Re: [PATCH -mm 7/8] slub: make dead caches discard free slabs immediately

From: Joonsoo Kim
Date: Wed Jun 04 2014 - 04:53:40 EST


2014-06-03 17:16 GMT+09:00 Vladimir Davydov <vdavydov@xxxxxxxxxxxxx>:
> On Mon, Jun 02, 2014 at 11:03:51PM +0900, Joonsoo Kim wrote:
>> 2014-06-02 20:47 GMT+09:00 Vladimir Davydov <vdavydov@xxxxxxxxxxxxx>:
>> > Hi Joonsoo,
>> >
>> > On Mon, Jun 02, 2014 at 01:24:36PM +0900, Joonsoo Kim wrote:
>> >> On Sat, May 31, 2014 at 03:04:58PM +0400, Vladimir Davydov wrote:
>> >> > On Fri, May 30, 2014 at 09:57:10AM -0500, Christoph Lameter wrote:
>> >> > > On Fri, 30 May 2014, Vladimir Davydov wrote:
>> >> > >
>> >> > > > (3) is a bit more difficult, because slabs are added to per-cpu partial
>> >> > > > lists lock-less. Fortunately, we only have to handle the __slab_free
>> >> > > > case, because, as there shouldn't be any allocation requests dispatched
>> >> > > > to a dead memcg cache, get_partial_node() should never be called. In
>> >> > > > __slab_free we use cmpxchg to modify kmem_cache_cpu->partial (see
>> >> > > > put_cpu_partial) so that setting ->partial to a special value, which
>> >> > > > will make put_cpu_partial bail out, will do the trick.
>> > [...]
>> >> I think that we can do (3) easily.
>> >> If we check memcg_cache_dead() in the end of put_cpu_partial() rather
>> >> than in the begin of put_cpu_partial(), we can avoid the race you
>> >> mentioned. If someone do put_cpu_partial() before dead flag is set,
>> >> it can be zapped by who set dead flag. And if someone do
>> >> put_cpu_partial() after dead flag is set, it can be zapped by who
>> >> do put_cpu_partial().
>> >
>> > After put_cpu_partial() adds a frozen slab to a per cpu partial list,
>> > the slab becomes visible to other threads, which means it can be
>> > unfrozen and freed. The latter can trigger cache destruction. Hence we
>> > shouldn't touch the cache, in particular call memcg_cache_dead() on it,
>> > after calling put_cpu_partial(), otherwise we can get use-after-free.
>> >
>> > However, what you propose makes sense if we disable irqs before adding a
>> > slab to a partial list and enable them only after checking if the cache
>> > is dead and unfreezing all partials if so, i.e.
>> >
>> > diff --git a/mm/slub.c b/mm/slub.c
>> > index d96faa2464c3..14b9e9a8677c 100644
>> > --- a/mm/slub.c
>> > +++ b/mm/slub.c
>> > @@ -2030,8 +2030,15 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
>> > struct page *oldpage;
>> > int pages;
>> > int pobjects;
>> > + unsigned long flags;
>> > + int irq_saved = 0;
>> >
>> > do {
>> > + if (irq_saved) {
>> > + local_irq_restore(flags);
>> > + irq_saved = 0;
>> > + }
>> > +
>> > pages = 0;
>> > pobjects = 0;
>> > oldpage = this_cpu_read(s->cpu_slab->partial);
>> > @@ -2062,8 +2069,16 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
>> > page->pobjects = pobjects;
>> > page->next = oldpage;
>> >
>> > + local_irq_save(flags);
>> > + irq_saved = 1;
>> > +
>> > } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
>> > != oldpage);
>> > +
>> > + if (memcg_cache_dead(s))
>> > + unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
>> > +
>> > + local_irq_restore(flags);
>> > #endif
>> > }
>> >
>> >
>> > That would be safe against possible cache destruction, because to remove
>> > a slab from a per cpu partial list we have to run on the cpu it was
>> > frozen on. Disabling irqs makes it impossible.
>>
>> Hmm... this is also a bit ugly.
>> How about following change?
>>
>> Thanks.
>>
>> diff --git a/mm/slub.c b/mm/slub.c
>> index 2b1ce69..6adab87 100644
>> --- a/mm/slub.c
>> +++ b/mm/slub.c
>> @@ -2058,6 +2058,21 @@ static void put_cpu_partial(struct kmem_cache
>> *s, struct page *page, int drain)
>>
>> } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
>> != oldpage);
>> +
>> + if (memcg_cache_dead(s)) {
>> + bool done = false;
>> + unsigned long flags;
>
> Suppose we are preempted here. In the meanwhile all objects are freed to
> the cache, all frozen pages are unfrozen and also freed. The cache
> destruction is then scheduled (patch 2 of this set). Then when this
> thread continues execution it will operate on the cache that was
> destroyed - use-after-free.
>
> I admit, this is very unlikely, but can we ignore this possibility?
>

Hello,

>From your comment, now, I realize that your cache destruction solution
has severe problem.

With you solution, kmem_cache can be destroyed before last kfree() caller
has returned. It means that we can't safely do anything related to
the kmem_cache after losing control about that slab where we try to free
object in free path.

Consider __slab_free(). After put_cpu_partial() in __slab_free() is called,
we attempt to update stat. There is possibility that this operation could be
use-after-free with your solution. Until now, we have just stat operation, but
it could be more. I don't like to impose this constraint to the slab free path.
So IMHO, it is better that we should defer to destroy kmem_cache
until last kfree() caller returns. Is it fair enough? :)

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/