Archive for the ‘Software’ Category


Sunday, November 11th, 2012

Hey guys

I’ve been pretty busy in the last few months, working, and working.

I am thinking about coming up with a diStorm-ARM version, it’s pretty self-explanatory. ;)
For the ones who are not familiar with diStorm, it is a disassembler that returns a binary structure that describes an x86/x64 instruction, instead of textual output.

Would love to hear your thoughts, as always.

diStorm Goes on Diet

Saturday, June 11th, 2011

I just wanted to share my happiness with you guys. After a long hard work (over a month in my free time, which ain’t much these days), I managed to refactor all the data-structures of the instructions DB in diStorm3. As the title says, I spared around 40kb in data! The original distorm3.dll file took around 130kb and currently it takes 90kb. I then went ahead and reconfigured the settings of the project in Visual Studio and instructed the compiler not to include the CRT shits. Then it bitched about “static constructors won’t be called and the like”, well duh. But since diStorm is written in C and I don’t have anything static to initialize (which is based on code) before the program starts, I didn’t mind it at all. And eventually I got the .dll file size to ~65kb. That’s really 50% of the original file size. This is sick.

I really don’t want to elaborate with the details of what I did, it’s really deep shit into diStorm. Hey, actually I can give you an simple example. Suppose I have around 900 mnemonics. Do not confuse mnemonic with opcodes – some opcodes share the same mnemonic although they are completely different in their behavior. You have so many variations of the instruction ‘ADD’, for instance. Just to clarify: mnemonic=display name of an opcode, opcode: the binary byte code which identifies the operation to do, instruction: all the bytes which represent the opcode and the operands, the whole.
Anyway, so there are 900 mnemonics, and the longest mnemonic by length takes 17 characters, some AVX mofo. Now since we want a quick look up in the mnemonics table, it was an array of [900][19], which means 900 mnemonics X 19 characters per mnemonic. Why 19? An extra character for the null terminating char, right? And another one for the Pascal string style – means there’s a leading length byte in front of the string data. Now you ask why I need them both: C string and Pascal string together. That’s because in diStorm all strings are concatenated very fast by using Pascal strings. And also because the guy who uses diStorm wants to use printf to display the mnemonic too, which uses C string, he will need a null terminating character at the end of the string, right?
So back to business, remember we have to allocate 19 bytes per mnemonic, even if the mnemonic is as short as ‘OR’, or ‘JZ’, we waste tons of space, right? 900×19=~17kb. And this is where you get the CPU vs. MEMORY issue once again, you get random access into the mnemonic, which is very important but it takes lots of space. Fortunately I came up with a cooler idea. I packed all the strings into a very long string, which looks something like this (copied from the source):

“\x09” “UNDEFINED\0” “\x03” “ADD\0” “\x04” “PUSH\0” “\x03” “POP\0” “\x02” “OR\0” \
“\x03” “ADC\0” “\x03” “SBB\0” “\x03” “AND\0” “\x03” “DAA\0” “\x03” “SUB\0”
and so on… 


You can see the leading length and the extra null terminating character for each mnemonic, and then it’s being followed by another mnemonic. And now it seems like we’re lost with random-access cause each string has a varying length and we can never get to the one we want… but lo and behold! Each instruction in the DB contains a field ‘opcodeId’ which denotes the index in the mnemonics array, the offset into the new mnemonics uber string. And now if you use the macro mnemonics.h supplies, you will get to the same mnemonic nevertheless. And all in all I spared around 10kb only on mnemonic strings!

FYI the macro is:

#define GET_MNEMONIC_NAME(m) ((_WMnemonic*)&_MNEMONICS[(m)])->p

As you can see, I access the mnemonics string with the given OpcodeId field which is taken from the decoded instruction and returns a WMnemonic structure, which is a Pascal string (char length; char bytes[1])…

The DB was much harder to compact, but one thing I can tell you when you serialize trees is that you can (and should) use integer-indices, rather than pointers! In x64, each pointer takes 8 bytes, for crying out loud! Now in the new layout, each index in the tree takes only 13 bits, the rest 5 bits talks about the type, where/what the index really points to… And it indirectly means that now the DB takes the same size both for x86 and x64 images, since it is not based on pointers.

Thanks for your time, I surely had pure fun :)

Memory Management

Tuesday, December 7th, 2010

In this post I decided to write about a few things you have to keep on mind while writing or designing a big application regarding memory management, I will also try to give you a new way for looking at memory allocations.

The problem is that one day you wake up and you have either memory leaks or memory fragmentation, you don’t know which or why. To eliminate the possibility of memory fragmentation, you will have to rule out memory leaks first.

Trying to find a solution to such a case you find that your dozens of DLLs work (normally but-)sharing the same CRT, so if you pass an allocated object from one DLL to be used and freed by another DLL, you’re screwed. Also, you can’t separate the CRTs, etc. The root of the problem is because each CRT has its own heap, thus if you statically link a DLL with its own CRT, it will have its own heap. If all DLLs use the same CRT, they share the same heap, or even the application’s global heap. Why is it bad? Because then if you have memory leaks in one DLL, it will dirty the whole global heap. I will talk about it soon. An advantage for separating heaps is that each thread that belongs to some DLL can access its own heap faster, without congestion on the global heap lock which all DLLs/threads need.
The idea is to separate the applications into logical groups of components, and each logical group will use its own heap, so you can pass pointers around. This is really hard to do if you have existing code and never thought about this issue. So you’re saying, “oh yeah, but if I don’t have leaks, and my code is doing alright, I don’t need it anyway”. Well I wouldn’t be so quick about thinking that. Keep on reading.

In C++ for instance, overriding new and delete for your class, is really useful, because even if it gets used in a different component, C++ is responsible to bring the “deletion” to the appropriate heap.

Now I wanna go back a bit to some background on how memory allocators work in general, which is usually in a naive fashion. It really means that when you ask for a block of memory, the memory manager (malloc for you) allocates some chunk and returns a pointer to you. Once another request for memory is being done, the next memory address is getting used to supply the memory request… (suppose the first chunk requested is still busy), and so on. What you can see here, is that the memory allocator works linearly (allocates a block after the other) and chronologically (depends on the time you asked for the allocation).

Let me give you some example so you understand what I’m talking about, in this example there are two threads: Thread A and Thread B.
Thread A allocates chunks of 0.5kb in a loop, and at some point shuts down after it’s done with its task.
Thread B allocates chunks of 0.5kb in a loop, simultaneously, and keeps on working.
And suppose they are using the same CRT, thus the same heap.

What we will see from the memory manager perspective is something like:
<0.5kb busy> <0.5kb busy> <0.5kb busy>…………………….and then something like ….

Each chunk might belong to either Thread A or Thread B. Don’t be picky about it, it’s really up to so many environmental issues (processors, locks, etc).

Suppose Thread A is finished with its memory because it was only up to do some initialization job(!) and decides to free everything it has allocated. The memory picture would be something like this now:
<0.5kb free><0.5kb busy><0.5kb free><0.5kb busy> …………………lots of <0.5kb busy belongs to ThreadB> and then ….

What we got now is a hole-y memory map, some blocks are free, the others are busy. Now if we want to allocate another 0.5kb, no problem, we can use the first free block, it will match perfectly, so everybody’s happy so far. As long as the block we want to allocate in such a memory state is smaller or equal to 0.5kb, we will get them for sure, right?
But what happens if we ask to allocate a block of 10kb, ha ah, now we have to scan the whole memory to find a free chunk whose size is bigger than 10kb, right? Can you see what’s wrong here?
What really is wrong is that you have a total free amount of X kbytes, which surely can contain 10kb, but they are all chunked, so the memory allocator has to find the next block which can satisfy such a request and it so happens that all these free kbytes are useless to us.
In short, this example is a classic memory fragmentation scenario. To check how bad a memory fragmentation is, you have to measure it by the total free memory, and largest free block size. This way, if you have 2 megs that are fragmented in tiny chunks over 200mb, and the biggest free block is 2 megs – it means you can’t even allocate a 3mb block while the total free size is 198mb. Memory is cruel.

This brings me to say that each memory-consumer has a specific behavior. A memory-consumer like Thread A, that was doing some initialization job and then got down, and freeing all its blocks; is just an init phase memory user.
A memory-consumer like Thread B, that still works after Thread A is done and uses its allocated memory, is an application life-time memory user.
Basically each memory-consumer can be classified into some behavior! Once you mix between those behaviors on the same heap, you’re sooner or later screwed.

The rule of thumb I would suggest you should follow is:
You should distinguish between each memory-consumer type, if you know something consumes memory and it really caches some objects, make sure it works as early as possible in your application. And only then allocate your application life-time objects. This causes its memory to be allocated first and not cause fragmentations to other consumers with different behaviors.

If you know you have to cache some objects while other application life-time objects have to be allocated from the same heap at the same time(!), this is where you should see the red light and create another heap for each behavior.
A practical example is when you open some application that has a workspace, now while you open a workspace in the first time, the application is designed to init its UI stuff only at that moment. Therefore while you’re loading and parsing the data from the file into memory, you’re also setting up the UI and do lots of allocations, the UI is going to stay there until you shutdown the application, and the data itself is going to be de-allocated once you close the workspace. This will create fragmentations if you don’t split the heaps. Can you see why? Think of loading and unloading workspaces and what happens to the memory. Don’t forget that each time you open a workspace, the memory chunks that need to get allocated have a different size. And that the UI still uses its memory which is probably fragmented, so it only gets worse to some extent. Another alternative would be to make sure the UI allocations happen first, no matter what. And only then you can allocate and de-allocate lots of memory blocks without worrying for fragmentations.
Even if you got the UI allocations done first, then there still might be a problem. Suppose you load the file’s data to memory and some logic works in a way that some of the data stays cached for the application life-time even after the workspace is going down. This is a big no no, it will leave you with fragmentations again, hence you should split your heap again.

That’s it mostly, I hope I managed to make it clear.
I didn’t want to talk about the implementation of the heaps, whether they are (LFH) low-fragmentation-heap or whatever, if you mix your memory-consumers, nothing will save you.

One note to mention, that if your cached-memory leaks, for whatever reason, with the new strategy and a standalone heap, you won’t hurt anybody with your leaks, and it will even be easier to track leaks per heap rather than dozens of DLLs on the same global heap. And if you already have a big application with a jungle of one heap and various memory-consumers types, there are some techniques to save the day…

diStorm3 is Ready

Monday, August 16th, 2010

diStorm3 is ready for the masses! :)
– if you want to maximize the information you get from a single instruction; Structure output rather than text, flow control analysis support and more!

Check it out now at its new google page.

Good luck!

Custom Kernel Debugging is Faster

Tuesday, July 20th, 2010

When you start to write a post you always get a problem with the headline for the post. You need to find something that will, in a few words, sum it up for the reader. I was wondering which one is better, “Boosting WinDbg”, “Faster Kernel Debugging in WinDbg”, “Hacking WinDbg” and so on. But they might be not accurate, and once you will read the post you won’t find them appropriate. But instead of talking about meta-post issues, let’s get going.

Two posts ago, I was talking about hunting a specific race condition bug we had in some software I work on. At last, I have free time to write this post and get into some interesting details about Windows Kernel and Debugging.

First I want to say that I got really pissed off that I couldn’t hunt the bug we had in the software like a normal human being, that Jond and I had to do it the lame old school way, which takes more time, lots of time. What really bothered me is that computers are fast and so is debugging, at least, should be. Why the heck do I have to sit down in front of the computer, not mentioning – trying to dupe the damned bug, and only then manage to debug it and see what’s going on wrong. Unacceptable. You might say, write a better code in the first place, I agree, but even then people have bugs, and will have, forever, and I was called to simply help.

Suppose we want to set a breakpoint on memory access this time, but something more complicated with conditions. The reason we need a condition, rather than a normal breakpoint is because the memory we want to monitor gets accessed thousands times per second, in my case with the race condition, for instance.
You’re even welcome to make the following test locally on your computer, fire up Visual Studio, and test the following code: unsigned int counter = 1; while (counter < 99999999+1) { counter++; }, set a memory access breakpoint on counter which stops when hit count reach 99999999, and time the whole process, and then time it without the bp set, and compare the result, what's the ratio you got? Isn't that just crazy? Here's an example in WinDbg's syntax, would be something like this: ba w4 0x491004 "j (poi(0x491004)==0) 'gc'" Which reads: break on write access for an integer at address 0x491004 only if its value is 0, otherwise continue execution. It will be tens-thousands times faster without the bp set, hence the debugging infrastructure, even locally (usermode), is slowing things down seriously. And think that you want to debug something similar on a remote machine, it's impossible, you are going to wait years in vain for something to happen on that machine. Think of all the COM/Pipe/USB/whatever-protocol messages that have to be transmitted back and forth the debugged machine to the debugger. And add to that the conditional breakpoint we set, someone has to see whether the condition is true or false and continue execution accordingly. And even if you use great tools like VirtualKD. Suppose you set a breakpoint on a given address, what really happens once the processor executes the instruction at that address? Obviously a lot, but I am going to talk about Windows Kernel point of view. Let's start bottom up, Interrupt #3 is being raised by the processor which ran that thread, which halts execution of the thread and transfers control _KiTrap3 in ntoskrnl. _KiTrap3 will build a context for the trapped thread, with all registers and this likely info and call CommonDispatchException with code 0x80000003 (to denote a breakpoint exception). Since the 'exception-raising' is common, everybody uses it, in other exceptions as well. CommonDispatchException calls _KiDispatchException. And _KiDispatchException is really the brain behind all the Windows-Exception mechanism. I'm not going to cover normal exception handling in Windows, which is very interesting in its own. So far nothing is new here. But we're getting to this function because it has something to do with debugging, it checks whether the _KdDebuggerEnabled is set and eventually it will call _KiDebugRoutine if it's set as well. Note that _KiDebugRoutine is a pointer to a function that gets set when the machine is debug-enabled. This is where we are going to get into business later, so as you can see the kernel has some minimal infrastructure to support kernel debugging with lots of functionality, many functions in ntoskrnl which start in "kdp", like KdpReadPhysicalMemory, KdpSetContext and many others. Eventually the controlling machine that uses WinDbg, has to speak to the remote machine using some protocol named KdCom, there's a KDCOM.DLL which is responsible for all of it. Now, once we set a breakpoint in WinDbg, I don't know exactly what happens, but I guess it’s something like this: it stores the bp in some internal table locally, then sends it to the debugged machine using this KdCom protocol, the other machine receives the command and sets the breakpoint locally. Then when the bp occurs, eventually WinDbg gets an event that describes the debug event from the other machine. Then it needs to know what to do with this bp according to the dude who debugs the machine. So much going on for what looks like a simple breakpoint. The process is very similar for single stepping as well, though sending a different exception code.

The problem with conditional breakpoints is that they are being tested for the condition locally, on the WinDbg machine, not on the server, so to speak. I agree it’s a fine design for Windows, after all, Windows wasn’t meant to be an uber debugging infrastructure, but an operating system. So having a kernel debugging builtin we should say thanks… So no complaints on the design, and yet something has to be done.

Custom Debugging to our call!

That’s the reason I decided to describe above how the debugging mechanism works in the kernel, so we know where we can intervene that process and do something useful. Since we want to do smart debugging, we have to use conditional breakpoints, otherwise in critical variables that get touched every now and then, we will have to hit F5 (‘go’) all the time, and the application we are debugging won’t get time to process. That’s clear. Next thing we realized is that the condition tests are being done locally on our machine, the one that runs WinDbg. That’s not ok, here’s the trick:
I wrote a driver that replaces (hooks) the _KiDebugRoutine with my own function, which checks for the exception code, then examines the context according to my condition and only then sends the event to WinDbg on the other machine, or simply “continues-execution”, thus the whole technique happens on the debugged machine without sending a single message outside (regarding the bp we set), unless that condition is true, and that’s why everything is thousands of times or so faster, which is now acceptable and usable. Luckily, we only need to replace a pointer to a function and using very simple tests we get the ability to filter exceptions on spot. Although we need to get our hands dirty with touching Debug-Registers and the context of the trapping thread, but that’s a win, after all.

Here’s the debug routine I used to experiment this issue (using constants tough):

int __stdcall my_debug(IN PVOID TrapFrame,
	IN PVOID Reserved,
	IN UCHAR LastChance)
	ULONG _dr6, _dr0;
	__asm {
		mov eax, dr6
		mov _dr6, eax
		mov eax, dr0
		mov _dr0, eax
	if ((ExceptionRecord->ExceptionCode == 0x80000003) &&
		(_dr6 & 0xf) &&
		(_dr0 == MY_WANTED_POINTER) &&
		(ExceptionRecord->ExceptionAddress != MY_WANTED_EIP))
		return 1;
	return old_debug_routine(TrapFrame, Reserved, ExceptionRecord, Context, PreviousMode, LastChance);

This routine checks when a breakpoint interrupt happened and stops the thread only if the pointer I wanted to monitor was accessed from a given address, else it would resume running that thread. This is where you go custom, and write whatever crazy condition you are up to. Using up to 4 breakpoints, that’s the processor limit for hardware breakpoints. Also checking out which thread or process trapped, etc. using the Kernel APIs… It just reminds me “compiled sprites” :)

I was assuming that there’s only one bp set on the machine which is the one I set through WinDbg, though this time, there was no necessity to set a conditional breakpoint in WinDbg itself, since we filter them using our own routine, and once WinDbg gets the event it will stop and let us act.

For some reason I had a problem with accessing the DRs from the Context structure, I didn’t try too hard, so I just backed to use them directly because I can.

Of course, doing what I did is not anything close to production quality, it was only a proof of concept, and it worked well. Next time that I will find myself in a weird bug hunting, I will know that I can draw this weapon.
I’m not sure how many people are interested in such things, but I thought it might help someone out there, I wish one day someone would write an open source WinDbg plugin that injects kernel code through WinDbg to the debugged machine that sets this routine with its custom runtime conditional breakpoints :)

I really wanted to paint some stupid pictures that show what’s going on between the two machines and everything, but my capabilities at doing that are aweful, so it’s up to you to imagine that, sorry.

For more related information you can see:

Race Condition From Hell, aren’t they all?

Monday, April 19th, 2010

Actually I had a trouble to come up with a good title for this post, at least one that I was satisfied with. Therefore I will start with a background story, as always.
The problem started when I had to debug a huge software which was mostly in Kernel mode. And there was this critical section (critsec from now on) synchronization object that wasn’t held always correctly. And eventually after 20 mins of trying to replicate the bug, we managed to crash the system with a NULL dereference. This variable was a global that everybody who after acquiring the critsec was its owner. Then how come we got a crash ? Simple, someone was touching the global out of it critsec scope. That’s why it was also very hard to replicate, or took very long.

The pseudo code was something like this:
Acquire Crit-Sec
g_ptr = “some structure we use”
do safe task with g_ptr

g_ptr = NULL
Release Crit-Sec

So you see, before the critsec was released the global pointer was NULLed again. Obvisouly this is totally fine, because it’s still in the scope of the acquired crit, so we can access it safely.

Looking at the crash dumps, we saw a very weird thing, but nothing surprising for those race conditions bugs. Also if you ask me, I think I would prefer dead-lock bugs to race conditions, since in dead lock, everything gets stuck and then you can examine which locks are held, and see why some thread (out of the two) is trying to acquire the lock, when it surely can’t… Not saying it’s easier, though.
Anyway, back to the crash dump, we saw that the g_ptr variable was accessed in some internal function after the critsec was acquired. So far so good. Then after a few instructions, in an inner function that referenced the variable again, suddenly it crashed. Traversing back to the point where we know by the disassembly listing of the function, where the g_ptr was touched first, we knew it worked there. Cause otherwise, it would have crashed there and then, before going on, right? I have to mention that between first time reading the variable and the second one where it crashed, we didn’t see any function calls.
This really freaked me out, because the conclusion was one – somebody else is tempering with our g_ptr in a different thread without locking the crit. If there were any function calls, might be that some of them, caused our thread to be in a Waitable state, which means we could accept APCs or other events, and then it could lead to a whole new execution path, that was hidden from the crash dump, which somehow zeroed the g_ptr variable. Also at the time of the crash, it’s important to note that the owner of the critsec was the crashing thread, no leads then to other problematic threads…

Next thing was to see that everybody touches the g_ptr only when the critsec is acquired. We surely know for now that someone is doing something very badly and we need to track the biatch down. Also we know the value that is written to the g_ptr variable is zero, so it limits the number of occurrences of such instruction (expression), which lead to two spots. Looking at both spots, everything looked fine. Of course, it looked fine, otherwise I would have spotted the bug easily, besides, we got a crash, which means, nothing is fine. Also, it’s time to admit, that part of the code was Windows itself, which made the problem a few times harder, because I couldn’t do whatever I wanted with it.

I don’t know how you guys would approach such a problem in order to solve it. But I had three ideas. Sometimes just like printf/OutputDebugPrint is your best friend, print logs when the critsec is acquired and released, who is waiting for it and just every piece of information we can gather about it. Mind you that part of it was Windows kernel itself, so we had to patch those functions too, to see, who’s acquiring the critsec and when. Luckily in debug mode, patchguard is down :) Otherwise, it would be bloody around the kernel. So looking at the log, everything was fine, again, damn. You can stare at the god damned thing for hours and tracking the acquiring and releasing pairs of the critsec, and nothing is wrong. So it means, this is not going to be the savior.

The second idea, was to comment out some code portions with #if 0 surrouding the potential problematic code. And starting to eliminate the possibilities of which function is the cause of this bug. This is not such a great idea. Since a race condition can happen in a few places, finding one of them is not enough usually. Though it can teach you something about the original bug’s characteristics, then you can look at the rest of the code to fix that same thing. It’s really old school technique but sometimes it is of a help as bad as it sounds. So guess what we did? Patched the g_ptr = NULL of the kernel and then everything went smooth, no crashes and nothing. But the problem still was around, now we knew for sure it’s our bug and not MS, duh. And there were only a few places in our code which set this g_ptr. Looking at all of them, again, seemed fine. This is where I started going crazy, seriously.

While you were reading the above ideas, didn’t you come up with the most banal idea, to put a dumb breakpoint – on memory access, on g_ptr with a condition of “who writes zero”. Of course you did, that what you should have done in the first place. I hope you know that. Why we couldn’t do that?
Because the breakpoint was fired tens of thousands times in a single second. Rendering the whole system almost to freeze. Assuming it took us 20 mins to replicate the bug, when we heavily loaded the system. Doing that with such a breakpoint set, would take days or so, no kidding. Which is out of question.

This will lead me to the next post. Stay tuned.

Don’t Wait, Shoot. (KeSetEvent)

Tuesday, November 3rd, 2009

Apparently when you call down a driver with IoCallDriver you can either wait ’till the operation is finished or not. If you wait, you will need somebody to tell you “hey dude, you can stop waiting”. But that’s trivial, you set up a completion routine that will be called once the lower driver is finished with your IRP. The problem is if in some cases you don’t check the return code from that driver, and you assume you should always wait. So you Wait. Now what? Now the lower driver suddenly returns immediately, but did you know that? Probably not, cause you’re not blocked forever, otherwise you would have noticed it immediately. However, there’s seemingly no problem, because the lower driver will call your completion routine anyway indirectly and there you will signal “hey dude, you can stop waiting”, right? Therefore, it turns out you waited for nothing and just consumed some resources (locks).
That’s why usually you will see a simple test to see if the return code from the IoCallDriver is STATUS_PENDING, and only then you will wait ’till the operation is finished, in order to make it synchronized, that’s all the talk about. The thing is that you still need to do that same check in the completion routine you supplied. It seems that if you simply call SetEvent (and remebmer, now we know nobody is waiting on the event anymore, so why signaling it from the beginning), you still cause some performance penalty. And when you’re in a filter driver, for instance, you shouldn’t. And it’s a bad practice programming anyway.

I think it’s quite clear why WaitForSingleObject is “slow”, though in our case it will be immediately satisfied and yet… But I didn’t realize SetEvent is also problematic at a first thought. I thought it was a matter of flagging a boolean. In some sense, it’s true, there’s more to it. You see, since somebody might be waiting on the event, you will have to wake up the waiting thread and for that you need to lock the dispatcher-lock, and to yield execution, etc. Now suddenly it becomes a pain, huh?

Actually it’s quite interesting the way the KeSetEvent works. They knew they had to satisfy waiters, so they acquire the dispatcher-lock in the first place, and then they can also safely touch the event-state.

Moral of the story, don’t wait if you don’t have to, just shoot!

Cleaning Resources Automatically

Tuesday, September 15th, 2009

HelllllloW everybody!!!

Finally I am back, after 7 months, from a crazy trip in South America! Where I got robbed at gun-point, and denied access to USA (wanted to consult there), saw amazing views and creatures, among other stories, but it’s not the place to talk about them :) Maybe if you insist I will post once about it.

<geek warning>
Honestly, I freaked out in the trip without a computer (bits/assembly/compiler),  so all I could do was reading and following many blogs and stuff.
</geek warning>

I even learned that my post about the kernel DoS in XPSP3 about the desktop wallpaper weakness became a CVE. It seems MS has fixed it already, yey.

And now I need to warm up a bit, and I decided to dive into C++ with an interesting and very useful example. Cleaning resources automatically when they go out of scope. I think that not many people are aware to the simplicity of writing such a feature in C++ (or any other language that supports templates).

The whole issue is about how you destroy resources, I will use Win32 resources for the sake of example. I already talked once about this issue in C.

Suppose we have the following snippet:

HBITMAP h = LoadBitmap(...);
if (h == NULL) return 0;
HBITMAP h2 = Loadbitmap(...);
if (h2 ==  NULL) {
   return 0;
char* p = (char*)malloc(1000);
if (p == NULL) {
   return 0;

And so on, every failure handling in the if statements we have to clean more and more resources. And you know what, even today people forget to clean all resources, and it might even lead to security problems. But that was C times, and now we are all cool and know C++, so why not use it? One book which talks about these issues is Effective C++, recommended.

Also, another problem with the above code is while an exception is being thrown in the middle or afterward, you still have to clean those resources and copy/paste some lines, which makes it prone to errors.

Basically, all we need is a nice small class which will be called AutoResource that holds the object itself and will manage it. Personally, it reminds me auto_ptr class, but it’s way less permissive. You will be only able to initialize and use the object. Of course, it will be destroyed automatically when it goes out of scope.
How about this code now:

AutoResource<HBITMAP> h(LoadBitmap(...));
AutoResource<HBITMAP> h2(LoadBitmap(...));
char* p = new char[1000]; // If an exception is thrown, all objects initialized prior to this line are automatically cleaned.

Now you can move on and be totally free of ugly failure testing code and not worry about leaking objects, etc. Let’s get into details. What we really need is a special class that we can change the behavior of the CleanUp()  method according to its object’s type, that’s easily possible in C++ by using a method specialization technique. We will not let to copy or assign to the class. We want a total control over the object, therefore we will let the user to get() it too. And as a type of defense programming, we will force the coder to implement a specialized CleanUp() in a case he uses the class for new types and forgets to implement the new CleanUp code; By using compile time assertion (I used this trick from Boost). Also, there might be a case where the constructor input is NULL, and therefore the constructor will have to inform the caller by throwing an exception, download and check out the complete code later.

template <class T> class AutoResource {
   AutoResource(T t) : m_obj(t) { }

   void CleanUp()
      // WARNING:
      // If the assertion occurred you will have to specialize the CleanUp with the new type.

      m_obj = NULL;
   T get() const
      return m_obj;
   T m_obj;

//Here we specialize the CleanUp() for the HICON resource.
template<> void AutoResource<HICON>::CleanUp()

You can easily add new types and enjoy the class.  If you have any suggestions/fixes please leave a comment!

Download complete code: AutoResource.cpp.

Thanks to Yan Michalevsky for helping with the code!

P.S – While compiling this stuff, I found a crash in the Optimization Compiler of VS2008 :)  lol.

VML + ANI ZERT Patches

Tuesday, February 3rd, 2009

It is time to release an old presentation about the VML and ANI vulnerabilities that were patched by ZERT. It explains the vulnerabilities and how they were closed. It is somewhat very technical, Assembly is required if you wanna really enjoy it. I also gave a talk using this presentation in CCC 2007. It so happened that I wrote the patches, with the extensive help of the team, of course.

ZERT Patches.ppt

Proxy Functions – The Right Way

Thursday, August 21st, 2008

As much as I am an Assembly freak, I try to avoid it whenever possible. It’s just something like “pick the right language for your project” and don’t use overqualified stuff. Actually, in the beginning, when I started my patch on the IPhone, I compiled a simple stub for my proxy and then fixed it manually and only then used that code for the patch. Just to be sure about something here – a proxy function is a function that gets called instead of the original function, and then when the control belongs to the proxy function it might call the original function or not.

The way most people do this proxy function technique is using detour patching, which simply means, that we patch the first instruction (or a few, depends on the architecture) and change it to branch into our code. Now mind you that I’m messing with ARM here – iphone… However, the most important difference is that the return address of a function is stored on a register rather than in the stack, which if you’re not used to it – will get you confused easily and experiencing some crashes.

So suppose my target function begins with something like:

SUB SP, SP, #4
ADD R7, SP, #0xC

This prologue is very equivalent to push ebp; mov ebp, esp thing on x86, plus storing a few registers so we can change their values without harming the caller, of course. And the last thing, we also store LR (link-register), the register which stores the return address of the caller.

Anyhow, in my case, I override (detour) the first instruction to branch into my code, wherever it is. Therefore, in order my proxy function to continue execution on the original function, I have to somehow emulate that overriden instruction and only then continue from the next instruction as if the original patched function wasn’t touched. Although, there are rare times when you cannot override some specific instructions, but then it means you only have to work harder and change the way your detour works (instructions that use the program counter as an operand or branches, etc).

Since the return address of the caller is stored onto a register, we can’t override the first instruction with a branch-link (‘call’ equivalent on x86). Because then we would have lost the original caller’s return address. Give it a thought for a second, it’s confusing in the first time, I know. Just an interesting point to note that it so happens that if there’s a function which don’t call internally to other functions, it doesn’t have to store LR on the stack and later pop the PC (program-counter, IP register) off the stack, because nobody touched that register, unless the function needs around 14 registers for optimizations, instead of using local stack variables… This way you can tell which of the functions are leaves on the call graph, although it is not guaranteed.

Once we understand how the ARM architecture works we can move on. However, I have to mention that the 4 first parameters are passed on registers (R0 to R3) and the rest on the stack, so in the proxy we will have to treat the parameters accordingly. The good thing is that this ABI (Application-Binary-Interface) is something known to the compiler (LLVM with GCC front-end in my case), so you don’t have to worry about it, unless you manually write the proxy function yourself.

My proxy function can be written fully in C, although it’s possible to use C++ as well, but then you can’t use all features…

int foo(int a, int b)
 if (a == 1000) b /= 2;

That’s my sample foo proxy function, which doesn’t do anything useful nor interesting, but usually in proxies, we want to change the arguments, before moving on to the original function.

Once it is compiled, we can rip the code from the object or executable file, doesn’t really matter, and put it inside our patched file, but we are still missing the glue code. The glue code is a sequence of manually crafted instructions that will allow you to use your C code within the rest of the binary file. And to be honest, this is what I really wanted to avoid in first place. Of course, you say, “but you could write it once and then copy paste that glue code and voila”. So in a way you’re right, I can do it. But it’s bothersome and takes too much time, even that simple copy paste. And besides it is enough that you have one or more data objects stored following your function that you have to relocate all the references to them. For instance, you might have a string that you use in the proxy function. Now the way ARM works it is all get compiled as PIC (Position-Independent-Code) for the good and bad of it, probably the good of it, in our case. But then if you want to put your glue code inside the function and before the string itself, you will have to change the offset from the current PC register to the string… Sometimes it’s just easier to see some code:

stmfd sp!, {lr} 
mov r0, #0
add r0, pc, r0
bl _strlen
ldmfd sp! {pc}
db “this function returns my length :)”, 0

 When you read the current PC, you get that current instruction’s address + 8, because of the way the pipeline works in ARM. So that’s why the offset to the string is 0. Trying to put another instruction at the end of the function, for the sake of glue code, you will have to change the offset to 4. This really gets complicated if you have more than one resource to read. Even 32 bits values are stored after the end of the function, rather than in the operand of the instruction itself, as we know it on the x86.

So to complete our proxy code in C, it will have to be:

int foo(int a, int b)

 int (*orig_code)(int, int) = (int (*)(int, int))<addr of orig_foo + 4>; 
// +4 = We skip the first instruction which branches into this code!
 if (a == 1000) b /= 2;
// Emulate the real instruction we overrode, so stack is balanced before we continue with original function.
 asm(“sub sp, sp, #4”);
 return orig_foo(a, b);

This code looks more complete than before but contains a potential bug, can you spot it? Ok, I will give you a hint, if you were to use this code for x86, it would blow, though for ARM it would work well to some extent.

The bug lies in the number of arguments the original function receives. And since on ARM, only the 5th argument is passed through the stack, our “sub sp, sp, #4” will make some things go wrong. The stack of the original function should be as if it were running without we touched that function. This means that we want to push the arguments on the stack, ONLY then, do the stack fix by 4, and afterwards branch to the second instruction of the original function. Sounds good, but this is not possible in C. :( cause it means we have to run ‘user-defined’ code between the ‘pushing-arguments’ phase and the ‘calling-function’ phase. Which is actually not possible in any language I’m aware of. Correct me if I’m wrong though. So my next sentence is going to be “except Assembly”. Saved again ;)

Since I don’t want to dirty my hands with editing the binary of my new proxy function after I compile it, we have to fix that problem I just desribed above. This is the way to do it, ladies and gentlemen:

int foo(int a, int b)
 if (a == 1000) b /= 2;
 return orig_foo(a, b);

void __attribute__((naked)) orig_foo(int a, int b)
// Emulate the real instruction we overrode, so stack is balanced before we continue with original function.
 asm(“sub sp, sp, #4\nldr r12, [pc]\n bx r12\n.long <FOO ADDR + 4>”);

The code simply fixes the stack, reads the address of the original absolute foo address, again skipping the first instruction, and branches into that code. Though, it won’t change the return address in LR, therefore when the original function is over, it will return straight to the caller of orig_foo, which is our proxy function, that way we can still control the return values, if we wish to do so.

We had to use the naked attribute (__declspec(naked) in VC) so that the compiler won’t put a prologue that will unbalance our stack again. In any way the epilogue wouldn’t get to run…

This technique will work on x86 the same way, though for branching into an absolute address, one should use: push <addr>; ret.

In the bottom line, I don’t mind to pay the price for a few code lines in Assembly, that’s perfectly ok with me. The problem was that I had to edit the binary after compilation in order to fix it so it’s becoming ready to be put in the original binary as a patch. Besides, the Assembly code is a must, if you wish to compile it without further a do, and as long as the first instruction of the function hasn’t changed, your code is good to go.

This code works well and just as I really wanted, so I thought so share it with you guys, for a better “infrastructure” to make proxy function patches.

However, it could have been perfect if the compiler would have stored the functions in the same order you write them in the source code, thus the first instruction of the block would be the first instruction you have to run. Now you might need to add another branch in the beginning of the code so it skips the non-entry code. This is really compiler dependent. GCC seems to be the best in preserving the functions’ order. VC and LLVM are more problematic when optimizations are enabled. I believe I will cover this topic in the future.

One last thing, if you use -O3, or functions inline, the orig_foo naked function gets to be part of the foo function, and then the way we assume the original function returns to our foo proxy function, won’t happen. So just be sure to peek at the code so everything is fine ;)