Re: [PATCH 1/3] mm, numa: rework do_pages_move

From: Andrew Morton
Date: Mon Jan 29 2018 - 18:48:10 EST


On Mon, 29 Jan 2018 18:39:01 -0500 "Zi Yan" <zi.yan@xxxxxxxxxxxxxx> wrote:

> On 29 Jan 2018, at 17:35, Andrew Morton wrote:
>
> > On Mon, 29 Jan 2018 17:06:14 -0500 "Zi Yan" <zi.yan@xxxxxxxxxxxxxx> wrote:
> >
> >> I discover that this patch does not hold mmap_sem while migrating pages in
> >> do_move_pages_to_node().
> >>
> >> A simple fix below moves mmap_sem from add_page_for_migration()
> >> to the outmost do_pages_move():
> >
> > I'm not surprised. Why does do_move_pages_to_node() need mmap_sem
> > and how is a reader to discover that fact???
>
> do_move_pages_to_node() calls migrate_pages(), which requires down_read(&mmap_sem).
>
> In the outmost do_pages_move(), both add_page_for_migration() and
> do_move_pages_to_node() inside it need to hold read lock of mmap_sem.
>
> Do we need to add comments for both functions?

Just for migrate_pages(), I guess. Let's include a description of
*why* mmap_sem is needed. What it is protecting.