[Libreoffice] New Tinderbox Linux-RHEL6-x86_64-check

[Stephan Bergmann]

I added a new tinderbox "Linux-RHEL6-x86_64-check" to 
<http://tinderbox.libreoffice.org/MASTER/status.html> (and it will spam 
suspects with messages about failing builds in the usual way).

The box does a rather complete build (--enable-binfilter, no 
--with-system switches) and especially enables the debug features 
--enable-dbgutil and --enable-werror.  Additionally, it calls tinbuild2 
with -t switch, i.e., it runs all the subsequent tests via a toplevel 
"make check".  The turnaround time is approximately four hours.

"make check" does still suffer from sporadic soffice.bin failures, but 
the given box appears to run rather stable now.  In the last couple of 
weeks, the vast majority of builds succeeded, and the few that failed 
consistently did so for one of two reasons.

One was the issue (hopefully) fixed now by 
<http://cgit.freedesktop.org/libreoffice/core/commit/?id=0e4b36269097f8865bdbc1b9d299a2776ae0cc06> 
"ScGlobal::GetAutoFormat not always required to create fresh instance."

The other is a sporadic failure of forms_unoapi, an soffice.bin crash 
somewhere underneath

> [...]
> #12 in SdrPageView::DrawLayer at svx/source/svdraw/svdpagv.cxx:398
> #13 in SwViewImp::PaintLayer at sw/source/core/view/vdraw.cxx:148
> #14 in SwRootFrm::Paint at sw/source/core/layout/paintfrm.cxx:2976
> #15 in ViewShell::Paint at sw/source/core/view/viewsh.cxx:1678
> #16 in SwCrsrShell::Paint at sw/source/core/crsr/crsrsh.cxx:1165
> #17 in SwEditWin::Paint at sw/source/ui/docvw/edtwin2.cxx:535
> #18 in Window::ImplCallPaint at vcl/source/window/window.cxx:2417
> #19 in Window::ImplCallPaint at vcl/source/window/window.cxx:2441
> #20 in Window::ImplCallPaint at vcl/source/window/window.cxx:2441
> #21 in Window::ImplCallPaint at vcl/source/window/window.cxx:2441
> #22 in Window::ImplCallPaint at vcl/source/window/window.cxx:2441
> #23 in Window::ImplCallPaint at vcl/source/window/window.cxx:2441
> #24 in Window::ImplCallPaint at vcl/source/window/window.cxx:2441
> #25 in Window::ImplCallOverlapPaint at vcl/source/window/window.cxx:2477
> #26 in Window::ImplHandlePaintHdl at vcl/source/window/window.cxx:2497
> #27 in Window::LinkStubImplHandlePaintHdl at vcl/source/window/window.cxx:2491
> #28 in Link::Call at solver/unxlngx6/inc/tools/link.hxx:140
> #29 in Timer::Timeout at vcl/source/app/timer.cxx:256
> [...]

on the main thread while an URP thread is simultaneously executing 
SwXTextDocument::close.  Looks like SwXTextDocument::close destroys data 
that is still accessed within SwEditWin::Paint, but I haven't yet come 
around to find the cause here.  (Problem can only be observed 
sporadically, and valgrind didn't help yet.)  I plan to follow up on 
this, however.  Please ignore tinderbox failure mails that are due to 
this crash in the meantime.

The one local change I did to tinbuild2 is

> diff --git a/bin/tinbuild_internals.sh b/bin/tinbuild_internals.sh
> index e9f4f3d..680aecb 100644
> --- a/bin/tinbuild_internals.sh
> +++ b/bin/tinbuild_internals.sh
> @@ -354,7 +354,8 @@ do_build()
>              if [ -n "${last_checkout_date}" ] ; then
>                 report_error committer "$last_checkout_date" `printf "${report_msgs?}:\n\n"` "$(cat build_error.
>  ======
> -$(tail -n50 ${report_log?} | grep -A25 'internal build errors' | grep 'ERROR:' )"
> +Last 400 lines of ${report_log?}:
> +$(tail -n400 ${report_log?} )"
>                 report_to_tinderbox "${last_checkout_date?}" "build_failed" "yes"
>              else
>                  log_msgs "Failed to primed branch '$TINDER_BRANCH'. see build_error.log"

so that a failing test produces meaningful log output in the failure 
mails (i.e., the last 400 lines of the log, which should be enough to 
include the backtraces from the crashed soffice.bin process, if any).

The state of "make check" at least for this given box looks rather 
promising, so I would encourage other tinderbox owners to try whether 
adding the -t tinbuild2 switch would work well for their boxes as well.

Stephan