Messages in this thread | | | Date | Mon, 9 Jun 2014 16:52:13 +0400 | From | Vladimir Davydov <> | Subject | Re: [PATCH -mm v2 5/8] slub: make slab_free non-preemptable |
| |
On Fri, Jun 06, 2014 at 09:46:57AM -0500, Christoph Lameter wrote: > On Fri, 6 Jun 2014, Vladimir Davydov wrote: > > > This patch makes SLUB's implementation of kmem_cache_free > > non-preemptable. As a result, synchronize_sched() will work as a barrier > > against kmem_cache_free's in flight, so that issuing it before cache > > destruction will protect us against the use-after-free. > > > Subject: slub: reenable preemption before the freeing of slabs from slab_free > > I would prefer to call the page allocator with preemption enabled if possible. > > Signed-off-by: Christoph Lameter <cl@linux.com> > > Index: linux/mm/slub.c > =================================================================== > --- linux.orig/mm/slub.c 2014-05-29 11:45:32.065859887 -0500 > +++ linux/mm/slub.c 2014-06-06 09:45:12.822480834 -0500 > @@ -1998,6 +1998,7 @@ > if (n) > spin_unlock(&n->list_lock); > > + preempt_enable();
The whole function (unfreeze_partials) is currently called with irqs off, so this is effectively a no-op. I guess we can restore irqs here though.
> while (discard_page) { > page = discard_page; > discard_page = discard_page->next; > @@ -2006,6 +2007,7 @@ > discard_slab(s, page);
If we just freed the last slab of the cache and then get preempted (suppose we restored irqs above), nothing will prevent the cache from destruction, which may result in use-after-free below. We need to be more cautious if we want to call for page allocator with preemption and irqs on.
However, I still don't understand what's the point in it. We *already* call discard_slab with irqs disabled, which is harder, and it haven't caused any problems AFAIK. Moreover, even if we enabled preemption/irqs, it wouldn't guarantee that discard_slab would always be called with preemption/irqs on, because the whole function - I mean kmem_cache_free - can be called with preemption/irqs disabled.
So my point it would only complicate the code.
Thanks.
> stat(s, FREE_SLAB); > } > + preempt_disable(); > #endif > } >
| |