Re: [PATCH]Fix broken VMI in 2.6.27-rc..

[Jeremy Fitzhardinge <jeremy@goop.org>]

Zachary Amsden wrote:
> On Thu, 2008-08-07 at 16:08 -0700, Linus Torvalds wrote:
>   
>> On Thu, 7 Aug 2008, H. Peter Anvin wrote:
>>     
>>> Just moving it down by 4 MB doesn't help, since the VMI guys want as much as
>>> 64 MB, which is half the standard vmalloc area and hence too much address
>>> space lost.  We can't put it at the bottom of the vmalloc area, since that
>>> boundary is not fixed, either.
>>>       
>> Yeah, ok. Since this is a 32-bit only issue, 64MB is actually a fair chunk
>> of our already limited virtual space.
>>
>>     
>>> The one remaining fixed boundary in the machine is the kernel-userspace
>>> boundary.  Hence moving the 1:1 area up by one PDE unit and sticking the
>>> fixmap area in that region.
>>>       
>> Yeah, ok, but I'd be more nervous about the validation issues there. There
>> might be a lot of code that assumes that TASK_SIZE is the start of the 1:1
>> area. It does sound like a good approach, it just makes me worry about the
>> test coverage.
>>     
>
> Well, here's an idea from outer space.  The fixmap can't possibly be
> used until it's got a backing page table and initial mappings installed.
> One can imagine a world where references to the fixmap are left as
> unresolved, and then those unresolved symbols are linked to the fixmap
> area when it gets set up in the kernel page table.  Voilla!
>
> The requisite foodling required to massage various gcci and lds into
> compliance with this scheme, not to mention the required module loading
> changes might be a bit of headache, and even then, I'm not sure that gcc
> will be smart enough to allow all the required relocations to generate
> optimal code.
>
> But the upshot would be the potential for dynamic registration of fixmap
> areas, yet still keeping direct pointers into the thing, and also
> removing all the ifdefs from the fixmap definitions for the various
> platform specific fixmap pages.  Just leave dangling references to some
> fixed bad address (fixmap_hole) for things unused.  And even allow
> kernel modules to register new fixmap types!
>
> All it requires is a well thought out strategy for naming fixmap pages
> and then two sprinkles of linker magic.  You could even randomize the
> non-randomized VDSO location at boot-time.  Whee!
>   

fixmap.ko, except backwards?

That said, isn't this exactly what the immediate values stuff is 
supposed to be able to do?

    J