Re: regression: 2f569af (CONFIG_HIGHPTE vs. sub-page page tables.)

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

Hi Uwe,

On Tue, 2008-02-26 at 15:38 +0100, Uwe Kleine-König wrote:
> Moreover a "grep PageTables /proc/meminfo" resulted in 
> 
> 	9x __dec_zone_state with item == NR_PAGETABLE
> 	7x __inc_zone_state with item == NR_PAGETABLE
> 
> All those 16 calles where done via __(inc|dec)_zone_page_state.
> 
> ... some time later ...
> 
> I got it.  This code is currently in free_pgd_slow():
> 
>         pte = pmd_page(*pmd);
>         pmd_clear(pmd);
>         dec_zone_page_state(virt_to_page((unsigned long *)pgd), NR_PAGETABLE);
>         pte_lock_deinit(pte);
>         pte_free(mm, pte);
>         pmd_free(mm, pmd);
>
> So because (since 2f569af) pte_free does dec_zone_page_state and
> pte_lock_deinit, these two happen twice here.

Ah, good. I wonder if there are any other pte_lock_deinit hidden in the
arch code... nope, arm is the only architecture that does it.

> Removing the calls to dec_zone_page_state() and pte_lock_deinit() in
> free_pgd_slow() fixed it for me.

That is imho the best way to fix it.

> For a complete fix we might want to change the type of pte?

You mean instead of a "struct page *" use a pgtable_t ? Yes, that would
be cleaner even if it is the same type.

-- 
blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.