[cpp-threads] Spurious failures of try_lock{_for}({rel_time}) vs. strong try_lock{_for}({rel_time})

[Alexander Terekhov]

(additional observation + correction)

On Tue, Dec 23, 2008 at 4:16 PM, Alexander Terekhov
<alexander.terekhov at gmail.com> wrote:
> On Tue, Dec 23, 2008 at 7:44 AM, Hans Boehm <Hans.Boehm at hp.com> wrote:
>>
>> On Mon, 22 Dec 2008, Alexander Terekhov wrote:
>>
>>> N2800's try_lock():
>>>
>>> "Effects: Attempts to obtain ownership of the mutex for the calling
>>> thread without blocking. If ownership is not obtained, there is no
>>> effect and try_lock() immediately returns. An implementation may fail
>>> to obtain the lock even if it is not held by any other thread."
>>>
>>> N2800's try_lock_for(rel_time):
>>>
>>> "Effects: The function attempts to obtain ownership of the mutex
>>> within the time specified by rel_time. If the time specified by
>>> rel_time is less than or equal to 0, the function attempts to obtain
>>> ownership without blocking (as if by calling try_lock())."
>>>
>>> seem to contradict POSIX:
>>>
>>> http://www.opengroup.org/onlinepubs/009695399/functions/pthread_mutex_trylock.html
>>>
>>> "The pthread_mutex_trylock() function shall be equivalent to
>>> pthread_mutex_lock(), except that if the mutex object referenced by
>>> mutex is currently locked (by any thread, including the current
>>> thread), the call shall return immediately."
>>>
>>> http://www.opengroup.org/onlinepubs/009695399/functions/pthread_mutex_timedlock.html
>>>
>>> "Under no circumstance shall the function fail with a timeout if the
>>> mutex can be locked immediately."
>>
>> I would argue that this is a mistake in the Posix standard.  Current
>
> I disagree. Current POSIX XBD 4.10 Memory Synchronization states that
>
> http://www.opengroup.org/onlinepubs/009695399/basedefs/xbd_chap04.html#tag_04_10
>
> "Unless explicitly stated otherwise, if one of the above functions
> returns an error, it is unspecified whether the invocation causes
> memory to be synchronized."
>
> So there isn't any release-acquire pairing involving failed
> try{timed}lock() in POSIX. See my interpretation of XBD 4.10 in terms
> of release-acquire pairings:
>
> http://www.decadentplace.org.uk/pipermail/cpp-threads/2005-April/000222.html
>
> I agree with N2800's note stating
>
> "[ Note: Since lock() does not synchronize with a failed subsequent
> try_lock(), the visibility rules are weak enough that little would be
> known about the state after a failure, even in the absence of spurious
> failures. —end note ]"
>
> The situation is the same in POSIX.
>
>> implementations on weakly ordered architectures are arguably often
>> incorrect with respect to this specification.  If you want to promise
>> sequential consistency for data-race-free programs in the presence of
>> pthread_mutex_timedlock(), you would need lock() to have release semantics
>> as well.
>
> I disagree, see below.
>
>>
>> The details are in my PPoPP 07 paper.  But the basic example is:
>>
>> Thread1: x = 42; pthread_mutex_lock(&l);

Note that as long as POSIX doesn't insist on having any
release-acquire pairing involving pthread_mutex_lock() on release side
(just like it is the case for mutex's lock() under proposed C++ memory
model... unless I'm just missing something), then in regards to
ordering with respect to other threads, Thread1's code above is
equivalent to

Thread1: pthread_mutex_lock(&l); x = 42;

so ...

>>
>> Thread2:
>>
>> while (pthread_mutex_timed_lock(&l, small) == 0) pthread_mutex_unlock(&l);
>> pthread_mutex_lock(&dummy); pthread_mutex_unlock(&dummy);
>> assert(x == 42);

... even C++'s thread_atomic_fence(memory_order_seq_cst) in Thread2:

while (pthread_mutex_timed_lock(&l, small) == 0) pthread_mutex_unlock(&l);
thread_atomic_fence(memory_order_seq_cst);
assert(x == 42);

won't help to make this example data-race-free.

>
> This example is not data-race-free. Consider that it doesn't prevent
> reordering of x's load above dummy's unlock() and completion of failed
> timedlock() after x's load resulting in a race on x.
>
>>
>> (The initial loop waits for thread 1 to acquire the lock.  Yes, this
>> is evil code.)
>>
>> This can result in non-sequentially-consistent behavior, and the
>> assertion may fail, if the assignment and lock acquisition in thread1
>> are not ordered.  I believe they often are not.  Fixing this can result
>> in appreciable, and completely useless, overhead for most lock
>> acquisitions.
>
> Well, consider lock operations on semaphores
>
> http://www.opengroup.org/onlinepubs/009695399/basedefs/xbd_chap04.html#tag_04_15_00_01
>
> sem_wait() (and XSI IPC sema's locking as well).
>
> The example would be
>
> (Initially: semaphore value == 1)
>
> Thread1:
>
> x = 42;
> sem_wait(&l); // decrement semaphore value
>
> Thread2:
>
> int v;
> sem_getvalue(&l, &v);
> if (l == 0) assert(x == 42);

Typo. I meant: if (v == 0) assert(x == 42);

>
> I think that this shall be data-race-free due to release-acquire
> pairing between sem_wait(&l) and sem_getvalue(&l, &v).
>
>> (The dummy lock acquisition isn't necessary to see that this is non-SC
>> behavior.  It does help to argue that the current behavior is actually
>> officially inconsistent with the current spec, which is unclear
>> about when SC behavior is intended.)
>
> I agree that POSIX XBD 4.10 needs some fixing to provide clear listing
> of all release-acquire pairings involving XBD.4.10 calls.
>
> http://www.decadentplace.org.uk/pipermail/cpp-threads/2005-April/000222.html

regards,
alexander.