[PATCH v5 3/3] arm64: hibernate: Support DEBUG_PAGEALLOC

[Catalin Marinas]

On Mon, Aug 22, 2016 at 06:35:19PM +0100, James Morse wrote:
> diff --git a/arch/arm64/kernel/hibernate.c b/arch/arm64/kernel/hibernate.c
> index b4082017c4cb..da4470de1807 100644
> --- a/arch/arm64/kernel/hibernate.c
> +++ b/arch/arm64/kernel/hibernate.c
> @@ -235,6 +235,7 @@ out:
>  	return rc;
>  }
>  
> +#define dcache_clean_range(start, end)	__flush_dcache_area(start, (end - start))
>  
>  int swsusp_arch_suspend(void)
>  {
> @@ -252,8 +253,14 @@ int swsusp_arch_suspend(void)
>  	if (__cpu_suspend_enter(&state)) {
>  		ret = swsusp_save();
>  	} else {
> -		/* Clean kernel to PoC for secondary core startup */
> -		__flush_dcache_area(LMADDR(KERNEL_START), KERNEL_END - KERNEL_START);
> +		/* Clean kernel core startup/idle code to PoC*/
> +		dcache_clean_range(__mmuoff_text_start, __mmuoff_text_end);
> +		dcache_clean_range(__mmuoff_data_start, __mmuoff_data_end);
> +		dcache_clean_range(__idmap_text_start, __idmap_text_end);
> +
> +		/* Clean kvm setup code to PoC? */
> +		if (el2_reset_needed())
> +			dcache_clean_range(__hyp_idmap_text_start, __hyp_idmap_text_end);
>  
>  		/*
>  		 * Tell the hibernation core that we've just restored
> @@ -269,6 +276,32 @@ int swsusp_arch_suspend(void)
>  	return ret;
>  }
>  
> +static void _copy_pte(pte_t *dst_pte, pte_t *src_pte, unsigned long addr)
> +{
> +	unsigned long pfn = virt_to_pfn(addr);

I assume this is only called on the kernel linear mapping (according to
the copy_page_tables() use in swsusp_arch_resume). Otherwise
virt_to_pfn() would not work.

Something I missed in the original hibernation support but it may look
better if you have something like:

	pte_t pte = *src_pte;

> +
> +	if (pte_valid(*src_pte)) {
> +		/*
> +		 * Resume will overwrite areas that may be marked
> +		 * read only (code, rodata). Clear the RDONLY bit from
> +		 * the temporary mappings we use during restore.
> +		 */
> +		set_pte(dst_pte, __pte(pte_val(*src_pte) & ~PTE_RDONLY));

and here:

		set_pte(dst_pte, pte_mkwrite(pte));

> +	} else if (debug_pagealloc_enabled()) {
> +		/*
> +		 * debug_pagealloc may have removed the PTE_VALID bit if
> +		 * the page isn't in use by the resume kernel. It may have
> +		 * been in use by the original kernel, in which case we need
> +		 * to put it back in our copy to do the restore.
> +		 *
> +		 * Check for mappable memory that gives us a translation
> +		 * like part of the linear map.
> +		 */
> +		if (pfn_valid(pfn) && pte_pfn(*src_pte) == pfn)

Is there a case where this condition is false?

> +			set_pte(dst_pte, __pte((pte_val(*src_pte) & ~PTE_RDONLY) | PTE_VALID));

With some more macros:

			set_pte(dst_pte, pte_mkwrite(pte_mkpresent(pte)))

(pte_mkpresent() needs to be added)

> +	}
> +}
> +
>  static int copy_pte(pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long start,
>  		    unsigned long end)
>  {
> @@ -284,13 +317,7 @@ static int copy_pte(pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long start,
>  
>  	src_pte = pte_offset_kernel(src_pmd, start);
>  	do {
> -		if (!pte_none(*src_pte))

You seem to no longer check for pte_none(). Is this not needed or
covered by the pte_pfn() != pfn check above?

> -			/*
> -			 * Resume will overwrite areas that may be marked
> -			 * read only (code, rodata). Clear the RDONLY bit from
> -			 * the temporary mappings we use during restore.
> -			 */
> -			set_pte(dst_pte, __pte(pte_val(*src_pte) & ~PTE_RDONLY));
> +		_copy_pte(dst_pte, src_pte, addr);
>  	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
>  
>  	return 0;
> diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
> index ca6d268e3313..b6c0da84258c 100644
> --- a/arch/arm64/mm/pageattr.c
> +++ b/arch/arm64/mm/pageattr.c
> @@ -139,4 +139,42 @@ void __kernel_map_pages(struct page *page, int numpages, int enable)
>  					__pgprot(0),
>  					__pgprot(PTE_VALID));
>  }
> -#endif
> +#ifdef CONFIG_HIBERNATION
> +/*
> + * When built with CONFIG_DEBUG_PAGEALLOC and CONFIG_HIBERNATION, this function
> + * is used to determine if a linear map page has been marked as not-present by
> + * CONFIG_DEBUG_PAGEALLOC. Walk the page table and check the PTE_VALID bit.
> + * This is based on kern_addr_valid(), which almost does what we need.
> + */
> +bool kernel_page_present(struct page *page)
> +{
> +	pgd_t *pgd;
> +	pud_t *pud;
> +	pmd_t *pmd;
> +	pte_t *pte;
> +	unsigned long addr = (unsigned long)page_address(page);
> +
> +	pgd = pgd_offset_k(addr);
> +	if (pgd_none(*pgd))
> +		return false;
> +
> +	pud = pud_offset(pgd, addr);
> +	if (pud_none(*pud))
> +		return false;
> +	if (pud_sect(*pud))
> +		return true;

This wouldn't normally guarantee "present" but I don't think we ever
have a non-present section mapping for the kernel (we do for user
though). You may want to add a comment.

> +
> +	pmd = pmd_offset(pud, addr);
> +	if (pmd_none(*pmd))
> +		return false;
> +	if (pmd_sect(*pmd))
> +		return true;
> +
> +	pte = pte_offset_kernel(pmd, addr);
> +	if (pte_none(*pte))
> +		return false;
> +
> +	return pte_valid(*pte);

You can return pte_valid() directly without the pte_none() check since
pte_none() implies !pte_valid().

-- 
Catalin