Re: [PATCH] slub: make sure struct kmem_cache_node is initialized before publication

From: Christoph Lameter
Date: Mon Jul 10 2017 - 11:32:44 EST


On Mon, 10 Jul 2017, Alexander Potapenko wrote:

> >> Could the slab maintainers please take a look at these and also have a
> >> think about Alexander's READ_ONCE/WRITE_ONCE question?
> >
> > Was I cced on these?
> I've asked Andrew about READ_ONCE privately.

Please post to a mailing list and cc the maintainers?

> Since unfreeze_partials() sees uninitialized value of n->list_lock, I
> was suspecting there's a data race between unfreeze_partials() and
> init_kmem_cache_nodes().

I have not seen the details but I would suspect that this is related to
early boot issues? The list lock is initialized upon slab creation and at
that time no one can get to the kmem_cache structure.

There are a couple of boot time slabs that will temporarily be available.
and released upon boot completion.

> If so, reads and writes to s->node[node] must be acquire/release
> atomics (not actually READ_ONCE/WRITE_ONCE, but
> smp_load_acquire/smp_store_release).

Can we figure the reason for these out before proposing fixes?