Re: [PATCH net v2] packet: fix use after free race in send path when dev is released

[Daniel Borkmann <dborkman@redhat.com>]

On 11/21/2013 04:27 PM, Eric Dumazet wrote:
> On Thu, 2013-11-21 at 16:02 +0100, Daniel Borkmann wrote:
>
>>
>> That was also my first thought, but Salam pointed out to me, that in case
>> we have a situation such as ...
>>
>> in packet_cached_dev_get():
>>       rcu_read_lock();
>>       dev = rcu_dereference(po->cached_dev);      in packet_notifier():
>>                                                   ---> CPU1: dev_put(po->prot_hook.dev);
>> ---> CPU0:
>>       if (dev)
>> 	dev_hold(dev);
>>       rcu_read_unlock();
>>
>> ... we could reach a refcount of 0, before we increase it back to 1. Not sure
>> if this can actually happen, maybe in preemptible RCU where read-side critical
>> sections to be preempted? So with this rather paranoid approach we make sure
>> to avoid such a situation as we wait a grace period when readers finished.
>
> There is no need, because we respect a rcu grace period at dismantle
> time already.
>
> Nothing bad can happen inside the rcu_read_lock()/rcu_read_unlock() pair
> in packed_cached_dev_get()
>
> Note that dev_put() does not take any immediate action, it only
> decrements the refcount.
>
> So if CPU1 does the dev_put(po->prot_hook.dev) after
> setting cached_dev to NULL, we should be safe.

Ok, then I'll update the patch and send out v3 with that removed.

Thanks for the input Eric!