[v3,0/7] binfmt_elf: Update READ_IMPLIES_EXEC logic for modern CPUs

Message

Kees Cook Feb. 10, 2020, 7:30 p.m. UTC

Hi,

This is a refresh of my earlier attempt to fix READ_IMPLIES_EXEC. I think
it incorporates the feedback from v2, and I've now added a selftest. This
series is for x86, arm, and arm64; I'd like it to go via -tip, though,
just to keep this change together with the selftest. To that end, I'd like
to collect Acks from the respective architecture maintainers. (Note that
most other architectures don't suffer from this problem. e.g. powerpc's
behavior appears to already be correct. MIPS may need adjusting but the
history of CPU features and toolchain behavior is very unclear to me.)

Repeating the commit log from later in the series:


The READ_IMPLIES_EXEC work-around was designed for old toolchains that
lacked the ELF PT_GNU_STACK marking under the assumption that toolchains
that couldn't specify executable permission flags for the stack may not
know how to do it correctly for any memory region.

This logic is sensible for having ancient binaries coexist in a system
with possibly NX memory, but was implemented in a way that equated having
a PT_GNU_STACK marked executable as being as "broken" as lacking the
PT_GNU_STACK marking entirely. Things like unmarked assembly and stack
trampolines may cause PT_GNU_STACK to need an executable bit, but they
do not imply all mappings must be executable.

This confusion has led to situations where modern programs with explicitly
marked executable stack are forced into the READ_IMPLIES_EXEC state when
no such thing is needed. (And leads to unexpected failures when mmap()ing
regions of device driver memory that wish to disallow VM_EXEC[1].)

In looking for other reasons for the READ_IMPLIES_EXEC behavior, Jann
Horn noted that glibc thread stacks have always been marked RWX (until
2003 when they started tracking the PT_GNU_STACK flag instead[2]). And
musl doesn't support executable stacks at all[3]. As such, no breakage
for multithreaded applications is expected from this change.

[1] https://lkml.kernel.org/r/[e-mail]
[2] https://sourceware.org/git/?p=glibc.git;a=commitdiff;h=54ee14b3882
[3] https://lkml.kernel.org/r/[e-mail]


-Kees


v3:
 - split steps in to distinct patches
 - include arm32 and arm64/compat
 - add selftests to validate behavior
v2: https://lore.kernel.org/lkml/20190424203408.GA11386@beast/
v1: https://lore.kernel.org/lkml/20190423181210.GA2443@beast/

Kees Cook (7):
  x86/elf: Add table to document READ_IMPLIES_EXEC
  x86/elf: Split READ_IMPLIES_EXEC from executable GNU_STACK
  x86/elf: Disable automatic READ_IMPLIES_EXEC for 64-bit address spaces
  arm32/64, elf: Add tables to document READ_IMPLIES_EXEC
  arm32/64, elf: Split READ_IMPLIES_EXEC from executable GNU_STACK
  arm64, elf: Disable automatic READ_IMPLIES_EXEC for 64-bit address
    spaces
  selftests/exec: Add READ_IMPLIES_EXEC tests

 arch/arm/kernel/elf.c                         |  27 +++-
 arch/arm64/include/asm/elf.h                  |  23 +++-
 arch/x86/include/asm/elf.h                    |  22 +++-
 fs/compat_binfmt_elf.c                        |   5 +
 tools/testing/selftests/exec/Makefile         |  42 +++++-
 .../selftests/exec/read_implies_exec.c        | 121 ++++++++++++++++++
 .../selftests/exec/strip-gnu-stack-bits.c     |  34 +++++
 .../testing/selftests/exec/strip-gnu-stack.c  |  69 ++++++++++
 8 files changed, 336 insertions(+), 7 deletions(-)
 create mode 100644 tools/testing/selftests/exec/read_implies_exec.c
 create mode 100644 tools/testing/selftests/exec/strip-gnu-stack-bits.c
 create mode 100644 tools/testing/selftests/exec/strip-gnu-stack.c

Comments

Jason Gunthorpe Feb. 11, 2020, 5:17 p.m. UTC | #1

On Mon, Feb 10, 2020 at 11:30:42AM -0800, Kees Cook wrote:
> Hi,
> 
> This is a refresh of my earlier attempt to fix READ_IMPLIES_EXEC. I think
> it incorporates the feedback from v2, and I've now added a selftest. This
> series is for x86, arm, and arm64; I'd like it to go via -tip, though,
> just to keep this change together with the selftest. To that end, I'd like
> to collect Acks from the respective architecture maintainers. (Note that
> most other architectures don't suffer from this problem. e.g. powerpc's
> behavior appears to already be correct. MIPS may need adjusting but the
> history of CPU features and toolchain behavior is very unclear to me.)
> 
> Repeating the commit log from later in the series:
> 
> 
> The READ_IMPLIES_EXEC work-around was designed for old toolchains that
> lacked the ELF PT_GNU_STACK marking under the assumption that toolchains
> that couldn't specify executable permission flags for the stack may not
> know how to do it correctly for any memory region.
> 
> This logic is sensible for having ancient binaries coexist in a system
> with possibly NX memory, but was implemented in a way that equated having
> a PT_GNU_STACK marked executable as being as "broken" as lacking the
> PT_GNU_STACK marking entirely. Things like unmarked assembly and stack
> trampolines may cause PT_GNU_STACK to need an executable bit, but they
> do not imply all mappings must be executable.
> 
> This confusion has led to situations where modern programs with explicitly
> marked executable stack are forced into the READ_IMPLIES_EXEC state when
> no such thing is needed. (And leads to unexpected failures when mmap()ing
> regions of device driver memory that wish to disallow VM_EXEC[1].)
> 
> In looking for other reasons for the READ_IMPLIES_EXEC behavior, Jann
> Horn noted that glibc thread stacks have always been marked RWX (until
> 2003 when they started tracking the PT_GNU_STACK flag instead[2]). And
> musl doesn't support executable stacks at all[3]. As such, no breakage
> for multithreaded applications is expected from this change.
> 
> [1] https://lkml.kernel.org/r/[e-mail]
> [2] https://sourceware.org/git/?p=glibc.git;a=commitdiff;h=54ee14b3882
> [3] https://lkml.kernel.org/r/[e-mail]

I'm happy to see this, I think it will help the situation.

While I'm not well versed in all the historical details, the general
approach makes sense to me and I've looked through the patches.

I would like to follow this up with a patch to again block VM_EXEC
from the RDMA related mmap of BAR paths..

Reviewed-by: Jason Gunthorpe <[e-mail]>

Jason