Re: [PATCH 06/10] migrate: add hugepage migration code tomove_pages()

[Michal Hocko]

On Tue 26-03-13 03:06:18, Naoya Horiguchi wrote:
> On Mon, Mar 25, 2013 at 02:36:44PM +0100, Michal Hocko wrote:
> > On Fri 22-03-13 16:23:51, Naoya Horiguchi wrote:
[...]
> > > @@ -1514,8 +1515,9 @@ struct page *follow_page_mask(struct vm_area_struct *vma,
> > >  	if (pmd_none(*pmd))
> > >  		goto no_page_table;
> > >  	if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
> > > -		BUG_ON(flags & FOLL_GET);
> > >  		page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
> > > +		if (flags & FOLL_GET && PageHead(page))
> > > +			get_page_foll(page);
> > 
> > Hmm, so the caller gets a non-null page without elevated ref counted
> > even when he asked for it. This means that all callers have to check
> > PageTail && hugetlb and put_page according to that. That is _really_
> > fragile.
> 
> I agree. And refcounting of tail pages are already very fragile,
> because get_page_foll() does something very tricky on tail pages,
> where we use page->_mapcount for refcount.
> This seems to be to handle some thp splitting problem,
> and is never intended to be used for hugepage.

yes this is THP thingy.

> So I just avoid calling it for tail pages of hugepage in caller's side.
> 
> > I think that returning NULL would make more sense in this case.
> 
> Sounds nice. I'll do this with some comment.
> 
> > >  		goto out;
> > >  	}
> > >  	if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
> > > @@ -1164,6 +1175,12 @@ static int do_move_page_to_node_array(struct mm_struct *mm,
> > [...]
> > >  				!migrate_all)
> > >  			goto put_and_set;
> > >  
> > > +		if (PageHuge(page)) {
> > > +			get_page(page);
> > > +			list_move_tail(&page->lru, &pagelist);
> > > +			goto put_and_set;
> > > +		}
> > 
> > Why do you take an additional reference here? You have one from
> > follow_page already.
> 
> For normal pages, follow_page(FOLL_GET) takes a refcount and
> isolate_lru_page() takes another one, so I think the same should
> be done for hugepages. Refcounting of this function looks tricky,
> and I'm not sure why existing code does like that.

Ohh, I see. But the whole reference is taken just to release it in goto
put_and_set because isolate_lru_page elevates reference count because
other users require that. I think you do not have to mimic this behavior
here and you can drop get_page and use goto set_status.

-- 
Michal Hocko
SUSE Labs
--
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/