hook_tests/README.md

Introduction
============
This project aims to give a simple overview on how good various x64 hooking
engines (on windows) are. I'll try to write various functions, that are hard to
patch and then see how each hooking engine does.

I'll test:

* [EasyHook](https://easyhook.github.io/)
* [PolyHook](https://github.com/stevemk14ebr/PolyHook)
* [MinHook](https://www.codeproject.com/Articles/44326/MinHook-The-Minimalistic-x-x-API-Hooking-Libra)
* [Mhook](http://codefromthe70s.org/mhook24.aspx)

(I'd like to test detours, but I'm not willing to pay for it. So that isn't
tested :( )

There are multiple things that make hooking difficult. Maybe you want to patch
while the application is running -- in that case you might get race conditions,
as the application is executing your half finished hook. Maybe the software has
some self protection features (or other software on the system provides that,
e.g. Trustee Rapport)

Evaluating how the hooking engines stack up against that is not the goal here.
Neither are non-functional criteria, like how fast it is or how much memory it
needs for each hook. This is just about the challenges the function to be
hooked itself poses.

Namely:

* Are jumps relocated?
* What about RIP adressing?
* If there's a loop at the beginning / if it's a tail recurisve function, does
  the hooking engine handle it?
* How good is the dissassembler, how many instructions does it know?
* Can it hook already hooked functions?

At first I will give a short walk through of the architecture, then quickly go
over the test cases. After that come the results and an evaluation for each
engine.

I think I found a flaw in all of them; I'll publish a small POC which should at
least detect the existence of problematic code.

**A word of caution**: my results are worse than expected, so do assume I have
made a mistake in using the libraries. I went into this expecting that some
engines at least would try to detect e.g. the loops back into the first few
bytes. But none did? That's gotta be wrong.

**Another word of caution**: parts of this are rushed and/or ugly. Please
double check parts that seem suspicious. And I'd love to get patches, even for
the most trivial things -- spelling mistakes? Yes please.

Architecture
============
This project is made up of two parts. A .DLL with the test cases and an .exe
that hooks those, tests whether they still work and prints the results.

(I could have done it all in the .exe but this makes it trivial to (at some
point) force the function to be hooked and the target function to be further
apart than 2GB. Just set fixed image bases in the project settings and you're
done)

My main concern was automatically identifying whether the hook worked. I
consider a hook to work if: a) the original function can still execute
successfully *and* b) the hook was called.

The criteria a) is really similar to a unit test. Verify that a function
returns what is expected. So for a) the .exe just runs unit tests after all the
hooks have been applied. Each failing function is reported (or the program
crashes and I can look at the callstack) so I can correlate that with which
hooking engine I'm currently testing and see where those fail. I've used
Catch2 for the unit tests, because I wanted to try it anyway.

From the get-to it was clear that I wanted to test multiple hooking engines.
And they all needed to do the same steps in the same order -- so I implemented
a basic AbstractHookingEngine with a boolean for every test case and make a
child class for each engine. The children classes have to overwrite `hook_all`
and `unhook_all`. Inbetween the calls to that, the unit tests run.

Test case: Small
================
This is just a very small function; it is smaller than the hook code will be -
so how does the library react?


	_small:
		xor eax, eax
		ret


Test case: Branch
=================
Instead of the FASM code I'll show the disassembled version, so you can see the
instruction lengths & offsets.


	0026 | 48 83 E0 01 | and rax,1
	002A | 74 17       | je test_cases.0043 --+
	002C | 48 31 C0    | xor rax,rax          |
	002F | 90          | nop                  |
	0030 | 90          | nop                  |
	...                                       |
	0041 | 90          | nop                  |
	0042 | 90          | nop                  |
	0043 | C3          | ret <----------------+


This function has a branch in the first 5 bytes. Hooking it detour-style isn't
possible without fixing that branch in the trampoline. The NOP sled is just so
the hooking engine can't cheat and just put the whole function into the
trampoline. Instead the jump in the trampoline needs to be modified so it jumps
back to the original destinations

Test case: RIP relative
=======================
One of the new things in AMD64 is RIP relative addressing. I guess the reason
to include it was to make it easier to generate PIC -- all references to data
can now be made relative, instead of absolute. So it doesn't matter anymore
where the program is loaded into memory and there's less need for the
relocation table.

A quick and dirty[1] test for this is re-implementing the well known C rand
function.


	public _rip_relative
	_rip_relative:
		mov rax, qword[seed]
		mov ecx, 214013
		mul ecx
		add eax, 2531011
		mov [seed], eax

		shr eax, 16
		and eax, 0x7FFF
		ret

	seed dd 1


The very first instruction uses rip relative addressing, thus it needs to be
fixed in the trampoline.

Test case: AVX & RDRAND
=======================

The AMD64 instruction set is extended with every CPU generation. Becayse the
hooking engines need to know the instruction lengths and their side effects to
properly apply their hooks, they need to keep up.

The actual code in the test case is boring and doesn't matter. I'm sure there
are disagreements on whether I've picked good candidates of "exotic" or new
instructions, but those were the first that came to mind.

Test case: loop and TailRec
===========================

My hypothesis before starting this evaluation was that those two cases would
make most hooking engines fail. Back in the good ol' days of x86 detour hooking
didn't require any special thought because the prologue was exactly as big as
the hook itself -- 5 bytes for `PUSH ESP; MOV EBP, ESP` and 5 bytes for `JMP +-
2GB`[2]. That isn't so easy for AMD64: a) the hook sometimes needs to be *way*
bigger b) due to changes in the calling convention and the general architecture
of AMD64 there just isn't a common prologue, used for almost all functions,
anymore.

Those by itself arn't a problem, since the hooking engines can fix all the
instructions they would overwrite. However I hypothesized that only a few would
check whether the function contained a loop that jumps back into the
instructions that have been overwritten. Consider this:

	public _loop
	_loop:
		mov rax, rcx
	@loop_loop:
		mul rcx
		nop
		nop
		nop
		loop @loop_loop ; lol
		ret

There's only 3 bytes that can be safely overwritten. Right after that is the
destination of the jump backwards. This is a very simple (and kinda pointless)
function so detecting that the loop might lead to problems shouldn't be a
problem. But consider what happens with MHook (and all the others):

_loop original:

	008C | 48 89 C8                 | mov rax,rcx
	008F | 48 F7 E1                 | mul rcx
	0092 | 90                       | nop
	0093 | 90                       | nop
	0094 | 90                       | nop
	0095 | E2 F8                    | loop test_cases.008F
	0097 | C3                       | ret

_loop hooked:

	008C | E9 0F 69 23 00           | jmp <MHook_Hooks::hookLoop>
	0091 | E1 90                    | loope test_cases.0023
	0093 | 90                       | nop
	0094 | 90                       | nop
	0095 | E2 F8                    | loop test_cases.008F
	0097 | C3                       | ret

trampoline:

	00007FFF7CD200C0 | 48 89 C8                 | mov rax,rcx
	00007FFF7CD200C3 | 48 F7 E1                 | mul rcx
	00007FFF7CD200C6 | E9 C7 96 DC FF           | jmp test_cases.0092

then executes:

	0092 | 90                       | nop
	0093 | 90                       | nop
	0094 | 90                       | nop
	0095 | E2 F8                    | loop test_cases.008F

But that jumps back into the middle of the jump and thus executes:

	008F | 23 00                    | and eax,dword ptr ds:[rax]
	0091 | E1 90                    | loope test_cases.0023

Which isn't right and will crash horribly.

(Preliminary) Results
=====================

+----------+-----+------+------------+---+------+----+-------+
|      Name|Small|Branch|RIP Relative|AVX|RDRAND|Loop|TailRec|
+----------+-----+------+------------+---+------+----+-------+
|  PolyHook|  X  |   X  |      X     | X |      |    |       |
|   MinHook|  X  |   X  |      X     |   |      |    |   X   |
|     MHook|     |      |      X     |   |      |    |       |
+----------+-----+------+------------+---+------+----+-------+

[1] This is one of the things that could easily be improved, but haven't been
because I just couldn't motivate myself. Putting the data right after the func
meant that a section containing code needed to be writable. Which is bad. Also
I load the seed DWORD as a QWORD -- which only works because the upper half is
then thrown away by the multiplication. It's shitty code is what I'm saying.

In retrospect I should have used a jump table like a switch-case could be
compiled into. That would be read only data. Oh well.

[2] And Microsoft decided at some point to make it even easier for their code
with the advent of hotpatching.