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

[Paul E. McKenney]

On Mon, Dec 22, 2008 at 10:44:40PM -0800, Hans Boehm 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 
> 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.
> 
> The details are in my PPoPP 07 paper.  But the basic example is:
> 
> Thread1: x = 42; pthread_mutex_lock(&l);
> 
> Thread2:
> 
> while (pthread_mutex_timed_lock(&l, small) == 0) pthread_mutex_unlock(&l);
> pthread_mutex_lock(&dummy); pthread_mutex_unlock(&dummy);
> assert(x == 42);
> 
> (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.
> 
> (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.)
> 
> >
> > Judging from the draft and also
> >
> > http://www.hpl.hp.com/techreports/2008/HPL-2008-56.pdf
> >
> > the primary intent behind introduction of spurious failures was to
> > make unlock()->try_lock{_for}({rel_time}) "synchronize with"
> > conditionally ("If try_lock() returns true, prior unlock() operations
> > on the same object synchronize with (1.10) this operation.").
> >
> > But why insist on spurious failures for try_lock{_for}({rel_time})
> > instead of simply saying that these operations provide the effect of
> > atomic_thread_fence(memory_order_acquire) if mutex ownership is
> > obtained, otherwise no fence effect
> > (atomic_thread_fence(memory_order_relaxed) if you like)?
> The problem isn't really primarily with the try_lock implementation.  It 
> has to do with the cost of the lock operation.

Interesting.  I have come across cases where spurious failure of a
try_lock primitive would be OK and others where it would be problematic.
However, algorithms relying on no-spurious-failure try_lock are a bit
tricky to get right, so I am not sure that I would scream too loudly at
their being excluded...

							Thanx, Paul