Re: [PATCH] mm: fix hard lockup in __split_huge_page

[Zhaoyang Huang <huangzhaoyang@gmail.com>]

On Tue, Jun 18, 2024 at 11:19 AM Matthew Wilcox <willy@infradead.org> wrote:
>
> On Tue, Jun 18, 2024 at 10:09:26AM +0800, zhaoyang.huang wrote:
> > Hard lockup[2] is reported which should be caused by recursive
> > lock contention of lruvec->lru_lock[1] within __split_huge_page.
> >
> > [1]
> > static void __split_huge_page(struct page *page, struct list_head *list,
> >                 pgoff_t end, unsigned int new_order)
> > {
> >         /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
> > //1st lock here
> >         lruvec = folio_lruvec_lock(folio);
> >
> >         for (i = nr - new_nr; i >= new_nr; i -= new_nr) {
> >                 __split_huge_page_tail(folio, i, lruvec, list, new_order);
> >                 /* Some pages can be beyond EOF: drop them from page cache */
> >                 if (head[i].index >= end) {
> >                         folio_put(tail);
> >                             __page_cache_release
> > //2nd lock here
> >                                folio_lruvec_relock_irqsave
>
> Why doesn't lockdep catch this?
It is reported by a regression test of the fix patch which aims at the
find_get_entry livelock issue as below. I don't know the details of
the kernel configuration.

https://lore.kernel.org/linux-mm/5f989315-e380-46aa-80d1-ce8608889e5f@marcinwanat.pl/

>
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 9859aa4f7553..ea504df46d3b 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -2925,7 +2925,9 @@ static void __split_huge_page(struct page *page, struct list_head *list,
> >                               folio_account_cleaned(tail,
> >                                       inode_to_wb(folio->mapping->host));
> >                       __filemap_remove_folio(tail, NULL);
> > +                     unlock_page_lruvec(lruvec);
> >                       folio_put(tail);
> > +                     folio_lruvec_lock(folio);
>
> Why is it safe to drop & reacquire this lock?  Is there nothing we need
> to revalidate?
My stupid. I will take that into consideration in the next version.
>