question on spinlocks

question on spinlocks

[Oliver Neukum]

Hi,

is the following sequence legal ?

spin_lock_irqsave(...);
...
spin_unlock(...);
schedule();
spin_lock(...);
...
spin_unlock_irqrestore(...);

	TIA
		Oliver

Re: question on spinlocks

[Ralf Baechle]

On Sun, Sep 01, 2002 at 07:27:53PM +0200, Oliver Neukum wrote:

> is the following sequence legal ?
> 
> spin_lock_irqsave(...);
> ...
> spin_unlock(...);
> schedule();
> spin_lock(...);
> ...
> spin_unlock_irqrestore(...);

No; spin_lock_irqsave/spin_unlock_irqrestore and spin_lock/spin_unlock
have to be used in matching pairs.

  Ralf

Re: question on spinlocks

[Thunder from the hill]

Hi,

On Sun, 1 Sep 2002, Ralf Baechle wrote:
> On Sun, Sep 01, 2002 at 07:27:53PM +0200, Oliver Neukum wrote:
> > is the following sequence legal ?
> > 
> > spin_lock_irqsave(...);
> > ...
> > spin_unlock(...);
> > schedule();
> > spin_lock(...);
> > ...
> > spin_unlock_irqrestore(...);
> 
> No; spin_lock_irqsave/spin_unlock_irqrestore and spin_lock/spin_unlock
> have to be used in matching pairs.

If it was his least problem! He'll run straight into a "schedule w/IRQs 
disabled" bug.

			Thunder
-- 
--./../...-/. -.--/---/..-/.-./..././.-../..-. .---/..-/.../- .-
--/../-./..-/-/./--..-- ../.----./.-../.-.. --./../...-/. -.--/---/..-
.- -/---/--/---/.-./.-./---/.--/.-.-.-
--./.-/-.../.-./.././.-../.-.-.-

Re: question on spinlocks

[Oliver Neukum]

> > No; spin_lock_irqsave/spin_unlock_irqrestore and spin_lock/spin_unlock
> > have to be used in matching pairs.
>
> If it was his least problem! He'll run straight into a "schedule w/IRQs
> disabled" bug.

OK, how do I drop an irqsave spinlock if I don't have flags?

	Regards
		Oliver

Re: question on spinlocks

[Robert Love]

On Sun, 2002-09-01 at 17:53, Thunder from the hill wrote:

> If it was his least problem! He'll run straight into a "schedule w/IRQs 
> disabled" bug.

No, the "schedule with irqs disabled" message is on involuntary
reschedule (e.g. kernel preemption) when interrupts are disabled.

It "safe" (maybe not sane) to call schedule() with interrupts disabled -
some system calls and scheduler code do it since interrupts will be
unconditionally enabled when rescheduled or upon returning to
user-space.  E.g., see sys_sched_yield().

The actual problem with the above is that when schedule() returns,
interrupts will be on and that is probably not the intention of the
author.  What Oliver probably wants to do is:

	spin_lock_irq(&lck);
	...
	spin_unlock(&lck);
	schedule();
	spin_lock_irq(&lck);
	...
	spin_unlock_irq(&lck);

The first unlock not having the irq-enable is an optimization as
described above.  Also note I did not use irqsave... if there is a
chance interrupts were previously disabled and you have who-knows-what
sort of call-chain behind you, it is probably not safe to go calling
schedule() and reenabling interrupts anyhow.

	Robert Love

Re: question on spinlocks

[Thunder from the hill]

Hi,

On Mon, 2 Sep 2002, Oliver Neukum wrote:
> > > No; spin_lock_irqsave/spin_unlock_irqrestore and spin_lock/spin_unlock
> > > have to be used in matching pairs.
> >
> > If it was his least problem! He'll run straight into a "schedule w/IRQs
> > disabled" bug.
> 
> OK, how do I drop an irqsave spinlock if I don't have flags?

IMHO you might even ask "How do I start a car when I don't have the keys?"

You might find a way, but it's not desired. Are you sure you want to 
reschedule in an interrupt handler? If it's none, are you sure you want to 
disable interrupts?

			Thunder
-- 
--./../...-/. -.--/---/..-/.-./..././.-../..-. .---/..-/.../- .-
--/../-./..-/-/./--..-- ../.----./.-../.-.. --./../...-/. -.--/---/..-
.- -/---/--/---/.-./.-./---/.--/.-.-.-
--./.-/-.../.-./.././.-../.-.-.-

Re: question on spinlocks

[Robert Love]

On Sun, 2002-09-01 at 18:02, Oliver Neukum wrote:
> 
> > > No; spin_lock_irqsave/spin_unlock_irqrestore and spin_lock/spin_unlock
> > > have to be used in matching pairs.
> >
> > If it was his least problem! He'll run straight into a "schedule w/IRQs
> > disabled" bug.
> 
> OK, how do I drop an irqsave spinlock if I don't have flags?

See my previous message.

Do not do what you are trying to do.  Dropping a lock and calling
schedule is fine.  Ditto with the interrupt part.

But note:

	- interrupts will be reenabled when you reschedule and still
	  enabled when your task is finally running again.

	- Since interrupts are going to magically restore, if you are
	  worried about the state of interrupts previous to your
	  function... you have a problem.

OK?

	Robert Love

Re: question on spinlocks

[Robert Love]

On Sun, 2002-09-01 at 18:09, Thunder from the hill wrote:

> IMHO you might even ask "How do I start a car when I don't have the keys?"
> 
> You might find a way, but it's not desired. Are you sure you want to 
> reschedule in an interrupt handler? If it's none, are you sure you want to 
> disable interrupts?

I do not think he is in an interrupt handler.

If he were, the system would die when he called schedule().

	Robert Love

Re: question on spinlocks

[Thunder from the hill]

Hi,

On 1 Sep 2002, Robert Love wrote:
> 	spin_lock_irq(&lck);
> 	...
> 	spin_unlock(&lck);
> 	schedule();
> 	spin_lock_irq(&lck);
> 	...
> 	spin_unlock_irq(&lck);

That makes me understand his intention a lot more. I must have got beaten 
up by the "irqsave".

			Thunder

Re: question on spinlocks

[Oliver Neukum]

Am Montag, 2. September 2002 00:09 schrieb Thunder from the hill:
> Hi,
>
> On Mon, 2 Sep 2002, Oliver Neukum wrote:
> > > > No; spin_lock_irqsave/spin_unlock_irqrestore and
> > > > spin_lock/spin_unlock have to be used in matching pairs.
> > >
> > > If it was his least problem! He'll run straight into a "schedule
> > > w/IRQs disabled" bug.
> >
> > OK, how do I drop an irqsave spinlock if I don't have flags?
>
> IMHO you might even ask "How do I start a car when I don't have the
> keys?"

Break off the lock, touch some cables ... ;-)

> You might find a way, but it's not desired. Are you sure you want to
> reschedule in an interrupt handler? If it's none, are you sure you want
> to disable interrupts?

I am not in an interrupt handler. It's not my fault that the scsi layer
calls queuecommand with a spinlock held. But I need to sleep,
I have to get rid of that spinlock's effects. If possible I even want
interrupts to be enabled.

	Regards
		Oliver

Re: question on spinlocks

[Jan Hudec]

On Mon, Sep 02, 2002 at 12:33:23AM +0200, Oliver Neukum wrote:
> Am Montag, 2. September 2002 00:09 schrieb Thunder from the hill:
> > Hi,
> >
> > On Mon, 2 Sep 2002, Oliver Neukum wrote:
> > > > > No; spin_lock_irqsave/spin_unlock_irqrestore and
> > > > > spin_lock/spin_unlock have to be used in matching pairs.
> > > >
> > > > If it was his least problem! He'll run straight into a "schedule
> > > > w/IRQs disabled" bug.
> > >
> > > OK, how do I drop an irqsave spinlock if I don't have flags?
> >
> > IMHO you might even ask "How do I start a car when I don't have the
> > keys?"
> 
> Break off the lock, touch some cables ... ;-)
> 
> > You might find a way, but it's not desired. Are you sure you want to
> > reschedule in an interrupt handler? If it's none, are you sure you want
> > to disable interrupts?
> 
> I am not in an interrupt handler. It's not my fault that the scsi layer
> calls queuecommand with a spinlock held. But I need to sleep,
> I have to get rid of that spinlock's effects. If possible I even want
> interrupts to be enabled.

If it's calling it with spinlock held, it probably relies on that it
won't be sleeping. So you either have to avoid sleeping (start now,
finish in tasklet) or patch the calling code.

-------------------------------------------------------------------------------
						 Jan 'Bulb' Hudec <bulb@ucw.cz>

Re: question on spinlocks

[Pete Zaitcev]

>> > No; spin_lock_irqsave/spin_unlock_irqrestore and spin_lock/spin_unlock
>> > have to be used in matching pairs.
>[...] 
> OK, how do I drop an irqsave spinlock if I don't have flags?

It was a good thing you didn't have flags, because everything
that passes flags as arguments blows up on sparc immediately.

Most likely answer is "restructure your locking".

-- Pete

Re: question on spinlocks

[george anzinger]

Oliver Neukum wrote:
> 
> Am Montag, 2. September 2002 00:09 schrieb Thunder from the hill:
> > Hi,
> >
> > On Mon, 2 Sep 2002, Oliver Neukum wrote:
> > > > > No; spin_lock_irqsave/spin_unlock_irqrestore and
> > > > > spin_lock/spin_unlock have to be used in matching pairs.
> > > >
> > > > If it was his least problem! He'll run straight into a "schedule
> > > > w/IRQs disabled" bug.
> > >
> > > OK, how do I drop an irqsave spinlock if I don't have flags?
> >
> > IMHO you might even ask "How do I start a car when I don't have the
> > keys?"
> 
> Break off the lock, touch some cables ... ;-)
> 
> > You might find a way, but it's not desired. Are you sure you want to
> > reschedule in an interrupt handler? If it's none, are you sure you want
> > to disable interrupts?
> 
> I am not in an interrupt handler. It's not my fault that the scsi layer
> calls queuecommand with a spinlock held. But I need to sleep,
> I have to get rid of that spinlock's effects. If possible I even want
> interrupts to be enabled.

I don't know scsi, but if the coder decided that the lock
and irq were needed, I suspect that messing with them will
get you in big trouble.  I think you need to rethink this
thing at the scsi layer...

-g
> 
>         Regards
>                 Oliver
> 
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/

-- 
George Anzinger   george@mvista.com
High-res-timers: 
http://sourceforge.net/projects/high-res-timers/
Preemption patch:
http://www.kernel.org/pub/linux/kernel/people/rml