Date: Sat, 14 Apr 2012 13:20:12 GMT From: Jeremie Le Hen <jeremie@le-hen.org> To: freebsd-bugs@FreeBSD.org Subject: Re: bin/166660: [libc] [patch] New util/shlib to change per-fd default stdio buffering mode Message-ID: <201204141320.q3EDKCSE076995@freefall.freebsd.org>
next in thread | raw e-mail | index | archive | help
The following reply was made to PR bin/166660; it has been noted by GNATS.
From: Jeremie Le Hen <jeremie@le-hen.org>
To: John Baldwin <jhb@freebsd.org>
Cc: Jeremie Le Hen <jeremie@le-hen.org>, bug-followup@freebsd.org
Subject: Re: bin/166660: [libc] [patch] New util/shlib to change per-fd
default stdio buffering mode
Date: Sat, 14 Apr 2012 15:11:31 +0200
On Tue, Apr 10, 2012 at 09:43:16AM -0400, John Baldwin wrote:
> On Monday, April 09, 2012 5:21:03 pm Jeremie Le Hen wrote:
> > Hi John,
> >
> > On Mon, Apr 09, 2012 at 11:30:08AM -0400, John Baldwin wrote:
> > > I think it would be fine to do this in libc directly rather than via
> > > LD_PRELOAD. That would let it work for static binaries as well as
> > > dynamic libraries. My understanding is that this is how stdbuf works on
> > > Linux (glibc honors the relevant magic environment variables). To that
> > > end, I think it would be ok to move this into libc directly.
> >
> > I thought it would be too expensive to check for three (actually up to
> > six, see below) in such a critical path. Moreover, this would have
> > lowered a lot my chances to see this committed simply because very few
> > committers would have taken the responsibility for this and the time to
> > handle the debates that would have sprouted.
> >
> > Your point for static binaries is very valid but aren't you afraid of
> > the performance impact? I'll try to spare some time this week to move
> > libstdbuf code into libc and do some benchmarks.
>
> Hmm, I hadn't considered the performance impact, but to be honest, this
> is stdio. :) If it only happens once when stdio is first used then I think
> this is fine to do in libc.
I looked in the stdio source to see how I could implement there
efficiently, but the problem is that there isn't a single entry point.
The best I can do I think is basically something like this:
int stdbuf_done = 0;
void
_stdbuf()
{
/* libstdbuf code */
stdbuf_done = 1;
}
#define STDBUF() if (!stdbuf_done) _stdbuf()
And scatter STDBUF() all around. What do you think of it?
(FWIW, I checked how Linux implemented this, they used an additional
shared library.)
> > > One more question, do you use the same environment variable as glibc for
> > > this, or do you use a different scheme?
> >
> > I didn't like the GNU variable names (_STDBUF_I, _STDBUF_O and
> > _STDBUF_E) so I used STDBUF_0, STDBUF_1 and STDBUF_2 instead. But the
> > former are supported for obvious compatibility reasons. To be honest I
> > don't really care about the names, we can use the GNU ones if you think
> > it's better to avoid doing to much strcmp(3), especially if we but the
> > code in the libc startup path.
>
> If the variable values have the same semantics, then I think it is best to
> simply use the same names as glibc.
Ok, I'll do this.
--
Jeremie Le Hen
Men are born free and equal. Later on, they're on their own.
Jean Yanne
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201204141320.q3EDKCSE076995>