Re: [PATCH] x86: x86_{phys,virt}_bits field also for i386 (v3)

[Jeremy Fitzhardinge <jeremy@goop.org>]

Jan Beulich wrote:
>> I would say the simplest thing to do here is be explicit:
>>
>> 	if (sizeof(addr) == sizeof(u64))
>> 		return !(addr >> boot_cpu_data.x86_phys_bits);
>> 	else
>> 		return 1; 
>>
>> That's not ideal, but I guess its good enough.  I assume x86_phys_bits
>> can never be less than 32?
>>     
>
> Yes, one could do it that way. But what's the point of having RESOURCES_64BIT
> set and resource_size_t nevertheless being a 32-bit quantity?

CONFIG_RESOURCES_64BIT was removed, so testing for it makes no sense. 
(Not being able to distinguish between non-existent and unset config
variables is an outstanding Kconfig problem.)

Directly testing the size of the type is the most robust approach,
though it would be simpler if shifting a variable by more bits than its
size had a guaranteed 0 result.

>  And why,
> independent of that, was ioremap() not changed to use phys_addr_t?

Well, ioremap is supposed to be used for IO mappings, so taking a
resource_size_t still makes sense.

The question of whether resource_size_t should be the same as a
phys_addr_t is still a bit undecided. Andrew's of the opinion that they
should be separate, and that it could make sense to have 32-bit resource
addresses in an otherwise 64-bit system.  I think that's a pretty narrow
special case (32-bit PAE system with no 64-bit IO devices), and its not
worth having the extra definition complexity for it.

    J