Re: [PATCH] kernel/timer.c: using spin_lock_irqsave instead of spin_lock + local_irq_save, especially when CONFIG_LOCKDEP not defined

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, 19 Jun 2013, Chen Gang wrote:

> On 06/19/2013 04:41 PM, Thomas Gleixner wrote:
> > On Wed, 19 Jun 2013, Chen Gang wrote:
> > 
> >> > 
> >> > When CONFIG_LOCKDEP is not defined, spin_lock_irqsave() is not equal to
> >> > spin_lock() + local_irq_save().
> >> > 
> >> > In __mod_timer(), After call spin_lock_irqsave() with 'base->lock' in
> >> > lock_timer_base(), it may use spin_lock() with the 'new_base->lock'.
> >> > 
> >> > It may let original call do_raw_spin_lock_flags() with 'base->lock',
> >> > but new  call LOCK_CONTENDED() with 'new_base->lock'.
> >> > 
> >> > In fact, we need both of them call do_raw_spin_lock_flags(), so use
> >> > spin_lock_irqsave() instead of spin_lock() + local_irq_save().
> > Why do we need to do that? There is no reason to do so and it's
> > totally irrelevant whether CONFIG_LOCKDEP is enabled or not.
> >
> 
> Please see include/linux/spinlock_api_smp.h (or see bottom of this mail)
> 
>  
> > The code is written intentionally this way.
> > 
> > What's the difference between:
> > 
> >        spin_lock_irqsave(&l1, flags);
> >        spin_unlock(&l1);
> >        spin_lock(l2);
> >        spin_unlock_irqrestore(&l2, flags);
> > 
> > and
> > 
> >        spin_lock_irqsave(&l1, flags);
> >        spin_unlock_irqrestore(&l1);
> >        spin_lock_irqsave(l2, flags);
> >        spin_unlock_irqrestore(&l2, flags);
> > 
> 
> Yes
> 
> 
> > The difference is that we avoid to touch the interrupt disable in the
> > cpu, which might be an expensive operation depending on the cpu model.
> > 
> > There is no point in reenabling interrupts just to disable them
> > again a few instruction cycles later.
> > 
> > And lockdep is perfectly fine with that code. All lockdep cares about
> > is whether the lock context (interrupts disabled) is correct or
> > not.
> 
> Please reference include/linux/spinlock_api_smp.h and include/linux/spinlock.h
>
>   spin_lock_irqsave() ->
>     raw_spin_lock_irqsave() ->
>       _raw_spin_lock_irqsave() ->
>         __raw_spin_lock_irqsave()
> 
>   spin_lock() ->
>     raw_spin_lock() ->
>       _raw_spin_lock() ->
>         __raw_spin_lock()
> 

I'm well aware how that works. And there is no difference whether you
do:

	local_irq_save(flags);
	spin_lock(&lock);
or
	spin_lock_irqsave(&lock, flags);


Lockdep tracks lock ordering and the context in which a lock is
taken. The timer base lock can be taken in interrupt context, so it
always needs to be taken with interrupts disabled. That's what lockdep
cares about.

And 
	spin_lock_irqsave(&l1, flags);
	spin_unlock(&l1);
	spin_lock(&l2);
	spin_unlock_irqrestore(&l2, flags);

fulfils that for both l1 and l2.

It does not matter whether the code pathes are different, what matters
is that they are semantically the same. And that's the case.

Thanks,

	tglx