RFR: 8218975: Bug in macOSX kernel's pthread support

[Kim Barrett]

> On Mar 17, 2019, at 9:43 PM, David Holmes <david.holmes at oracle.com> wrote:
> 
> Hi Kim and Patricio
> 
> On 16/03/2019 9:03 am, Kim Barrett wrote:
>> Please review this fix for intermittent Mac-only crashes involving the
>> OWSTTaskTerminator.
>> This is joint work with Patricio Chilano Mateo.
> 
> Fantastic job in pinning this one down!

For the record, Patricio did the heavy lifting on getting to the
bottom of this, including building and running patched versions of the
macOSX kernel and pthread library, developing the standalone reproducer,
and analyzing its behavior.

> The fix generally looks good though I agree with Thomas that X_ptr() should just be X().

OK, _ptr suffix dropped.  In my defence, that convention is used by
the C++ standard, both in standard APIs and in examples.  Though
following "standard" conventions would probably use references in a
lot of places where HotSpot code uses pointers.

> There is one minor issue:
> 
> + pthread_mutex_t os::PlatformMonitor::_freelist_lock = PTHREAD_MUTEX_INITIALIZER;
> 
> on FreeBSD the default mutex initialization doesn't use PTHREAD_MUTEX_NORMAL and so we get a slow debug version IIRC. That's why we explicitly use non-static pthread_mutex initialization using a pthread_mutex_attr. You should be able to move the _freelist_lock initialization to pthread_init_common() with the PLATFORM_MONITOR_IMPL_INDIRECT guard.

The workaround isn't in FreeBSD code, it is macOSX-specific, where the
implemented default is "normal" rather than "errorcheck".  But the
documentation seems to be vague on that (as with FreeBSD and POSIX
documentation, but not Linux, where the defaults are spelled out).
Since the cost of explicitly ensuring we use "normal" mutexes is
small, let's be paranoid and do so.

New webrevs:
full: http://cr.openjdk.java.net/~kbarrett/8218975/open.02/
incr: http://cr.openjdk.java.net/~kbarrett/8218975/open.02.inc/