Re: [rfc/patch] mm/slub: restore/expand unfreeze_partials() local exclusion scope

[Mike Galbraith <efault@gmx.de>]

On Sat, 2021-07-24 at 00:39 +0200, Vlastimil Babka wrote:
> On 7/21/21 11:33 AM, Mike Galbraith wrote:
> > On Wed, 2021-07-21 at 10:44 +0200, Vlastimil Babka wrote:
> > >
> > > So this doesn't look like our put_cpu_partial() preempted a
> > > __slab_alloc() on the same cpu, right?
> >
> > No, likely it was the one preempted by someone long gone, but we'll
> > never know without setting a trap.
> >
> > > BTW did my ugly patch work?
> >
> > Nope.  I guess you missed my reporting it to have been a -ENOBOOT, and
>
> Indeed, I misunderstood it as you talking about your patch.
>
> > that cutting it in half, ie snagging only __slab_free() does boot, and
> > seems to cure all of the RT fireworks.
>
> OK, so depending on drain=1 makes this apply only to put_cpu_partial()
> called from __slab_free and not get_partial_node(). One notable
> difference is that in __slab_free we don't have n->list_lock locked and
> in get_partial_node() we do. I guess in case your list_lock is made raw
> again by another patch, that explains a local_lock can't nest under it.
> If not, then I would expect this to work (I don't think they ever nest
> in the opposite order, also lockdep should tell us instead of
> -ENOBOOT?), but might be missing something...

RT used to convert list_lock to raw_spinlock_t, but no longer does.
Whatever is going on, box does not emit a single sign of life with the
full patch.

> I'd rather not nest those locks in any case. I just need to convince
> myself that the scenario the half-fix fixes is indeed the only one
> that's needed and we're not leaving there other races that are just
> harder to trigger...

Yup.  I can only state with confidence that the trouble I was able to
easily reproduce was fixed up by serializing __slab_free(). Hopefully
you'll find that's the only hole in need of plugging.

	-Mike