Many core dumps in pthread_getspecific.

[Gary Palmer]

On Sat, Jun 06, 2015 at 06:08:21PM +0300, Konstantin Belousov wrote:
> On Sat, Jun 06, 2015 at 09:50:22AM +0200, Andre Meiser wrote:
> > On Wed, Jun 03, 2015 at 16:58 +0200, Konstantin Belousov wrote:
> > > You should recompile both libc and libthr with debugging symbols, like
> > > cd /usr/src
> > > (cd lib/libc && make all install DEBUG_FLAGS=-g)
> > > (cd lib/libthr && make all install DEBUG_FLAGS=-g)
> > > then obtain the core dump and post backtraces.
> > 
> > Thank you, that was exactly what I was looking for. Now I like FreeBSD
> > even more. ;)
> >
> > I made this short after and also rebooted the entire system to make
> > all programmes use those debug libs. Since than I had not a single
> > core dump.
> Which means, most likely, that your installation was corrupted.
> 
> >
> > I experienced something similar in the past, that with activated
> > debugging some errors can't be triggered any longer.
> >
> > At the moment I'm happy without crashes and I can work with this
> > system. As soon as I'm getting a new core dump, I'll post the
> > backtrace. If this won't happen for weeks, I may recompile the libs
> > again, try to find a way to trigger the bug on purpose, enable the
> > debug flag again and than provide the info.
> >
> > In the meantime, maybe a core developer may think about the lines
> > of code I'd provided. Why is _get_curthread() compared to NULL at
> > thr_kern.c (line 601) but not at thr_spec.c (line 224)? Either
> > _get_curthread() never ever returns NULL, than it's pointless to test
> > it or it's missing [not only] at thr_spec.c.
> Whole thr_kern.c file is 213 lines long in the HEAD.
> Might be, a core developer has a 601-lines long thr_kern.c.

lib/libkse/thread/thr_kern.c is longer and has a 
_get_curthread() call around line 601.

Regards,

Gary
 
> The only way for _get_curthread() to return NULL or, more likely, a
> garbage, is for the call to _get_curthread() to happens before the
> threading machinery was initialized. WRT the thr_spec.c, correct code
> must call pthread_key_create(3) at least once before calling any other
> pthread_key_*(3) functions. The implementation of pthread_key_create(3)
> interface, the function _pthread_key_create(), calls _thr_check_init()
> as the first action, and this helper initializes threading. So if
> _get_curthread() in any pthread_key_*() function returns NULL or
> garbage, this is an application' bug.