Re: [PATCH] mm/hugetlbfs: Unmap pages if page fault raced with hole punch

[Andrew Morton]

On Mon, 11 Jan 2016 15:38:40 -0800 Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:

> On 01/11/2016 02:35 PM, Andrew Morton wrote:
> > On Wed,  6 Jan 2016 14:37:04 -0800 Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
> > 
> >> Page faults can race with fallocate hole punch.  If a page fault happens
> >> between the unmap and remove operations, the page is not removed and
> >> remains within the hole.  This is not the desired behavior.  The race
> >> is difficult to detect in user level code as even in the non-race
> >> case, a page within the hole could be faulted back in before fallocate
> >> returns.  If userfaultfd is expanded to support hugetlbfs in the future,
> >> this race will be easier to observe.
> >>
> >> If this race is detected and a page is mapped, the remove operation
> >> (remove_inode_hugepages) will unmap the page before removing.  The unmap
> >> within remove_inode_hugepages occurs with the hugetlb_fault_mutex held
> >> so that no other faults will be processed until the page is removed.
> >>
> >> The (unmodified) routine hugetlb_vmdelete_list was moved ahead of
> >> remove_inode_hugepages to satisfy the new reference.
> >>
> >> ...
> >>
> >> --- a/fs/hugetlbfs/inode.c
> >> +++ b/fs/hugetlbfs/inode.c
> >>
> >> ...
> >>
> >> @@ -395,37 +431,43 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
> >>  							mapping, next, 0);
> >>  			mutex_lock(&hugetlb_fault_mutex_table[hash]);
> >>  
> >> -			lock_page(page);
> >> -			if (likely(!page_mapped(page))) {
> > 
> > hm, what are the locking requirements for page_mapped()?
> 
> page_mapped is just reading/evaluating an atomic within the struct page
> which we have a referene on from the pagevec_lookup.  But, I think the
> question is what prevents page_mapped from changing after we check it?
> 
> The patch takes the hugetlb_fault_mutex_table lock before checking
> page_mapped.  If the page is unmapped and the hugetlb_fault_mutex_table
> is held, it can not be faulted in and change from unmapped to mapped.
> 
> The new comment in the patch about taking hugetlb_fault_mutex_table is
> right before the check for page_mapped.

OK, thanks.

> > 
> >> -				bool rsv_on_error = !PagePrivate(page);
> >> -				/*
> >> -				 * We must free the huge page and remove
> >> -				 * from page cache (remove_huge_page) BEFORE
> >> -				 * removing the region/reserve map
> >> -				 * (hugetlb_unreserve_pages).  In rare out
> >> -				 * of memory conditions, removal of the
> >> -				 * region/reserve map could fail.  Before
> >> -				 * free'ing the page, note PagePrivate which
> >> -				 * is used in case of error.
> >> -				 */
> >> -				remove_huge_page(page);
> > 
> > And remove_huge_page().
> 
> The page must be locked before calling remove_huge_page, since it will
> call delete_from_page_cache.  It currently is locked.  Would you prefer
> a comment stating this before the call?

No, that doesn't seem nevessary.

I'll mark this patch as "pending, awaiting Mike's go-ahead".