[PATCH] elf: Check invalid hole in PT_LOAD segments [BZ #28838]

[H.J. Lu]

On Mon, Jan 31, 2022 at 7:40 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu:
>
> > commit 163f625cf9becbb82dfec63a29e566324129c0cd
> > Author: H.J. Lu <hjl.tools@gmail.com>
> > Date:   Tue Dec 21 12:35:47 2021 -0800
> >
> >     elf: Remove excessive p_align check on PT_LOAD segments [BZ #28688]
> >
> > removed the p_align check against the page size.  It caused the loader
> > crash in shared objects with the invalid p_align.  Update _dl_map_segments
> > to detect invalid holes.  This fixes BZ #28838.
>
> Commit message should reference commit ID and mention the failing
> test/module name.
>
> > ---
> >  elf/dl-map-segments.h | 3 +++
> >  1 file changed, 3 insertions(+)
> >
> > diff --git a/elf/dl-map-segments.h b/elf/dl-map-segments.h
> > index 172692b120..fd24cf5d01 100644
> > --- a/elf/dl-map-segments.h
> > +++ b/elf/dl-map-segments.h
> > @@ -113,6 +113,9 @@ _dl_map_segments (struct link_map *l, int fd,
> >               unallocated.  Then jump into the normal segment-mapping loop to
> >               handle the portion of the segment past the end of the file
> >               mapping.  */
> > +       if (__glibc_unlikely (loadcmds[nloadcmds - 1].mapstart <
> > +                             c->mapend))
> > +         return N_("ELF load command address/offset not page-aligned");
> >            if (__glibc_unlikely
> >                (__mprotect ((caddr_t) (l->l_addr + c->mapend),
> >                             loadcmds[nloadcmds - 1].mapstart - c->mapend,
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
If loadcmds[nloadcmds - 1].mapstart < c->mapend, the length argument
passed to __mprotect is a huge number and invalid.

>
> This seems to be fairly risky because I don't think that so far, we
> enforce increasing LOAD segment addresses (although required by te EHF
> specification).
>
> Given
>
> LDFLAGS-tst-p_alignmod3.so += -Wl,-z,max-page-size=0x100,-z,common-page-size=0x100
>
> and RELRO construction for that
>
>   .../elf/tst-p_alignmod3.so: cannot change memory protections
>
> seems to be a valid failure string for this test.  However, worst case,
> there could be a different kind of failure, if the RELRO mprotect start
> is page-aligned by chance, and the kernel rounds up the end address to a
> page boundary.  The RELRO protection then covers more than what the link
> editor expected, and this can cause crashes later on.  But this isn't
> something we can detect easily, I think.
>
> Thanks,
> Florian
>


-- 
H.J.