Re: kernel BUG at fs/ext4/inode.c:LINE!

[Linus Torvalds]

On Tue, Nov 24, 2020 at 10:33 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> We could fix this by turning that 'if' into a 'while' in
> write_cache_pages().

That might be the simplest patch indeed.

At the same time, I do worry about other cases like this: while
spurious wakeup events are normal and happen in other places, this is
a bit different.

This is literally a wakeup that leaks from a previous use of a page,
and makes us think that something could have happened to the new use.

The unlock_page() case presumably never hits that, because even if we
have some unlock without a page ref (which I don't think can happen,
but whatever..), the exclusive nature of "lock_page()" means that no
locker can care - once you get the lock, you own the page./

The writeback code is special in that the writeback bit isn't some
kind of exclusive bit, but this code kind of expected it to be that.

So I'd _like_ to have something like

        WARN_ON_ONCE(!page_count(page));

in the wake_up_page_bit() function, to catch things that wake up a
page that has already been released and might be reused..

And that would require the "get_page()" to be done when we set the
writeback bit and queue the page up for IO (so that then
end_page_writeback() would clear the bit, do the wakeup, and then drop
the ref).

Hugh's second patch isn't pretty - I think the "get_page()" is
conceptually in the wrong place - but it "works" in that it keeps that
"implicit page reference" being kept by the PG_writeback bit, and then
it takes an explicit page reference before it clears the bit.

So while I don't love the whole "PG_writeback is an implicit reference
to the page" model, Hugh's patch at least makes that model much more
straightforward: we really either have that PG_writeback, _or_ we have
a real ref to the page, and we never have that odd "we could actually
lose the page" situation.

So I think I prefer Hugh's two-liner over your one-liner suggestion.

But your one-liner is technically not just smaller, it obviously also
avoids the whole mucking with the atomic page ref.

I don't _think_ that the extra get/put overhead could possibly really
matter: doing the writeback is going to be a lot more expensive
anyway. And an atomic access to a 'struct page' sounds expensive, but
that cacheline is already likely dirty in the L1 cache because we've
touch page->flags and done other things to it).

So I'd personally be inclined to go with Hugh's patch. Comments?

                 Linus