On Mon, Aug 17, 2015 at 10:42:56PM -0400, Rich Felker wrote:
> On Mon, Aug 17, 2015 at 02:19:34PM -0700, H.J. Lu wrote:
> > On Tue, Jun 23, 2015 at 9:18 PM, Rich Felker <dal...@libc.org> wrote:
> > > For background on the static PIE model I'm working with, see the
> > > following post to the GCC list:
> > >
> > > https://gcc.gnu.org/ml/gcc/2015-06/msg00008.html
> > >
> > > So far, I've been prototyping static PIE support by having GCC pass
> > > the following options to ld instead of -static -pie:
> > >
> > > -static -shared -Bsymbolic
> > >
> > > This partly works, but since ld does not know it's producing a main
> > > executable, it misses important details, including the ability to link
> > > initial-exec and local-exec model TLS code correctly, as well as
> > > various linking optimizations. So I think the right way forward is
> > > making ld accept -static and -pie together to do the right thing.
> > >
> > > In elflink.c, _bfd_elf_link_create_dynamic_sections assumes that
> > > executables should always have a .interp section.
> > > bfd_elf_size_dynamic_sections asserts this assumption again, and the
> > > individual elf??-*.c files also do so in *_elf_size_dynamic_sections
> > > where they set a default interpreter. (Is this even useful? Most of
> > > the names are out of touch with reality, and GCC always passes an
> > > explicit -dynamic-linker anyway, so I think this code should just be
> > > removed.)
> > >
> > > Now I have a working prototype by changing the info->executable
> > > condition to info->executable && info->dynamic, and having lexsup.c
> > > store the value of input_flags.dynamic in link_info.dynamic after
> > > processing the command line, but I'm not sure if this is the right
> > > approach.
> >
> > It is OK to use -static/-Bstatic/-non_shared with -shared and -pie.
> > I think you want --no-dynamic-linker.
>
> I see two overall approaches to making the option to omit .interp:
>
> 1. In elflink.c, make the creation of the .interp section conditional
> on a new field in link_info.
>
> 2. In ld code (ldlang.c? elf32.em?), check the command line option and
> remove the .interp section before it can be processed.
>
> I think option 1 is a lot cleaner, but it's also going to be a lot
> more invasive, because every single target arch (elf32-*.c and
> elf64-*.c) has its own ASSERT that the .interp section exists. These
> would also need to be updated to either check the new field in
> link_info, or to replace the ASSERT with a conditional.
>
> Before I spend a lot of time implementing one or the other, do you
> have any feelings on which way would be appropriate?
I went ahead and did option 1 modulo all the target code except sh
which is where I'm testing it. My work-in-progress patch is attached.
This is obviously not ready to submit but I would appreciate any
feedback that's possible at this stage.
Rich
--- binutils-2.25.1.orig/bfd/elf32-sh.c
+++ binutils-2.25.1/bfd/elf32-sh.c
@@ -3344,7 +3344,7 @@
if (htab->root.dynamic_sections_created)
{
/* Set the contents of the .interp section to the interpreter. */
- if (info->executable)
+ if (info->executable && !info->nointerp)
{
s = bfd_get_linker_section (dynobj, ".interp");
BFD_ASSERT (s != NULL);
--- binutils-2.25.1.orig/bfd/elflink.c
+++ binutils-2.25.1/bfd/elflink.c
@@ -206,7 +206,7 @@
/* A dynamically linked executable has a .interp section, but a
shared library does not. */
- if (info->executable)
+ if (info->executable && !info->nointerp)
{
s = bfd_make_section_anyway_with_flags (abfd, ".interp",
flags | SEC_READONLY);
@@ -5620,7 +5620,7 @@
bfd_boolean all_defined;
*sinterpptr = bfd_get_linker_section (dynobj, ".interp");
- BFD_ASSERT (*sinterpptr != NULL || !info->executable);
+ BFD_ASSERT (*sinterpptr != NULL || !info->executable || info->nointerp);
if (soname != NULL)
{
--- binutils-2.25.1.orig/include/bfdlink.h
+++ binutils-2.25.1/include/bfdlink.h
@@ -426,6 +426,9 @@
/* TRUE if BND prefix in PLT entries is always generated. */
unsigned int bndplt: 1;
+ /* TRUE if generation of .interp/PT_INTERP should be suppressed. */
+ unsigned int nointerp: 1;
+
/* Char that may appear as the first char of a symbol, but should be
skipped (like symbol_leading_char) when looking up symbols in
wrap_hash. Used by PowerPC Linux for 'dot' symbols. */
--- binutils-2.25.1.orig/ld/ldlex.h
+++ binutils-2.25.1/ld/ldlex.h
@@ -33,6 +33,7 @@
OPTION_DEFSYM,
OPTION_DEMANGLE,
OPTION_DYNAMIC_LINKER,
+ OPTION_NO_DYNAMIC_LINKER,
OPTION_SYSROOT,
OPTION_EB,
OPTION_EL,
--- binutils-2.25.1.orig/ld/lexsup.c
+++ binutils-2.25.1/ld/lexsup.c
@@ -138,6 +138,9 @@
{ {"dynamic-linker", required_argument, NULL, OPTION_DYNAMIC_LINKER},
'I', N_("PROGRAM"), N_("Set PROGRAM as the dynamic linker to use"),
TWO_DASHES },
+ { {"no-dynamic-linker", no_argument, NULL, OPTION_NO_DYNAMIC_LINKER},
+ '\0', NULL, N_("Produce an executable with no program interpreter header"),
+ TWO_DASHES },
{ {"library", required_argument, NULL, 'l'},
'l', N_("LIBNAME"), N_("Search for library LIBNAME"), TWO_DASHES },
{ {"library-path", required_argument, NULL, 'L'},
@@ -747,6 +750,11 @@
case 'I': /* Used on Solaris. */
case OPTION_DYNAMIC_LINKER:
command_line.interpreter = optarg;
+ link_info.nointerp = 0;
+ break;
+ case OPTION_NO_DYNAMIC_LINKER:
+ command_line.interpreter = NULL;
+ link_info.nointerp = 1;
break;
case OPTION_SYSROOT:
/* Already handled in ldmain.c. */
