Re: [PATCH RFC 2/6] mm, slab: always maintain per-node slab and object count
From: David Rientjes
Date: Sat Jul 20 2024 - 22:37:58 EST
On Mon, 15 Jul 2024, Vlastimil Babka wrote:
> Currently SLUB counts per-node slabs and total objects only with
> CONFIG_SLUB_DEBUG, in order to minimize overhead. However, the detection
> in __kmem_cache_shutdown() whether there are no outstanding object
> relies on the per-node slab count (node_nr_slabs()) so it may be
> unreliable without CONFIG_SLUB_DEBUG. Thus we might be failing to warn
> about such situations, and instead destroy a cache while leaving its
> slab(s) around (due to a buggy slab user creating such a scenario, not
> in normal operation).
>
> We will also need node_nr_slabs() to be reliable in the following work
> to gracefully handle kmem_cache_destroy() with kfree_rcu() objects in
> flight. Thus make the counting of per-node slabs and objects
> unconditional.
>
> Note that CONFIG_SLUB_DEBUG is the default anyway, and the counting is
> done only when allocating or freeing a slab page, so even in
> !CONFIG_SLUB_DEBUG configs the overhead should be negligible.
>
> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
Acked-by: David Rientjes <rientjes@xxxxxxxxxx>