Re: [PATCH 9/9] mm, slab/slub: move and improve cache_from_obj()

[Roman Gushchin <guro@fb.com>]

On Thu, Jun 11, 2020 at 11:56:53AM +0200, Vlastimil Babka wrote:
> On 6/11/20 12:46 AM, Roman Gushchin wrote:
> > On Wed, Jun 10, 2020 at 06:31:35PM +0200, Vlastimil Babka wrote:
> >> @@ -3672,6 +3672,14 @@ void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
> >>  }
> >>  EXPORT_SYMBOL(__kmalloc_track_caller);
> >>  
> >> +static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
> >> +{
> >> +	if (memcg_kmem_enabled())
> >> +		return virt_to_cache(x);
> >> +	else
> >> +		return s;
> >> +}
> > 
> > Hm, it looks like all the SLAB version doesn't perform any sanity checks anymore.
> > Is it intended?
> 
> Yes, it was the same before commit b9ce5ef49f00. The commit could have been more
> precise - kmemcg needs virt_to_cache(), but not the sanity check. The SLUB
> version also shouldn't really be doing the sanity check if only
> memcg_kmem_enabled() is true (and not the debugging/hardening), but the code
> then looks ugly and I hope this will just fix itself with your kmemcg slab rework.

Got it.

> 
> > Also, Is it ever possible that s != virt_to_cache(x) if there are no bugs?
> 
> Well, only in the kmemcg case it should be possible.
> 
> > kmem_cache_free_bulk() in slab.c does contain the following:
> > 	if (!orig_s) /* called via kfree_bulk */
> > 		s = virt_to_cache(objp);
> > 	else
> > 		s = cache_from_obj(orig_s, objp);
> > which looks a bit strange with the version above.
> 
> Looks fine to me. If we are called with non-NULL s, and kmemcg is not enabled,
> we can just trust s. If we are called with NULL s (via kfree_bulk()) we need to
> get cache from the object, even if kmemcg is not enabled, so we do
> virt_to_cache() unconditionally.
> Once your series is fully accepted, we can remove SLAB's cache_from_obj() and
> the whole 'else' part in the hunk above. Or am I missing something?

Right. I guess there will be even more cleanups possible, let's see where we'll end up.
It looks like nothing prevents it from being queued for 5.9 after 5.8-rc1 will be out,
right?

> 
> 
> >> @@ -3175,6 +3179,23 @@ void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
> >>  }
> >>  #endif
> >>  
> >> +static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
> >> +{
> >> +	struct kmem_cache *cachep;
> >> +
> >> +	if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
> >> +	    !memcg_kmem_enabled() &&
> >> +	    !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
> >> +		return s;
> >> +
> >> +	cachep = virt_to_cache(x);
> >> +	if (WARN(cachep && !slab_equal_or_root(cachep, s),
> >> +		  "%s: Wrong slab cache. %s but object is from %s\n",
> >> +		  __func__, s->name, cachep->name))
> >> +		print_tracking(cachep, x);
> >> +	return cachep;
> >> +}
> > 
> > Maybe we can define a trivial SLAB version of kmem_cache_debug_flags()
> > and keep a single version of cache_from_obj()?
> 
> I think the result would be more obfuscated than just making it plain that SLAB
> doesn't have those SLUB features. And I still hope SLAB's version will go away
> completely. If your series is accepted first, then this patch based in that will
> not introduce slab.c cache_from_obj() at all.

Ok, makes sense to me.

Thank you!