Re: kvm: use-after-free in process_srcu

[Dmitry Vyukov]

On Tue, Jan 17, 2017 at 10:47 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> On Mon, Jan 16, 2017 at 10:48 PM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>>
>>
>> ----- Original Message -----
>>> From: "Dmitry Vyukov" <dvyukov@xxxxxxxxxx>
>>> To: "Steve Rutherford" <srutherford@xxxxxxxxxx>
>>> Cc: "syzkaller" <syzkaller@xxxxxxxxxxxxxxxx>, "Paolo Bonzini" <pbonzini@xxxxxxxxxx>, "Radim KrÄmÃÅ"
>>> <rkrcmar@xxxxxxxxxx>, "KVM list" <kvm@xxxxxxxxxxxxxxx>, "LKML" <linux-kernel@xxxxxxxxxxxxxxx>
>>> Sent: Monday, January 16, 2017 10:34:26 PM
>>> Subject: Re: kvm: use-after-free in process_srcu
>>>
>>> On Sun, Jan 15, 2017 at 6:11 PM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>>> > On Fri, Jan 13, 2017 at 10:19 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>>> >> On Fri, Jan 13, 2017 at 4:30 AM, Steve Rutherford
>>> >> <srutherford@xxxxxxxxxx> wrote:
>>> >>> I'm not that familiar with the kernel's workqueues, but this seems
>>> >>> like the classic "callback outlives the memory it references"
>>> >>> use-after-free, where the process_srcu callback is outliving struct
>>> >>> kvm (which contains the srcu_struct). If that's right, then calling
>>> >>> srcu_barrier (which should wait for all of the call_srcu callbacks to
>>> >>> complete, which are what enqueue the process_srcu callbacks) before
>>> >>> cleanup_srcu_struct in kvm_destroy_vm probably fixes this.
>>> >>>
>>> >>> The corresponding patch to virt/kvm/kvm_main.c looks something like:
>>> >>> static void kvm_destroy_vm(struct kvm *kvm)
>>> >>> {
>>> >>> ...
>>> >>>         for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++)
>>> >>>                 kvm_free_memslots(kvm, kvm->memslots[i]);
>>> >>> +      srcu_barrier(&kvm->irq_srcu);
>>> >>>         cleanup_srcu_struct(&kvm->irq_srcu);
>>> >>> +      srcu_barrier(&kvm->srcu);
>>> >>>         cleanup_srcu_struct(&kvm->srcu);
>>> >>> ...
>>> >>>
>>> >>>
>>> >>> Since we don't have a repro, this obviously won't be readily testable.
>>> >>> I find srcu subtle enough that I don't trust my reasoning fully (in
>>> >>> particular, I don't trust that waiting for all of the call_srcu
>>> >>> callbacks to complete also waits for all of the process_srcu
>>> >>> callbacks). Someone else know if that's the case?
>>> >>
>>> >>
>>> >> From the function description it looks like it should do the trick:
>>> >>
>>> >> 514 /**
>>> >> 515  * srcu_barrier - Wait until all in-flight call_srcu() callbacks
>>> >> complete.
>>> >> 516  * @sp: srcu_struct on which to wait for in-flight callbacks.
>>> >> 517  */
>>> >> 518 void srcu_barrier(struct srcu_struct *sp)
>>> >>
>>> >> I see this failure happening several times per day. I've applied your
>>> >> patch locally and will check if I see these failures happening.
>>> >
>>> >
>>> > I have not seen the crash in 3 days, when usually I see several
>>> > crashes per night. So I think we can consider that the patch fixes the
>>> > crash:
>>> >
>>> > Tested-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
>>>
>>>
>>> Unfortunately I hit it again with the patch applied. It definitely
>>> happens less frequently now, but still happens:
>>
>> Try this:
>>
>> diff --git a/kernel/rcu/srcu.c b/kernel/rcu/srcu.c
>> index 9b9cdd549caa..ef5599c65299 100644
>> --- a/kernel/rcu/srcu.c
>> +++ b/kernel/rcu/srcu.c
>> @@ -283,6 +283,7 @@ void cleanup_srcu_struct(struct srcu_struct *sp)
>>  {
>>         if (WARN_ON(srcu_readers_active(sp)))
>>                 return; /* Leakage unless caller handles error. */
>> +       flush_delayed_work(&sp->work);
>>         free_percpu(sp->per_cpu_ref);
>>         sp->per_cpu_ref = NULL;
>>  }
>>
>> I think it should subsume Steve's patch, but I'm not 101% sure.  We
>> will have to run this through Paul.
>
>
> Hi +Pual,
>
> I am seeing use-after-frees in process_srcu as struct srcu_struct is
> already freed. Before freeing struct srcu_struct, code does
> cleanup_srcu_struct(&kvm->irq_srcu). We also tried to do:
>
> +      srcu_barrier(&kvm->irq_srcu);
>          cleanup_srcu_struct(&kvm->irq_srcu);
>
> It reduced rate of use-after-frees, but did not eliminate them
> completely. The full threaded is here:
> https://groups.google.com/forum/#!msg/syzkaller/i48YZ8mwePY/0PQ8GkQTBwAJ
>
> Does Paolo's fix above make sense to you? Namely adding
> flush_delayed_work(&sp->work) to cleanup_srcu_struct()?


I am not sure about interaction of flush_delayed_work and
srcu_reschedule... flush_delayed_work probably assumes that no work is
queued concurrently, but what if srcu_reschedule queues another work
concurrently... can't it happen that flush_delayed_work will miss that
newly scheduled work?

Meanwhile I will apply this change instead of Steve's change and see
what happens.