Re: slub-i386-support.patch

[Jeremy Fitzhardinge <jeremy@goop.org>]

Christoph Lameter wrote:
>> This is wrong; pgd's are smaller than PAGE_SIZE on PAE. Burning lowmem
>> like this is very, very bad for such systems. pmd_cache is rather
>> trivial to convert to quicklists, since all it does is zero pages. I
>> still don't approve of even the !SHARED_KERNEL_PMD case using PAGE_SIZE
>> -sized pgd's. Xen should really be fixed to avoid requiring guests to
>> have recursive pagetables or whatever it's doing.
>>     
>
> But the Xen guys need the full sized pgd?
>   

Yes, Xen needs a page-sized PGD.  There are two reasons:

   1. Xen inherently needs a whole page for the pgd, because it
      classifies pages by type, and the pgd is a pagetable-typed page.
   2. When the kernel mappings are not being shared between processes,
      as in non-PAE or PAE with !SHARED_KERNEL_PMD, it maintains a
      pgd_list linked though the index field in the page structure's
      index field.  For this to work, you can't have more than one pgd
      per page.  The pgd_list is needed to sync the vmalloc mappings
      across all the pagetables.

Xen needs to set !SHARED_KERNEL_PMD for PAE when running a 32-bit kernel
under a 32-bit hypervisor; when running under a 64-bit hypervisor,
there's no need to steal any guest address space for the hypervisor.  (I
haven't implemented this on the Linux side yet.)

    J