diff mbox

V2 [PATCH] Check multiple NT_GNU_PROPERTY_TYPE_0 notes [BZ #23509]

Message ID CAMe9rOoXP7fz+XaM5hFnzDa=qSzz=8r3HkoddaTscrh7HUt03g@mail.gmail.com
State New, archived
Headers show

Commit Message

H.J. Lu Aug. 26, 2018, 2:52 p.m. UTC
On Tue, Aug 14, 2018 at 10:32 AM, H.J. Lu <hongjiu.lu@intel.com> wrote:
> The older linker treats .note.gnu.property section as a generic note
> section and just concatenates all .note.gnu.property sections from the
> inputs to the output.  When the older linker is used to created the
> program on CET-enabled OS, the generated output has .note.gnu.property
> section with multiple NT_GNU_PROPERTY_TYPE_0 notes whose IBT and SHSTK
> enable bits are set even if the program isn't CET enabled.  Such program
> will crash on CET-enabled machines.  This patch updates the note parser:
>
> 1. Skip note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
> 2. Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
>
> OK for master?
>
> H.J.
> ---
>         [BZ #23509]
>         * sysdeps/x86/dl-prop.h (_dl_process_cet_property_note): Skip
>         note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
>         Update the l_cet field when processing NT_GNU_PROPERTY_TYPE_0 note.
>         Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
>         * sysdeps/x86/link_map.h (l_cet): Expand to 3 bits,  Add
>         lc_unknown.

Here is the updated patch which passed CET smoke test:

https://github.com/hjl-tools/cet-smoke-test

Comments

Florian Weimer Sept. 20, 2018, 8:54 a.m. UTC | #1
* H. J. Lu:

> From e1c6cdeb7530c37ddb6688b17435c1c9a373a09c Mon Sep 17 00:00:00 2001
> From: "H.J. Lu" <hjl.tools@gmail.com>
> Date: Fri, 10 Aug 2018 18:43:22 -0700
> Subject: [PATCH] Check multiple NT_GNU_PROPERTY_TYPE_0 notes [BZ #23509]
>
> The older linker treats .note.gnu.property section as a generic note
> section and just concatenates all .note.gnu.property sections from the
> inputs to the output.  When the older linker is used to created the
> program on CET-enabled OS, the generated output has .note.gnu.property
> section with multiple NT_GNU_PROPERTY_TYPE_0 notes whose IBT and SHSTK
> enable bits are set even if the program isn't CET enabled.  Such program
> will crash on CET-enabled machines.  This patch updates the note parser:
>
> 1. Skip note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
> 2. Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
>
> 	[BZ #23509]
> 	* sysdeps/x86/dl-prop.h (_dl_process_cet_property_note): Skip
> 	note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
> 	Update the l_cet field when processing NT_GNU_PROPERTY_TYPE_0 note.
> 	Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
> 	* sysdeps/x86/link_map.h (l_cet): Expand to 3 bits,  Add
> 	lc_unknown.

As far as I can tell, this patch catches only the case where a single
PT_NOTE segment contains multiple notes.  If there are multiple PT_NOTE
segments with notes, and each one contains a single note, the patch will
still enable CET.

Is my summary accurate?  Do you think it is safe to treat PT_NOTE
segments in isolation?

If yes, this should be mentioned in the commit message.

This seems to be an unrelated change?

+	      /* Property type must be in ascending order.  */
+	      if (type < last_type)
+		return;

Thanks,
Florian
H.J. Lu Sept. 20, 2018, 12:36 p.m. UTC | #2
On Thu, Sep 20, 2018 at 1:54 AM, Florian Weimer <fweimer@redhat.com> wrote:
> * H. J. Lu:
>
>> From e1c6cdeb7530c37ddb6688b17435c1c9a373a09c Mon Sep 17 00:00:00 2001
>> From: "H.J. Lu" <hjl.tools@gmail.com>
>> Date: Fri, 10 Aug 2018 18:43:22 -0700
>> Subject: [PATCH] Check multiple NT_GNU_PROPERTY_TYPE_0 notes [BZ #23509]
>>
>> The older linker treats .note.gnu.property section as a generic note
>> section and just concatenates all .note.gnu.property sections from the
>> inputs to the output.  When the older linker is used to created the
>> program on CET-enabled OS, the generated output has .note.gnu.property
>> section with multiple NT_GNU_PROPERTY_TYPE_0 notes whose IBT and SHSTK
>> enable bits are set even if the program isn't CET enabled.  Such program
>> will crash on CET-enabled machines.  This patch updates the note parser:
>>
>> 1. Skip note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
>> 2. Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
>>
>>       [BZ #23509]
>>       * sysdeps/x86/dl-prop.h (_dl_process_cet_property_note): Skip
>>       note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
>>       Update the l_cet field when processing NT_GNU_PROPERTY_TYPE_0 note.
>>       Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
>>       * sysdeps/x86/link_map.h (l_cet): Expand to 3 bits,  Add
>>       lc_unknown.
>
> As far as I can tell, this patch catches only the case where a single
> PT_NOTE segment contains multiple notes.  If there are multiple PT_NOTE
> segments with notes, and each one contains a single note, the patch will
> still enable CET.

I believe this is covered by CET smoke test:

https://github.com/hjl-tools/cet-smoke-test

My patch should work in all cases.  If it misses some cases, please add it
to master branch at

https://github.com/hjl-tools/cet-smoke-test

under the "note" directory.

> Is my summary accurate?  Do you think it is safe to treat PT_NOTE
> segments in isolation?

What do you mean by that?  The only assumption I made is that
property note must follow the spec.  Any violation makes it invalid.

> If yes, this should be mentioned in the commit message.
>
> This seems to be an unrelated change?
>
> +             /* Property type must be in ascending order.  */
> +             if (type < last_type)
> +               return;
>

This is the part of property spec.  If properties aren't properly
sorted, it is invalid.
Florian Weimer Sept. 21, 2018, 2:49 p.m. UTC | #3
* H. J. Lu:

>> As far as I can tell, this patch catches only the case where a single
>> PT_NOTE segment contains multiple notes.  If there are multiple PT_NOTE
>> segments with notes, and each one contains a single note, the patch will
>> still enable CET.
>
> I believe this is covered by CET smoke test:
>
> https://github.com/hjl-tools/cet-smoke-test
>
> My patch should work in all cases.  If it misses some cases, please add it
> to master branch at
>
> https://github.com/hjl-tools/cet-smoke-test
>
> under the "note" directory.
>
>> Is my summary accurate?  Do you think it is safe to treat PT_NOTE
>> segments in isolation?
>
> What do you mean by that?  The only assumption I made is that
> property note must follow the spec.  Any violation makes it invalid.

I'm attaching a patched executable which has two PT_NOTE segments with
separately valid GNU property notes (if I did my patching correctly).

I think your code will still accept it.  It is conceivable that a linker
which is not aware of the required merging for property notes would
create such segments, I think.

>> If yes, this should be mentioned in the commit message.
>>
>> This seems to be an unrelated change?
>>
>> +             /* Property type must be in ascending order.  */
>> +             if (type < last_type)
>> +               return;
>>
>
> This is the part of property spec.  If properties aren't properly
> sorted, it is invalid.

Okay, that part makes sense because it's internal to the notes AFAICS.
And note merging without GNU property note awareness will likely violate
this constraint.  Maybe you could add a comment to this effect?

Thanks,
Florian
H.J. Lu Sept. 21, 2018, 3:31 p.m. UTC | #4
On Fri, Sep 21, 2018 at 7:49 AM, Florian Weimer <fweimer@redhat.com> wrote:
> * H. J. Lu:
>
>>> As far as I can tell, this patch catches only the case where a single
>>> PT_NOTE segment contains multiple notes.  If there are multiple PT_NOTE
>>> segments with notes, and each one contains a single note, the patch will
>>> still enable CET.
>>
>> I believe this is covered by CET smoke test:
>>
>> https://github.com/hjl-tools/cet-smoke-test
>>
>> My patch should work in all cases.  If it misses some cases, please add it
>> to master branch at
>>
>> https://github.com/hjl-tools/cet-smoke-test
>>
>> under the "note" directory.
>>
>>> Is my summary accurate?  Do you think it is safe to treat PT_NOTE
>>> segments in isolation?
>>
>> What do you mean by that?  The only assumption I made is that
>> property note must follow the spec.  Any violation makes it invalid.
>
> I'm attaching a patched executable which has two PT_NOTE segments with
> separately valid GNU property notes (if I did my patching correctly).
>
> I think your code will still accept it.  It is conceivable that a linker
> which is not aware of the required merging for property notes would
> create such segments, I think.

My goal is to allow older linkers which don't understand property notes
to generate working executables on CET enabled OS.  It isn't tended to
handle all hand-crafted binaries.  The worst case is that such binaries
won't run on CET enabled OS.  I don't believe it is an issue.  It may even
be a bonus.

Do you have a testcase which may be generated by an old linker on
CET enabled OS?

>>> If yes, this should be mentioned in the commit message.
>>>
>>> This seems to be an unrelated change?
>>>
>>> +             /* Property type must be in ascending order.  */
>>> +             if (type < last_type)
>>> +               return;
>>>
>>
>> This is the part of property spec.  If properties aren't properly
>> sorted, it is invalid.
>
> Okay, that part makes sense because it's internal to the notes AFAICS.
> And note merging without GNU property note awareness will likely violate
> this constraint.  Maybe you could add a comment to this effect?

I updated comments to

              /* Property types must be sorted in ascending order.
                 Ignore property note with the wrong type order.  */
              if (type < last_type)
                return;

Thanks.
Florian Weimer Sept. 22, 2018, 11:07 a.m. UTC | #5
* H. J. Lu:

> My goal is to allow older linkers which don't understand property notes
> to generate working executables on CET enabled OS.  It isn't tended to
> handle all hand-crafted binaries.  The worst case is that such binaries
> won't run on CET enabled OS.  I don't believe it is an issue.  It may even
> be a bonus.
>
> Do you have a testcase which may be generated by an old linker on
> CET enabled OS?

My position is that ELF is an open format and that we don't really know
what kind of link editors are out there and are used in practice.

Other views are certainly possible.  I mainly wanted to understand the
rationale behind the code.  If you do not expect multiple PT_NOTE
segments with properly aligned property notes, I think that's okay.  We
can adjust the dynamic linker again if such binaries show up in the
future.

>> Okay, that part makes sense because it's internal to the notes AFAICS.
>> And note merging without GNU property note awareness will likely violate
>> this constraint.  Maybe you could add a comment to this effect?
>
> I updated comments to
>
>               /* Property types must be sorted in ascending order.
>                  Ignore property note with the wrong type order.  */
>               if (type < last_type)
>                 return;

Thanks.

Florian
Florian Weimer Nov. 7, 2018, 6:27 p.m. UTC | #6
* Florian Weimer:

>>> If yes, this should be mentioned in the commit message.
>>>
>>> This seems to be an unrelated change?
>>>
>>> +             /* Property type must be in ascending order.  */
>>> +             if (type < last_type)
>>> +               return;
>>>
>>
>> This is the part of property spec.  If properties aren't properly
>> sorted, it is invalid.
>
> Okay, that part makes sense because it's internal to the notes AFAICS.
> And note merging without GNU property note awareness will likely violate
> this constraint.  Maybe you could add a comment to this effect?

Hmm, isn't the point that this is the internal type in the note (not the
tag), and there isn't any *note* merging, only note *section* merging in
an old linker.  The latter leaves the note boundaries intact, so the
internal layout will be unmodified.

Or am I confused about this?

Thanks,
Florian
Florian Weimer Nov. 7, 2018, 6:32 p.m. UTC | #7
* H. J. Lu:

> The older linker treats .note.gnu.property section as a generic note
> section and just concatenates all .note.gnu.property sections from the
> inputs to the output.  When the older linker is used to created the
> program on CET-enabled OS, the generated output has .note.gnu.property
> section with multiple NT_GNU_PROPERTY_TYPE_0 notes whose IBT and SHSTK
> enable bits are set even if the program isn't CET enabled.  Such program
> will crash on CET-enabled machines.  This patch updates the note parser:
>
> 1. Skip note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
> 2. Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
>
> 	[BZ #23509]
> 	* sysdeps/x86/dl-prop.h (_dl_process_cet_property_note): Skip
> 	note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
> 	Update the l_cet field when processing NT_GNU_PROPERTY_TYPE_0 note.
> 	Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
> 	* sysdeps/x86/link_map.h (l_cet): Expand to 3 bits,  Add
> 	lc_unknown.

I would like to move this forward.

I think the code is okay, as long as we are confident that older linkers
will not produce multiple PT_NOTE segments, each containing one
NT_GNU_PROPERTY_TYPE_0 note.  Maybe this should be mentioned in the
commit message (assuming that you agree that my understanding of the
code is correct).

Thanks,
Florian
H.J. Lu Nov. 7, 2018, 8:24 p.m. UTC | #8
On Wed, Nov 7, 2018 at 10:27 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * Florian Weimer:
>
> >>> If yes, this should be mentioned in the commit message.
> >>>
> >>> This seems to be an unrelated change?
> >>>
> >>> +             /* Property type must be in ascending order.  */
> >>> +             if (type < last_type)
> >>> +               return;
> >>>
> >>
> >> This is the part of property spec.  If properties aren't properly
> >> sorted, it is invalid.
> >
> > Okay, that part makes sense because it's internal to the notes AFAICS.
> > And note merging without GNU property note awareness will likely violate
> > this constraint.  Maybe you could add a comment to this effect?
>
> Hmm, isn't the point that this is the internal type in the note (not the
> tag), and there isn't any *note* merging, only note *section* merging in
> an old linker.  The latter leaves the note boundaries intact, so the
> internal layout will be unmodified.
>

Older linkers will concatenate all input GNU property note sections
into a single output GNU property note section without merging them.
diff mbox

Patch

From e1c6cdeb7530c37ddb6688b17435c1c9a373a09c Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Fri, 10 Aug 2018 18:43:22 -0700
Subject: [PATCH] Check multiple NT_GNU_PROPERTY_TYPE_0 notes [BZ #23509]

The older linker treats .note.gnu.property section as a generic note
section and just concatenates all .note.gnu.property sections from the
inputs to the output.  When the older linker is used to created the
program on CET-enabled OS, the generated output has .note.gnu.property
section with multiple NT_GNU_PROPERTY_TYPE_0 notes whose IBT and SHSTK
enable bits are set even if the program isn't CET enabled.  Such program
will crash on CET-enabled machines.  This patch updates the note parser:

1. Skip note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
2. Check multiple NT_GNU_PROPERTY_TYPE_0 notes.

	[BZ #23509]
	* sysdeps/x86/dl-prop.h (_dl_process_cet_property_note): Skip
	note parsing if a NT_GNU_PROPERTY_TYPE_0 note has been processed.
	Update the l_cet field when processing NT_GNU_PROPERTY_TYPE_0 note.
	Check multiple NT_GNU_PROPERTY_TYPE_0 notes.
	* sysdeps/x86/link_map.h (l_cet): Expand to 3 bits,  Add
	lc_unknown.
---
 sysdeps/x86/dl-prop.h  | 51 +++++++++++++++++++++++++++++++++---------
 sysdeps/x86/link_map.h |  9 ++++----
 2 files changed, 46 insertions(+), 14 deletions(-)

diff --git a/sysdeps/x86/dl-prop.h b/sysdeps/x86/dl-prop.h
index 26c3131ac5..9ab890d12b 100644
--- a/sysdeps/x86/dl-prop.h
+++ b/sysdeps/x86/dl-prop.h
@@ -49,6 +49,10 @@  _dl_process_cet_property_note (struct link_map *l,
 			      const ElfW(Addr) align)
 {
 #if CET_ENABLED
+  /* Skip if we have seen a NT_GNU_PROPERTY_TYPE_0 note before.  */
+  if (l->l_cet != lc_unknown)
+    return;
+
   /* The NT_GNU_PROPERTY_TYPE_0 note must be aliged to 4 bytes in
      32-bit objects and to 8 bytes in 64-bit objects.  Skip notes
      with incorrect alignment.  */
@@ -57,6 +61,9 @@  _dl_process_cet_property_note (struct link_map *l,
 
   const ElfW(Addr) start = (ElfW(Addr)) note;
 
+  unsigned int feature_1 = 0;
+  unsigned int last_type = 0;
+
   while ((ElfW(Addr)) (note + 1) - start < size)
     {
       /* Find the NT_GNU_PROPERTY_TYPE_0 note.  */
@@ -64,10 +71,18 @@  _dl_process_cet_property_note (struct link_map *l,
 	  && note->n_type == NT_GNU_PROPERTY_TYPE_0
 	  && memcmp (note + 1, "GNU", 4) == 0)
 	{
+	  /* Stop if we see more than one GNU property note which may
+	     be generated by the older linker.  */
+	  if (l->l_cet != lc_unknown)
+	    return;
+
+	  /* Check CET status now.  */
+	  l->l_cet = lc_none;
+
 	  /* Check for invalid property.  */
 	  if (note->n_descsz < 8
 	      || (note->n_descsz % sizeof (ElfW(Addr))) != 0)
-	    break;
+	    return;
 
 	  /* Start and end of property array.  */
 	  unsigned char *ptr = (unsigned char *) (note + 1) + 4;
@@ -78,9 +93,15 @@  _dl_process_cet_property_note (struct link_map *l,
 	      unsigned int type = *(unsigned int *) ptr;
 	      unsigned int datasz = *(unsigned int *) (ptr + 4);
 
+	      /* Property type must be in ascending order.  */
+	      if (type < last_type)
+		return;
+
 	      ptr += 8;
 	      if ((ptr + datasz) > ptr_end)
-		break;
+		return;
+
+	      last_type = type;
 
 	      if (type == GNU_PROPERTY_X86_FEATURE_1_AND)
 		{
@@ -89,14 +110,18 @@  _dl_process_cet_property_note (struct link_map *l,
 		     we stop the search regardless if its size is correct
 		     or not.  There is no point to continue if this note
 		     is ill-formed.  */
-		  if (datasz == 4)
-		    {
-		      unsigned int feature_1 = *(unsigned int *) ptr;
-		      if ((feature_1 & GNU_PROPERTY_X86_FEATURE_1_IBT))
-			l->l_cet |= lc_ibt;
-		      if ((feature_1 & GNU_PROPERTY_X86_FEATURE_1_SHSTK))
-			l->l_cet |= lc_shstk;
-		    }
+		  if (datasz != 4)
+		    return;
+
+		  feature_1 = *(unsigned int *) ptr;
+
+		  /* Keep searching for the next GNU property note
+		     generated by the older linker.  */
+		  break;
+		}
+	      else if (type > GNU_PROPERTY_X86_FEATURE_1_AND)
+		{
+		  /* Stop since property type is in ascending order.  */
 		  return;
 		}
 
@@ -112,6 +137,12 @@  _dl_process_cet_property_note (struct link_map *l,
 	      + ELF_NOTE_NEXT_OFFSET (note->n_namesz, note->n_descsz,
 				      align));
     }
+
+  /* We get here only if there is one or no GNU property note.  */
+  if ((feature_1 & GNU_PROPERTY_X86_FEATURE_1_IBT))
+    l->l_cet |= lc_ibt;
+  if ((feature_1 & GNU_PROPERTY_X86_FEATURE_1_SHSTK))
+    l->l_cet |= lc_shstk;
 #endif
 }
 
diff --git a/sysdeps/x86/link_map.h b/sysdeps/x86/link_map.h
index ef1206a9d2..9367ed0889 100644
--- a/sysdeps/x86/link_map.h
+++ b/sysdeps/x86/link_map.h
@@ -19,8 +19,9 @@ 
 /* If this object is enabled with CET.  */
 enum
   {
-    lc_none = 0,			 /* Not enabled with CET.  */
-    lc_ibt = 1 << 0,			 /* Enabled with IBT.  */
-    lc_shstk = 1 << 1,			 /* Enabled with STSHK.  */
+    lc_unknown = 0,			 /* Unknown CET status.  */
+    lc_none = 1 << 0,			 /* Not enabled with CET.  */
+    lc_ibt = 1 << 1,			 /* Enabled with IBT.  */
+    lc_shstk = 1 << 2,			 /* Enabled with STSHK.  */
     lc_ibt_and_shstk = lc_ibt | lc_shstk /* Enabled with both.  */
-  } l_cet:2;
+  } l_cet:3;
-- 
2.17.1