Machine State - memory, registers, and so on

So far, we’ve only used angr’s simulated program states (SimState objects) in the barest possible way in order to demonstrate basic concepts about angr’s operation. Here, you’ll learn about the structure of a state object and how to interact with it in a variety of useful ways.

Review: Reading and writing memory and registers

If you’ve been reading this book in order (and you should be, at least for this first section), you already saw the basics of how to access memory and registers. state.regs provides read and write access to the registers through attributes with the names of each register, and state.mem provides typed read and write access to memory with index-access notation to specify the address followed by an attribute access to specify the type you would like to interpret the memory as.

Additionally, you should now know how to work with ASTs, so you can now understand that any bitvector-typed AST can be stored in registers or memory.

Here are some quick examples for copying and performing operations on data from the state:

>>> import angr, claripy
>>> proj = angr.Project('/bin/true')
>>> state = proj.factory.entry_state()

# copy rsp to rbp
>>> state.regs.rbp = state.regs.rsp

# store rdx to memory at 0x1000
>>> state.mem[0x1000].uint64_t = state.regs.rdx

# dereference rbp
>>> state.regs.rbp = state.mem[state.regs.rbp].uint64_t.resolved

# add rax, qword ptr [rsp + 8]
>>> state.regs.rax += state.mem[state.regs.rsp + 8].uint64_t.resolved

Basic Execution

Earlier, we showed how to use a Simulation Manager to do some basic execution. We’ll show off the full capabilities of the simulation manager in the next chapter, but for now we can use a much simpler interface to demonstrate how symbolic execution works: state.step(). This method will perform one step of symbolic execution and return an object called angr.engines.successors.SimSuccessors. Unlike normal emulation, symbolic execution can produce several successor states that can be classified in a number of ways. For now, what we care about is the .successors property of this object, which is a list containing all the “normal” successors of a given step.

Why a list, instead of just a single successor state? Well, angr’s process of symbolic execution is just the taking the operations of the individual instructions compiled into the program and performing them to mutate a SimState. When a line of code like if (x > 4) is reached, what happens if x is a symbolic bitvector? Somewhere in the depths of angr, the comparison x > 4 is going to get performed, and the result is going to be <Bool x_32_1 > 4>.

That’s fine, but the next question is, do we take the “true” branch or the “false” one? The answer is, we take both! We generate two entirely separate successor states - one simulating the case where the condition was true and simulating the case where the condition was false. In the first state, we add x > 4 as a constraint, and in the second state, we add !(x > 4) as a constraint. That way, whenever we perform a constraint solve using either of these successor states, the conditions on the state ensure that any solutions we get are valid inputs that will cause execution to follow the same path that the given state has followed.

To demonstrate this, let’s use a fake firmware image <../examples/fauxware/fauxware> as an example. If you look at the source code <../examples/fauxware/fauxware.c> for this binary, you’ll see that the authentication mechanism for the firmware is backdoored; any username can be authenticated as an administrator with the password “SOSNEAKY”. Furthermore, the first comparison against user input that happens is the comparison against the backdoor, so if we step until we get more than one successor state, one of those states will contain conditions constraining the user input to be the backdoor password. The following snippet implements this:

>>> proj = angr.Project('examples/fauxware/fauxware')
>>> state = proj.factory.entry_state(stdin=angr.SimFile)  # ignore that argument for now - we're disabling a more complicated default setup for the sake of education
>>> while True:
...     succ = state.step()
...     if len(succ.successors) == 2:
...         break
...     state = succ.successors[0]

>>> state1, state2 = succ.successors
>>> state1
<SimState @ 0x400629>
>>> state2
<SimState @ 0x400699

Don’t look at the constraints on these states directly - the branch we just went through involves the result of strcmp, which is a tricky function to emulate symbolically, and the resulting constraints are very complicated.

The program we emulated took data from standard input, which angr treats as an infinite stream of symbolic data by default. To perform a constraint solve and get a possible value that input could have taken in order to satisfy the constraints, we’ll need to get a reference to the actual contents of stdin. We’ll go over how our file and input subsystems work later on this very page, but for now, just use state.posix.stdin.load(0, state.posix.stdin.size) to retrieve a bitvector representing all the content read from stdin so far.

>>> input_data = state1.posix.stdin.load(0, state1.posix.stdin.size)

>>> state1.solver.eval(input_data, cast_to=bytes)
b'\x00\x00\x00\x00\x00\x00\x00\x00\x00SOSNEAKY\x00\x00\x00'

>>> state2.solver.eval(input_data, cast_to=bytes)
b'\x00\x00\x00\x00\x00\x00\x00\x00\x00S\x00\x80N\x00\x00 \x00\x00\x00\x00'

As you can see, in order to go down the state1 path, you must have given as a password the backdoor string “SOSNEAKY”. In order to go down the state2 path, you must have given something besides “SOSNEAKY”. z3 has helpfully provided one of the billions of strings fitting this criteria.

Fauxware was the first program angr’s symbolic execution ever successfully worked on, back in 2013. By finding its backdoor using angr you are participating in a grand tradition of having a bare-bones understanding of how to use symbolic execution to extract meaning from binaries!

State Presets

So far, whenever we’ve been working with a state, we’ve created it with project.factory.entry_state(). This is just one of several state constructors available on the project factory:

  • .blank_state() constructs a “blank slate” blank state, with most of its data left uninitialized. When accessing uninitialized data, an unconstrained symbolic value will be returned.

  • .entry_state() constructs a state ready to execute at the main binary’s entry point.

  • .full_init_state() constructs a state that is ready to execute through any initializers that need to be run before the main binary’s entry point, for example, shared library constructors or preinitializers. When it is finished with these it will jump to the entry point.

  • .call_state() constructs a state ready to execute a given function.

You can customize the state through several arguments to these constructors:

  • All of these constructors can take an addr argument to specify the exact address to start.

  • If you’re executing in an environment that can take command line arguments or an environment, you can pass a list of arguments through args and a dictionary of environment variables through env into entry_state and full_init_state. The values in these structures can be strings or bitvectors, and will be serialized into the state as the arguments and environment to the simulated execution. The default args is an empty list, so if the program you’re analyzing expects to find at least an argv[0], you should always provide that!

  • If you’d like to have argc be symbolic, you can pass a symbolic bitvector as argc to the entry_state and full_init_state constructors. Be careful, though: if you do this, you should also add a constraint to the resulting state that your value for argc cannot be larger than the number of args you passed into args.

  • To use the call state, you should call it with .call_state(addr, arg1, arg2, ...), where addr is the address of the function you want to call and argN is the Nth argument to that function, either as a Python integer, string, or array, or a bitvector. If you want to have memory allocated and actually pass in a pointer to an object, you should wrap it in an PointerWrapper, i.e. angr.PointerWrapper("point to me!"). The results of this API can be a little unpredictable, but we’re working on it.

  • To specify the calling convention used for a function with call_state, you can pass a SimCC instance as the cc argument.:raw-html-m2r:<br> We try to pick a sane default, but for special cases you will need to help angr out.

There are several more options that can be used in any of these constructors! See the docs on the project.factory object (an angr.factory.AngrObjectFactory) for more details.

Low level interface for memory

The state.mem interface is convenient for loading typed data from memory, but when you want to do raw loads and stores to and from ranges of memory, it’s very cumbersome. It turns out that state.mem is actually just a bunch of logic to correctly access the underlying memory storage, which is just a flat address space filled with bitvector data: state.memory. You can use state.memory directly with the .load(addr, size) and .store(addr, val) methods:

>>> s = proj.factory.blank_state()
>>> s.memory.store(0x4000, s.solver.BVV(0x0123456789abcdef0123456789abcdef, 128))
>>> s.memory.load(0x4004, 6) # load-size is in bytes
<BV48 0x89abcdef0123>

As you can see, the data is loaded and stored in a “big-endian” fashion, since the primary purpose of state.memory is to load an store swaths of data with no attached semantics. However, if you want to perform a byteswap on the loaded or stored data, you can pass a keyword argument endness - if you specify little-endian, byteswap will happen. The endness should be one of the members of the Endness enum in the archinfo package used to hold declarative data about CPU architectures for angr. Additionally, the endness of the program being analyzed can be found as arch.memory_endness - for instance state.arch.memory_endness.

>>> import archinfo
>>> s.memory.load(0x4000, 4, endness=archinfo.Endness.LE)
<BV32 0x67452301>

There is also a low-level interface for register access, state.registers, that uses the exact same API as state.memory, but explaining its behavior involves a dive into the abstractions that angr uses to seamlessly work with multiple architectures. The short version is that it is simply a register file, with the mapping between registers and offsets defined in archinfo.

State Options

There are a lot of little tweaks that can be made to the internals of angr that will optimize behavior in some situations and be a detriment in others. These tweaks are controlled through state options.

On each SimState object, there is a set (state.options) of all its enabled options. Each option (really just a string) controls the behavior of angr’s execution engine in some minute way. A listing of the full domain of options, along with the defaults for different state types, can be found in the appendix. You can access an individual option for adding to a state through angr.options. The individual options are named with CAPITAL_LETTERS, but there are also common groupings of objects that you might want to use bundled together, named with lowercase_letters.

When creating a SimState through any constructor, you may pass the keyword arguments add_options and remove_options, which should be sets of options that modify the initial options set from the default.

# Example: enable lazy solves, an option that causes state satisfiability to be checked as infrequently as possible.
# This change to the settings will be propagated to all successor states created from this state after this line.
>>> s.options.add(angr.options.LAZY_SOLVES)

# Create a new state with lazy solves enabled
>>> s = proj.factory.entry_state(add_options={angr.options.LAZY_SOLVES})

# Create a new state without simplification options enabled
>>> s = proj.factory.entry_state(remove_options=angr.options.simplification)

State Plugins

With the exception of the set of options just discussed, everything stored in a SimState is actually stored in a plugin attached to the state. Almost every property on the state we’ve discussed so far is a plugin - memory, registers, mem, regs, solver, etc. This design allows for code modularity as well as the ability to easily implement new kinds of data storage for other aspects of an emulated state, or the ability to provide alternate implementations of plugins.

For example, the normal memory plugin simulates a flat memory space, but analyses can choose to enable the “abstract memory” plugin, which uses alternate data types for addresses to simulate free-floating memory mappings independent of address, to provide state.memory. Conversely, plugins can reduce code complexity: state.memory and state.registers are actually two different instances of the same plugin, since the registers are emulated with an address space as well.

The globals plugin

state.globals is an extremely simple plugin: it implements the interface of a standard Python dict, allowing you to store arbitrary data on a state.

The history plugin

state.history is a very important plugin storing historical data about the path a state has taken during execution. It is actually a linked list of several history nodes, each one representing a single round of execution—you can traverse this list with state.history.parent.parent etc.

To make it more convenient to work with this structure, the history also provides several efficient iterators over the history of certain values. In general, these values are stored as history.recent_NAME and the iterator over them is just history.NAME. For example, for addr in state.history.bbl_addrs: print hex(addr) will print out a basic block address trace for the binary, while state.history.recent_bbl_addrs is the list of basic blocks executed in the most recent step, state.history.parent.recent_bbl_addrs is the list of basic blocks executed in the previous step, etc. If you ever need to quickly obtain a flat list of these values, you can access .hardcopy, e.g. state.history.bbl_addrs.hardcopy. Keep in mind though, index-based accessing is implemented on the iterators.

Here is a brief listing of some of the values stored in the history:

  • history.descriptions is a listing of string descriptions of each of the rounds of execution performed on the state.

  • history.bbl_addrs is a listing of the basic block addresses executed by the state. There may be more than one per round of execution, and not all addresses may correspond to binary code - some may be addresses at which SimProcedures are hooked.

  • history.jumpkinds is a listing of the disposition of each of the control flow transitions in the state’s history, as VEX enum strings.

  • history.jump_guards is a listing of the conditions guarding each of the branches that the state has encountered.

  • history.events is a semantic listing of “interesting events” which happened during execution, such as the presence of a symbolic jump condition, the program popping up a message box, or execution terminating with an exit code.

  • history.actions is usually empty, but if you add the angr.options.refs options to the state, it will be populated with a log of all the memory, register, and temporary value accesses performed by the program.

The callstack plugin

angr will track the call stack for the emulated program. On every call instruction, a frame will be added to the top of the tracked callstack, and whenever the stack pointer drops below the point where the topmost frame was called, a frame is popped. This allows angr to robustly store data local to the current emulated function.

Similar to the history, the callstack is also a linked list of nodes, but there are no provided iterators over the contents of the nodes - instead you can directly iterate over state.callstack to get the callstack frames for each of the active frames, in order from most recent to oldest. If you just want the topmost frame, this is state.callstack.

  • callstack.func_addr is the address of the function currently being executed

  • callstack.call_site_addr is the address of the basic block which called the current function

  • callstack.stack_ptr is the value of the stack pointer from the beginning of the current function

  • callstack.ret_addr is the location that the current function will return to if it returns

More about I/O: Files, file systems, and network sockets

Please refer to Working with File System, Sockets, and Pipes for a more complete and detailed documentation of how I/O is modeled in angr.

Copying and Merging

A state supports very fast copies, so that you can explore different possibilities:

>>> proj = angr.Project('/bin/true')
>>> s = proj.factory.blank_state()
>>> s1 = s.copy()
>>> s2 = s.copy()

>>> s1.mem[0x1000].uint32_t = 0x41414141
>>> s2.mem[0x1000].uint32_t = 0x42424242

States can also be merged together.

# merge will return a tuple. the first element is the merged state
# the second element is a symbolic variable describing a state flag
# the third element is a boolean describing whether any merging was done
>>> (s_merged, m, anything_merged) = s1.merge(s2)

# this is now an expression that can resolve to "AAAA" *or* "BBBB"
>>> aaaa_or_bbbb = s_merged.mem[0x1000].uint32_t

Todo

describe limitations of merging