Re: [PATCH 00 of 10] x86: unify asm/pgtable.h

[Jeremy Fitzhardinge <jeremy@goop.org>]

Andi Kleen wrote:
>> Or to put it another way, what's the underlying rationale for making 
>> __PAGE_KERNEL_* not include the GLOBAL flag, but including it in the pgprot 
>> versions?  It means that code like this in ioremap_64.c:
>>     
>
> There is none, but that is not what change_page_attr() cares about.
> It just cares that you pass in the correct bits and you don't currently.
>
> I think you're missing the forrest because of all the trees currently.
>   

Yeah, that may be true, but this particular tree is weird, and I'm 
trying to understand what's going on here.  Specifically, 64-bit 
ioremap()s *don't* set _PAGE_GLOBAL, which appears to be an accident 
resulting from the strange definitions of __PAGE_KERNEL_* vs 
PAGE_KERNEL_*. 

For example, ioremap_64.c:__ioremap() creates a vma for the io mapping, 
and explicitly sets _PAGE_GLOBAL in the vma's version of pgprot - but 
then it calls ioremap_page_range() to actually create the mapping, which 
ends up making a non-global mapping, because its rolling its own version 
of PAGE_KERNEL by using pgprot(__PAGE_KERNEL) - which is not the actual 
definition of PAGE_KERNEL.

I think there's a bug around here, but I think its currently being 
hidden by accident.  I think my changes are correct, and they're 
exposing some other bug.  But I don't really understand how all this 
stuff is supposed to fit together, so I'm looking for an explanation of 
what's supposed to be happening - and ideally - why the current code 
isn't actually buggy.

For example: is ioremap_change_attr() actually *deliberately* creating 
non-global mappings?  Or is it an accident?  And if it really intends to 
create non-global mappings, why?  And why is it buggy for it to create 
global mappings?

    J