[lldb-dev] dlsym() and RTLD_FIRST

[Zachary Turner]

The review is up on the LLVM side.  One point which was raised, and which I
agree with, is that the presence of the string makes the class much
heavier.  This string is only needed to mimic MacOSX's RTLD_FIRST behavior
on other posix platforms.  However, going back through the history of when
this was added, I never actually saw a use case from anyone saying "we
*need* this on Linux".  See the full original thread at [1].  But the TL;DR
is that the flag is nice to have on MacOSX, and the filename comparison was
added to Linux to maintain parity.

If nobody actually knows of a specific example of why this is necessary on
Linux, can we just remove this behavior on Linux?  My understanding is that
the only thing which will change by removing this for Linux is the
following: Imagine a plugin X is loaded, and X has a library dependency on
Y and Z.  X doesn't contain the plugin Initialize or Terminate symbol, but
Y or Z does.  With the filename comparison code, LoadPlugin would fail, and
without it, it would succeed and use the symbol found in Y or Z.  I can
understand that with the comparison the algorithm is a bit better, but it
seems such an extremely unusual edge case that I don't think it's a big
deal to remove it from the Linux side.

Thoughts?

[1] - http://thread.gmane.org/gmane.comp.debugging.lldb.devel/300/focus=302


On Thu, Aug 21, 2014 at 5:27 PM, Greg Clayton <gclayton at apple.com> wrote:

> Sounds good to me. Hopefully if they don't want that they might accept an
> extra boolean argument that can specify to only look in the current shared
> library and then we can switch over to using LLVM's DynamicLibrary.
>
> > On Aug 21, 2014, at 4:22 PM, Zachary Turner <zturner at google.com> wrote:
> >
> > This seems like the only case we ever want, so I'm going to post a patch
> to LLVM's DynamicLibrary class to use RTLD_FIRST on Apple, and a similar
> method of checking the module filespec on other platforms, and see if they
> accept it.  If so, I will convert our Plugin code to use LLVM's
> DynamicLibrary and then delete our DynamicLibrary
> >
> >
> > On Thu, Aug 21, 2014 at 4:08 PM, Greg Clayton <gclayton at apple.com>
> wrote:
> >
> > > On Aug 21, 2014, at 3:31 PM, Zachary Turner <zturner at google.com>
> wrote:
> > >
> > > Can someone explain this flag to me?
> >
> > It says "only look in this binary, don't look in any others. We are
> looking for a plug-in initialization function and we don't want to get one
> back from another dylib.
> >
> > As Enrico said, the email from a while back details this:
> >
> > http://comments.gmane.org/gmane.comp.debugging.lldb.devel/305
> >
> > >  I've read the documentation, but it's still not clear to me.  If you
> ask dlsym() to search some module X, why would it ever search modules other
> than X?
> >
> > I don't know but it does.
> >
> > >
> > > The reason I ask about this is that llvm support library already has a
> DynamicLibrary class whose purpose almost exactly matches what we're using
> the Host::DynamicLibrary related functions for.  However, it doesn't use
> the RTLD_FIRST flag, and so I'm not sure what the implications are of us
> using it and deleting our own DynamicLibrary code.
> >
> > It would be nice if we could specify this flag so we either find the
> symbol from libx.dylib or we don't. We don't want to find the symbol in
> liby.dylib and call it in our case.
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/lldb-dev/attachments/20140826/144c6961/attachment.html>