Re: Process-shared futexes on hugepages puts the kernel in an infinite loop in 2.6.32.11; is this fixed now?

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2010-04-19 at 12:43 +0100, Mel Gorman wrote:

> > Right, so I had a quick chat with Mel, and it appears MAP_PRIVATE
> > hugetlb pages don't have their page->mapping set.
> > 
> > I guess something like the below might work, but I'd really rather not
> > add hugetlb knowledge to futex.c. Does anybody else have a better idea?
> > Maybe create something similar to an anon_vma for hugetlb pages?
> > 
> 
> anon_vma for hugetlb pages sounds overkill, what would it gain? In this
> context, futex only appears to distinguish between whether the
> references are private or shared.
> 
> Looking at the hugetlbfs code, I can't see a place where it actually cares
> about the mapping as such. It's used to find shared pages in the page cache
> (but not in the LRU) that are backed by the hugetlbfs file. For hugetlbfs
> though, the mapping is mostly kept in page->private for reservation accounting
> purposes.
> 
> I can't think of other parts of the VM that touch the mapping if the
> page is managed by hugetlbfs so the following patch should also work but
> without futex having hugetlbfs-awareness. What do you think? Maybe for
> safety, it would be better to make the mapping some obvious poison bytes
> or'd with PAGE_MAPPING_ANON so an oops will be more obvious?

Yes, this seems perfectly adequate to me, that poison idea might be
worthwhile too :-)

Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>

> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 6034dc9..57a5faa 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -546,6 +546,7 @@ static void free_huge_page(struct page *page)
>  
>  	mapping = (struct address_space *) page_private(page);
>  	set_page_private(page, 0);
> +	page->mapping = NULL;
>  	BUG_ON(page_count(page));
>  	INIT_LIST_HEAD(&page->lru);
>  
> @@ -2447,8 +2448,10 @@ retry:
>  			spin_lock(&inode->i_lock);
>  			inode->i_blocks += blocks_per_huge_page(h);
>  			spin_unlock(&inode->i_lock);
> -		} else
> +		} else {
>  			lock_page(page);
> +			page->mapping = (struct address_space *)PAGE_MAPPING_ANON;
> +		}
>  	}
>  
>  	/*