This is the mail archive of the gcc-help@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Segfault calling dlopen()'ed executable built with -pie and using TLS


On 2016-04-14 at 09:41 -0700, Cary Coutant wrote:
> > 
> > 1. a "payload" executable built with "-fPIC -pie" (or "-fPIE -pie")
> > ÂÂÂwhich uses threads and thread-local storage;
> > 
> > 2. a "loader" executable, which dlopen()s the payload binary and
> > calls
> > ÂÂÂits main().
> > 
> > Starting the loader results in a segfault at thread creation time.
> > Building payload with "-fPIC -shared" results in normally working
> > loader, but I need to be able to launch payload both directly and
> > via
> > the loader.
> > 
> > Is this a bug or I'm doing something wrong?
> 
> When you build a binary with -pie, the linker makes an executable,
> not
> a shared library, and in general you can't dlopen an executable, even
> if it's position-independent.

When I build a binary with -pie, the linker produces an ELF file with
ET_DYN in header. So it is pretty much explicitly "a shared library",
I guess...

> 
> There are a few reasons for this, but the one I suspect is causing
> the
> segfault is that the thread model used in an executable is different
> from the model used in a shared library. In an executable (including
> position-independent executables), all the TLS variables are
> allocated
> at fixed tp-relative offsets, and the code is bound to those offsets.
> When you load a shared library, the dynamic loader expects to be able
> to allocate the library's TLS variables at dynamic offsets, and it
> expects that the code will use the GOT to find these offsets.
> 
> Even when you compile your payload with -fPIC, which generates TLS
> code compatible with a shared library, the linker will rewrite the
> code where it can to take advantage of the fixed offsets, knowing
> that
> it's building an executable.

Well, I don't have any knowledge of ELF and toolchain's internals, so I
won't argue too much, but see above.
Either way, is there any way to tell the linker not to do these things
and instead produce just a shared library with an entry point?

-- 
Ivan Shapovalov / intelfx /

Attachment: signature.asc
Description: This is a digitally signed message part


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]