From: | Noah Misch <noah(at)leadboat(dot)com> |
---|---|
To: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
Cc: | pgsql-hackers(at)postgresql(dot)org |
Subject: | Re: Obsolete coding in fork_process.c |
Date: | 2014-05-02 01:24:56 |
Message-ID: | 20140502012456.GB1191730@tornado.leadboat.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
On Thu, May 01, 2014 at 08:44:46PM -0400, Tom Lane wrote:
> Noah Misch <noah(at)leadboat(dot)com> writes:
> > On Thu, May 01, 2014 at 05:59:08PM -0400, Tom Lane wrote:
> >> However ... after looking around I notice that fflush(NULL) is already
> >> being used in parallel pg_dump and pg_upgrade; and at least in the latter
> >> case I'm afraid to change that because it looks like there are probably
> >> other stdio output files open in the process.
>
> > Do those programs, operating in those modes, read from stdin or some other
> > long-lived, pipe-backed FILE*?
>
> Probably not.
Then the use of fflush(NULL) in those programs tells us nothing about the
status of the aforementioned HP-UX bug.
> However, if that's your criterion, I'd say that I'd much
> rather assume that the postmaster doesn't have a pipe connected to stdin
> than that it has no stdio outputs besides stdout/stderr.
Either way, the beneficiary is theoretical; who can tell which one deserves
priority? Let's do what we usually do in the absence of a clear improvement:
keep the longstanding behavior.
> >> BTW, while working on this I noticed that there are a boatload of places
> >> where we use system() or popen() without a prior fflush. I suspect most
> >> of them are safe, or we'd have heard more complaints --- but shouldn't
> >> we clamp down on that?
>
> > You need the fflush() when forking and then using stdio in the child before
> > any exec(). Have you caught wind of any system() or popen() implementation
> > having that property?
>
> You're only considering one aspect of the problem. Yeah, you might not
> get duplicated output unless system() prints something before exec(),
> but we would also like to have causality: that is, whatever we sent to
> stdout before calling system() should appear there before anything the
> child process sends to stdout.
Good point. I suppose a couple of fflush() calls have negligible cost next to
a system() or popen(). Introduce pg_popen()/pg_system(), and adopt a rule
that they are [almost] our only callers of raw popen()/system()?
--
Noah Misch
EnterpriseDB http://www.enterprisedb.com
From | Date | Subject | |
---|---|---|---|
Next Message | Alvaro Herrera | 2014-05-02 01:33:02 | Re: includedir_internal headers are not self-contained |
Previous Message | Tom Lane | 2014-05-02 00:44:46 | Re: Obsolete coding in fork_process.c |