Re: Opinion on ordering of writel vs. stores to RAM

[Michael Buesch <mb@bu3sch.de>]

On Sunday 10 September 2006 19:19, Jesse Barnes wrote:
> On Saturday, September 09, 2006 8:09 am, Alan Cox wrote:
> > Ar Sad, 2006-09-09 am 17:23 +1000, ysgrifennodd Benjamin 
> Herrenschmidt:
> > > The problem is that very few people have any clear idea of what
> > > mmiowb is :) In fact, what you described is not the definition of
> > > mmiowb according to Jesse
> >
> > Some of us talked a little about this at Linux Kongress and one
> > suggestion so people did understand it was
> >
> > 	spin_lock_io();
> > 	spin_unlock_io();
> >
> > so that it can be expressed not as a weird barrier op but as part of
> > the locking.
> 
> That's what IRIX had.  It would let us get rid of mmiowb and avoid doing 
> a full sync in writeX, so may be the best option.

Last time I suggested that, people did not want it.
Probably about 9 months ago. Don't remember exactly.
We came to the decision that if a driver depends on some weak
ordering, it should either directly use mmiowb() or have its
own locking wrapper which wraps spin_unlock() and mmiowb().

There is one little problem in practice with something
like spin_unlock_io().

spin_lock_io(&lock);
foovalue = new_foovalue;
if (device_is_fooing)
	writel(foovalue, REGISTER);
spin_unlock_io(&lock);

That would be an unneccessary sync in case device is not fooing.
In contrast to the explicit version:

spin_lock(&lock);
foovalue = new_foovalue;
if (device_is_fooing) {
	writel(foovalue, REGISTER);
	mmiowb();
}
spin_unlock(&lock);

-- 
Greetings Michael.