While I share my opinion here, I don't mean that to block Madhavan's
work. I'd rather defer to people more knowledgeable in current userland
and ABI issues/limitations and plans on dealing with those, especially
to Florian Weimer. I haven't seen Florian say anything specific for or
against Madhavan's proposal, and I'd like to. (Have I missed that?)
There was a previous discussion, where I provided feedback (not much
different from the feedback here, given that the mechanism is mostly the
same).
I think it's unnecessary for the libffi use case. Precompiled code can
be loaded from disk because the libffi trampolines are so regular. On
most architectures, it's not even the code that's patched, but some of
the data driving it, which happens to be located on the same page due to
a libffi quirk.
The libffi use case is a bit strange anyway: its trampolines are
type-generic, and the per-call adjustment is data-driven. This means
that once you have libffi in the process, you have a generic
data-to-function-call mechanism available that can be abused (it's even
fully CET compatible in recent versions). And then you need to look at
the processes that use libffi. A lot of them contain bytecode
interpreters, and those enable data-driven arbitrary code execution as
well. I know that there are efforts under way to harden Python, but
it's going to be tough to get to the point where things are still
difficult for an attacker once they have the ability to make mprotect
calls.
It was pointed out to me that libffi is doing things wrong, and the
trampolines should not be type-generic, but generated so that they match
the function being called. That is, the marshal/unmarshal code would be
open-coded in the trampoline, rather than using some generic mechanism
plus run-time dispatch on data tables describing the function type.
That is a very different design (and typically used by compilers (JIT or
not JIT) to implement native calls). Mapping some code page with a
repeating pattern would no longer work to defeat anti-JIT measures
because it's closer to real JIT. I don't know if kernel support could
make sense in this context, but it would be a completely different
patch.
Thanks,
Florian
--
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill
From: Andy Lutomirski <luto@kernel.org> Date: 2020-09-23 18:09:48
On Wed, Sep 23, 2020 at 7:39 AM Florian Weimer [off-list ref] wrote:
* Solar Designer:
quoted
While I share my opinion here, I don't mean that to block Madhavan's
work. I'd rather defer to people more knowledgeable in current userland
and ABI issues/limitations and plans on dealing with those, especially
to Florian Weimer. I haven't seen Florian say anything specific for or
against Madhavan's proposal, and I'd like to. (Have I missed that?)
There was a previous discussion, where I provided feedback (not much
different from the feedback here, given that the mechanism is mostly the
same).
I think it's unnecessary for the libffi use case. Precompiled code can
be loaded from disk because the libffi trampolines are so regular. On
most architectures, it's not even the code that's patched, but some of
the data driving it, which happens to be located on the same page due to
a libffi quirk.
The libffi use case is a bit strange anyway: its trampolines are
type-generic, and the per-call adjustment is data-driven. This means
that once you have libffi in the process, you have a generic
data-to-function-call mechanism available that can be abused (it's even
fully CET compatible in recent versions). And then you need to look at
the processes that use libffi. A lot of them contain bytecode
interpreters, and those enable data-driven arbitrary code execution as
well. I know that there are efforts under way to harden Python, but
it's going to be tough to get to the point where things are still
difficult for an attacker once they have the ability to make mprotect
calls.
It was pointed out to me that libffi is doing things wrong, and the
trampolines should not be type-generic, but generated so that they match
the function being called. That is, the marshal/unmarshal code would be
open-coded in the trampoline, rather than using some generic mechanism
plus run-time dispatch on data tables describing the function type.
That is a very different design (and typically used by compilers (JIT or
not JIT) to implement native calls). Mapping some code page with a
repeating pattern would no longer work to defeat anti-JIT measures
because it's closer to real JIT. I don't know if kernel support could
make sense in this context, but it would be a completely different
patch.
I would very much like to see a well-designed kernel facility for
helping userspace do JIT in a safer manner, but designing such a thing
is likely to be distinctly nontrivial. To throw a half-backed idea
out there, suppose a program could pre-declare a list of JIT
verifiers:
static bool ffi_trampoline_verifier(void *target_address, size_t
target_size, void *source_data, void *context);
struct jit_verifier {
.magic = 0xMAGIC_HERE,
.verifier = ffi_trampoline_verifier,
} my_verifier __attribute((section("something special here?)));
and then a system call something like:
instantiate_jit_code(target, source, size, &my_verifier, context);
The idea being that even an attacker that can force a call to
instantiate_jit_code() can only create code that passes verification
by one of the pre-declared verifiers in the process.
From: Solar Designer <hidden> Date: 2020-09-23 18:11:46
On Wed, Sep 23, 2020 at 04:39:31PM +0200, Florian Weimer wrote:
* Solar Designer:
quoted
While I share my opinion here, I don't mean that to block Madhavan's
work. I'd rather defer to people more knowledgeable in current userland
and ABI issues/limitations and plans on dealing with those, especially
to Florian Weimer. I haven't seen Florian say anything specific for or
against Madhavan's proposal, and I'd like to. (Have I missed that?)
[...]
I think it's unnecessary for the libffi use case.
[...]
I don't know if kernel support could
make sense in this context, but it would be a completely different
patch.
Thanks. Are there currently relevant use cases where the proposed
trampfd would be useful and likely actually made use of by userland -
e.g., specific userland project developers saying they'd use it, or
Madhavan intending to develop and contribute userland patches?
Alexander
On Wed, Sep 23, 2020 at 08:11:36PM +0200, Solar Designer wrote:
On Wed, Sep 23, 2020 at 04:39:31PM +0200, Florian Weimer wrote:
quoted
* Solar Designer:
quoted
While I share my opinion here, I don't mean that to block Madhavan's
work. I'd rather defer to people more knowledgeable in current userland
and ABI issues/limitations and plans on dealing with those, especially
to Florian Weimer. I haven't seen Florian say anything specific for or
against Madhavan's proposal, and I'd like to. (Have I missed that?)
[...]
quoted
I think it's unnecessary for the libffi use case.
[...]
quoted
I don't know if kernel support could
make sense in this context, but it would be a completely different
patch.
Thanks. Are there currently relevant use cases where the proposed
trampfd would be useful and likely actually made use of by userland -
e.g., specific userland project developers saying they'd use it, or
Madhavan intending to develop and contribute userland patches?
Alexander
The trampoline it provides in this version can be implemented completely
in userspace. The kernel part of it is essentially just providing a way
to do text relocations without needing a WX mapping, but the text
relocations would be unnecessary in the first place if the trampoline
was position-independent code.
From: Madhavan T. Venkataraman <hidden> Date: 2020-09-23 23:53:29
On 9/23/20 9:39 AM, Florian Weimer wrote:
* Solar Designer:
quoted
While I share my opinion here, I don't mean that to block Madhavan's
work. I'd rather defer to people more knowledgeable in current userland
and ABI issues/limitations and plans on dealing with those, especially
to Florian Weimer. I haven't seen Florian say anything specific for or
against Madhavan's proposal, and I'd like to. (Have I missed that?)
There was a previous discussion, where I provided feedback (not much
different from the feedback here, given that the mechanism is mostly the
same).
I think it's unnecessary for the libffi use case. Precompiled code can
be loaded from disk because the libffi trampolines are so regular. On
most architectures, it's not even the code that's patched, but some of
the data driving it, which happens to be located on the same page due to
a libffi quirk.
The libffi use case is a bit strange anyway: its trampolines are
type-generic, and the per-call adjustment is data-driven. This means
that once you have libffi in the process, you have a generic
data-to-function-call mechanism available that can be abused (it's even
fully CET compatible in recent versions). And then you need to look at
the processes that use libffi. A lot of them contain bytecode
interpreters, and those enable data-driven arbitrary code execution as
well. I know that there are efforts under way to harden Python, but
it's going to be tough to get to the point where things are still
difficult for an attacker once they have the ability to make mprotect
calls.
It was pointed out to me that libffi is doing things wrong, and the
trampolines should not be type-generic, but generated so that they match
the function being called. That is, the marshal/unmarshal code would be
open-coded in the trampoline, rather than using some generic mechanism
plus run-time dispatch on data tables describing the function type.
That is a very different design (and typically used by compilers (JIT or
not JIT) to implement native calls). Mapping some code page with a
repeating pattern would no longer work to defeat anti-JIT measures
because it's closer to real JIT. I don't know if kernel support could
make sense in this context, but it would be a completely different
patch.
Thanks,
Florian
Hi Florian,
I am making myself familiar with anti-JIT measures before I can respond
to this comment. Bear with me. I will also respond to the above
libffi comment.
Madhavan