Re: linux-next: BUG: Bad page state in process ip6tables-save pfn:1499f4

[Punit Agrawal <punit.agrawal@arm.com>]

"Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> writes:

> On Tue, Jun 27, 2017 at 09:18:15AM +0200, Vlastimil Babka wrote:
>> On 06/24/2017 05:08 PM, Andrei Vagin wrote:
>> > On Fri, Jun 23, 2017 at 05:17:44PM -0700, Andrei Vagin wrote:
>> >> On Thu, Jun 22, 2017 at 11:21:03PM -0700, Andrei Vagin wrote:
>> >>> Hello,
>> >>>
>> >>> We run CRIU tests for linux-next and today they triggered a kernel
>> >>> bug. I want to mention that this kernel is built with kasan. This bug
>> >>> was triggered in travis-ci. I can't reproduce it on my host. Without
>> >>> kasan, kernel crashed but it is impossible to get a kernel log for
>> >>> this case.
>> >>
>> >> We use this tree
>> >> https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/
>> >>
>> >> This issue isn't reproduced on the akpm-base branch and
>> >> it is reproduced each time on the akpm branch. I didn't
>> >> have time today to bisect it, will do on Monday.
>> > 
>> > c3aab7b2d4e8434d53bc81770442c14ccf0794a8 is the first bad commit
>> > 
>> > commit c3aab7b2d4e8434d53bc81770442c14ccf0794a8
>> > Merge: 849c34f 93a7379
>> > Author: Stephen Rothwell
>> > Date:   Fri Jun 23 16:40:07 2017 +1000
>> > 
>> >     Merge branch 'akpm-current/current'
>> 
>> Hm is it really the merge of mmotm itself and not one of the patches in
>> mmotm?
>> Anyway smells like THP, adding Kirill.
>
> Okay, it took a while to figure it out.

I'm sorry you had to go chasing for this one again.

I'd found the same issue while investigating an ltp failure on arm64[0] and
sent a fix[1]. The fix is effectively the same as your patch below.

Andrew picked up the patch from v5 posting and I can see it in today's
next[2].


[0] http://lists.infradead.org/pipermail/linux-arm-kernel/2017-June/510318.html
[1] https://patchwork.kernel.org/patch/9766193/
[2] https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commit/mm/gup.c?h=next-20170627&id=d31945b5d4ab4490fb5f961dd5b066cc9f560eb3

>
> The bug is in patch "mm, gup: ensure real head page is ref-counted when
> using hugepages". We should look for a head *before* the loop. Otherwise
> 'page' may point to the first page beyond the compound page.
>
> The patch below should help.
>
> If no objections, Andrew, could you fold it into the problematic patch?
>
> diff --git a/mm/gup.c b/mm/gup.c
> index d8db6e5016a8..6f9ca86b3d03 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -1424,6 +1424,7 @@ static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
>  
>  	refs = 0;
>  	page = pmd_page(orig) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
> +	head = compound_head(page);
>  	do {
>  		pages[*nr] = page;
>  		(*nr)++;
> @@ -1431,7 +1432,6 @@ static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
>  		refs++;
>  	} while (addr += PAGE_SIZE, addr != end);
>  
> -	head = compound_head(page);
>  	if (!page_cache_add_speculative(head, refs)) {
>  		*nr -= refs;
>  		return 0;
> @@ -1462,6 +1462,7 @@ static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
>  
>  	refs = 0;
>  	page = pud_page(orig) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
> +	head = compound_head(page);
>  	do {
>  		pages[*nr] = page;
>  		(*nr)++;
> @@ -1469,7 +1470,6 @@ static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
>  		refs++;
>  	} while (addr += PAGE_SIZE, addr != end);
>  
> -	head = compound_head(page);
>  	if (!page_cache_add_speculative(head, refs)) {
>  		*nr -= refs;
>  		return 0;
> @@ -1499,6 +1499,7 @@ static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
>  	BUILD_BUG_ON(pgd_devmap(orig));
>  	refs = 0;
>  	page = pgd_page(orig) + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
> +	head = compound_head(page);
>  	do {
>  		pages[*nr] = page;
>  		(*nr)++;
> @@ -1506,7 +1507,6 @@ static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
>  		refs++;
>  	} while (addr += PAGE_SIZE, addr != end);
>  
> -	head = compound_head(page);
>  	if (!page_cache_add_speculative(head, refs)) {
>  		*nr -= refs;
>  		return 0;