Re: [PATCH] ksm: Revert "use GET_KSM_PAGE_NOLOCK to get ksm page in remove_rmap_item_from_tree()"

From: Hugh Dickins
Date: Mon May 10 2021 - 19:43:10 EST


On Mon, 10 May 2021, Miaohe Lin wrote:
> On 2021/5/10 13:59, Hugh Dickins wrote:
> > This reverts commit 3e96b6a2e9ad929a3230a22f4d64a74671a0720b.
> > General Protection Fault in rmap_walk_ksm() under memory pressure:
> > remove_rmap_item_from_tree() needs to take page lock, of course.
> >
>
> I'am really sorry about it! And many thanks for this bugfix!
> It seems rmap_walk_ksm() relies on the page lock to protect against
> concurrent modifications to that page's node of the stable tree.
> Could you please add a comment in remove_rmap_item_from_tree() to
> clarify this in case similar trouble again? Many thanks!

Sorry, no. Page lock is held by callers of stable_tree_append() when
adding an rmap_item to the tree, and held by callers of rmap_walk_ksm()
(see VM_BUG_ON_PAGE there) when walking the tree: you would surely
expect some kind of locking when removing an rmap_item from the tree,
and the appropriate page lock is what GET_KSM_PAGE_LOCK provided.

I do not want us to go through the kernel source adding a comment
/* We really mean to take this lock: it protects against concurrency */
every time we take a lock in the kernel: you should generally assume
that if a lock is taken, then the writer intended it to be taken.

There are sure to be some exceptions, where a lock is taken pointlessly:
but please look deeper before assuming that is the case.

Hugh

>
> > Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> > ---
> >
> > mm/ksm.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > --- 5.13-rc1/mm/ksm.c 2021-05-09 17:03:44.010422188 -0700
> > +++ linux/mm/ksm.c 2021-05-09 22:12:39.403008350 -0700
> > @@ -776,11 +776,12 @@ static void remove_rmap_item_from_tree(s
> > struct page *page;
> >
> > stable_node = rmap_item->head;
> > - page = get_ksm_page(stable_node, GET_KSM_PAGE_NOLOCK);
> > + page = get_ksm_page(stable_node, GET_KSM_PAGE_LOCK);
> > if (!page)
> > goto out;
> >
> > hlist_del(&rmap_item->hlist);
> > + unlock_page(page);
> > put_page(page);
> >
> > if (!hlist_empty(&stable_node->hlist))