Fingerpointing about broken Ada tasking starting with FreeBSD 9.0 threading

[John Marino]

Hi guys, it's been a while.

Before FreeBSD 9.0 was released, we made sure lang/gnat-aux built on it, 
and it did.  It wasn't until after it was released that we realized that 
the Ada compiler GNAT tasking was broken.  lang/gnat-aux is based on 
gcc-4.6.

I brought this to Adacore's attention as they maintain the Ada compiler 
in GCC.  They blew it off and said, "try gcc 4.7 first".  It was true 
that a lot of tasking code had changed, but I ported some of it to 
gnat-aux and it didn't fix a thing, so I wasn't too hopeful.

I finally got around to building gcc-4.7 and sure enough, it's tasking 
is just as broken.  The pending port lang/gcc-aux is here:
http://www.freebsd.org/cgi/query-pr.cgi?pr=ports/169951

The problem appears to stem from this commit:
"Convert thread list lock from mutex to rwlock." by davidxu

https://github.com/freebsd/freebsd/commit/c741b41fb2a88f630e77ada2888443e3fea358c9

When an Ada task exits (not sure if it's all tasks or just some of them, 
but it seems to be all), it aborts with the message "thread exits with 
resources held!"

I brought this back to Adacore and said that message was caused by 
either a locklevel > 0 or a critical_count > 0.  At that point, they 
said they needed a lot more convincing that it was a posix-wide problem 
that FreeBSD just happened to detect. (it turns out that critical_count 
is the culprit by the way.)

The DragonFly BSD threading library has the same locklevel and 
critical_count thread properties, but putting the same THR_IN_CRITICAL 
panic in that library did NOT result in thread panics on DragonFly.  All 
the Ada task tests passed completely.  That strengthens Adacore's 
position that FreeBSD is "fishy" and coming up with false positives. 
NetBSD's thread library is too different to try to check there.

If I remove
> #if defined(_PTHREADS_INVARIANTS)
> if (THR_IN_CRITICAL(curthread))
>    PANIC("thread exits with resources held!");	
> #endif
then tasking tests pass on FreeBSD as well.  Unfortunately there doesn't 
seem to be any way to disable the check without modifying and rebuilding 
the libpthread

So what can I do?
My first blame was aimed towards GCC and GNAT, but now I'm not so sure. 
  And if, in the worst case, this is actually a FreeBSD threading bug, 
how can it be worked around for FreeBSD 9?

I haven't been very successful using GDB to troubleshoot this thing.

Regards,
John