On Wed, 2021-07-28 at 18:59 +0200, Vlastimil Babka wrote:

On 7/27/21 6:09 AM, Mike Galbraith wrote:

quoted

On Mon, 2021-07-26 at 23:26 +0200, Vlastimil Babka wrote:

quoted

On 7/26/21 7:00 PM, Mike Galbraith wrote:

quoted

Why not do something like the below?...

Yep, sounds like a good approach, thanks. Percpu partial is not *the*
SLUB fast path, so it should be sufficient without the lockless cmpxchg
tricks. Will incorporate in updated series.

The updated series incorporating hopefully all fixes from Mike and
bigeasy, and rebased to 5.14-rc3 (Thomas told me RT is moving to it), is
here:

https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/linux.git/log/?h=slub-local-lock-v3r0

I had to resurrect the hunk below to build with lockdep, but modulo
dinky speedbump, the same RT testdrive that previously exploded was as
entertainment free as such testing is supposed to be.

---
 mm/slub.c |    4 ++++
 1 file changed, 4 insertions(+)

--- a/mm/slub.c
+++ b/mm/slub.c

@@ -2890,7 +2890,11 @@ static void *___slab_alloc(struct kmem_c

 load_freelist:

+#ifdef CONFIG_PREEMPT_RT
+	lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock.lock));
+#else
 	lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
+#endif

 	/*
 	 * freelist is pointing to the list of objects to be used.