On 6/5/20 8:46 PM, Kees Cook wrote:
>>
>> Hmm I have a different idea. The whole cache_from_obj() was added because of
>> kmemcg (commit b9ce5ef49f00d) where per-memcg cache can be different from the
>> root one. And I just realized this usecase can go away with Roman's series [1].
>> But cache_from_obj() also kept the original SLUB consistency check case, and you
>> added the freelist hardening case. If kmemcg use case went away it would be nice
>> to avoid the virt_to_cache() and check completely again, unless in debugging or
>> hardened kernel.
>
> Is it that expensive? (I'm fine with it staying behind debug/hardening,
> but if we can make it on by default, that'd be safer.)
Well, it's fast path and e.g. networking guys did a lot of work to optimize
SLUB. If we decide to stop trusting the supplied cache pointer completely, we
can deprecate kmem_cache_free() and use kfree() everywhere (SLOB would need some
adjustments to store size with each object like for kmalloc) but it would have
to be a conscious decision.
>> Furthermore, the original SLUB debugging case was an unconditional pr_err() plus
>> WARN_ON_ONCE(1), which was kept by commit b9ce5ef49f00d. With freelist
>> hardening this all changed to WARN_ONCE. So the second and later cases are not
>> reported at all for hardening and also not for explicitly enabled debugging like
>> in this case, which is IMHO not ideal.
>
> Oh, I have no problem with WARN vs WARN_ONCE -- there's no reason to
> split this. And I'd love the hardening side to gain the tracking call
> too, if it's available.
>
> I had just used WARN_ONCE() since sometimes it can be very noisy to keep
> warning for some condition that might not be correctable.
OK.
>> So I propose the following - the freelist hardening case keeps the WARN_ONCE,
>> but also a one-line pr_err() for each case so they are not silent. The SLUB
>> debugging case is always a full warning, and printing the tracking info if
>> enabled and available. Pure kmemcg case does virt_to_cache() for now (until
>> hopefully removed by Roman's series) but no checking at all. Would that work for
>> everyone?
>> [...]
>> @@ -520,9 +528,18 @@ static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
>> return s;
>>
>> cachep = virt_to_cache(x);
>> - WARN_ONCE(cachep && !slab_equal_or_root(cachep, s),
>> - "%s: Wrong slab cache. %s but object is from %s\n",
>> - __func__, s->name, cachep->name);
>> + if (unlikely(s->flags & SLAB_CONSISTENCY_CHECKS)) {
>> + 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))
>> + slab_print_tracking(cachep, x);
>> + } else if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED)) {
>> + if (unlikely(cachep && !slab_equal_or_root(cachep, s))) {
>> + pr_err("%s: Wrong slab cache. %s but object is from %s\n",
>> + __func__, s->name, cachep->name);
>> + WARN_ON_ONCE(1);
>> + }
>> + }
>
> How about just this (in addition to your slab_print_tracking() refactor):
That could work, I will send a proper patch.
> diff --git a/mm/slab.h b/mm/slab.h
> index 207c83ef6e06..107b7f6db3c3 100644
> --- a/mm/slab.h
> +++ b/mm/slab.h
> @@ -520,9 +520,10 @@ static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
> return s;
>
> cachep = virt_to_cache(x);
> - WARN_ONCE(cachep && !slab_equal_or_root(cachep, s),
> + 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);
> + __func__, s->name, cachep->name))
> + slab_print_tracking(cachep, x);
> return cachep;
> }
>
>
