Re: [mmotm] BUG: Bad page state in process khugepaged ?

[Hugh Dickins]

On Fri, 11 Feb 2011, Andrea Arcangeli wrote:
> On Thu, Feb 10, 2011 at 11:02:50PM -0800, Hugh Dickins wrote:
> > There is a separate little issue here, Andrea.
> > 
> > Although we went to some trouble for bad_page() to take the page out
> > of circulation yet let the system continue, your VM_BUG_ON(!PageBuddy)
> > inside __ClearPageBuddy(page), from two callsites in bad_page(), is
> > turning it into a fatal error when CONFIG_DEBUG_VM.
> 
> I see what you mean. Of course it is only a problem after bad_page
> already triggered.... but then it trigger an BUG_ON instead of only a
> bad_page.
> 
> > You could that only MM developers switch CONFIG_DEBUG_VM=y, and they
> > would like bad_page() to be fatal; maybe, but if so we should do that
> > as an intentional patch, rather than as an unexpected side-effect ;)
> 
> Fedora kernels are built with CONFIG_DEBUG_VM, all my kernels runs
> with CONFIG_DEBUG_VM too, so we want it to be as "production" as
> possible, and we don't want DEBUG_VM to decrease any reliability (only
> to increase it of course).

Oh, I hadn't realized Fedora use it.  I wonder if that's wise, I thought
Nick introduced it partly for the more expensive checks, and there might
be one or two of those around - those bad_range()s in page_alloc.c?

> 
> > I noticed this a few days ago, but hadn't quite decided whether just to
> > remove the VM_BUG_ON, or move it to __ClearPageBuddy's third callsite,
> > or... doesn't matter much.
> >
> > I do also wonder if PageBuddy would better be _mapcount -something else:
> > if we've got a miscounted page (itself unlikely of course), there's a
> > chance that its _mapcount will be further decremented after it has been
> > freed: whereupon it will go from -1 to -2, PageBuddy at present.  The
> > special avoidance of PageBuddy being that it can pull a whole block of
> > pages into misuse if its mistaken.
> 
> Agreed. What about the below?
> 
> =====
> Subject: mm: PageBuddy cleanups
> 
> From: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> 
> bad_page could VM_BUG_ON(!PageBuddy(page)) inside __ClearPageBuddy().
> I prefer to keep the VM_BUG_ON for safety and to add a if to solve it.

Too much iffery: I ended up preferring it in rmv_page_order() myself.

> 
> Change the _mapcount value indicating PageBuddy from -2 to -1024 for more
> robusteness against page_mapcount() undeflows.

But the patch actually says -1024*1024: either would do.

> 
> Signed-off-by: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> Reported-by: Hugh Dickins <hughd@xxxxxxxxxx>
> ---
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index f6385fc..fa16ba0 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -402,16 +402,22 @@ static inline void init_page_count(struct page *page)
>  /*
>   * PageBuddy() indicate that the page is free and in the buddy system
>   * (see mm/page_alloc.c).
> + *
> + * PAGE_BUDDY_MAPCOUNT_VALUE must be <= -2 but better not too close to
> + * -2 so that an underflow of the page_mapcount() won't be mistaken
> + * for a genuine PAGE_BUDDY_MAPCOUNT_VALUE.

Yes, good to comment that, thanks.

>   */
> +#define PAGE_BUDDY_MAPCOUNT_VALUE (-1024*1024)
> +
>  static inline int PageBuddy(struct page *page)
>  {
> -	return atomic_read(&page->_mapcount) == -2;
> +	return atomic_read(&page->_mapcount) == PAGE_BUDDY_MAPCOUNT_VALUE;
>  }
>  
>  static inline void __SetPageBuddy(struct page *page)
>  {
>  	VM_BUG_ON(atomic_read(&page->_mapcount) != -1);
> -	atomic_set(&page->_mapcount, -2);
> +	atomic_set(&page->_mapcount, PAGE_BUDDY_MAPCOUNT_VALUE);
>  }
>  
>  static inline void __ClearPageBuddy(struct page *page)

Yes, that's fine, 0xfff00000 looks unlikely enough (and my
imagination for "deadbeef"-like magic is too drowsy today).

> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index a873e61..8aac134 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -286,7 +286,9 @@ static void bad_page(struct page *page)
>  
>  	/* Don't complain about poisoned pages */
>  	if (PageHWPoison(page)) {
> -		__ClearPageBuddy(page);
> +		/* __ClearPageBuddy VM_BUG_ON(!PageBuddy(page)) */
> +		if (PageBuddy(page))
> +			__ClearPageBuddy(page);
>  		return;
>  	}
>  
> @@ -317,7 +319,8 @@ static void bad_page(struct page *page)
>  	dump_stack();
>  out:
>  	/* Leave bad fields for debug, except PageBuddy could make trouble */
> -	__ClearPageBuddy(page);
> +	if (PageBuddy(page)) /* __ClearPageBuddy VM_BUG_ON(!PageBuddy(page)) */
> +		__ClearPageBuddy(page);
>  	add_taint(TAINT_BAD_PAGE);
>  }
>  

Okay I suppose: it seems rather laboured to me, I think I'd have just
moved the VM_BUG_ON into rmv_page_order() if I'd done the patch; but
since I was too lazy to do it, I'd better be grateful for yours!

Hugh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/