Bernd Edlinger <bernd.edlinger@xxxxxxxxxx> writes:
> On 3/11/20 1:15 AM, Eric W. Biederman wrote:
>> Jann Horn <jannh@xxxxxxxxxx> writes:
>>
>>> On Tue, Mar 10, 2020 at 10:33 PM Eric W. Biederman
>>> <ebiederm@xxxxxxxxxxxx> wrote:
>>>> Jann Horn <jannh@xxxxxxxxxx> writes:
>>>>> On Sun, Mar 8, 2020 at 10:41 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
>>>>>> The cred_guard_mutex is problematic. The cred_guard_mutex is held
>>>>>> over the userspace accesses as the arguments from userspace are read.
>>>>>> The cred_guard_mutex is held of PTRACE_EVENT_EXIT as the the other
>>>>>> threads are killed. The cred_guard_mutex is held over
>>>>>> "put_user(0, tsk->clear_child_tid)" in exit_mm().
>>>>>>
>>>>>> Any of those can result in deadlock, as the cred_guard_mutex is held
>>>>>> over a possible indefinite userspace waits for userspace.
>>>>>>
>>>>>> Add exec_update_mutex that is only held over exec updating process
>>>>>> with the new contents of exec, so that code that needs not to be
>>>>>> confused by exec changing the mm and the cred in ways that can not
>>>>>> happen during ordinary execution of a process.
>>>>>>
>>>>>> The plan is to switch the users of cred_guard_mutex to
>>>>>> exec_udpate_mutex one by one. This lets us move forward while still
>>>>>> being careful and not introducing any regressions.
>>>>> [...]
>>>>>> @@ -1034,6 +1035,11 @@ static int exec_mmap(struct mm_struct *mm)
>>>>>> return -EINTR;
>>>>>> }
>>>>>> }
>>>>>> +
>>>>>> + ret = mutex_lock_killable(&tsk->signal->exec_update_mutex);
>>>>>> + if (ret)
>>>>>> + return ret;
>>>>>
>>>>> We're already holding the old mmap_sem, and now nest the
>>>>> exec_update_mutex inside it; but then while still holding the
>>>>> exec_update_mutex, we do mmput(), which can e.g. end up in ksm_exit(),
>>>>> which can do down_write(&mm->mmap_sem) from __ksm_exit(). So I think
>>>>> at least lockdep will be unhappy, and I'm not sure whether it's an
>>>>> actual problem or not.
>>>>
>>>> Good point. I should double check the lock ordering here with mmap_sem.
>>>> It doesn't look like mmput takes mmap_sem
>>>
>>> You sure about that? mmput() -> __mmput() -> ksm_exit() ->
>>> __ksm_exit() -> down_write(&mm->mmap_sem)
>>>
>>> Or also: mmput() -> __mmput() -> khugepaged_exit() ->
>>> __khugepaged_exit() -> down_write(&mm->mmap_sem)
>>>
>>> Or is there a reason why those paths can't happen?
>>
>> Clearly I didn't look far enough.
>>
>> I will adjust this so that exec_update_mutex is taken before mmap_sem.
>> Anything else is just asking for trouble.
>>
>
> Note that vm_access does also mmput under the exec_update_mutex.
> So I don't see a huge problem here.
> But maybe I missed something.
The issue is that to prevent deadlock locks must always be taken
in the same order.
Taking mmap_sem then exec_update_mutex at the start of the function,
then taking exec_update_mutex then mmap_sem in mmput, takes the
two locks in two different orders. Which means that in the right
set or circumstances:
thread1: thread2:
obtain mmap_sem optain exec_update_mutex
wait for exec_update_mutex wait for mmap_sem
Which guarantees that neither thread will make progress.
The fix is easy I just need to take exec_update_mutex a few lines
earlier.
Eric
