use of volatile in iounmap()?

use of volatile in iounmap()?

[Sam Ravnborg]

While reviewing some CAN driver I stumbled on iounmap
which has following prototype on x86:

extern void iounmap(volatile void __iomem *addr);

I argued that the driver should not use volatile
but then I cannot explain why the argument to
iounmap takes a volatile.

The same goes for many other functions in
the io*.h headers.

Grepping the other archs they mostly follow
same pattern.

Can anyone explain the rational for volatile in this case.

Thanks,
	Sam

Re: use of volatile in iounmap()?

[H. Peter Anvin]

Sam Ravnborg wrote:
> While reviewing some CAN driver I stumbled on iounmap
> which has following prototype on x86:
> 
> extern void iounmap(volatile void __iomem *addr);
> 
> I argued that the driver should not use volatile
> but then I cannot explain why the argument to
> iounmap takes a volatile.
> 
> The same goes for many other functions in
> the io*.h headers.
> 
> Grepping the other archs they mostly follow
> same pattern.
> 
> Can anyone explain the rational for volatile in this case.
> 

Yes.  The use of volatile in a function prototype like this means that 
it is valid to pass a volatile pointer to that function -- in other 
words, we're telling gcc that we're not going to do anything with the 
pointer that is invalid for a volatile pointer.

A lot of the "volatile considered harmful" stuff that has been bandied 
about is explicitly about marking *data* items volatile (it does have 
its uses, but it's easy to get wrong); Linus has explicitly made the 
distinction between volatile *data* and volatile *operations*.

	-hpa

Re: use of volatile in iounmap()?

[Sam Ravnborg]

On Fri, Mar 28, 2008 at 01:51:12PM -0700, H. Peter Anvin wrote:
> Sam Ravnborg wrote:
> >While reviewing some CAN driver I stumbled on iounmap
> >which has following prototype on x86:
> >
> >extern void iounmap(volatile void __iomem *addr);
> >
> >I argued that the driver should not use volatile
> >but then I cannot explain why the argument to
> >iounmap takes a volatile.
> >
> >The same goes for many other functions in
> >the io*.h headers.
> >
> >Grepping the other archs they mostly follow
> >same pattern.
> >
> >Can anyone explain the rational for volatile in this case.
> >
> 
> Yes.  The use of volatile in a function prototype like this means that 
> it is valid to pass a volatile pointer to that function -- in other 
> words, we're telling gcc that we're not going to do anything with the 
> pointer that is invalid for a volatile pointer.
If I understand you correct then it is then not wrong to say
that we have the argument volatile to avoid warnings from gcc
when we pass a volatile pointer.

And then having the pointer marked volatile put a few restrictions
on iounmap().

> 
> A lot of the "volatile considered harmful" stuff that has been bandied 
> about is explicitly about marking *data* items volatile (it does have 
> its uses, but it's easy to get wrong); Linus has explicitly made the 
> distinction between volatile *data* and volatile *operations*.

Yes - but unfortunately the volatile-considered-harmful.txt
does many deal with the data part.

Thanks,
	Sam

Re: use of volatile in iounmap()?

[H. Peter Anvin]

Sam Ravnborg wrote:
>>>
>> Yes.  The use of volatile in a function prototype like this means that 
>> it is valid to pass a volatile pointer to that function -- in other 
>> words, we're telling gcc that we're not going to do anything with the 
>> pointer that is invalid for a volatile pointer.

> If I understand you correct then it is then not wrong to say
> that we have the argument volatile to avoid warnings from gcc
> when we pass a volatile pointer.
> 
> And then having the pointer marked volatile put a few restrictions
> on iounmap().

Correct.

>> A lot of the "volatile considered harmful" stuff that has been bandied 
>> about is explicitly about marking *data* items volatile (it does have 
>> its uses, but it's easy to get wrong); Linus has explicitly made the 
>> distinction between volatile *data* and volatile *operations*.
> 
> Yes - but unfortunately the volatile-considered-harmful.txt
> does many deal with the data part.

Yes, it does.

	-hpa

Re: use of volatile in iounmap()?

[Al Viro]

On Fri, Mar 28, 2008 at 09:34:28PM +0100, Sam Ravnborg wrote:
> While reviewing some CAN driver I stumbled on iounmap
> which has following prototype on x86:
> 
> extern void iounmap(volatile void __iomem *addr);
> 
> I argued that the driver should not use volatile
> but then I cannot explain why the argument to
> iounmap takes a volatile.
> 
> The same goes for many other functions in
> the io*.h headers.
> 
> Grepping the other archs they mostly follow
> same pattern.
> 
> Can anyone explain the rational for volatile in this case.

"Passing a pointer to volatile is allowed, along with passing pointers
to unqualified".