ebiederm@xxxxxxxxxxxx (Eric W. Biederman) writes:
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>
>> On Mon, May 29, 2017 at 04:11:13AM -0500, Eric W. Biederman wrote:
>>
>>> Kees I I have a concern:
>>>
>>> __must_check bool refcount_add_not_zero(unsigned int i, refcount_t *r)
>>> {
>>> unsigned int new, val = atomic_read(&r->refs);
>>>
>>> do {
>>> if (!val)
>>> return false;
>>>
>>> if (unlikely(val == UINT_MAX))
>>> return true;
>>>
>>> new = val + i;
>>> if (new < val)
>>> new = UINT_MAX;
>>>
>>> } while (!atomic_try_cmpxchg_relaxed(&r->refs, &val, new));
>>>
>>> WARN_ONCE(new == UINT_MAX, "refcount_t: saturated; leaking memory.\n");
>>>
>>> return true;
>>> }
>>>
>>> Why in the world do you succeed when you the value saturates????
>>
>> Why not? On saturation the object will leak and returning a reference to
>> it is always good.
>>
>>> From a code perspective that is bizarre. The code already has to handle
>>> the case when the counter does not increment.
>>
>> I don't see it as bizarre, we turned an overflow/use-after-free into a
>> leak. That's the primary mechanism here.
>>
>> As long as we have a reference to a leaked object, we might as well use
>> it, its not going anywhere.
>>
>>> Fixing the return value would move refcount_t into the realm of
>>> something that is desirable because it has bettern semantics and
>>> is more useful just on a day to day correctness point of view. Even
>>> ignoring the security implications.
>>
>> It changes the semantics between inc_not_zero() and inc(). It also
>> complicates the semantics of inc_not_zero(), where currently the failure
>> implies the count is 0 and means no-such-object, you complicate matters
>> by basically returning 'busy'.
>
> Busy is not a state of a reference count.
>
> It is true I am suggesting treating something with a saturated reference
> as not available. If that is what you mean by busy. But if it's
> reference is zero it is also not available. So there is no practical
> difference.
>
>> That is a completely new class of failure that is actually hard to deal
>> with, not to mention that it completely destroys refcount_inc_not_zero()
>> being a 'simple' replacement for atomic_inc_not_zero().
>>
>> In case of the current failure, the no-such-object, we can fix that by
>> creating said object. But what to do on 'busy' ? Surely you don't want
>> to create another. You'd have to somehow retrofit something to wait on
>> in every user.
>
> Using little words.
>
> A return of true from inc_not_zero means we took a reference.
> A return of false means we did not take a reference.
>
> The code already handles I took a reference or I did not take a
> reference.
>
> Therefore lying with refcount_t is not helpful. It takes failures
> the code could easily handle and turns them into leaks.
>
> At least that is how I have seen reference counts used. And those
> are definitely the plane obivous semantics.
>
> Your changes are definitely not drop in replacements for atomic_t in my
> code.
To clarify.
If my code uses atomic_inc it does not expect a failure of any sort
and saturate semantics are a fine replacement.
If my code uses atomic_inc_not_zero it knows how to handle a failure
to take a reference count. Making hiding the failure really bizarre.
A must check function that hides a case I can handle and requires
checking in a case where my code is built not to check is a drop in
replacement for neither.
So anyone who is proposing a refcount_t change as a drop in replacement
for any code I maintain I will nack on sight because refcount_t is not
currently a no-brain drop in replacement.
Eric
