Re: [PATCH 0/5] Rewrite Motorola MMU page-table layout

[Peter Zijlstra]

On Fri, Jan 31, 2020 at 09:38:13AM +0000, Will Deacon wrote:

> > This series breaks compilation for the ColdFire (with MMU) variant of
> > the m68k family:

That's like the same I had reported by the build robots for sun3, which
I fixed by frobbing pgtable_t. That said, this is probably a more
consistent change.

One note below:


> -static inline struct page *pte_alloc_one(struct mm_struct *mm)
> +static inline pgtable_t pte_alloc_one(struct mm_struct *mm)
>  {
>  	struct page *page = alloc_pages(GFP_DMA, 0);
>  	pte_t *pte;
> @@ -54,20 +55,19 @@ static inline struct page *pte_alloc_one(struct mm_struct *mm)
>  		return NULL;
>  	}
>  
> -	pte = kmap(page);
> -	if (pte) {
> -		clear_page(pte);
> -		__flush_page_to_ram(pte);
> -		flush_tlb_kernel_page(pte);
> -		nocache_page(pte);
> -	}
> -	kunmap(page);
> +	pte = page_address(page);
> +	clear_page(pte);
> +	__flush_page_to_ram(pte);
> +	flush_tlb_kernel_page(pte);
> +	nocache_page(pte);

See how it does the nocache dance ^

>  
> -	return page;
> +	return pte;
>  }
>  
> -static inline void pte_free(struct mm_struct *mm, struct page *page)
> +static inline void pte_free(struct mm_struct *mm, pgtable_t pgtable)
>  {
> +	struct page *page = virt_to_page(pgtable);
> +

but never sets it cached again!

>  	pgtable_pte_page_dtor(page);
>  	__free_page(page);
>  }

Also, the alloc_one_kernel() also suspicioudly doesn't do the nocache
thing.

So either, alloc_one() shouldn't either, or it's all buggered.