readme & issues

issues.md

@@ -0,0 +1,21 @@
++ unimportant
++++ bug
+
++++ fix relative calls in build_trampoline()
+
+++Detect loops to the start of the function to be hooked
+    start:
+    	jmp hook_function
+    	bla
+    	jXX start+3
+    How to fix that:
+        Also fix that jump
+        Copy more into the trampoline
+
+
+
++ fix disassemble_func()
++ fix is_end_of_function()

readme.md

@@ -0,0 +1,36 @@
+x64hook is a hooking library for x64 code that aims to be able to patch everything!
+
+The general problems with hooking general functions are:
+1) The code to be overwritten with the jump to your code has relative jumps (forward) or calls.
+   Those need to be fixed in the trampoline. As the trampoline is probably out of reach for 32bit
+   jumps all the fixed jumps need to be x64 RIP relative jumps.
+
+   Conditional jumps are also handled - there are no jcc rel64 instructions though.
+   That means that the condition is inverted as to jump, over the fixup code, further
+   in the trampoline. The fixup code then is a unconditional 64bit jump.
+
+2) The functions loops back into the code that has to be overwritten - that happens for functions
+   with a loop (duh!) but also in case of tail-recursion optimization.
+
+   Currently hooking such functions is not supported and in fact this is not even detected.
+   Detection: Search for the end of the function (which has its own problems), keep track of all jumps into the
+   first few bytes.
+
+3) Other threads might try to execute the code as it is overwritten - if that happens shit hits the fan
+   and the normal working of the application *will* be disrupted.
+   There are two methods of overcoming that problem:
+   a) "The stop-the-world approach" stops all threads to hinder them from executing.
+      That makes the hooking time intensive and easily detectable.
+   b) Writing all the problematic modifications atomically. That's not doable for > 8bytes.
+      How is that write reflected in the instruction cache?
+      Instruction fetches follow a different memory model (no lock/unlock & µOp Caches)
+   c) The approach described in "Living on the edge: Rapid-toggline probes with cross modification on x86"
+      Firstly writing a INT3 to the code which lets all other threads who execute that code jump into an newly
+      installed execution handler with a spin loop until the hooking is done - i.e. execution can be safely
+      resumed.
+      Then the code is overwritten atomically with respect to cache lines.
+      Lastly the exception handler is deactived.
+
+
+As this is a x64 hooking library running on x64 the overwritten code is fairly long, making the occurence of
+the problems detailed above very likely.