[PATCH] Initialise the machine specific part of rtld bootstrap map

[PATCH] Initialise the machine specific part of rtld bootstrap map

[Matthew Fortune]

Hi,

This is a long delayed follow up to a patch by Sandra:
https://sourceware.org/ml/libc-alpha/2015-03/msg00704.html

I've reimplemented it to be MIPS specific and lightweight to just
initialise the necessary fields. Tested using mips64el-linux-gnu (n64)
with some manual verification to make sure I saw the loader failures if I
initialised the fpabi to an illegal value instead of zero.

MIPS requires two fields in the machine specific part of the link map to
be zero initialised.  This is natually achieved except when the map is
allocated on the stack.  The only map allocated on the stack is the
bootstrap map which is often the first use of the stack space following
kernel allocation and is therefore zero.  However, if rtld is invoked
such that the stack has already been used then there may be non-zero
data and ABI checks which use the affected fields will spuriously fail.

	* elf/rtld.c (_dl_start) [ifndef DONT_USE_BOOTSTRAP_MAP]: Call
	ELF_MACHINE_INIT_MAP.
	* sysdeps/mips/dl-machine.h (ELF_MACHINE_INIT_MAP): Define macro.

Thanks,
Matthew

---
 elf/rtld.c                | 3 +++
 sysdeps/mips/dl-machine.h | 7 +++++++
 2 files changed, 10 insertions(+)

diff --git a/elf/rtld.c b/elf/rtld.c
index 647661c..31539a4 100644
--- a/elf/rtld.c
+++ b/elf/rtld.c
@@ -364,6 +364,9 @@ _dl_start (void *arg)
      do not have to use the temporary bootstrap_map.  Global variables
      are initialized to zero by default.  */
 #ifndef DONT_USE_BOOTSTRAP_MAP
+# ifdef ELF_MACHINE_INIT_MAP
+  ELF_MACHINE_INIT_MAP (bootstrap_map);
+# endif
 # ifdef HAVE_BUILTIN_MEMSET
   __builtin_memset (bootstrap_map.l_info, '\0', sizeof (bootstrap_map.l_info));
 # else
diff --git a/sysdeps/mips/dl-machine.h b/sysdeps/mips/dl-machine.h
index 8c0b40e..d929477 100644
--- a/sysdeps/mips/dl-machine.h
+++ b/sysdeps/mips/dl-machine.h
@@ -93,6 +93,13 @@ do { if ((l)->l_info[DT_MIPS (RLD_MAP_REL)]) \
 # define ELF_MACHINE_NAN2008 0
 #endif
 
+/* Initialise the machine dependent parts of a map.  This is not normally
+   required unless the map is allocated on the stack.  */
+#define ELF_MACHINE_INIT_MAP(MAP) \
+do { (MAP)->l_mach.fpabi = 0; \
+     (MAP)->l_mach.odd_spreg = 0; \
+   } while (0)
+
 /* Return nonzero iff ELF header is compatible with the running host.  */
 static inline int __attribute_used__
 elf_machine_matches_host (const ElfW(Ehdr) *ehdr)
-- 
2.2.1

Re: [PATCH] Initialise the machine specific part of rtld bootstrap map

[Joseph Myers]

On Wed, 26 Oct 2016, Matthew Fortune wrote:

> +# ifdef ELF_MACHINE_INIT_MAP
> +  ELF_MACHINE_INIT_MAP (bootstrap_map);
> +# endif

We don't encourage use of #ifdef like that.  It's better to have an inline 
function defined everywhere and used unconditionally, for which most 
systems have a dummy definition (see dl-machine-reject-phdr.h and 
elf_machine_reject_phdr_p for an example - if you have a header for a 
single function, you don't need to update lots of dl-machine.h headers, 
just add a generic version - which has the comments detailing the 
semantics of the function and when it's needed - and a MIPS version).

-- 
Joseph S. Myers
joseph@codesourcery.com

RE: [PATCH] Initialise the machine specific part of rtld bootstrap map

[Matthew Fortune]

Joseph Myers <joseph@codesourcery.com> writes:
> On Wed, 26 Oct 2016, Matthew Fortune wrote:
> 
> > +# ifdef ELF_MACHINE_INIT_MAP
> > +  ELF_MACHINE_INIT_MAP (bootstrap_map); # endif
> 
> We don't encourage use of #ifdef like that.  It's better to have an
> inline function defined everywhere and used unconditionally, for which
> most systems have a dummy definition (see dl-machine-reject-phdr.h and
> elf_machine_reject_phdr_p for an example - if you have a header for a
> single function, you don't need to update lots of dl-machine.h headers,
> just add a generic version - which has the comments detailing the
> semantics of the function and when it's needed - and a MIPS version).

Thanks Joseph. It's been a while since I did a glibc patch and couldn't
remember the recommended approach.

Do you think I should add a whole new header for this? Or, since this
is directly related to the reject_phdr feature for MIPS and only MIPS
is affected then I could just add it to dl-machine-reject-phdr.h?

Matthew

Re: [PATCH] Initialise the machine specific part of rtld bootstrap map

[Sandra Loosemore]

On 10/26/2016 03:01 PM, Matthew Fortune wrote:
> Joseph Myers <joseph@codesourcery.com> writes:
>> On Wed, 26 Oct 2016, Matthew Fortune wrote:
>>
>>> +# ifdef ELF_MACHINE_INIT_MAP
>>> +  ELF_MACHINE_INIT_MAP (bootstrap_map); # endif
>>
>> We don't encourage use of #ifdef like that.  It's better to have an
>> inline function defined everywhere and used unconditionally, for which
>> most systems have a dummy definition (see dl-machine-reject-phdr.h and
>> elf_machine_reject_phdr_p for an example - if you have a header for a
>> single function, you don't need to update lots of dl-machine.h headers,
>> just add a generic version - which has the comments detailing the
>> semantics of the function and when it's needed - and a MIPS version).
>
> Thanks Joseph. It's been a while since I did a glibc patch and couldn't
> remember the recommended approach.
>
> Do you think I should add a whole new header for this? Or, since this
> is directly related to the reject_phdr feature for MIPS and only MIPS
> is affected then I could just add it to dl-machine-reject-phdr.h?

Wouldn't it be easier and more maintainable just to unconditionally 
zero-initialize the structure, as I did in the original patch?

-Sandra

RE: [PATCH] Initialise the machine specific part of rtld bootstrap map

[Matthew Fortune]

Sandra Loosemore <sandra@codesourcery.com> writes:
> On 10/26/2016 03:01 PM, Matthew Fortune wrote:
> > Joseph Myers <joseph@codesourcery.com> writes:
> >> On Wed, 26 Oct 2016, Matthew Fortune wrote:
> >>
> >>> +# ifdef ELF_MACHINE_INIT_MAP
> >>> +  ELF_MACHINE_INIT_MAP (bootstrap_map); # endif
> >>
> >> We don't encourage use of #ifdef like that.  It's better to have an
> >> inline function defined everywhere and used unconditionally, for
> >> which most systems have a dummy definition (see
> >> dl-machine-reject-phdr.h and elf_machine_reject_phdr_p for an example
> >> - if you have a header for a single function, you don't need to
> >> update lots of dl-machine.h headers, just add a generic version -
> >> which has the comments detailing the semantics of the function and
> when it's needed - and a MIPS version).
> >
> > Thanks Joseph. It's been a while since I did a glibc patch and
> > couldn't remember the recommended approach.
> >
> > Do you think I should add a whole new header for this? Or, since this
> > is directly related to the reject_phdr feature for MIPS and only MIPS
> > is affected then I could just add it to dl-machine-reject-phdr.h?
> 
> Wouldn't it be easier and more maintainable just to unconditionally
> zero-initialize the structure, as I did in the original patch?

I was just trying to factor in the comments from Roland on the original
patch.  I don't know if the initialisation cost outweighs the drawback
of not simply initialising the whole structure.  I'm happy to be led
by other opinions though.

Matthew

RE: [PATCH] Initialise the machine specific part of rtld bootstrap map

[Joseph Myers]

On Wed, 26 Oct 2016, Matthew Fortune wrote:

> Do you think I should add a whole new header for this? Or, since this
> is directly related to the reject_phdr feature for MIPS and only MIPS
> is affected then I could just add it to dl-machine-reject-phdr.h?

If you can argue that there is no possible need for this new function 
other than with the reject_phdr feature and include that explanation in 
the comments in the header, sharing the header seems reasonable.  
Otherwise it should be a separate header.

-- 
Joseph S. Myers
joseph@codesourcery.com

RE: [PATCH] Initialise the machine specific part of rtld bootstrap map

[Matthew Fortune]

Sandra Loosemore <sandra@codesourcery.com> writes:
> On 10/26/2016 03:01 PM, Matthew Fortune wrote:
> > Joseph Myers <joseph@codesourcery.com> writes:
> >> On Wed, 26 Oct 2016, Matthew Fortune wrote:
> >>
> >>> +# ifdef ELF_MACHINE_INIT_MAP
> >>> +  ELF_MACHINE_INIT_MAP (bootstrap_map); # endif
> >>
> >> We don't encourage use of #ifdef like that.  It's better to have an
> >> inline function defined everywhere and used unconditionally, for
> >> which most systems have a dummy definition (see
> >> dl-machine-reject-phdr.h and elf_machine_reject_phdr_p for an example
> >> - if you have a header for a single function, you don't need to
> >> update lots of dl-machine.h headers, just add a generic version -
> >> which has the comments detailing the semantics of the function and
> when it's needed - and a MIPS version).
> >
> > Thanks Joseph. It's been a while since I did a glibc patch and
> > couldn't remember the recommended approach.
> >
> > Do you think I should add a whole new header for this? Or, since this
> > is directly related to the reject_phdr feature for MIPS and only MIPS
> > is affected then I could just add it to dl-machine-reject-phdr.h?
> 
> Wouldn't it be easier and more maintainable just to unconditionally
> zero-initialize the structure, as I did in the original patch?

I'd like to get some consensus on the best solution here before I do
another implementation.

We either go for:

1) The bare minimum to initialise just the fields that must be zero for
successful execution on a per architecture basis (and live with the
associated risk of missing some)

or

2) Unconditionally zero the whole l_mach link_map_machine structure

Or

3) Unconditionally zero the entire link_map

Given these I'd actually go for (2) as a good balance.

Thanks,
Matthew