Re: Handling of SHELL by GNU make: not POSIX compliant

[David Boyce]

At 09:56 AM 10/8/2002 -0400, Paul D. Smith wrote:
> Have you looked at the proof-of-concept for this that Han-Wen Nienhuys
> developed?  We talked about it a bit on make-alpha a few years ago and
> he whipped up a very early sample--works only on Linux.  There's a link
> at the bottom of my "Advanced Auto-Dependencies" web page:
>
>   http://make.paulandlesley.org/autodep.html

As a matter of fact I just ran across this URL a couple of days ago and 
noticed the reference to his implementation. Unfortunately the link 404's. 
I'd love to look at it, so if anyone know where it went to please let me 
know. And if there are other tools in the same space I'd be happy to hear 
of them too, of course. I've looked into truss (that's SVR4/Solaris only), 
LD_AUDIT (same thing), strace (my understanding is this slows down 
processes quite a bit, it would be analogous to running your builds under 
control of a debugger).

Pros for the LD_PRELOAD technique appear to be that (a) it's relatively 
portable and (b) it's cheap. I've been building Perl 5.8.0 with my auditor 
turned on for a stress test and it seems to cost only a few percentage 
points over the unaudited build. This is for the raw list of files opened: 
if you were going to generate checksums for them it would take whatever 
time checksums take but that's orthogonal. Cons for LD_PRELOAD include (a) 
it won't audit a statically-linked binary, but those are pretty rare 
nowadays and I believe I can at least give a "Warning: unable to audit 
static binary 'foo'" message in that case, (b) it won't audit anything that 
uses syscall() directly; I don't think this is a significant concern and 
(c) since it's hooking into an undocumented user-space layer there's no way 
to be absolutely sure I've captured all the entry points which are capable 
of opening files. I think this last just requires some trial and error. For 
a few minutes I thought I couldn't handle mmap-ed files but then I realized 
mmap() takes a file descriptor which would have to have been gotten via 
open() anyway.

It currently works (where "works" is defined as "passes my suite of exactly 
1 test") on Linux and Solaris. I believe the LD_PRELOAD concept is portable 
to modern versions of HPUX and AIX as well, though I haven't tried and 
shared library implementations are pretty different there. I'm sure it will 
*not* be portable to many more marginal and/or older Unix implementations. 
On Windows there are equivalent ways of intercepting system calls and I 
haven't ruled out playing with them but it would be a completely orthogonal 
piece of code.

If this discussion should come off the list just let me know.

-David