Potential regression?

[H. S. Teoh]

On Sun, May 05, 2019 at 11:18:43AM +0200, Rainer Schuetze via Digitalmars-d wrote:
> On 03/05/2019 01:57, H. S. Teoh wrote:
> > After updating my dmd toolchain to the latest git master today, I
> > started getting strange random program lockups.  At first I thought
> > it was caused by passing the wrong druntime options (--DRT-*) that's
> > incompatible with the latest druntime, but then I started seeing
> > messages of this sort:
> > 
> > 	core.sync.exception.SyncError@(0): Unable to unlock mutex.
> > 
> > The program in question uses std.parallelism.parallel to parallelize
> > a bunch of I/O-heavy tasks.  It has been working fine for many
> > months now.  Did something break recently in std.parallelism or the
> > core.sync.* utilities that it presumably uses under the hood?
> 
> Parallel marking by the GC has recently been merged into master which
> might create a couple of extra threads. Is this causing the failures?
> Please try disabling it with --DRT-gcopt=parallel:0

Turns out, this problem is quite elusive.  Running it with
--DRT-gcopt=parallel:0 *seems* to make the problem go away, but it's
hard to tell because sometimes it takes 10-20 trials before a lockup
happens, so I'm not 100% confident --DRT-gcopt=parallel:0 fixes the
problem -- I might have just gotten lucky so far.  But at least, I've
yet to see a lockup with it, whereas I've seen a few lockups when
running without it, so perhaps this is the problem?

I wish I could narrow it down, but it's going to be hard because it's so
unpredictable -- I couldn't use dustmite to reduce the code since the
problem can't be reliably reproduced.


T

-- 
It is not the employer who pays the wages. Employers only handle the money. It is the customer who pays the wages. -- Henry Ford