Re: [PATCH v20 08/20] mm: page_idle_get_page() does not need lru_lock

[Johannes Weiner]

On Wed, Nov 04, 2020 at 07:27:21PM +0800, Alex Shi wrote:
> 在 2020/11/3 上午4:20, Johannes Weiner 写道:
> > On Mon, Nov 02, 2020 at 02:49:27PM +0000, Matthew Wilcox wrote:
> >> On Mon, Nov 02, 2020 at 09:41:10AM -0500, Johannes Weiner wrote:
> >>> On Thu, Oct 29, 2020 at 06:44:53PM +0800, Alex Shi wrote:
> >>>> From: Hugh Dickins <hughd@xxxxxxxxxx>
> >>>>
> >>>> It is necessary for page_idle_get_page() to recheck PageLRU() after
> >>>> get_page_unless_zero(), but holding lru_lock around that serves no
> >>>> useful purpose, and adds to lru_lock contention: delete it.
> >>>>
> >>>> See https://lore.kernel.org/lkml/20150504031722.GA2768@blaptop for the
> >>>> discussion that led to lru_lock there; but __page_set_anon_rmap() now
> >>>> uses WRITE_ONCE(),
> >>>
> >>> That doesn't seem to be the case in Linus's or Andrew's tree. Am I
> >>> missing a dependent patch series?
> >>>
> >>>> and I see no other risk in page_idle_clear_pte_refs() using
> >>>> rmap_walk() (beyond the risk of racing PageAnon->PageKsm, mostly but
> >>>> not entirely prevented by page_count() check in ksm.c's
> >>>> write_protect_page(): that risk being shared with page_referenced()
> >>>> and not helped by lru_lock).
> >>>
> >>> Isn't it possible, as per Minchan's description, for page->mapping to
> >>> point to a struct anon_vma without PAGE_MAPPING_ANON set, and rmap
> >>> thinking it's looking at a struct address_space?
> >>
> >> I don't think it can point to an anon_vma without the ANON bit set.
> >> Minchan's concern in that email was that it might still be NULL.
> > 
> > Hm, no, the thread is a lengthy discussion about whether the store
> > could be split such that page->mapping is actually pointing to
> > something invalid (anon_vma without the PageAnon bit).
> > 
> > From his email:
> > 
> >         CPU 0                                                                           CPU 1
> > 
> > do_anonymous_page
> >   __page_set_anon_rmap
> >   /* out of order happened so SetPageLRU is done ahead */
> >   SetPageLRU(page)
> 
> This SetPageLRU done in __pagevec_lru_add_fn() which under the lru_lock
> protection, so the original memory barrier or lock concern isn't
> correct. that means, the SetPageLRU isn't possible to be here.
> And then no warry on right side 'CPU 1' problem.

The SetPageLRU is done under lru_lock, but the store to page->mapping
is not, so what ensures ordering between them? And what prevents the
compiler from tearing the store to page->mapping?

The writer does this:

	CPU 0
	page_add_new_anon_rmap()
	  page->mapping = anon_vma + PAGE_MAPPING_ANON
	lru_cache_add_inactive_or_unevictable()
	  spin_lock(lruvec->lock)
	  SetPageLRU()
	  spin_unlock(lruvec->lock)

The concern is what CPU 1 will observe at page->mapping after
observing PageLRU set on the page.

1. anon_vma + PAGE_MAPPING_ANON

   That's the in-order scenario and is fine.

2. NULL

   That's possible if the page->mapping store gets reordered to occur
   after SetPageLRU. That's fine too because we check for it.

3. anon_vma without the PAGE_MAPPING_ANON bit

   That would be a problem and could lead to all kinds of undesirable
   behavior including crashes and data corruption.

   Is it possible? AFAICT the compiler is allowed to tear the store to
   page->mapping and I don't see anything that would prevent it.

That said, I also don't see how the reader testing PageLRU under the
lru_lock would prevent that in the first place. AFAICT we need that
WRITE_ONCE() around the page->mapping assignment that's referenced in
the changelog of this patch but I cannot find in any tree or patch.