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 18:51, Linus Torvalds wrote:
> On Wed, Aug 11, 2021 at 10:45 PM David Hildenbrand <david@redhat.com> wrote:
>>
>>          /* No ordering required: file already has been exposed. */
>> -       RCU_INIT_POINTER(mm->exe_file, get_mm_exe_file(oldmm));
>> +       exe_file = get_mm_exe_file(oldmm);
>> +       RCU_INIT_POINTER(mm->exe_file, exe_file);
>> +       if (exe_file)
>> +               deny_write_access(exe_file);
> 
> Can we make a helper function for this, since it's done in two different places?

Sure, no compelling reason not to (except finding a suitable name, but 
I'll think about that tomorrow).

> 
>> -       if (new_exe_file)
>> +       if (new_exe_file) {
>>                  get_file(new_exe_file);
>> +               /*
>> +                * exec code is required to deny_write_access() successfully,
>> +                * so this cannot fail
>> +                */
>> +               deny_write_access(new_exe_file);
>> +       }
>>          rcu_assign_pointer(mm->exe_file, new_exe_file);
> 
> And the above looks positively wrong. The comment is also nonsensical,
> in that it basically says "we thought this cannot fail, so we'll just
> rely on it".

Well, it documents the expectation towards the caller, but in a 
suboptimal way, I agree.

> 
> If it truly cannot fail, then the comment should give the reason, not
> the "we depend on this not failing".

Right, "We depend on the caller already have done a deny_write_access() 
successfully first such that this call cannot fail." combined with

if (deny_write_access(new_exe_file))
	pr_warn("Unexpected failure of deny_write_access() in %s",
                  __func__);

suggestions welcome.

> 
> And honestly, I don't see why it couldn't fail. And if it *does* fail,
> we cannot then RCU-assign the exe_file pointer with this, because
> you'll get a counter imbalance when you do the allow_write_access()
> later.

Anyone calling set_mm_exe_file() (-> begin_new_exec()) is expected to 
successfully triggered a deny_write_access() upfront such that we won't 
fail at that point.

Further, on the dup_mmap() path we are sure the previous oldmm exe_file 
properly saw a successful deny_write_access() already, because that's 
now guaranteed for any exe_file.

> 
> Anyway, do_open_execat() does do deny_write_access() with proper error
> checking. I think that is the existing reference that you depend on -
> so that it doesn't fail. So the comment could possibly say that the
> only caller has done this, but can we not just use the reference
> deny_write_access() directly, and not do a new one here?

I think that might over-complicate the exec code where we would see a 
allow_write_access() on error paths, but not on success paths. This here 
looks cleaner to me, agreeing that the comment and the error check has 
to be improved.

We handle all allow_write_access()/deny_write_access() regarding 
exe_file completely in kernel/fork.c, which is IMHO quite nice.

> 
> IOW, maybe there's an extraneous 'allow_write_access()' somewhere that
> should be dropped when we do the whole binprm dance in execve()?

fs/exec.c: free_bprm() and exec_binprm() to be precise.

Thanks!

-- 
Thanks,

David / dhildenb