[PATCH 00/14] replace call_rcu by kfree_rcu for simple kmem_cache_free callback

[Vlastimil Babka]

On 6/14/24 9:33 PM, Jason A. Donenfeld wrote:
> On Fri, Jun 14, 2024 at 02:35:33PM +0200, Uladzislau Rezki wrote:
>> +	/* Should a destroy process be deferred? */
>> +	if (s->flags & SLAB_DEFER_DESTROY) {
>> +		list_move_tail(&s->list, &slab_caches_defer_destroy);
>> +		schedule_delayed_work(&slab_caches_defer_destroy_work, HZ);
>> +		goto out_unlock;
>> +	}
> 
> Wouldn't it be smoother to have the actual kmem_cache_free() function
> check to see if it's been marked for destruction and the refcount is
> zero, rather than polling every one second? I mentioned this approach
> in: https://lore.kernel.org/all/Zmo9-YGraiCj5-MI@zx2c4.com/ -
> 
>     I wonder if the right fix to this would be adding a `should_destroy`
>     boolean to kmem_cache, which kmem_cache_destroy() sets to true. And
>     then right after it checks `if (number_of_allocations == 0)
>     actually_destroy()`, and likewise on each kmem_cache_free(), it
>     could check `if (should_destroy && number_of_allocations == 0)
>     actually_destroy()`. 

I would prefer not to affect the performance of kmem_cache_free() by doing
such checks, if possible. Ideally we'd have a way to wait/poll for the
kfree_rcu() "grace period" expiring even with the batching that's
implemented there. Even if it's pesimistically long to avoid affecting
kfree_rcu() performance. The goal here is just to print the warnings if
there was a leak and the precise timing of them shouldn't matter. The owning
module could be already unloaded at that point? I guess only a kunit test
could want to be synchronous and then it could just ask for
kmem_cache_free() to wait synchronously.

> Jason