

> sure if fs/proc/array.c should be updated to display frozen tasks; I
> couldn't see how that was useful, but thought I'd mention it anyway.

Yeah, I considered it too, but I figured that if we're all frozen
there's noone left to observe us being frozen, so I didn't bother.

> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index 2982cfab1ae9..bfadc1dbcf24 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -95,7 +95,12 @@ struct task_group;
> >  #define TASK_WAKING			0x0200
> >  #define TASK_NOLOAD			0x0400
> >  #define TASK_NEW			0x0800
> > -#define TASK_STATE_MAX			0x1000
> > +#define TASK_FREEZABLE			0x1000
> > +#define __TASK_FREEZABLE_UNSAFE		0x2000
> 
> Give that this is only needed to avoid lockdep checks, maybe we should avoid
> allocating the bit if lockdep is not enabled? Otherwise, people might start
> to use it for other things.

Something like

#define __TASK_FREEZABLE_UNSAFE			(0x2000 * IS_ENABLED(CONFIG_LOCKDEP))

?

> > +#define TASK_FROZEN			0x4000
> > +#define TASK_STATE_MAX			0x8000
> > +
> > +#define TASK_FREEZABLE_UNSAFE		(TASK_FREEZABLE | __TASK_FREEZABLE_UNSAFE)
> 
> We probably want to preserve the "DO NOT ADD ANY NEW CALLERS OF THIS STATE"
> comment for the unsafe stuff.

Done.

> > +/* Recursion relies on tail-call optimization to not blow away the stack */
> > +static bool __frozen(struct task_struct *p)
> > +{
> > +	if (p->state == TASK_FROZEN)
> > +		return true;
> 
> READ_ONCE()?

task_struct::state is volatile -- for now. I've got other patches to
deal with that.

> > +
> > +	/*
> > +	 * If stuck in TRACED, and the ptracer is FROZEN, we're frozen too.
> > +	 */
> > +	if (task_is_traced(p))
> > +		return frozen(rcu_dereference(p->parent));
> > +
> > +	/*
> > +	 * If stuck in STOPPED and the parent is FROZEN, we're frozen too.
> > +	 */
> > +	if (task_is_stopped(p))
> > +		return frozen(rcu_dereference(p->real_parent));
> 
> This looks convincing, but I really can't tell if we're missing anything.

Yeah, Oleg would be the one to tell us I suppose.

> > +static bool __freeze_task(struct task_struct *p)
> > +{
> > +	unsigned long flags;
> > +	unsigned int state;
> > +	bool frozen = false;
> > +
> > +	raw_spin_lock_irqsave(&p->pi_lock, flags);
> > +	state = READ_ONCE(p->state);
> > +	if (state & TASK_FREEZABLE) {
> > +		/*
> > +		 * Only TASK_NORMAL can be augmented with TASK_FREEZABLE,
> > +		 * since they can suffer spurious wakeups.
> > +		 */
> > +		WARN_ON_ONCE(!(state & TASK_NORMAL));
> > +
> > +#ifdef CONFIG_LOCKDEP
> > +		/*
> > +		 * It's dangerous to freeze with locks held; there be dragons there.
> > +		 */
> > +		if (!(state & __TASK_FREEZABLE_UNSAFE))
> > +			WARN_ON_ONCE(debug_locks && p->lockdep_depth);
> > +#endif
> > +
> > +		p->state = TASK_FROZEN;
> > +		frozen = true;
> > +	}
> > +	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
> > +
> > +	return frozen;
> > +}
> > +
> >  /**
> >   * freeze_task - send a freeze request to given task
> >   * @p: task to send the request to
> > @@ -116,20 +173,8 @@ bool freeze_task(struct task_struct *p)
> >  {
> >  	unsigned long flags;
> >  
> >  	spin_lock_irqsave(&freezer_lock, flags);
> > +	if (!freezing(p) || frozen(p) || __freeze_task(p)) {
> >  		spin_unlock_irqrestore(&freezer_lock, flags);
> >  		return false;
> >  	}
> 
> I've been trying to figure out how this serialises with ttwu(), given that
> frozen(p) will go and read p->state. I suppose it works out because only the
> freezer can wake up tasks from the FROZEN state, but it feels a bit brittle.

p->pi_lock; both ttwu() and __freeze_task() (which is essentially a
variant of set_special_state()) take ->pi_lock. I'll put in a comment.

> > @@ -137,7 +182,7 @@ bool freeze_task(struct task_struct *p)
> >  	if (!(p->flags & PF_KTHREAD))
> >  		fake_signal_wake_up(p);
> >  	else
> > -		wake_up_state(p, TASK_INTERRUPTIBLE);
> > +		wake_up_state(p, TASK_INTERRUPTIBLE); // TASK_NORMAL ?!?
> >  
> >  	spin_unlock_irqrestore(&freezer_lock, flags);
> >  	return true;
> > @@ -148,8 +193,8 @@ void __thaw_task(struct task_struct *p)
> >  	unsigned long flags;
> >  
> >  	spin_lock_irqsave(&freezer_lock, flags);
> > -	if (frozen(p))
> > -		wake_up_process(p);
> > +	WARN_ON_ONCE(freezing(p));
> > +	wake_up_state(p, TASK_FROZEN | TASK_NORMAL);
> 
> Why do we need TASK_NORMAL here?

It's a left-over from hacking, but I left it in because anything
TASK_NORMAL should be able to deal with spuriuos wakeups, something
try_to_freeze() now also relies on.