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

[Catalin Marinas]

On Tue, Aug 23, 2016 at 02:33:04PM +0100, James Morse wrote:
> On 22/08/16 19:51, Catalin Marinas wrote:
> > 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);
[...]
> > Something I missed in the original hibernation support but it may look
> > better if you have something like:
> > 
> > 	pte_t pte = *src_pte;
> 
> Sure,
> 
> >> +
> >> +	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));
> 
> pte_write() doesn't clear the PTE_RDONLY flag.
> Should it be changed?

I missed this. The PTE_RDONLY flag is supposed to be cleared by
set_pte_at() but we don't call this function here. Well, I guess it's
better as you originally wrote it, so no need to change.

> >> +	} 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?
> 
> Hopefully not, but I tried to avoid marking whatever happens to be there as
> valid. This is as paranoid as I can make it: checking the pfn should be mapped,
> and that the output-address part of the record is correct.
> 
> If you're happy with the assumption that only valid records ever appear in the
> linear map page tables, (and that anything marked not-valid is a result of
> debug_pagealloc), then we can change this to !pte_none().

I think we can go for pte_none() but with a
BUG_ON(!pfn_valid(pte_pfn(*src_pte))).

> >> +			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?
> 
> A bit of both:
> Previously this copied over any values it found. _copy_pte() now copies valid
> values, and if debug_pagealloc is turned on, tries to guess whether
> the non-valid values should be copied and marked valid.

I haven't looked in detail at debug_pagealloc but if it only ever clears
the valid bit, we shouldn't have an issue.

> >> 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.
> 
> Sure.
> 
> Just in case I've totally miss-understood:
> > * Because this is only called on the kernel linar map we don't need to
> > * use p?d_present() to check for PROT_NONE regions, as these don't occur
> > * in the linear map.

Something simpler not to confuse the reader with PTE_PROT_NONE:

	/*
	 * Because this is only called on the kernel linear map,
	 * p?d_sect() implies p?d_present(). When debug_pagealloc is
	 * enabled, sections mappings are disabled.
	 */

-- 
Catalin