Re: [PATCH v1 3/7] kernel/fork: always deny write access to current MM exe_file

[David Hildenbrand <david@redhat.com>]

On 12.08.21 14:32, Christian Brauner wrote:
> On Thu, Aug 12, 2021 at 12:13:44PM +0200, David Hildenbrand wrote:
>> On 12.08.21 12:05, Christian Brauner wrote:
>>> [+Cc Andrei]
>>>
>>> On Thu, Aug 12, 2021 at 10:43:44AM +0200, David Hildenbrand wrote:
>>>> We want to remove VM_DENYWRITE only currently only used when mapping the
>>>> executable during exec. During exec, we already deny_write_access() the
>>>> executable, however, after exec completes the VMAs mapped
>>>> with VM_DENYWRITE effectively keeps write access denied via
>>>> deny_write_access().
>>>>
>>>> Let's deny write access when setting the MM exe_file. With this change, we
>>>> can remove VM_DENYWRITE for mapping executables.
>>>>
>>>> This represents a minor user space visible change:
>>>> sys_prctl(PR_SET_MM_EXE_FILE) can now fail if the file is already
>>>> opened writable. Also, after sys_prctl(PR_SET_MM_EXE_FILE), the file
>>>
>>> Just for completeness, this also affects PR_SET_MM_MAP when exe_fd is
>>> set.
>>
>> Correct.
>>
>>>
>>>> cannot be opened writable. Note that we can already fail with -EACCES if
>>>> the file doesn't have execute permissions.
>>>>
>>>> Signed-off-by: David Hildenbrand <david@redhat.com>
>>>> ---
>>>
>>> The biggest user I know and that I'm involved in is CRIU which heavily
>>> uses PR_SET_MM_MAP (with a fallback to PR_SET_MM_EXE_FILE on older
>>> kernels) during restore. Afair, criu opens the exe fd as an O_PATH
>>> during dump and thus will use the same flag during restore when
>>> opening it. So that should be fine.
>>
>> Yes.
>>
>>>
>>> However, if I understand the consequences of this change correctly, a
>>> problem could be restoring workloads that hold a writable fd open to
>>> their exe file at dump time which would mean that during restore that fd
>>> would be reopened writable causing CRIU to fail when setting the exe
>>> file for the task to be restored.
>>
>> If it's their exe file, then the existing VM_DENYWRITE handling would have
>> forbidden these workloads to open the fd of their exe file writable, right?
> 
> Yes.
> 
>> At least before doing any PR_SET_MM_MAP/PR_SET_MM_EXE_FILE. But that should
>> rule out quite a lot of cases we might be worried about, right?
> 
> Yes, it rules out the most obvious cases. The problem is really just
> that we don't know how common weirder cases are. But that doesn't mean
> we shouldn't try and risk it. This is a nice cleanup and playing
> /proc/self/exe games isn't super common.
> 

Right, and having the file your executing opened writable isn't 
something very common as well.

If we really run into problems, we could not protect the new file when 
issuing PR_SET_MM_MAP/PR_SET_MM_EXE_FILE. But I'd like to avoid that, if 
possible, because it feels like working around something that never 
should have worked that way and is quite inconsistent.

-- 
Thanks,

David / dhildenb