Re: reply: [PATCH] mm: fix a race scenario in folio_isolate_lru

From: Zhaoyang Huang
Date: Mon Mar 18 2024 - 04:01:38 EST


On Mon, Mar 18, 2024 at 2:15 PM Zhaoyang Huang <huangzhaoyang@gmailcom> wrote:
>
> On Mon, Mar 18, 2024 at 11:28 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > On Mon, Mar 18, 2024 at 01:37:04AM +0000, 黄朝阳 (Zhaoyang Huang) wrote:
> > > >On Sun, Mar 17, 2024 at 12:07:40PM +0800, Zhaoyang Huang wrote:
> > > >> Could it be this scenario, where folio comes from pte(thread 0), local
> > > >> fbatch(thread 1) and page cache(thread 2) concurrently and proceed
> > > >> intermixed without lock's protection? Actually, IMO, thread 1 also
> > > >> could see the folio with refcnt==1 since it doesn't care if the page
> > > >> is on the page cache or not.
> > > >>
> > > >> madivise_cold_and_pageout does no explicit folio_get thing since the
> > > >> folio comes from pte which implies it has one refcnt from pagecache
> > > >
> > > >Mmm, no. It's implicit, but madvise_cold_or_pageout_pte_range()
> > > >does guarantee that the folio has at least one refcount.
> > > >
> > > >Since we get the folio from vm_normal_folio(vma, addr, ptent); we know that
> > > >there is at least one mapcount on the folio. refcount is always >= mapcount.
> > > >Since we hold pte_offset_map_lock(), we know that mapcount (and therefore
> > > >refcount) cannot be decremented until we call pte_unmap_unlock(), which we
> > > >don't do until we have called folio_isolate_lru().
> > > >
> > > >Good try though, took me a few minutes of looking at it to convince myself that
> > > >it was safe.
> > > >
> > > >Something to bear in mind is that if the race you outline is real, failing to hold a
> > > >refcount on the folio leaves the caller susceptible to the
> > > >VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio); if the other thread calls
> > > >folio_put().
> > > Resend the chart via outlook.
> > > I think the problem rely on an special timing which is rare, I would like to list them below in timing sequence.
> > >
> > > 1. thread 0 calls folio_isolate_lru with refcnt == 1
> >
> > (i assume you mean refcnt == 2 here, otherwise none of this makes sense)
> >
> > > 2. thread 1 calls release_pages with refcnt == 2.(IMO, it could be 1 as release_pages doesn't care if the folio is used by page cache or fs)
> > > 3. thread 2 decrease refcnt to 1 by calling filemap_free_folio.(as I mentioned in 2, thread 2 is not mandatary here)
> > > 4. thread 1 calls folio_put_testzero and pass.(lruvec->lock has not been take here)
> >
> > But there's already a bug here.
> >
> > Rearrange the order of this:
> >
> > 2. thread 1 calls release_pages with refcount == 2 (decreasing refcount to 1)
> > 3. thread 2 decrease refcount to 0 by calling filemap_free_folio
> > 1. thread 0 calls folio_isolate_lru() and hits the BUG().
> >
> > > 5. thread 0 clear folio's PG_lru by calling folio_test_clear_lru. The folio_get behind has no meaning there.
> > > 6. thread 1 failed in folio_test_lru and leave the folio on the LRU.
> > > 7. thread 1 add folio to pages_to_free wrongly which could break the LRU's->list and will have next folio experience list_del_invalid
> > >
> > > #thread 0(madivise_cold_and_pageout) #1(lru_add_drain->fbatch_release_pages) #2(read_pages->filemap_remove_folios)
> > > refcnt == 1(represent page cache) refcnt==2(another one represent LRU) folio comes from page cache
> >
> > This is still illegible. Try it this way:
> >
> > Thread 0 Thread 1 Thread 2
> > madvise_cold_or_pageout_pte_range
> > lru_add_drain
> > fbatch_release_pages
> > read_pages
> > filemap_remove_folio
> Thread 0 Thread 1 Thread 2
> madvise_cold_or_pageout_pte_range
> truncate_inode_pages_range
> fbatch_release_pages
> truncate_inode_pages_range
> filemap_remove_folio
> Sorry for the confusion. Rearrange the timing chart like above
> according to the real panic's stacktrace. Thread 1&2 are all from
> truncate_inode_pages_range(I think thread2(read_pages) is not
> mandatory here as thread 0&1 could rely on the same refcnt==1).
> >
> > Some accuracy in your report would also be appreciated. There's no
> > function called madivise_cold_and_pageout, nor is there a function called
> > filemap_remove_folios(). It's a little detail, but it's annoying for
> > me to try to find which function you're actually referring to. I have
> > to guess, and it puts me in a bad mood.
> >
> > At any rate, these three functions cannot do what you're proposing.
> > In read_page(), when we call filemap_remove_folio(), the folio in
> > question will not have the uptodate flag set, so can never have been
> > put in the page tables, so cannot be found by madvise().
> >
> > Also, as I said in my earlier email, madvise_cold_or_pageout_pte_range()
> > does guarantee that the refcount on the folio is held and can never
> > decrease to zero while folio_isolate_lru() is running. So that's two
> > ways this scenario cannot happen.
> The madivse_xxx comes from my presumption which has any proof.
> Whereas, It looks like truncate_inode_pages_range just cares about
> page cache refcnt by folio_put_testzero without noticing any task's VM
> stuff. Furthermore, I notice that move_folios_to_lru is safe as it
> runs with holding lruvec->lock.
> >
BTW, I think we need to protect all
folio_test_clear_lru/folio_test_lru by moving them into lruvec->lock
in such as __page_cache_release and folio_activate functions.
Otherwise, there is always a race window between judging PG_lru and
following actions.