[Bug ld/19249] New: ld.so not properly linked in with ld --no-allow-shlib-undefined

[pjones at redhat dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=19249

            Bug ID: 19249
           Summary: ld.so not properly linked in with ld
                    --no-allow-shlib-undefined
           Product: binutils
           Version: unspecified
            Status: NEW
          Severity: normal
          Priority: P2
         Component: ld
          Assignee: unassigned at sourceware dot org
          Reporter: pjones at redhat dot com
  Target Milestone: ---

Some time ago I decided to try to enforce some hygiene on one
of the libraries I maintain, and started linking it with
-Wl,--fatal-warnings,--no-allow-shlib-undefined .  Then yesterday I
tried to build this tree on an aarch64 machine, and I got this error
while building a .so:

gcc -O2 -g -Werror -Wall -Wsign-compare -Wstrict-aliasing -std=gnu11 \
        -fshort-wchar -fPIC -flto -fvisibility=hidden -D_GNU_SOURCE \
        -I/root/efivar/src/include/efivar/  -Wmaybe-uninitialized \
        -fno-merge-constants \
        -Wl,--fatal-warnings,--no-allow-shlib-undefined,--default-symver \
        -Wl,-O2 -Wl,--no-undefined-version   -shared \
        -Wl,-soname,libefivar.so.0 \
        -o libefivar.so dp.o dp-acpi.o dp-hw.o dp-media.o dp-message.o \
           efivarfs.o export.o guid.o guids.o guid-symbols.o lib.o vars.o \
        -ldl
/lib64/libc.so.6: undefined reference to address@hidden'
/lib64/libc.so.6: undefined reference to address@hidden'
/lib64/libc.so.6: undefined reference to address@hidden'
/lib64/libc.so.6: undefined reference to address@hidden'
/usr/lib/gcc/aarch64-redhat-linux/4.8.5/../../../../lib64/libdl.so: undefined
reference to address@hidden'
/usr/lib/gcc/aarch64-redhat-linux/4.8.5/../../../../lib64/libdl.so: undefined
reference to address@hidden'
/usr/lib/gcc/aarch64-redhat-linux/4.8.5/../../../../lib64/libdl.so: undefined
reference to address@hidden'
collect2: error: ld returned 1 exit status

Now, I'm not using those symbols directly, but I am using
dlopen()/dlsym()/dlclose(), and no other libraries.  To me it looks like
a couple of things are happening - one is that --no-allow-shlib-undefined
is being applied to all the symbols in the link map, rather than just
the symbols going in to the target.  That's a bit unexpected, but not
normally insurmountable.  But the other thing, I eventually traced down
to this:

address@hidden ~]# for x in /usr/lib/debug/lib64/*.so.debug ; do \
                      echo -n "${x}: " ; grep -c _dl_argv ${x} ; done
/usr/lib/debug/lib64/ld-2.17.so.debug: 3
/usr/lib/debug/lib64/libBrokenLocale-2.17.so.debug: 0
/usr/lib/debug/lib64/libSegFault.so.debug: 0
/usr/lib/debug/lib64/libanl-2.17.so.debug: 0
/usr/lib/debug/lib64/libc-2.17.so.debug: 2
/usr/lib/debug/lib64/libcidn-2.17.so.debug: 0
/usr/lib/debug/lib64/libcrypt-2.17.so.debug: 0
/usr/lib/debug/lib64/libdl-2.17.so.debug: 0
/usr/lib/debug/lib64/libm-2.17.so.debug: 0
/usr/lib/debug/lib64/libnsl-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_compat-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_db-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_dns-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_files-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_hesiod-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_nis-2.17.so.debug: 0
/usr/lib/debug/lib64/libnss_nisplus-2.17.so.debug: 0
/usr/lib/debug/lib64/libpthread-2.17.so.debug: 0
/usr/lib/debug/lib64/libresolv-2.17.so.debug: 0
/usr/lib/debug/lib64/librt-2.17.so.debug: 0
/usr/lib/debug/lib64/libthread_db-1.0.so.debug: 0
/usr/lib/debug/lib64/libutil-2.17.so.debug: 0

address@hidden ~]# nm /usr/lib/debug/lib64/ld-2.17.so.debug | grep _dl_argv
0000000000031060 B _dl_argv
0000000000031060 b _dl_argv_internal
address@hidden ~]# nm /usr/lib/debug/lib64/libc-2.17.so.debug | grep _dl_argv
                 U _dl_argv@@GLIBC_PRIVATE

(The same basic outcome applies to e.g. __tls_get_addr as well.)

So that, I assume, is because large amounts of libdl are really shared
code from ld.so's version.  If I add --add-needed, this all works - but I
wouldn't think it would be necessary, since libc.so has "AS_NEEDED (
/lib/ld-linux-aarch64.so.1 )".

So to me it looks like libdl gets those symbols from the dl
implementation in ld.so, and libc.so tells it to add them automatically,
but unless I manually specify --add-needed that doesn't happen.  The result is
that using --no-allow-shlib-undefined to guarantee hygiene in a DSO
being built won't work unless you also use --add-needed, which doesn't
seem right.  If I use --add-needed, DT_NEEDED just winds up with:

  NEEDED            Shared library: [libdl.so.2]
  NEEDED            Shared library: [libc.so.6]

which is the same thing it would be without
-Wl,--add-needed,--no-allow-shlib-undefined

Somewhat confusingly, if I add /lib/ld-linux-aarch64.so.1 to my gcc
command line it works fine (as you'd expect), but changing libdl.so to
be:

OUTPUT_FORMAT(elf64-littleaarch64)
GROUP ( /lib64/libdl.so.2 AS_NEEDED ( /lib/ld-linux-aarch64.so.1 ) )

it doesn't help.  Are the AS_NEEDED()'s there and in libc.so being
ignored when --add-needed isn't there?

Anyway - I can make it work, but I don't really like using --add-needed ,
and the result there isn't really obviously meaningful either.

I asked Carlos O'Donell and Nick Clifton about this, and they said:

> On Fri, Nov 13, 2015 at 01:26:16PM -0500, Carlos O'Donell wrote:
>> On 11/13/2015 11:56 AM, Peter Jones wrote:
>>> So that, I assume, is because large amounts of libdl are really shared
>>> code from ld.so's version.  If I add --add-needed, this all works - but I
>>> wouldn't think it would be necessary, since libc.so has "AS_NEEDED (
>>> /lib/ld-linux-aarch64.so.1 )".
>>> Your analysis is correct. I don't know why AS_NEEDED is not enough.
>> 
>> In general ld.so is always treated as a "special case" and doesn't always
>> filter well through the various options to the static linker. Similarly
>> debugging ld.so is always a "special case", again, because it's not
>> "normal" and gdb developers rarely debug it.
>> 
>> I think you need to file a bug here because your use case is very sane
>> and the behaviour is not sane and we should fix that.
>
> I would definitely recommend doing this.  Although you found the problem on
> RHEL, I suspect that it is generic (at least as far as the AArch64 is
> concerned).  So I would recommend filing a bug report with the FSF binutils
> project here:
> 
>   https://sourceware.org/bugzilla/
> 
> It also helps that there are quite a few ARM engineers who examine these
> bugs and take a special interest in AArch64 problems...

So here I am, filing the bug.

All that said, I'm not at all sure this is Aarch64-specific - it may well be
that it's just the first platform on which I've hit the particular set of
configuration defaults that leads to this.

FYI, the versions of binutils and gcc in question:

binutils-2.23.52.0.1-55.el7.aarch64
gcc-4.8.5-4.el7.aarch64

-- 
You are receiving this mail because:
You are on the CC list for the bug.