Message-ID: <4227F258.9040709@samurai.com>
Date: Fri, 04 Mar 2005 16:30:00 +1100
From: Neil Conway <neilc@samurai.com>
User-Agent: Debian Thunderbird 1.0 (X11/20050116)
MIME-Version: 1.0
To: pgsql-hackers <pgsql-hackers@postgresql.org>
Subject: refactoring fork() and EXEC_BACKEND
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

While going through the usual motions needed to fork a child process of 
the postmaster, it occurred to me that there's a fair bit of duplicated 
code involved. There are also #ifdef for various situations (BeOS, 
LINUX_PROFILE, and EXEC_BACKEND), which makes the code yet more ugly. I 
think we could make this a lot cleaner.

I'd like to define an API like so:

pid_t fork_process(int proc_type);
pid_t fork_backend(Port *port);

If the process needs to add a lot of private information to the argv in 
the case of EXEC_BACKEND, they could invoke a third variant:

#ifdef EXEC_BACKEND
pid_t forkexec_process(int proc_type, int argc, char **argv);
#endif

(Or possibly using varargs, if that is cleaner for most call-sites). 
Hopefully most call sites could just use fork_process().

These functions would then take care of all the necessary 
platform-specific judo:

- flush stdout, stderr
- invoke BeOS hooks as necessary
- save and restore profiling timer, if necessary
- if EXEC_BACKEND, use proc_type to lay out the argv for the new process 
and then invoke internal_forkexec()
- otherwise, just invoke fork()
- return result to client

So, most call sites would be quite nice:

pid_t result = fork_process(PROC_TYPE_FOO);
if (result == -1) { /* fork failed, in parent */ }
else if (result == 0) { /* in child */ }
else { /* in parent, `result' is pid of child */ }

I'd also like to move the implementation of fork_process() and friends, 
as well as internal_forkexec(), into a separate file -- I'd rather not 
clutter up postmaster.c with it.

Comments?

-Neil