spin_is_locked() broken for uniprocessor?

spin_is_locked() broken for uniprocessor?

[Kumar Gala]

I just want to validate that what I'm seeing (for UP, non-debug  
features):

spin_is_locked() is defined as:

include/linux/spinlock.h:#define spin_is_locked(lock)    
__raw_spin_is_locked(&(lock)->raw_lock)

for UP that should get us:

include/linux/spinlock_up.h:#define __raw_spin_is_locked(lock)  ((void) 
(lock), 0)

which implies to me that spin_is_locked() will always return false.   
Is this expected behavior.

- k

Re: spin_is_locked() broken for uniprocessor?

[Thomas Gleixner]

On Tue, 18 Aug 2009, Kumar Gala wrote:
> I just want to validate that what I'm seeing (for UP, non-debug features):
> 
> spin_is_locked() is defined as:
> 
> include/linux/spinlock.h:#define spin_is_locked(lock)
> __raw_spin_is_locked(&(lock)->raw_lock)
> 
> for UP that should get us:
> 
> include/linux/spinlock_up.h:#define __raw_spin_is_locked(lock)  ((void)(lock),
> 0)
> 
> which implies to me that spin_is_locked() will always return false.  Is this
> expected behavior.

That's wrong. spin_is_locked should always return true on UP.

Thanks,

	tglx

Re: spin_is_locked() broken for uniprocessor?

[David Howells]

Thomas Gleixner <tglx@linutronix.de> wrote:

> > which implies to me that spin_is_locked() will always return false.  Is this
> > expected behavior.
> 
> That's wrong. spin_is_locked should always return true on UP.

Surely it's not that simple?  Maybe spin_is_lock() should be undefined on UP.

David

Re: spin_is_locked() broken for uniprocessor?

[Peter Zijlstra]

On Wed, 2009-08-19 at 10:38 +0100, David Howells wrote:
> Thomas Gleixner <tglx@linutronix.de> wrote:
> 
> > > which implies to me that spin_is_locked() will always return false.  Is this
> > > expected behavior.
> > 
> > That's wrong. spin_is_locked should always return true on UP.
> 
> Surely it's not that simple?  Maybe spin_is_lock() should be undefined on UP.

#define spin_is_locked(l) panic()

should sort things out quickly ;-)

Re: spin_is_locked() broken for uniprocessor?

[Alan Cox]

On Wed, 19 Aug 2009 10:38:06 +0100
David Howells <dhowells@redhat.com> wrote:

> Thomas Gleixner <tglx@linutronix.de> wrote:
> 
> > > which implies to me that spin_is_locked() will always return false.  Is this
> > > expected behavior.
> > 
> > That's wrong. spin_is_locked should always return true on UP.
> 
> Surely it's not that simple?  Maybe spin_is_lock() should be undefined on UP.

That would lead to a lot of

#ifdef CONFIG_SMP

#endif

in drivers because there is driver code that uses spin_is_locked() in
fairly sensible fashion when dealing with locking.

Re: spin_is_locked() broken for uniprocessor?

[Leon Woestenberg]

Hello,

On Wed, Aug 19, 2009 at 12:53 PM, Alan Cox<alan@lxorguk.ukuu.org.uk> wrote:
> On Wed, 19 Aug 2009 10:38:06 +0100
>
> in drivers because there is driver code that uses spin_is_locked() in
> fairly sensible fashion when dealing with locking.
>
One use is to measure lock contention hits on a particular spin lock.


However I wonder if there are tracing capabilities to measure lock
contention on a particular lock?

Currently I have inserted code much like this to get a feeling on the
contention:

this_cpu = get_cpu();
put_cpu();
contention = spin_is_locked(&lock);
spin_lock*(&lock);
if (contention) {
  /* spin lock was contended, prev_cpu, this_cpu */
  /* no hard guarantee, as we had a possible race inbetween
is_locked() and lock(), but works for driver/irq spin lock */
}

/* critical section */

prev_cpu = this_cpu;
spin_unlock*(&lock);

Regards,
-- 
Leon

Re: spin_is_locked() broken for uniprocessor?

[Peter Zijlstra]

On Wed, 2009-08-19 at 13:16 +0200, Leon Woestenberg wrote:
> Hello,
> 
> On Wed, Aug 19, 2009 at 12:53 PM, Alan Cox<alan@lxorguk.ukuu.org.uk> wrote:
> > On Wed, 19 Aug 2009 10:38:06 +0100
> >
> > in drivers because there is driver code that uses spin_is_locked() in
> > fairly sensible fashion when dealing with locking.
> >
> One use is to measure lock contention hits on a particular spin lock.
> 
> 
> However I wonder if there are tracing capabilities to measure lock
> contention on a particular lock?

lock_stat no good for you?

Re: spin_is_locked() broken for uniprocessor?

[Benjamin Herrenschmidt]

On Wed, 2009-08-19 at 11:41 +0200, Peter Zijlstra wrote:
> > > > which implies to me that spin_is_locked() will always return
> false.  Is this
> > > > expected behavior.
> > > 
> > > That's wrong. spin_is_locked should always return true on UP.
> > 
> > Surely it's not that simple?  Maybe spin_is_lock() should be
> undefined on UP.
> 
> #define spin_is_locked(l) panic()
> 
> should sort things out quickly ;-)

Not defining it would sort things out even faster and avoid nasty case
of system panic'ing when a module is loaded :-)

Cheers,
Ben.