Re: [RFC PATCH 2/2] x86/ibpb: Prevent missed IBPB flush

[Tim Chen]

On 01/25/2018 12:20 AM, David Woodhouse wrote:
> On Wed, 2018-01-24 at 16:36 -0800, Tim Chen wrote:
>> It is possible that the last uesr mm that we recorded for a cpu was
>> released, and a new mm with identical address was allocated when we
>> check it again.  We could skip IBPB flush here for the process with
>> the new mm.
>>
>> It is a difficult to exploit case as we have to exit() a process on a
>> cpu, free the mm, and fork() the victim to use the mm pointer on that
>> cpu. The exploiter needs the old mm to get recycled to the
>> newly forked process and no other processes run on the target cpu.
> 
> That's what it takes to have the victim process leak information into
> the cache. In order to *harvest* that information, the attacker must
> then get run on the same CPU again? And since her first process had to
> exits, as described above, she needs a new process for that?
> 
> I confess, with all the other wildly theoretical loopholes that exist,
> I wasn't losing much sleep over this one.
> 
>> Nevertheless, the patch below is one way to close this hole by
>> adding a ref count to prevent the last user mm from being released.
>> It does add ref counting overhead, and extra memory cost of keeping an mm
>> (though not the VMAs and most of page tables) around longer than we will
>> otherwise need to. Any better solutions are welcomed.
> 
> This has no upper bound on the amount of time the user mm gets held,
> right? If a given CPU remains idle for ever (and what happens if it's
> taken offline?) we'll never do that mmdrop() ?
> 

The downside with this approach is we do hold on to the mm longer
than we needs to.

Yes, the offline path does need to be fixed up.

Tim