Re: [PATCH v3 3/3] elf: Also try DT_RUNPATH for LD_AUDIT dlopen [BZ #28455]

["H.J. Lu" <hjl.tools@gmail.com>]

On Mon, Dec 13, 2021 at 2:07 PM Carlos O'Donell <carlos@redhat.com> wrote:
>
> On 12/8/21 10:25, H.J. Lu wrote:
> > DT_RUNPATH is only used to find the immediate dependencies of the
> > executable or shared object containing the DT_RUNPATH entry.  Update
> > the glibc LD_AUDIT dlopen call to try the DT_RUNPATH entry of the
> > executable.  This partially fixes BZ #28455.
>
> Please post v4 split out.
>
> > ---
> >  elf/Makefile       |  9 +++++++--
> >  elf/dl-load.c      | 15 +++++++++++++++
> >  elf/tst-audit14a.c |  1 +
> >  3 files changed, 23 insertions(+), 2 deletions(-)
> >  create mode 100644 elf/tst-audit14a.c
> >
> > diff --git a/elf/Makefile b/elf/Makefile
> > index 4b5bdcb08e..aa6b24d790 100644
> > --- a/elf/Makefile
> > +++ b/elf/Makefile
> > @@ -247,10 +247,10 @@ ifneq ($(selinux-enabled),1)
> >  tests-execstack-yes = tst-execstack tst-execstack-needed tst-execstack-prog
> >  endif
> >  ifeq ($(have-depaudit),yes)
> > -tests += tst-audit14 tst-audit15 tst-audit16
> > +tests += tst-audit14 tst-audit15 tst-audit16 tst-audit14a
>
> Please split one test per line and sort.

Fixed.

> >  ifeq ($(run-built-tests),yes)
> >  tests-special += $(objpfx)tst-audit14-cmp.out $(objpfx)tst-audit15-cmp.out \
> > -              $(objpfx)tst-audit16-cmp.out
> > +              $(objpfx)tst-audit16-cmp.out $(objpfx)tst-audit14a-cmp.out
>
> Please split one test per line and sort.

Fixed.

> >  endif
> >  endif
> >  endif
> > @@ -1521,6 +1521,8 @@ tst-auditmany-ENV = \
> >  LDFLAGS-tst-audit14 = -Wl,--audit=tst-auditlogmod-1.so
> >  $(objpfx)tst-auditlogmod-1.so: $(libsupport)
> >  $(objpfx)tst-audit14.out: $(objpfx)tst-auditlogmod-1.so
> > +LDFLAGS-tst-audit14a = -Wl,--audit=tst-auditlogmod-1.so,--enable-new-dtags
> > +$(objpfx)tst-audit14a.out: $(objpfx)tst-auditlogmod-1.so
>
> OK.
>
> >  LDFLAGS-tst-audit15 = \
> >    -Wl,--audit=tst-auditlogmod-1.so,--depaudit=tst-auditlogmod-2.so
> >  $(objpfx)tst-auditlogmod-2.so: $(libsupport)
> > @@ -1547,6 +1549,9 @@ tst-audit17-ENV = LD_AUDIT=$(objpfx)tst-auditmod17.so
> >  $(objpfx)tst-audit14-cmp.out: tst-audit14.exp $(objpfx)tst-audit14.out
> >       cmp $^ > $@; \
> >       $(evaluate-test)
> > +$(objpfx)tst-audit14a-cmp.out: tst-audit14.exp $(objpfx)tst-audit14a.out
> > +     cmp $^ > $@; \
> > +     $(evaluate-test)
>
> OK.
>
> >  $(objpfx)tst-audit15-cmp.out: tst-audit15.exp $(objpfx)tst-audit15.out
> >       cmp $^ > $@; \
> >       $(evaluate-test)
> > diff --git a/elf/dl-load.c b/elf/dl-load.c
> > index bf8957e73c..1f424ea7c8 100644
> > --- a/elf/dl-load.c
> > +++ b/elf/dl-load.c
>
> The preceding block is like this:
>
> 2146       /* When the object has the RUNPATH information we don't use any
> 2147          RPATHs.  */
> 2148       if (loader == NULL || loader->l_info[DT_RUNPATH] == NULL)
> 2149         {
>
> The entire design of this block is to look into DT_RPATH specifically
> under two conditions:
>
> (a) We don't know anything about the loading module, it is NULL, and
>     therefore we don't know if DT_RUNPATH was present, which if it
>     was present, would prevent the use of DT_RPATH.
>
> (b) We have a loader, and we know it doesn't have DT_RUNPATH and so
>     we will look for DT_RPATH.
>
> Your changes adds (c) to this case in which DT_RUNPATH is used, and
> this seems to be logically distinct from the above two cases.
>
> My suggestion is that the code you are adding should be in a distinct
> block *after* the above block starting around line 2189.
>
> We already have a block like this:
> 2189       /* Look at the RUNPATH information for this binary.  */
> 2190       if (fd == -1 && loader != NULL
> 2191           && cache_rpath (loader, &loader->l_runpath_dirs,
> 2192                           DT_RUNPATH, "RUNPATH"))
> 2193         fd = open_path (name, namelen, mode,
> 2194                         &loader->l_runpath_dirs, &realname, &fb, loader,
> 2195                         LA_SER_RUNPATH, &found_other_class);
>
> I think it is logically more correct to add to this DT_RUNPATH handling
> than ato add to the DT_RPATH handling above.
>
> So after this block you might add:
>
> /* When processing the lookup we may need to additionally try DT_RUNPATH
>    in the executable for a glibc internal dlopen call when looking for
>    audit modules.  */
> if (__glibc_unlikely (mode & __RTLD_AUDIT)
>     && fd == -1)
>   {
>     ...
>   }

Fixed.

> The hard part is that you may want to know a few things:
> (a) main_map (unavailable, but computable via GL(dl_ns)[LM_ID_BASE]._ns_loaded)
> (b) did_main_map (unavailable)
>     * This seems like an optimization because if you already looked up via
>       main_map and it failed, doing it twice won't help, so you can refactor
>       did_main_map to function scope and retain the fact that you did this
>       check already if you want to keep the optimization.

Fixed.

> Otherwise everything else is there to implement the fix.
>

Sent out the v4 patch.

Thanks.

-- 
H.J.