RE: [PATCH 0/1] Fix hang in t5562, introduced in v2.21.0-rc1

["Randall S. Becker" <rsbecker@nexbridge.com>]

On February 14, 2019 17:39, Junio C Hamano wrote:
> To: Randall S. Becker <rsbecker@nexbridge.com>
> Cc: 'Johannes Schindelin via GitGitGadget' <gitgitgadget@gmail.com>;
> git@vger.kernel.org; 'Max Kirillov' <max@max630.net>
> Subject: Re: [PATCH 0/1] Fix hang in t5562, introduced in v2.21.0-rc1
> 
> "Randall S. Becker" <rsbecker@nexbridge.com> writes:
> 
> > Unfortunately, subtest 13 still hangs on NonStop, even with this
> > patch, so our Pipeline still hangs. I'm glad it's better on Azure, but
> > I don't think this actually addresses the root cause of the hang.
> 
> Sigh.
> 
> > possible this is not the test that is failing, but actually the
> > git-http-backend? The code is not in a loop, if that helps. It is not
> > consuming any significant cycles. I don't know that part of the code
> > at all, sadly. The code is here:
> >
> > * in the operating system from here up *
> >   cleanup_children + 0x5D0 (UCr)
> >   cleanup_children_on_exit + 0x70 (UCr)
> >   git_atexit_dispatch + 0x200 (UCr)
> >   __process_atexit_functions + 0xA0 (DLL zcredll)
> >   CRE_TERMINATOR_ + 0xB50 (DLL zcredll)
> >   exit + 0x2A0 (DLL zcrtldll)
> >   die_webcgi + 0x240 (UCr)
> >   die_errno + 0x360 (UCr)
> >   write_or_die + 0x1C0 (UCr)
> >   end_headers + 0x1A0 (UCr)
> >   die_webcgi + 0x220 (UCr)
> >   die + 0x320 (UCr)
> >   inflate_request + 0x520 (UCr)
> >   run_service + 0xC20 (UCr)
> >   service_rpc + 0x530 (UCr)
> >   cmd_main + 0xD00 (UCr)
> >   main + 0x190 (UCr)
> >
> > Best guess is that a signal (SIGCHLD?) is possibly getting eaten or
> > neglected somewhere between the test, perl, and git-http-backend.
> 
> So we are trying to die(), which actually happens in die_webcgi(), and
then try
> to write some message _but_ notice an error inside
> write_or_dir() and try to exit because we do not want to recurse forever
> trying to die, giving a message to say how/why we died, and die because
> failing to give that message, forever.
> 
> But in our attempt to exit(), we try to "cleanup children" and that is
what gets
> stuck.
> 
> One big difference before and after the /dev/zero change is that the
process
> is now on a downstream of the pipe.  If we prepare a large file with a
finite
> size full of NULs and replace /dev/null with it, instead of feeding NULs
from
> the pipe, would it change the equation?

Doubtful. The processes are still around, and are waiting on read but not
actively reading (CPU time is not going up, so we're not reading an infinite
stream). To me, this is a pipe situation where there is simply nothing
waiting on the pipe (maybe a flush missing?). I'm grasping are straws
without knowing the actual process architecture of the test to debug it.