Re: One year since last libtool release

[Bob Friesenhahn]

On Wed, 4 Nov 2009, Ralf Wildenhues wrote:

> * Bob Friesenhahn wrote on Wed, Nov 04, 2009 at 07:45:56PM CET:
> > I am not sure how one would test for this in an automated fashion
> > since the only effect is a performance reduction.  It requires
> > system-dependent tools in order to watch the system calls and see
> > what is actually going on.
>
> Well, as a first approximation, both the measuring and the portability
> aspects can simply be ignored.  Just having something that reliably
> shows the issue for some setup already helps.  You can even require root
> rights; that may prevent most users from ever running the test (which is
> good!) but will still enable developers to try it out in their sandbox.

I really hate to hijack this discussion thread since there were 
already two discussion threads on this topic.  Regardless, what we 
should expect to see when loading a module which has no additional 
dependencies is:

 1) read somepath/module.la
 2) stat somepath/module.so (or whatever dlname was parsed from
    module.la)
 3) dlopen somepath/module.so (or supporting OS calls like open, mmap,
    etc.)

but instead between steps 1 & 2, we see 'stat lostsapaths/module.a' 
for every path in the system's shared library search path.  The reason 
for this is pretty clear, since libltdl speculatively tries searching 
for a preloaded module first (necessary for security), but 
unfortunately, supplies the same argument to all of the available 
module loaders.

Using an existing test case like lt_dlopen.at, when we are loading 
"./good-plugin.la", the output of a tool like 'truss' (SVR4, Solaris, 
FreeBSD) or 'strace' (Linux) should not show any attempt to access a 
file called "good-plugin.a", except for possibly on AIX where a 
loadable library or module might use the .a extension.

It is not normally necessary to require root access to run tools like 
'truss' or 'strace' on a process that you own, but special access 
might be required to run a DTrace-based tool like dtruss.

The alternative to the fancy fragile test is to simply fix the 
problem, which should not be difficult since it is a simple problem. 
The solution may be to add a flag which indicates that the desired 
load should be static linkage so that only the preloader processes the 
request.

Bob
--
Bob Friesenhahn
address@hidden, http://www.simplesystems.org/users/bfriesen/
GraphicsMagick Maintainer,    http://www.GraphicsMagick.org/