xen-devel.lists.xenproject.org archive mirror
 help / color / mirror / Atom feed
* [PATCH 4 00/16] XSA55 libelf fixes for unstable
@ 2013-06-04 17:59 Ian Jackson
  2013-06-04 17:59 ` [PATCH 01/16] libelf: abolish libelf-relocate.c Ian Jackson
                   ` (17 more replies)
  0 siblings, 18 replies; 47+ messages in thread
From: Ian Jackson @ 2013-06-04 17:59 UTC (permalink / raw)
  To: xen-devel; +Cc: andrew.cooper3, mattjd, security

This is version 4 of my (prematurely-released) series to try to fix
libelf.  This version deals better with some possibly-out-of-control
loops, fixes the three so-far-known regressions, and should fix the
32-bit ARM build.

Here is a summary of the series:

a  01/16 libelf: abolish libelf-relocate.c
A  02/16 libxc: introduce xc_dom_seg_to_ptr_pages
A  03/16 libelf: add `struct elf_binary*' parameter to elf_load_image
a  04/16 libelf: abolish elf_sval and elf_access_signed
A  05/16 libelf: move include of <asm/guest_access.h> to top of file
a  06/16 libelf/xc_dom_load_elf_symtab: Do not use "syms" uninitialised
 * 07/16 libxl: introduce macros for memory access and pointer handling
a  08/16 tools/xcutils/readnotes: adjust print_l1_mfn_valid_note
A  09/16 libelf: check nul-terminated strings properly
 * 10/16 libxl: check all pointer accesses
A  11/16 libxl: Check pointer references in elf_is_elfbinary
 * 12/16 libelf: Make all callers call elf_check_broken
 + 13/16 libelf: use C99 bool for booleans
 / 14/16 libelf: use only unsigned integers
 * 15/16 libelf: check loops for running away
a  16/16 libelf: abolish obsolete macros

Key to symbols:
 *   Updated in this version of the series.
 +   New patch in this version.
 /   Updated but only to remove changes into a separate patch.
 -   Updated with style changes only.
 a   Acked/reviwed by one reviewer.
 A   Acked/reviwed by more than one reviewer.
Also in every patch:
Updated commit msgs to correct email address for me.


libelf, and some of its callers, did not do nearly enough checking on
their input.  Invalid inputs could cause signed integer arithmetic
overflows and wild pointer dereferences.

In this series we try to systematically eliminate this problem in a
way which has a reasonable chance of (i) still accepting all
previously-accepted ELF images (ii) not having remaining security
bugs, in a form which can be reviewied to verify (i) and (ii).

The approach is:

(i) Remove all uses of signed integers (of any kind).  That
    elmininates all integer overflows as sources of undefined
    behaviour.  Of course it still means that we can get incorrect
    values throughout the code.

(ii) Replace all uses of pointers, both pointers into the supplied
    ELF image, and pointers into the output (where we are loading)
    by uintptr_t.  That eliminates all pointer arithmetic overflows as
    sources of undefined behaviour.  Of course it still means that we
    can get incorrect and unreasonable "pointer" values.

(iii) But these pointer values will be in uintptr_t, which cannot be
    simply dereferenced by mistake.  We will replace all dereferences
    by macros which do range checking; out of range reads will read 0
    and out of range writes will be ignored.  Happily most (but not
    all) of the reads in the code already go through macros which
    abstract endianness[1] and/or 32/64bitness.

    [1] Although not all the accesses use endian-aware techniques so
    in fact the code can't cope with foreign-endian ELFs.  This is a
    problem for another day.

(iv) Look for all loops and check that they are guaranteed to
    terminate.

To enable verification of correctness of these changes I provide them
as a series roughly as follows:

1-6:
   Pre-patches which make a few semantically neutral or semantically
   correct changes.  For human review.

7: Introduces a set of macros to abstract away pointer arithmetic and
   input and output image memory accesses in libelf.  Use these macros
   everwhere they are applicable.  However, define the macros in a way
   that corresponds to the existing code.  That this patch has no
   functional change can be verified by comparing the before-and-after
   assembler output from the compiler.

9. Introduce some macros for dealing with nul-terminated strings,
   defined so as not to have any functional change at this stage.

10. Change the macro definitions, and introduce the new pseudopointer
   types, pointer range checking, etc.  For close human review.  Each
   macro change is justified in the commit message.  This patch
   eliminates most of the potential wild pointer accesses.

8,11-12,14-15:
   Smaller patches for human review, fixing some leftover bugs,
   including ensuring that all loops terminate.

13: Eliminate signed integers.  Replace every "int", "int*_t",
   "long" and most "char"s by corresponding unsigned types.  This
    eliminates all integer arithmetic overflows.

After this patch, libelf should be safe against hostile input:

 * All arithmetic operations on values from the input file use
   unsigned arithmetic which is guaranteed to be defined (although
   it may of course result in wrong answers);

 * All pointer accesses based on pointers to locations which depend on
   the input file go via our range-checking accessors; accesses which
   are not to the input or output regions are ignored (reads returning
   0).

 * The loops have been checked to ensure that they terminate and are at
   worst O(image_size).

 * Whenever an array variable was declared, the code has been manually
   reviewed looking for possible out-of-bounds accesses.

This is XSA-55.

^ permalink raw reply	[flat|nested] 47+ messages in thread
* [PATCH v3.1 00/16] XSA55 libelf fixes for unstable
@ 2013-06-03 15:41 Ian Jackson
  2013-06-03 15:41 ` [PATCH 11/16] libelf: Check pointer references in elf_is_elfbinary Ian Jackson
  0 siblings, 1 reply; 47+ messages in thread
From: Ian Jackson @ 2013-06-03 15:41 UTC (permalink / raw)
  To: xen-devel; +Cc: mattjd

This is version 3.1 of my series to try to fix libelf.  (v3 was
published only via xsa.git.)

These are all in xsa.git.  Additionally, I have prepared backports for
Xen 4.2 and 4.1 which are there too.  I'm going to send all three off
to osstest for another once-over.

Review, please, particularly of the ones listed with only zero or one
ack/review.

Thanks,
Ian.

Here is a summary of the series:

a  01/16 libelf: abolish libelf-relocate.c
A  02/16 libxc: introduce xc_dom_seg_to_ptr_pages
A  03/16 libelf: add `struct elf_binary*' parameter to elf_load_image
a  04/16 libelf: abolish elf_sval and elf_access_signed
A  05/16 libelf: move include of <asm/guest_access.h> to top of file
a  06/16 libelf/xc_dom_load_elf_symtab: Do not use "syms" uninitialised
 * 07/16 libxl: introduce macros for memory access and pointer handling
a  08/16 tools/xcutils/readnotes: adjust print_l1_mfn_valid_note
A  09/16 libelf: check nul-terminated strings properly
 * 10/16 libxl: check all pointer accesses
A  11/16 libxl: Check pointer references in elf_is_elfbinary
 * 12/16 libelf: Make all callers call elf_check_broken
 + 13/16 libelf: use C99 bool for booleans
 / 14/16 libelf: use only unsigned integers
 - 15/16 libelf: check loops for running away
a  16/16 libelf: abolish obsolete macros

Key to symbols:
 *   Updated in this version of the series.
 +   New patch in this version.
 /   Updated but only to remove changes into a separate patch.
 -   Updated with style changes only.
 a   Acked/reviwed by one reviewer.
 A   Acked/reviwed by more than one reviewer.
Also in every patch:
Updated commit msgs to correct email address for me.


libelf, and some of its callers, did not do nearly enough checking on
their input.  Invalid inputs could cause signed integer arithmetic
overflows and wild pointer dereferences.

In this series we try to systematically eliminate this problem in a
way which has a reasonable chance of (i) still accepting all
previously-accepted ELF images (ii) not having remaining security
bugs, in a form which can be reviewied to verify (i) and (ii).

The approach is:

(i) Remove all uses of signed integers (of any kind).  That
    elmininates all integer overflows as sources of undefined
    behaviour.  Of course it still means that we can get incorrect
    values throughout the code.

(ii) Replace all uses of pointers, both pointers into the supplied
    ELF image, and pointers into the output (where we are loading)
    by uintptr_t.  That eliminates all pointer arithmetic overflows as
    sources of undefined behaviour.  Of course it still means that we
    can get incorrect and unreasonable "pointer" values.

(iii) But these pointer values will be in uintptr_t, which cannot be
    simply dereferenced by mistake.  We will replace all dereferences
    by macros which do range checking; out of range reads will read 0
    and out of range writes will be ignored.  Happily most (but not
    all) of the reads in the code already go through macros which
    abstract endianness[1] and/or 32/64bitness.

    [1] Although not all the accesses use endian-aware techniques so
    in fact the code can't cope with foreign-endian ELFs.  This is a
    problem for another day.

(iv) Look for all loops and check that they are guaranteed to
    terminate.

To enable verification of correctness of these changes I provide them
as a series roughly as follows:

1-6:
   Pre-patches which make a few semantically neutral or semantically
   correct changes.  For human review.

7: Introduces a set of macros to abstract away pointer arithmetic and
   input and output image memory accesses in libelf.  Use these macros
   everwhere they are applicable.  However, define the macros in a way
   that corresponds to the existing code.  That this patch has no
   functional change can be verified by comparing the before-and-after
   assembler output from the compiler.

9. Introduce some macros for dealing with nul-terminated strings,
   defined so as not to have any functional change at this stage.

10. Change the macro definitions, and introduce the new pseudopointer
   types, pointer range checking, etc.  For close human review.  Each
   macro change is justified in the commit message.  This patch
   eliminates most of the potential wild pointer accesses.

8,11-12,14-15:
   Smaller patches for human review, fixing some leftover bugs,
   including ensuring that all loops terminate.

13: Eliminate signed integers.  Replace every "int", "int*_t",
   "long" and most "char"s by corresponding unsigned types.  This
    eliminates all integer arithmetic overflows.

After this patch, libelf should be safe against hostile input:

 * All arithmetic operations on values from the input file use
   unsigned arithmetic which is guaranteed to be defined (although
   it may of course result in wrong answers);

 * All pointer accesses based on pointers to locations which depend on
   the input file go via our range-checking accessors; accesses which
   are not to the input or output regions are ignored (reads returning
   0).

 * The loops have been verified to terminate.

 * Whenever an array variable was declared, the code has been manually
   reviewed looking for possible out-of-bounds accesses.

This is XSA-55.

^ permalink raw reply	[flat|nested] 47+ messages in thread

end of thread, other threads:[~2013-06-07 15:50 UTC | newest]

Thread overview: 47+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2013-06-04 17:59 [PATCH 4 00/16] XSA55 libelf fixes for unstable Ian Jackson
2013-06-04 17:59 ` [PATCH 01/16] libelf: abolish libelf-relocate.c Ian Jackson
2013-06-04 17:59 ` [PATCH 02/16] libxc: introduce xc_dom_seg_to_ptr_pages Ian Jackson
2013-06-04 17:59 ` [PATCH 03/16] libelf: add `struct elf_binary*' parameter to elf_load_image Ian Jackson
2013-06-05 10:32   ` George Dunlap
2013-06-05 11:01     ` Andrew Cooper
2013-06-05 11:54     ` Ian Jackson
2013-06-04 17:59 ` [PATCH 04/16] libelf: abolish elf_sval and elf_access_signed Ian Jackson
2013-06-04 17:59 ` [PATCH 05/16] libelf: move include of <asm/guest_access.h> to top of file Ian Jackson
2013-06-04 17:59 ` [PATCH 06/16] libelf/xc_dom_load_elf_symtab: Do not use "syms" uninitialised Ian Jackson
2013-06-04 17:59 ` [PATCH 07/16] libelf: introduce macros for memory access and pointer handling Ian Jackson
2013-06-04 17:59 ` [PATCH 08/16] tools/xcutils/readnotes: adjust print_l1_mfn_valid_note Ian Jackson
2013-06-04 17:59 ` [PATCH 09/16] libelf: check nul-terminated strings properly Ian Jackson
2013-06-04 17:59 ` [PATCH 10/16] libelf: check all pointer accesses Ian Jackson
2013-06-06 11:19   ` George Dunlap
2013-06-06 14:51     ` Ian Jackson
2013-06-06 16:20       ` George Dunlap
2013-06-06 18:11         ` Ian Jackson
2013-06-06 12:25   ` Matthew Daley
2013-06-06 14:59     ` Ian Jackson
2013-06-07  3:44       ` Matthew Daley
2013-06-06 15:30     ` Ian Campbell
2013-06-07  4:03       ` Matthew Daley
2013-06-04 17:59 ` [PATCH 11/16] libelf: Check pointer references in elf_is_elfbinary Ian Jackson
2013-06-04 17:59 ` [PATCH 12/16] libelf: Make all callers call elf_check_broken Ian Jackson
2013-06-05 14:51   ` Andrew Cooper
2013-06-05 15:31   ` Andrew Cooper
2013-06-06 14:08   ` George Dunlap
2013-06-06 18:41     ` Ian Jackson
2013-06-04 18:00 ` [PATCH 13/16] libelf: use C99 bool for booleans Ian Jackson
2013-06-06 14:28   ` George Dunlap
2013-06-06 14:46     ` Ian Jackson
2013-06-04 18:00 ` [PATCH 14/16] libelf: use only unsigned integers Ian Jackson
2013-06-06 16:07   ` George Dunlap
2013-06-06 18:14     ` Ian Jackson
2013-06-07  7:14       ` Jan Beulich
2013-06-07 14:35         ` Ian Jackson
2013-06-07 15:50           ` Jan Beulich
2013-06-04 18:00 ` [PATCH 15/16] libelf: check loops for running away Ian Jackson
2013-06-04 18:00 ` [PATCH 16/16] libelf: abolish obsolete macros Ian Jackson
2013-06-04 18:08 ` [PATCH 4 00/16] XSA55 libelf fixes for unstable Ian Jackson
2013-06-04 21:39   ` Andrew Cooper
2013-06-05 11:53     ` Ian Jackson
2013-06-06 14:04 ` Matthew Daley
2013-06-06 18:39   ` Ian Jackson
2013-06-07  3:35     ` Matthew Daley
  -- strict thread matches above, loose matches on Subject: below --
2013-06-03 15:41 [PATCH v3.1 " Ian Jackson
2013-06-03 15:41 ` [PATCH 11/16] libelf: Check pointer references in elf_is_elfbinary Ian Jackson

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).