I can build glibc master branch with binutils 2.25.51.20151006 on
x86 and x86-64. What target were you building for?

(In reply to H.J. Lu from comment #1)
> I can build glibc master branch with binutils 2.25.51.20151006 on
> x86 and x86-64. What target were you building for?

s390x - I definitely should have mentioned this - sorry

What went wrong?

(gdb) p name
$1 = 0x8024eab7 "setjmp"
(gdb) p shortname
$2 = 0x3ff77b15020 "setjmp"
(gdb) p shortlen
$3 = 18446744071559648585
(gdb) l
1722
1723      shortlen = p - name;
1724      shortname = (char *) bfd_hash_allocate (&info->hash->table, shortlen + 1);
1725      if (shortname == NULL)
1726        return FALSE;
1727      memcpy (shortname, name, shortlen);
1728      shortname[shortlen] = '\0';
1729
1730      /* We are going to create a new symbol.  Merge it with any existing
1731         symbol with this name.  For the purposes of the merge, act as
(gdb) p p
$4 = 0x0
(gdb) p name
$5 = 0x8024eab7 "setjmp"
(gdb) p h->versioned
$6 = versioned

The problem is that there might be symbols marked as versioned but without having @ in its names. p is NULL then.

We have that with setjmp in Glibc:
234157: 0000000000015030     8 FUNC    WEAK   DEFAULT    2 setjmp@@GLIBC_2.2
234509: 0000000000015030     8 FUNC    WEAK   DEFAULT    2 setjmp

(In reply to Andreas Krebbel from comment #4)
> 
> The problem is that there might be symbols marked as versioned but without
> having @ in its names. p is NULL then.
> 
> We have that with setjmp in Glibc:
> 234157: 0000000000015030     8 FUNC    WEAK   DEFAULT    2 setjmp@@GLIBC_2.2
> 234509: 0000000000015030     8 FUNC    WEAK   DEFAULT    2 setjmp

Please provide a testcase.

Created attachment 8701 [details]
Testcase - reduced input file from glibc

With this testcase the failing symbol is getcontext:

 16488: 0000000000000000   106 FUNC    WEAK   DEFAULT    2 __v1__getcontext
 16534: 0000000000000000   106 FUNC    GLOBAL DEFAULT    2 __getcontext
 16562: 0000000000000000   106 FUNC    WEAK   DEFAULT    2 getcontext@@GLIBC_2.2
 16570: 0000000000000000   106 FUNC    WEAK   DEFAULT    2 getcontext@GLIBC_2.19
 16589: 0000000000000000   106 FUNC    WEAK   DEFAULT    2 __v2__getcontext
 16660: 0000000000000000   106 FUNC    WEAK   DEFAULT    2 getcontext

getcontext (as well as setjmp and others) on S/390 default to an older symbol version in order to avoid the buggy versions in Glibc 2.19.

Please provide a testcase I can reproduce it on x86-64 with
only cross binutils.

(In reply to H.J. Lu from comment #7)
> Please provide a testcase I can reproduce it on x86-64 with
> only cross binutils.

The testcase I've attached fails the same way with cross binutils:

install-s390x/bin/s390x-ibm-linux-ld -melf64_s390 b.os
Segmentation fault (core dumped)

The problem is that when processing "getcontext" in elf_link_add_object_symbols it is matched by the hash entry for "getcontext@@GLIBC_2.2" here:

      /* We need to make sure that indirect symbol dynamic flags are
	 updated.  */
      hi = h;
      while (h->root.type == bfd_link_hash_indirect
	     || h->root.type == bfd_link_hash_warning)
	h = (struct elf_link_hash_entry *) h->root.u.i.link;

While the name stays "getcontext" the new `h' has the `versioned' flag set. That's what confuses _bfd_elf_add_default_symbol.

This hack mimics the old behavior of exiting for all symbols not having @ in their name and fixes the problem for me:

diff --git a/bfd/elflink.c b/bfd/elflink.c
index 94bb710..2a56fd6 100644
--- a/bfd/elflink.c
+++ b/bfd/elflink.c
@@ -4374,8 +4374,9 @@ error_free_dyn:
 
          /* Check to see if we need to add an indirect symbol for
             the default name.  */
-         if (definition
-             || (!override && h->root.type == bfd_link_hash_common))
+         if (strchr (name, ELF_VER_CHR) != NULL
+             && (definition
+                 || (!override && h->root.type == bfd_link_hash_common)))
            if (!_bfd_elf_add_default_symbol (abfd, info, h, name, isym,
                                              sec, value, &old_bfd, &dynsym))
              goto error_free_vers;

(In reply to Andreas Krebbel from comment #8)
> (In reply to H.J. Lu from comment #7)
> > Please provide a testcase I can reproduce it on x86-64 with
> > only cross binutils.
> 
> The testcase I've attached fails the same way with cross binutils:
> 
> install-s390x/bin/s390x-ibm-linux-ld -melf64_s390 b.os
> Segmentation fault (core dumped)

[hjl@gnu-6 pr19073]$ cat xxx.c
void
__foo ()
{
}

asm (".weak foo_v1");
asm (".globl foo_v1");
asm (".set foo_v1, __foo");
asm (".weak foo_v2");
asm (".globl foo_v2");
asm (".set foo_v2, __foo");
asm (".symver foo_v2,foo@VERS.2");
asm (".symver foo_v1,foo@@VERS.1");
asm (".globl foo");
asm (".weak foo");
asm (".set foo, __foo");
[hjl@gnu-6 pr19073]$ make libfoo.so
gcc -B./ -fPIC   -c -o xxx.o xxx.c
./ld -o foo.o -r xxx.o
./ld -o libfoo.so -shared foo.o --version-script foo.v
Makefile:12: recipe for target 'libfoo.so' failed
make: *** [libfoo.so] Segmentation fault
make: *** Deleting file 'libfoo.so'
[hjl@gnu-6 pr19073]$ 

There is a bug in sysdeps/unix/sysv/linux/s390/s390-64/getcontext.S
which defines both getcontext@@GLIBC_2.2 and getcontext.  But linker
shouldn't crash.

The master branch has been updated by H.J. Lu <hjl@sourceware.org>:

https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=4373f8af3ddbc301227e883d5cbed8302c107e59

commit 4373f8af3ddbc301227e883d5cbed8302c107e59
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Mon Oct 12 04:55:24 2015 -0700

    Skip the unversioned definition after the default version
    
    We may see an unversioned definition after the default version.  We
    should skip the unversioned definition in this case.
    
    bfd/
    
    	PR ld/19073
    	* elflink.c (_bfd_elf_add_default_symbol): Skip the unversioned
    	definition after the default version.
    
    ld/testsuite/
    
    	PR ld/19073
    	* ld-elf/pr19073.map: New file.
    	* ld-elf/pr19073.rd: Likewise.
    	* ld-elf/pr19073.s: Likewise.
    	* ld-elf/shared.exp (build_tests): Add tests for PR ld/19073.

Fixed.

(In reply to H.J. Lu from comment #10)
> (In reply to Andreas Krebbel from comment #8)
> > (In reply to H.J. Lu from comment #7)
> > > Please provide a testcase I can reproduce it on x86-64 with
> > > only cross binutils.
> > 
> > The testcase I've attached fails the same way with cross binutils:
> > 
> > install-s390x/bin/s390x-ibm-linux-ld -melf64_s390 b.os
> > Segmentation fault (core dumped)
> 
> [hjl@gnu-6 pr19073]$ cat xxx.c
> void
> __foo ()
> {
> }
> 
> asm (".weak foo_v1");
> asm (".globl foo_v1");
> asm (".set foo_v1, __foo");
> asm (".weak foo_v2");
> asm (".globl foo_v2");
> asm (".set foo_v2, __foo");
> asm (".symver foo_v2,foo@VERS.2");
> asm (".symver foo_v1,foo@@VERS.1");
> asm (".globl foo");
> asm (".weak foo");
> asm (".set foo, __foo");
> [hjl@gnu-6 pr19073]$ make libfoo.so
> gcc -B./ -fPIC   -c -o xxx.o xxx.c
> ./ld -o foo.o -r xxx.o
> ./ld -o libfoo.so -shared foo.o --version-script foo.v
> Makefile:12: recipe for target 'libfoo.so' failed
> make: *** [libfoo.so] Segmentation fault
> make: *** Deleting file 'libfoo.so'
> [hjl@gnu-6 pr19073]$ 
> 
> There is a bug in sysdeps/unix/sysv/linux/s390/s390-64/getcontext.S
> which defines both getcontext@@GLIBC_2.2 and getcontext.  But linker
> shouldn't crash.

I agree that this does not look correct but it is similiar to e.g. setjmp on x86_64:

readelf -s libc.so.6 | grep " setjmp"
   622: 0000000000032ce0    10 FUNC    GLOBAL DEFAULT   13 setjmp@@GLIBC_2.2.5
  6015: 0000000000032ce0    10 FUNC    GLOBAL DEFAULT   13 setjmp

One difference appears to be that the symbol without version information is weak on s390.

(In reply to Andreas Krebbel from comment #13)
> > There is a bug in sysdeps/unix/sysv/linux/s390/s390-64/getcontext.S
> > which defines both getcontext@@GLIBC_2.2 and getcontext.  But linker
> > shouldn't crash.
> 
> I agree that this does not look correct but it is similiar to e.g. setjmp on
> x86_64:
> 
> readelf -s libc.so.6 | grep " setjmp"
>    622: 0000000000032ce0    10 FUNC    GLOBAL DEFAULT   13
> setjmp@@GLIBC_2.2.5
>   6015: 0000000000032ce0    10 FUNC    GLOBAL DEFAULT   13 setjmp

This is expected.

> One difference appears to be that the symbol without version information is
> weak on s390.

I was referring to the input object files, not output shared libraries.
There is no need to define both getcontext@@GLIBC_2.2 and getcontext in
sysdeps/unix/sysv/linux/s390/s390-64/getcontext.S.