There used to be a rcu_barrier() for SLAB_TYPESAFE_BY_RCU caches in
kmem_cache_destroy() until commit 657dc2f9 ("slab: remove
synchronous rcu_barrier() call in memcg cache release path") moved it to
an asynchronous work that finishes the destroying of such caches.

The motivation for that commit was the MEMCG_KMEM integration that at
the time created and removed clones of the global slab caches together
with their cgroups, and blocking cgroups removal was unwelcome. The
implementation later changed to per-object memcg tracking using a single
cache, so there should be no more need for a fast non-blocking
kmem_cache_destroy(), which is typically only done when a module is
unloaded etc.

Going back to synchronous barrier has the following advantages:

- simpler implementation
- it's easier to test the result of kmem_cache_destroy() in a kunit test

Thus effectively revert commit 657dc2f9. It is not a 1:1 revert as
the code has changed since. The main part is that kmem_cache_release(s)
is always called from kmem_cache_destroy(), but for SLAB_TYPESAFE_BY_RCU
caches there's a rcu_barrier() first.
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Reviewed-by: Jann Horn <jannh@google.com>
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>