On Fri, Jul 19, 2019 at 04:21:53PM +0200, Mark wrote:
> On Fri, 2019-07-19 at 16:43 +0300, Dmitry V. Levin wrote:
> > On Fri, Jul 19, 2019 at 02:47:09PM +0200, Mark Wielaard wrote:
> > [...]
> > > +static bool
> > > +is_shared (void)
> > > +{
> > > +  if (!is_loadable ())
> > > +    return false;
> > > +
> > > +  /* The ELF type is very clear: this is an executable.  */
> > > +  if (elf_type == ET_EXEC)
> > > +    return false;
> > > +
> > > +  /* If the object is marked as PIE, it is definitely an
> > > executable,
> > > +     and not a loadlable shared object.  */
> > > +  if (has_pie_flag)
> > > +    return false;
> > > +
> > > +  /* Treat a DT_SONAME tag as a strong indicator that this is a
> > > shared
> > > +     object.  */
> > > +  if (has_soname)
> > > +    return true;
> > 
> > I'm not sure DT_SONAME is a reliable indicator.
> > 
> > I've seen many cases of DT_SONAME being erroneously applied to 
> > non-libraries, e.g. lib.so was used as soname in openjdk executables.
> 
> I didn't know. Is this really common?

I don't think it is very common, but the mistake is very easy to make
(-Wl,-soname,lib.so) and it doesn't really break anything.  Apparently,
some projects apply the same linker flags that add DT_SONAME to all
generated files.

> I did find one java binary on my system that indeed has this problem.
> $ eu-readelf -d /usr/lib/jvm/java-1.8.0-openjdk-1.8.0.212.b04-
> 0.el7_6.x86_64/jre/bin/policytool
> 
> Dynamic segment contains 39 entries:
>  Addr: 0x0000000000600d88  Offset: 0x000d88  Link to section: [ 7]
> '.dynstr'
>   Type              Value
>   NEEDED            Shared library: [libpthread.so.0]
>   NEEDED            Shared library: [libz.so.1]
>   NEEDED            Shared library: [libX11.so.6]
>   NEEDED            Shared library: [libjli.so]
>   NEEDED            Shared library: [libdl.so.2]
>   NEEDED            Shared library: [libc.so.6]
>   SONAME            Library soname: [lib.so]
>   RPATH             Library rpath:
> [$ORIGIN/../lib/amd64/jli:$ORIGIN/../lib/amd64]
> [...]
> 
> But even so eu-elfclassify still doesn't treat it as a shared library,
> because:
> $ eu-elfclassify -v --shared policytool; echo $?
> info: policytool: ELF kind: ELF_K_ELF (0x3)
> info: policytool: ELF type: ET_EXEC (0x2)
> info: policytool: PT_LOAD found
> info: policytool: allocated PROGBITS section found
> info: policytool: program interpreter found
> info: policytool: dynamic segment found
> info: policytool: soname found
> info: policytool: DT_DEBUG found
> 1
> 
> So other characteristics like it being ET_EXEC mark it as an
> executable. And I assume if it was PIE (ET_DYN) the PIE DT_FLAGS would
> have caught it.

Yes, the checks above has_soname are much more definitive.

> So, I don't think the code is wrong. We might want to tweak the comment
> a bit though, to make it less definitive?

What I'm saying is that has_soname is just a hint which is probably even
less reliable than has_program_interpreter.


-- 
ldv