[Intel-gfx] SLAB_TYPESAFE_BY_RCU without constructors (was Re: [PATCH v4 13/17] khwasan: add hooks implementation)

[Eric Dumazet]

On Tue, Jul 31, 2018 at 10:10 AM Florian Westphal <fw at strlen.de> wrote:
>
> Andrey Ryabinin <aryabinin at virtuozzo.com> wrote:
> > Guys, it seems that we have a lot of code using SLAB_TYPESAFE_BY_RCU cache without constructor.
> > I think it's nearly impossible to use that combination without having bugs.
> > It's either you don't really need the SLAB_TYPESAFE_BY_RCU, or you need to have a constructor in kmem_cache.
> >
> > Could you guys, please, verify your code if it's really need SLAB_TYPSAFE or constructor?
> >
> > E.g. the netlink code look extremely suspicious:
> >
> >       /*
> >        * Do not use kmem_cache_zalloc(), as this cache uses
> >        * SLAB_TYPESAFE_BY_RCU.
> >        */
> >       ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
> >       if (ct == NULL)
> >               goto out;
> >
> >       spin_lock_init(&ct->lock);
> >
> > If nf_conntrack_cachep objects really used in rcu typesafe manner, than 'ct' returned by kmem_cache_alloc might still be
> > in use by another cpu. So we just reinitialize spin_lock used by someone else?
>
> That would be a bug, nf_conn objects are reference counted.
>
> spinlock can only be used after object had its refcount incremented.
>
> lookup operation on nf_conn object:
> 1. compare keys
> 2. attempt to obtain refcount (using _not_zero version)
> 3. compare keys again after refcount was obtained
>
> if any of that fails, nf_conn candidate is skipped.


Yes, the key here is the refcount, this is only what we need to clear
after kmem_cache_alloc()

By definition, if an object is being freed/reallocated, the refcount
should be already 0, and clearing it again is a NOP.