Re: [PATCH 1/3] io_uring: add support for async work inheriting files table

[Jens Axboe <axboe@kernel.dk>]

On 10/18/19 10:20 AM, Jann Horn wrote:
> On Fri, Oct 18, 2019 at 5:55 PM Jens Axboe <axboe@kernel.dk> wrote:
>> On 10/18/19 9:00 AM, Jens Axboe wrote:
>>> On 10/18/19 8:52 AM, Jann Horn wrote:
>>>> On Fri, Oct 18, 2019 at 4:43 PM Jens Axboe <axboe@kernel.dk> wrote:
>>>>>
>>>>> On 10/18/19 8:40 AM, Jann Horn wrote:
>>>>>> On Fri, Oct 18, 2019 at 4:37 PM Jens Axboe <axboe@kernel.dk> wrote:
>>>>>>>
>>>>>>> On 10/18/19 8:34 AM, Jann Horn wrote:
>>>>>>>> On Fri, Oct 18, 2019 at 4:01 PM Jens Axboe <axboe@kernel.dk> wrote:
>>>>>>>>> On 10/17/19 8:41 PM, Jann Horn wrote:
>>>>>>>>>> On Fri, Oct 18, 2019 at 4:01 AM Jens Axboe <axboe@kernel.dk> wrote:
>>>>>>>>>>> This is in preparation for adding opcodes that need to modify files
>>>>>>>>>>> in a process file table, either adding new ones or closing old ones.
>>>>>>>> [...]
>>>>>>>>> Updated patch1:
>>>>>>>>>
>>>>>>>>> http://git.kernel.dk/cgit/linux-block/commit/?h=for-5.5/io_uring-test&id=df6caac708dae8ee9a74c9016e479b02ad78d436
>>>>>>>>
>>>>>>>> I don't understand what you're doing with old_files in there. In the
>>>>>>>> "s->files && !old_files" branch, "current->files = s->files" happens
>>>>>>>> without holding task_lock(), but current->files and s->files are also
>>>>>>>> the same already at that point anyway. And what's the intent behind
>>>>>>>> assigning stuff to old_files inside the loop? Isn't that going to
>>>>>>>> cause the workqueue to keep a modified current->files beyond the
>>>>>>>> runtime of the work?
>>>>>>>
>>>>>>> I simply forgot to remove the old block, it should only have this one:
>>>>>>>
>>>>>>> if (s->files && s->files != cur_files) {
>>>>>>>             task_lock(current);
>>>>>>>             current->files = s->files;
>>>>>>>             task_unlock(current);
>>>>>>>             if (cur_files)
>>>>>>>                     put_files_struct(cur_files);
>>>>>>>             cur_files = s->files;
>>>>>>> }
>>>>>>
>>>>>> Don't you still need a put_files_struct() in the case where "s->files
>>>>>> == cur_files"?
>>>>>
>>>>> I want to hold on to the files for as long as I can, to avoid unnecessary
>>>>> shuffling of it. But I take it your worry here is that we'll be calling
>>>>> something that manipulates ->files? Nothing should do that, unless
>>>>> s->files is set. We didn't hide the workqueue ->files[] before this
>>>>> change either.
>>>>
>>>> No, my worry is that the refcount of the files_struct is left too
>>>> high. From what I can tell, the "do" loop in io_sq_wq_submit_work()
>>>> iterates over multiple instances of struct sqe_submit. If there are
>>>> two sqe_submit instances with the same ->files (each holding a
>>>> reference from the get_files_struct() in __io_queue_sqe()), then:
>>>>
>>>> When processing the first sqe_submit instance, current->files and
>>>> cur_files are set to $user_files.
>>>> When processing the second sqe_submit instance, nothing happens
>>>> (s->files == cur_files).
>>>> After the loop, at the end of the function, put_files_struct() is
>>>> called once on $user_files.
>>>>
>>>> So get_files_struct() has been called twice, but put_files_struct()
>>>> has only been called once. That leaves the refcount too high, and by
>>>> repeating this, an attacker can make the refcount wrap around and then
>>>> cause a use-after-free.
>>>
>>> Ah now I see what you are getting at, yes that's clearly a bug! I wonder
>>> how we best safely can batch the drops. We can track the number of times
>>> we've used the same files, and do atomic_sub_and_test() in a
>>> put_files_struct_many() type addition. But that would leave us open to
>>> the issue you describe, where someone could maliciously overflow the
>>> files ref count.
>>>
>>> Probably not worth over-optimizing, as long as we can avoid the
>>> current->files task lock/unlock and shuffle.
>>>
>>> I'll update the patch.
>>
>> Alright, this incremental on top should do it. And full updated patch
>> here:
>>
>> http://git.kernel.dk/cgit/linux-block/commit/?h=for-5.5/io_uring-test&id=40449c5a3d3b16796fa13e9469c69d62986e961c
>>
>> Let me know what you think.
> 
> Ignoring the locking elision, basically the logic is now this:
> 
> static void io_sq_wq_submit_work(struct work_struct *work)
> {
>          struct io_kiocb *req = container_of(work, struct io_kiocb, work);
>          struct files_struct *cur_files = NULL, *old_files;
>          [...]
>          old_files = current->files;
>          [...]
>          do {
>                  struct sqe_submit *s = &req->submit;
>                  [...]
>                  if (cur_files)
>                          /* drop cur_files reference; borrow lifetime must
>                           * end before here */
>                          put_files_struct(cur_files);
>                  /* move reference ownership to cur_files */
>                  cur_files = s->files;
>                  if (cur_files) {
>                          task_lock(current);
>                          /* current->files borrows reference from cur_files;
>                           * existing borrow from previous loop ends here */
>                          current->files = cur_files;
>                          task_unlock(current);
>                  }
> 
>                  [call __io_submit_sqe()]
>                  [...]
>          } while (req);
>          [...]
>          /* existing borrow ends here */
>          task_lock(current);
>          current->files = old_files;
>          task_unlock(current);
>          if (cur_files)
>                  /* drop cur_files reference; borrow lifetime must
>                   * end before here */
>                  put_files_struct(cur_files);
> }
> 
> If you run two iterations of this loop, with a first element that has
> a ->files pointer and a second element that doesn't, then in the
> second run through the loop, the reference to the files_struct will be
> dropped while current->files still points to it; current->files is
> only reset after the loop has ended. If someone accesses
> current->files through procfs directly after that, AFAICS you'd get a
> use-after-free.

Amazing how this is still broken. You are right, and it's especially
annoying since that's exactly the case I originally talked about (not
flipping current->files if we don't have to). I just did it wrong, so
we'll leave a dangling pointer in ->files.

The by far most common case is if one sqe has a files it needs to
attach, then others that also have files will be the same set. So I want
to optimize for the case where we only flip current->files once when we
see the files, and once when we're done with the loop.

Let me see if I can get this right...

-- 
Jens Axboe