Re: [PATCH 0/5] x86: another attempt at x86 pagetable unification

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
> * Jeremy Fitzhardinge <jeremy@goop.org> wrote:
>
>   
>>> found a couple of bugs.
>>>
>>> firstly, 64-bit wasnt so lucky, you broke 
>>> iounmap()/change_page_attr()
>>> :-)
>>>       
>> Crap.  Worked for me.  I'll look into it.
>>     
>
> well, there's an easy solution for unification patches: the resulting 
> object files must have _exactly the same_ content as without the 
> unification patches. (Modulo strings as WARN_ON()s referring to 
> include-file names.)
>
> If they differ then the unification did something wrong. With your 
> patchset and the config i sent, the difference is visible in the image 
> size already:
>
>    text     data   bss     dec              hex    filename
>    7763766  967330 5812328 14543424         ddea40 vmlinux.after
>    7763811  967330 5812328 14543469         ddea6d vmlinux.before
>
> also, reducing the size and scope of changes helps as well - because 
> that way it can be bisected down to specific changes. Mistakes 
> inevitably happen, especially if you do not enforce a rigid 
> byte-for-byte correctness along the way. You did 5 rather large patches, 
> and it's not testable because your unification steps were too coarse.
>   

But byte-for-byte identity isn't (necessarily) possible when actually
unifying. If the same function exists in different forms on 32- and
64-bit, then unifying requires I pick one of them (or perhaps a new
superset) to use in the unified form. That function may generate
different code compared to the one that it replaced...

But you're right, I can do the patches in a more piecemeal form. I'll
see if I can rework them.

J