More spawn & Windows processes

[Pierre A. Humblet]

At 01:23 PM 12/27/2004 -0500, Christopher Faylor wrote:
>On Mon, Dec 27, 2004 at 12:56:55PM -0500, Pierre A. Humblet wrote:
>>At 11:18 AM 12/27/2004 -0500, Christopher Faylor wrote:
>>>On Mon, Dec 27, 2004 at 11:07:49AM -0500, Pierre A. Humblet wrote:
>>>>At 11:05 AM 12/27/2004 -0500, Christopher Faylor wrote:
>>>>>On Mon, Dec 27, 2004 at 10:51:58AM -0500, Pierre A. Humblet wrote:
>>>>>>  should now be able to save microseconds by simply closing the pipe.
>>>>>
>>>>>Did you happen to notice that it *used* to be a CloseHandle?
>>>>
>>>>Yes. I think you changed it in an attempt to make P_DETACH work.
>>>
>>>No.  I remember that you floated that theory but it really had nothing
>>>to do with that.  I wanted to guarantee that the parent process saw a
>>>"close" regardless of whether there were other processes with the pipe
>>>open.  It's a paranoia thing as I round onto releasing 1.5.13.
>>
>>OK. At the time you made it, the change was in fact necessary because
>>there could be another process with the pipe open (a detached process).
>
>No, it wasn't necessary because a detached process didn't get the pipe.
>
>2004-12-23  Christopher Faylor  <cgf@timesys.com>
>
>        * pinfo.cc (_pinfo::exit): Signal exit more forcibly.
>        * sigproc.cc (proc_subproc): Add support for PROC_DETACHED_CHILD.
>
>Although the two things were checked in at the same time, they were
>unrelated.  I was working on a problem that Corinna had reported in
>private email.

You were lucky! The detached process did not get a pipe from
its physical parent, but it was still inheriting the pipe from the
logical parent of its physical parent (as an exec would, at the time).

>>The mindboggling line is 
>>  _pinfo::dup_proc_pipe: closed wr_proc_pipe 0x4 for pid 1014051(4293972685)
>>It indicates that it's apparently possible to successfully duplicate a
pipe into
>>a process that has exited! That open pipe is keeping the original sh
waiting for
>>the "termination" of its child. 
>
>!!!!!!!!!!!!!!!!!!
>
>You know, I was wondering if that was going to be a problem when I wrote that
>code and I thought that it couldn't possibly be a problem.  No OS could be so
>brain dead as to allow what you've apparently discovered.
>
>I was just looking into this, too, and the strace didn't make much sense.
Now,
>I guess I see why.
>
>I'll add a workaround.  I do so love to slow down code for the sake of the
Win9x
>family.

You could also set the pipe as inheritable for execs, as proposed earlier.
That makes sync_proc_pipe unnecessary.  Having a "set_proc_pipe_inheritance"
would not be any more complicated (code wise). It would also not be slower
than duplicating the pipe into the exec'ed process. 

Pierre
 
>cgf
>