Re: [PATCH] mm: Remove bogus VM_BUG_ON

[Hugh Dickins <hughd@google.com>]

On Wed, 18 Aug 2021, Matthew Wilcox wrote:
> On Wed, Aug 18, 2021 at 09:34:51AM -0700, Hugh Dickins wrote:
> > On Wed, 18 Aug 2021, Matthew Wilcox (Oracle) wrote:
> > 
> > > It is not safe to check page->index without holding the page lock.
> > > It can be changed if the page is moved between the swap cache and the
> > > page cache for a shmem file, for example.  There is a VM_BUG_ON below
> > > which checks page->index is correct after taking the page lock.
> > > 
> > > Cc: stable@vger.kernel.org
> > > Fixes: 5c211ba29deb ("mm: add and use find_lock_entries")
> > 
> > I don't mind that VM_BUG_ON_PAGE() being removed, but question whether
> > this Fixes anything, and needs to go to stable. Or maybe it's just that
> > the shmem example is wrong - moving shmem from page to swap cache does
> > not change page->index. Or maybe you have later changes in your tree
> > which change that and do require this. Otherwise, I'll have to worry
> > why my testing has missed it for six months.
> 
> I'm sorry, I think you're going to have to worry :-(

Indeed, it seems that way; or maybe I can leave this testing to syzbot.

> Syzbot found it initially:
> 
> https://lore.kernel.org/linux-mm/0000000000009cfcda05c926b34b@google.com/

Ah, that's useful info.  Though I can quite see why you didn't mention
that originally: it looks as if syzbot hit a find_lock_entries() crash
and an irqstate warning about the same time, and its bisection went off
and found the commit that introduced those irqstate warnings: neither
the root cause of the irqstate warning, nor the cause of the
find_lock_entries() crash which it claims in the Subject.

I have briefly tried the C reproducer, but didn't get anything out of it;
and suspect it may be a reproducer of the irqstate warning rather than
the crash which interests you and me.  And I can't tell more from the
dump, no dump_page() info is shown, and the "Code:" just points into a
function epilog of assorted ud2s.

> 
> and then I hit it today during my testing (which is definitely due to
> further changes in my tree).

Okay, and it's perfectly reasonable for your tree to make changes which
require that VM_BUG_ON_PAGE to be removed. But I do not yet understand
why it needs to be removed from the current or stable tree.

I don't believe it has anything to do with swap cache. The reproducer
is mounting with "huge=within_size", and doing lots of truncation: my
supposition is that a shmem THP is being collapsed or split,
concurrently with that find_lock_entries().

I don't actually see how that would lead to this VM_BUG_ON_PAGE:
I imagine find_get_entry()'s xas_reload check after get_speculative
should be good enough - but don't know my way around XArray,
so mention this in case it triggers an Aha from you.

While there's certainly a sense in which removing the VM_BUG_ON_PAGE
removes the root cause of the crash, I don't think we understand
what is going on here yet: and therefore I'm reluctant to remove it.

But I have not given this issue much time, busy with other stuff.

> 
> I should have added:
> 
> Reported-by: syzbot+c87be4f669d920c76330@syzkaller.appspotmail.com

That's fair, it did report it, if confusingly.

Hugh

(p.s. in parentheses, to minimize confusion from going slightly
off-topic, but I think I'd be wrong not to mention a separate
issue in this area, with mmotm and linux-next since your folios
went in: doesn't happen easily, but I have twice hit the
include/linux/pagemap.h:258 VM_BUG_ON_PAGE(PageTail(page), page),
in page_cache_add_speculative() - both times when serving
filemap_map_pages(). I have not thought about it at all, but
expect that when you do, you'll simply decide that one is unsafe
and has to be deleted.)