[PATCH] x86: Make protected symbols local for -shared

Fangrui Song i@maskray.me
Tue Jul 19 03:13:10 GMT 2022


On 2022-07-18, H.J. Lu wrote:
>On Mon, Jun 27, 2022 at 09:18:27PM -0700, Fangrui Song wrote:
>> On 2022-06-27, H.J. Lu wrote:
>> > On Mon, Jun 27, 2022 at 8:44 PM Fangrui Song <i@maskray.me> wrote:
>> > >
>> > > On 2022-06-27, H.J. Lu wrote:
>> > > >On Mon, Jun 27, 2022 at 8:07 PM Fangrui Song <i@maskray.me> wrote:
>> > > >>
>> > > >> On 2022-06-27, H.J. Lu wrote:
>> > > >> >On Mon, Jun 27, 2022 at 11:46 AM Fangrui Song <i@maskray.me> wrote:
>> > > >> >>
>> > > >> >> On 2022-06-27, H.J. Lu wrote:
>> > > >> >> >On Mon, Jun 27, 2022 at 10:53 AM Fangrui Song <i@maskray.me> wrote:
>> > > >> >> >>
>> > > >> >> >> On 2022-06-27, H.J. Lu wrote:
>> > > >> >> >> >On Mon, Jun 27, 2022 at 10:09 AM Fangrui Song <i@maskray.me> wrote:
>> > > >> >> >> >>
>> > > >> >> >> >> On 2022-06-27, H.J. Lu wrote:
>> > > >> >> >> >> >On Sun, Jun 26, 2022 at 12:03 PM Fangrui Song <i@maskray.me> wrote:
>> > > >> >> >> >> >>
>> > > >> >> >> >> >> On 2022-06-26, H.J. Lu wrote:
>> > > >> >> >> >> >> >On Sat, Jun 25, 2022 at 10:44 AM Fangrui Song <i@maskray.me> wrote:
>> > > >> >> >> >> >> >>
>> > > >> >> >> >> >> >> Call _bfd_elf_symbol_refs_local_p with local_protected==true.  This has
>> > > >> >> >> >> >> >> 2 noticeable effects for -shared:
>> > > >> >> >> >> >> >>
>> > > >> >> >> >> >> >> * GOT-generating relocations referencing a protected data symbol no
>> > > >> >> >> >> >> >>   longer lead to a GLOB_DAT (similar to a hidden symbol).
>> > > >> >> >> >> >> >> * Direct access relocations (e.g. R_X86_64_PC32) no longer has the
>> > > >> >> >> >> >> >>   confusing diagnostic below.
>> > > >> >> >> >> >> >>
>> > > >> >> >> >> >> >>     __attribute__((visibility("protected"))) void *foo() {
>> > > >> >> >> >> >> >>       return (void *)foo;
>> > > >> >> >> >> >> >>     }
>> > > >> >> >> >> >> >>
>> > > >> >> >> >> >> >>     // gcc -fpic -shared -fuse-ld=bfd
>> > > >> >> >> >> >> >>     relocation R_X86_64_PC32 against protected symbol `foo' can not be used when making a shared object
>> > > >> >> >> >> >> >>
>> > > >> >> >> >> >> >> The new behavior matches arm, aarch64 (commit
>> > > >> >> >> >> >> >> 83c325007c5599fa9b60b8d5f7b84842160e1d1b), and powerpc ports, and other
>> > > >> >> >> >> >> >> linkers: gold and ld.lld.
>> > > >> >> >> >> >> >>
>> > > >> >> >> >> >> >> Note: if some code tries to use direct access relocations to take the
>> > > >> >> >> >> >> >> address of foo, the pointer equality will break, but the error should be
>> > > >> >> >> >> >> >> reported on the executable link, not on the innocent shared object link.
>> > > >> >> >> >> >> >> glibc 2.36 will give a warning at relocation resolving time.
>> > > >> >> >> >> >> >
>> > > >> >> >> >> >> >It should be controlled by -z [no]indirect-extern-access.   Can you enable
>> > > >> >> >> >> >> >-z  indirect-extern-access with -shared by default instead?
>> > > >> >> >> >> >>
>> > > >> >> >> >> >> If I set `link_info.indirect_extern_access = 1;` in ld/ldmain.c,
>> > > >> >> >> >> >> bfd/elf-properties.c:654 will create a
>> > > >> >> >> >> >> GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS note.
>> > > >> >> >> >> >> This will probably be unexpected (and check-ld will have 280+ failures).
>> > > >> >> >> >> >
>> > > >> >> >> >> >This is normal when the default behavior is changed.  You can pass
>> > > >> >> >> >> >-z noindirect-extern-access to these testcases.
>> > > >> >> >> >>
>> > > >> >> >> >> Adding GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS will be a
>> > > >> >> >> >> significant behavior change and may unnecessarily break user programs
>> > > >> >> >> >> (glibc will report an error instead of a warning).
>> > > >> >> >> >
>> > > >> >> >> >If glibc reports an error, it is a real bug with unknown consequences
>> > > >> >> >> >when the copy in the executable is out of sync with the protected
>> > > >> >> >> >symbol in the shared library,
>> > > >> >> >>
>> > > >> >> >> Not necessary.
>> > > >> >> >>
>> > > >> >> >> In glibc, GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS has two effects,
>> > > >> >> >> 1 (copy relocations) and 2 (non-zero value of an undefined function
>> > > >> >> >> symbol) on
>> > > >> >> >> https://sourceware.org/pipermail/libc-alpha/2022-June/139552.html
>> > > >> >> >>
>> > > >> >> >> 2 does not necessarily cause a problem. In many cases it doesn't as
>> > > >> >> >> function pointer equality is not an invariant a program relies upon
>> > > >> >> >> (at least, for many functions, the property is not used). My previous
>> > > >> >> >> comment has mentioned two cases.
>> > > >> >> >>
>> > > >> >> >> 1 likely causes a problem, but technically the shared object can define
>> > > >> >> >> a protected data symbol without accessing it..
>> > > >> >> >
>> > > >> >> >These are unknown consequences.   We don't know what the worst
>> > > >> >> >cases are.
>> > > >> >>
>> > > >> >> They are, just like when a shared object is linked with -Bsymbolic.
>> > > >> >
>> > > >> >They have to deal with it since it is done on purpose.
>> > > >> >
>> > > >> >> This patch focuses on changing the x86 default to a sane value (matching
>> > > >> >> aarch64/arm/powerpc64/riscv/etc) and enabling future removal of
>> > > >> >> `extern_protected_data`.  If you want to switch to
>> > > >> >> indirect-extern-access default for x86, while I think unnecessary, I will not object.
>> > > >> >
>> > > >> >extern_protected_data can be safely removed only when
>> > > >> >direct access to external symbols are disallowed.   We can't
>> > > >> >have both ways.
>> > > >>
>> > > >> Just define has_no_copy_on_protected to 1 to catch the usage at link
>> > > >
>> > > >This is the same as using -z indirect-extern-access on executable.
>> > > >
>> > > >> time.  ld's aarch64 port has such an error by default.  gold and ld.lld
>> > > >> has such an error for a long time now.
>> > > >>
>> > > >> We don't need to worry about whether this stricter behavior breaks user
>> > > >> programs.  As is, protected symbol using GCC+binutils provides no
>> > > >> benefit.  Programs just avoid protected data symbols.
>> > > >
>> > > >Then there should be no problems with
>> > > >GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS.
>> > > >I'd like to disallow copy relocation on protected symbols at run-time
>> > > >when there are unknown consequences.
>> > >
>> > > Enabling GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS for x86 by default
>> > > has these effects:
>> > >
>> > > * 280+ check-ld tests will fail
>> >
>> > They should be updated.
>>
>> That will be a huge effort and may not be so necessary. See below.
>>
>> > > * The GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS note appears
>> >
>> > It will disallow copy relocation on protected symbols at run-time.
>> >
>> > >    redundant. It encodes an intention explicit but the intention is
>> > >    in ld aarch64, gold (all ports), and lld (all ports) with no extra option.
>> > >
>> > > IMO, we should do these:
>> > >
>> > > * push this commit
>> > > * treat elf_has_no_copy_on_protected as always true and remove all GNU_PROPERTY_NO_COPY_ON_PROTECTED
>> > >
>> > > Again, I understand that there is concern about protected data symbols
>> > > in shared object.  But as is, nobody uses protected symbols in shared objects.
>> > > My
>> > >
>> > >    // gcc -fpic -shared -fuse-ld=bfd
>> > >    __attribute__((visibility("protected"))) void *foo() {
>> > >      return (void *)foo;
>> > >    }
>> > >
>> > > example indicates that protected future symbol is also broken.
>> >
>> > To get protected symbol to work properly on x86-64, copy relocation on protected
>> > symbols should be disallowed at run-time.
>>
>> Yes that GNU_PROPERTY_1_NEEDED_INDIRECT_EXTERN_ACCESS will change the
>> glibc warning to an error, but we don't need to hurry making the cases
>> an error.  Since protected symbols do not have performance benefits (in
>> gcc's many ports and GNU ld's x86 port), people avoid using it.  My
>> advise is to just let ld stop producing executables which will trigger
>> glibc warning/error (this has precedent in gold and ld.lld and FreeBSD's
>> adoption of ld.lld means that this goes actually very well).  Projects
>> will gradually fix their builds to enable indirect external access in
>> the rare case they encounter protected symbols in shared objects.  Then
>> in a few years, the glibc warning can naturally upgrade to an error,
>> with possibly a method (e.g. similar to LD_DYNAMIC_WEAK) to downgrade to
>> a warning.  Finally, remove the opt-out method.
>>
>> With this scheme no GNU property is needed.
>
>Then, linker should disallow copy relocation against protected symbols
>and non-canonical reference to canonical protected functions.
>
>Something like this.
>
>
>H.J.
>----
>x86: Disallow invalid relocations against protected symbols
>
>Since glibc 2.36 will issue warnings for copy relocation against
>protected symbols and non-canonical reference to canonical protected
>functions, change the linker to always disallow such relocations.

Thanks.  When reporting relocation diagnostics, making the condition
stricter by removing elf_has_indirect_extern_access is the right
direction.


Your patch alone isn't sufficient to make -fpic -shared below work:

__attribute__((visibility("protected"))) void *foo() {
   return (void *)foo;
}


>bfd/
>
>	* elf32-i386.c (elf_i386_scan_relocs): Remove check for
>	elf_has_indirect_extern_access.
>	* elf64-x86-64.c (elf_x86_64_scan_relocs): Likewise.
>	(elf_x86_64_relocate_section): Remove check for
>	elf_has_no_copy_on_protected.
>	* elfxx-x86.c (elf_x86_allocate_dynrelocs): Check for building
>	executable instead of elf_has_no_copy_on_protected.
>	(_bfd_x86_elf_adjust_dynamic_symbol): Disallow copy relocation
>	against non-copyable protected symbol.
>	* elfxx-x86.h (SYMBOL_NO_COPYRELOC): Remove check for
>	elf_has_no_copy_on_protected.
>
>ld/
>
>	* testsuite/ld-i386/i386.exp: Expect linker error for PR ld/17709
>	test.
>	* testsuite/ld-i386/pr17709.rd: Removed.
>	* testsuite/ld-i386/pr17709.err: New file.
>	* testsuite/ld-x86-64/pr17709.rd: Removed.
>	* testsuite/ld-x86-64/pr17709.err: New file.
>	* testsuite/ld-x86-64/pr28875-func.err: Updated.
>	* testsuite/ld-x86-64/x86-64.exp: Expect linker error for PR
>	ld/17709 test.  Add tests for function pointer against protected
>	function.
>---
> bfd/elf32-i386.c                        |  3 +--
> bfd/elf64-x86-64.c                      | 10 +++-------
> bfd/elfxx-x86.c                         | 21 +++++++++++++++++++--
> bfd/elfxx-x86.h                         |  3 +--
> ld/testsuite/ld-i386/i386.exp           |  2 +-
> ld/testsuite/ld-i386/pr17709.err        |  2 ++
> ld/testsuite/ld-i386/pr17709.rd         |  4 ----
> ld/testsuite/ld-x86-64/pr17709.err      |  2 ++
> ld/testsuite/ld-x86-64/pr17709.rd       |  4 ----
> ld/testsuite/ld-x86-64/pr28875-func.err |  2 +-
> ld/testsuite/ld-x86-64/x86-64.exp       | 18 +++++++++++++++++-
> 11 files changed, 47 insertions(+), 24 deletions(-)
> create mode 100644 ld/testsuite/ld-i386/pr17709.err
> delete mode 100644 ld/testsuite/ld-i386/pr17709.rd
> create mode 100644 ld/testsuite/ld-x86-64/pr17709.err
> delete mode 100644 ld/testsuite/ld-x86-64/pr17709.rd
>
>diff --git a/bfd/elf32-i386.c b/bfd/elf32-i386.c
>index 04a972e646d..cfb0085b245 100644
>--- a/bfd/elf32-i386.c
>+++ b/bfd/elf32-i386.c
>@@ -1812,8 +1812,7 @@ elf_i386_scan_relocs (bfd *abfd,
> 		      && h->type == STT_FUNC
> 		      && eh->def_protected
> 		      && !SYMBOL_DEFINED_NON_SHARED_P (h)
>-		      && h->def_dynamic
>-		      && elf_has_indirect_extern_access (h->root.u.def.section->owner))
>+		      && h->def_dynamic)
> 		    {
> 		      /* Disallow non-canonical reference to canonical
> 			 protected function.  */
>diff --git a/bfd/elf64-x86-64.c b/bfd/elf64-x86-64.c
>index 3abc68a4127..62a9a22317a 100644
>--- a/bfd/elf64-x86-64.c
>+++ b/bfd/elf64-x86-64.c
>@@ -2255,8 +2255,7 @@ elf_x86_64_scan_relocs (bfd *abfd, struct bfd_link_info *info,
> 		      && h->type == STT_FUNC
> 		      && eh->def_protected
> 		      && !SYMBOL_DEFINED_NON_SHARED_P (h)
>-		      && h->def_dynamic
>-		      && elf_has_indirect_extern_access (h->root.u.def.section->owner))
>+		      && h->def_dynamic)
> 		    {
> 		      /* Disallow non-canonical reference to canonical
> 			 protected function.  */
>@@ -3156,8 +3155,7 @@ elf_x86_64_relocate_section (bfd *output_bfd,
> 	       || (h != NULL
> 		   && !h->root.linker_def
> 		   && !h->root.ldscript_def
>-		   && eh->def_protected
>-		   && elf_has_no_copy_on_protected (h->root.u.def.section->owner)));
>+		   && eh->def_protected));
>
> 	  if ((input_section->flags & SEC_ALLOC) != 0
> 	      && (input_section->flags & SEC_READONLY) != 0
>@@ -4097,9 +4095,7 @@ elf_x86_64_relocate_section (bfd *output_bfd,
> 	    {
> 	    case R_X86_64_32S:
> 	      sec = h->root.u.def.section;
>-	      if ((info->nocopyreloc
>-		   || (eh->def_protected
>-		       && elf_has_no_copy_on_protected (h->root.u.def.section->owner)))
>+	      if ((info->nocopyreloc || eh->def_protected)
> 		  && !(h->root.u.def.section->flags & SEC_CODE))
> 		return elf_x86_64_need_pic (info, input_bfd, input_section,
> 					    h, NULL, NULL, howto);
>diff --git a/bfd/elfxx-x86.c b/bfd/elfxx-x86.c
>index 18f3d335458..7fb972752b3 100644
>--- a/bfd/elfxx-x86.c
>+++ b/bfd/elfxx-x86.c
>@@ -524,8 +524,7 @@ elf_x86_allocate_dynrelocs (struct elf_link_hash_entry *h, void *inf)
>     {
>       asection *sreloc;
>
>-      if (eh->def_protected
>-	  && elf_has_no_copy_on_protected (h->root.u.def.section->owner))
>+      if (eh->def_protected && bfd_link_executable (info))
> 	{
> 	  /* Disallow copy relocation against non-copyable protected
> 	     symbol.  */
>@@ -3041,6 +3040,24 @@ _bfd_x86_elf_adjust_dynamic_symbol (struct bfd_link_info *info,
>     }
>   if ((h->root.u.def.section->flags & SEC_ALLOC) != 0 && h->size != 0)
>     {
>+      if (eh->def_protected && bfd_link_executable (info))
>+	for (p = h->dyn_relocs; p != NULL; p = p->next)
>+	  {
>+	    /* Disallow copy relocation against non-copyable protected
>+	       symbol.  */
>+	    s = p->sec->output_section;
>+	    if (s != NULL && (s->flags & SEC_READONLY) != 0)
>+	      {
>+		info->callbacks->einfo
>+		  /* xgettext:c-format */
>+		  (_("%F%P: %pB: copy relocation against non-copyable "
>+		     "protected symbol `%s' in %pB\n"),
>+		   p->sec->owner, h->root.root.string,
>+		   h->root.u.def.section->owner);
>+		return false;
>+	      }
>+	  }
>+
>       srel->size += htab->sizeof_reloc;
>       h->needs_copy = 1;
>     }
>diff --git a/bfd/elfxx-x86.h b/bfd/elfxx-x86.h
>index 77fb1ad72bc..7d23893938c 100644
>--- a/bfd/elfxx-x86.h
>+++ b/bfd/elfxx-x86.h
>@@ -135,12 +135,11 @@
>
> /* Should copy relocation be generated for a symbol.  Don't generate
>    copy relocation against a protected symbol defined in a shared
>-   object with GNU_PROPERTY_NO_COPY_ON_PROTECTED.  */
>+   object.  */
> #define SYMBOL_NO_COPYRELOC(INFO, EH) \
>   ((EH)->def_protected \
>    && ((EH)->elf.root.type == bfd_link_hash_defined \
>        || (EH)->elf.root.type == bfd_link_hash_defweak) \
>-   && elf_has_no_copy_on_protected ((EH)->elf.root.u.def.section->owner) \
>    && ((EH)->elf.root.u.def.section->owner->flags & DYNAMIC) != 0 \
>    && ((EH)->elf.root.u.def.section->flags & SEC_CODE) == 0)
>
>diff --git a/ld/testsuite/ld-i386/i386.exp b/ld/testsuite/ld-i386/i386.exp
>index b4f7de49fd5..0ab9c001336 100644
>--- a/ld/testsuite/ld-i386/i386.exp
>+++ b/ld/testsuite/ld-i386/i386.exp
>@@ -224,7 +224,7 @@ set i386tests {
>      "--32 -mx86-used-note=yes" {pr17709a.s} {} "libpr17709.so"}
>     {"PR ld/17709 (2)" "-melf_i386 tmpdir/libpr17709.so" ""
>      "--32 -mx86-used-note=yes"
>-     {pr17709b.s} {{readelf -r pr17709.rd}} "pr17709"}
>+     {pr17709b.s} {{ld "pr17709.err"}} "pr17709"}
>     {"Build pr19827a.o" "" ""
>      "--32 -mx86-used-note=yes" { pr19827a.S }}
>     {"Build pr19827b.so" "-melf_i386 -shared" ""
>diff --git a/ld/testsuite/ld-i386/pr17709.err b/ld/testsuite/ld-i386/pr17709.err
>new file mode 100644
>index 00000000000..fa6a4bacce3
>--- /dev/null
>+++ b/ld/testsuite/ld-i386/pr17709.err
>@@ -0,0 +1,2 @@
>+.*: tmpdir/pr17709b.o: copy relocation against non-copyable protected symbol `foo' in tmpdir/libpr17709.so
>+#...
>diff --git a/ld/testsuite/ld-i386/pr17709.rd b/ld/testsuite/ld-i386/pr17709.rd
>deleted file mode 100644
>index 8414784b736..00000000000
>--- a/ld/testsuite/ld-i386/pr17709.rd
>+++ /dev/null
>@@ -1,4 +0,0 @@
>-
>-Relocation section '.rel\..*' at offset .* contains 1 entry:
>- Offset     Info    Type            Sym\.Value  Sym\. Name
>-[0-9a-f ]+R_386_COPY +[0-9a-f]+ +foo
>diff --git a/ld/testsuite/ld-x86-64/pr17709.err b/ld/testsuite/ld-x86-64/pr17709.err
>new file mode 100644
>index 00000000000..fa6a4bacce3
>--- /dev/null
>+++ b/ld/testsuite/ld-x86-64/pr17709.err
>@@ -0,0 +1,2 @@
>+.*: tmpdir/pr17709b.o: copy relocation against non-copyable protected symbol `foo' in tmpdir/libpr17709.so
>+#...
>diff --git a/ld/testsuite/ld-x86-64/pr17709.rd b/ld/testsuite/ld-x86-64/pr17709.rd
>deleted file mode 100644
>index beffd3cb34c..00000000000
>--- a/ld/testsuite/ld-x86-64/pr17709.rd
>+++ /dev/null
>@@ -1,4 +0,0 @@
>-
>-Relocation section '.rela\..*' at offset .* contains 1 entry:
>- +Offset +Info +Type +Symbol's Value +Symbol's Name \+ Addend
>-[0-9a-f ]+R_X86_64_COPY+[0-9a-f ]+ +foo \+ 0
>diff --git a/ld/testsuite/ld-x86-64/pr28875-func.err b/ld/testsuite/ld-x86-64/pr28875-func.err
>index 64e961cb3d4..f6f4658deaf 100644
>--- a/ld/testsuite/ld-x86-64/pr28875-func.err
>+++ b/ld/testsuite/ld-x86-64/pr28875-func.err
>@@ -1,2 +1,2 @@
>-.*: tmpdir/protected-func-1b.o: non-canonical reference to canonical protected function `protected_func_1a' in tmpdir/libprotected-func-2b.so
>+.*: tmpdir/protected-func-1b.o: non-canonical reference to canonical protected function `protected_func_1a' in tmpdir/libprotected-func-2..so
> #...
>diff --git a/ld/testsuite/ld-x86-64/x86-64.exp b/ld/testsuite/ld-x86-64/x86-64.exp
>index a096c0b9d0f..e6a834a2a61 100644
>--- a/ld/testsuite/ld-x86-64/x86-64.exp
>+++ b/ld/testsuite/ld-x86-64/x86-64.exp
>@@ -177,7 +177,7 @@ set x86_64tests {
>     {"PR ld/17709 (1)" "-melf_x86_64 -shared" ""
>      "--64" {pr17709a.s} {} "libpr17709.so"}
>     {"PR ld/17709 (2)" "-melf_x86_64 tmpdir/libpr17709.so" ""
>-     "--64" {pr17709b.s} {{readelf -rW pr17709.rd}} "pr17709"}
>+     "--64" {pr17709b.s} {{ld "pr17709.err"}} "pr17709"}
>     {"Build pr19827a.o" "" ""
>      "--64" { pr19827a.S }}
>     {"Build pr19827b.so" "-melf_x86_64 -shared" ""
>@@ -1383,6 +1383,22 @@ if { [isnative] && [check_compiler_available] } {
> 	    {{error_output "pr28875-func.err"}} \
> 	    "protected-func-2" \
> 	] \
>+	[list \
>+	    "Build libprotected-func-2c.so" \
>+	    "-shared" \
>+	    "-fPIC -Wa,-mx86-used-note=yes" \
>+	    { protected-func-2c.c } \
>+	    {}  \
>+	    "libprotected-func-2c.so" \
>+	] \
>+	[list \
>+	    "Build protected-func-2a without PIE" \
>+	    "$NOPIE_LDFLAGS -Wl,--no-as-needed tmpdir/libprotected-func-2c.so" \
>+	    "$NOPIE_CFLAGS -Wa,-mx86-used-note=yes" \
>+	    { protected-func-1b.c } \
>+	    {{error_output "pr28875-func.err"}} \
>+	    "protected-func-2a" \
>+	] \
> 	[list \
> 	    "Build libprotected-data-1a.so" \
> 	    "-shared -z noindirect-extern-access" \
>-- 
>2.36.1
>


More information about the Binutils mailing list