Native posix_spawn() in Cygwin?

[Roland Mainz]

On Wed, Mar 6, 2024 at 1:08 AM Mark Geisert via Cygwin
<cygwin@cygwin.com> wrote:
> On 3/5/2024 2:42 PM, Dan Shelton via Cygwin wrote:
> > On Mon, 4 Mar 2024 at 07:45, Mark Geisert via Cygwin <cygwin@cygwin.com> wrote:
> >> On 3/3/2024 7:27 PM, Dan Shelton via Cygwin wrote:
> [...]
> >>> strace does not help, as I need the Win32 calls BELOW posix_spawn(),
> >>> to see the implementation details.
> >>
> >> Check the source code, then. It's at:
> >>       https://cygwin.com/cgit/newlib-cygwin/tree/winsup/cygwin/fork.cc
> >>
> >> Look at line 587; there's the static function dofork(). Look at the
> >> thirty or so lines above that; there's both fork() and
> >> __posix_spawn_fork() calling dofork(). So both those user-level
> >> functions call into the exact same internals. (BTW __posix_spawn_fork()
> >> is called from posix_spawn(); the latter is in newlib and not Cygwin.)
> >>
> >> You can even see the reason it's done this way by reading the comment.
> >
> > Yes, but it is as I feared, Cygwin posix_spawn() does not use Win32
> > spawn() at all, and instead uses a rather inefficient vfork()
> > solution.
>
> Cygwin's vfork() is just a wrapper around fork(), so yes. But anyway...
>
> > posix_spawn() was added to POSIX so a Win32 implementation can use Win32 spawn()
>
> ...now I see what you're getting at:
>
> If posix_spawn() is intended to launch truly unrelated processes, with
> minimal or no coordination with the launching process, why can't it just
> use Windows' CreateProcess? I assume here that's what Win32 spawn() does.
>
> That's an interesting research question for somebody. If somebody steps
> up for that, great, otherwise as usual PTC.
> Regards,

Just one note (which applies to UNIX/Linux):
When I was at SUN we had severe performance problems with large
JAVA&&database processes launching little helper apps. It turned out
to be a |fork()|+|exec()| problem - the |fork()| was basically
harmless, but the |exec()| syscall required to tear down all address
space, which involved cross-calls to all other CPUs. The performance
penalty on a 64 CPU Enterprise 10000 or 72 CPU (up to 144 threads)
SF25k was *DEVASTATING* ( <--- understatement).

This was fixed for Solaris 11 in several ways:
1. Solaris 11 got a native |posix_spawn()| syscall, and which avoids
both |fork()| and the (in this context) dreaded |exec()|, and just
sets up a new process with copying only the bare minimum of data from
the parent process.
2. Solaris's ksh93 (which is used as /sbin/sh+/bin/sh) got support for
|posix_spawn()| (thanks for David Korn and Glenn Fowler for doing
that)
3. JAVA was modified to use |posix_spawn()|
4. Oracle was asked to use |posix_spawn()| too

For Cygwin I think it would be good to implement a |posix_spawn()|
using the native Win32 API too...

----

Bye,
Roland
-- 
  __ .  . __
 (o.\ \/ /.o) roland.mainz@nrubsig.org
  \__\/\/__/  MPEG specialist, C&&JAVA&&Sun&&Unix programmer
  /O /==\ O\  TEL +49 641 3992797
 (;O/ \/ \O;)