Control-flow Graph Recovery (CFG)¶

angr includes analyses to recover the control-flow graph of a binary program. This also includes recovery of function boundaries, as well as reasoning about indirect jumps and other useful metadata.

General ideas¶

A basic analysis that one might carry out on a binary is a Control Flow Graph. A CFG is a graph with (conceptually) basic blocks as nodes and jumps/calls/rets/etc as edges.

In angr, there are two types of CFG that can be generated: a static CFG (CFGFast) and a dynamic CFG (CFGEmulated).

CFGFast uses static analysis to generate a CFG. It is significantly faster, but is theoretically bounded by the fact that some control-flow transitions can only be resolved at execution-time. This is the same sort of CFG analysis performed by other popular reverse-engineering tools, and its results are comparable with their output.

CFGEmulated uses symbolic execution to capture the CFG. While it is theoretically more accurate, it is dramatically slower. It is also typically less complete, due to issues with the accuracy of emulation (system calls, missing hardware features, and so on)

If you are unsure which CFG to use, or are having problems with CFGEmulated, try CFGFast first.

A CFG can be constructed by doing:

>>> import angr
# load your project
>>> p = angr.Project('/bin/true', load_options={'auto_load_libs': False})

# Generate a static CFG
>>> cfg = p.analyses.CFGFast()

# generate a dynamic CFG
>>> cfg = p.analyses.CFGEmulated(keep_state=True)

Using the CFG¶

The CFG, at its core, is a NetworkX di-graph. This means that all of the normal NetworkX APIs are available:

>>> print("This is the graph:", cfg.graph)
>>> print("It has %d nodes and %d edges" % (len(cfg.graph.nodes()), len(cfg.graph.edges())))

The nodes of the CFG graph are instances of class CFGNode. Due to context sensitivity, a given basic block can have multiple nodes in the graph (for multiple contexts).

# this grabs *any* node at a given location:
>>> entry_node = cfg.get_any_node(p.entry)

# on the other hand, this grabs all of the nodes
>>> print("There were %d contexts for the entry block" % len(cfg.get_all_nodes(p.entry)))

# we can also look up predecessors and successors
>>> print("Predecessors of the entry point:", entry_node.predecessors)
>>> print("Successors of the entry point:", entry_node.successors)
>>> print("Successors (and type of jump) of the entry point:", [ jumpkind + " to " + str(node.addr) for node,jumpkind in cfg.get_successors_and_jumpkind(entry_node) ])

Viewing the CFG¶

Control-flow graph rendering is a hard problem. angr does not provide any built-in mechanism for rendering the output of a CFG analysis, and attempting to use a traditional graph rendering library, like matplotlib, will result in an unusable image.

One solution for viewing angr CFGs is found in axt’s angr-utils repository.

Shared Libraries¶

The CFG analysis does not distinguish between code from different binary objects. This means that by default, it will try to analyze control flow through loaded shared libraries. This is almost never intended behavior, since this will extend the analysis time to several days, probably. To load a binary without shared libraries, add the following keyword argument to the Project constructor: load_options={'auto_load_libs': False}

Function Manager¶

The CFG result produces an object called the Function Manager, accessible through cfg.kb.functions. The most common use case for this object is to access it like a dictionary. It maps addresses to Function objects, which can tell you properties about a function.

>>> entry_func = cfg.kb.functions[p.entry]

Functions have several important properties!

entry_func.block_addrs is a set of addresses at which basic blocks belonging to the function begin.
entry_func.blocks is the set of basic blocks belonging to the function, that you can explore and disassemble using capstone.
entry_func.string_references() returns a list of all the constant strings that were referred to at any point in the function. They are formatted as (addr, string) tuples, where addr is the address in the binary’s data section the string lives, and string is a Python string that contains the value of the string.
entry_func.returning is a boolean value signifying whether or not the function can return. False indicates that all paths do not return.
entry_func.callable is an angr Callable object referring to this function. You can call it like a Python function with Python arguments and get back an actual result (may be symbolic) as if you ran the function with those arguments!
entry_func.transition_graph is a NetworkX DiGraph describing control flow within the function itself. It resembles the control-flow graphs IDA displays on a per-function level.
entry_func.name is the name of the function.
entry_func.has_unresolved_calls and entry.has_unresolved_jumps have to do with detecting imprecision within the CFG. Sometimes, the analysis cannot detect what the possible target of an indirect call or jump could be. If this occurs within a function, that function will have the appropriate has_unresolved_* value set to True.
entry_func.get_call_sites() returns a list of all the addresses of basic blocks which end in calls out to other functions.
entry_func.get_call_target(callsite_addr) will, given callsite_addr from the list of call site addresses, return where that callsite will call out to.
entry_func.get_call_return(callsite_addr) will, given callsite_addr from the list of call site addresses, return where that callsite should return to.

and many more !

CFGFast details¶

CFGFast performs a static control-flow and function recovery. Starting with the entry point (or any user-defined points) roughly the following procedure is performed:

The basic block is lifted to VEX IR, and all its exits (jumps, calls, returns, or continuation to the next block) are collected
For each exit, if this exit is a constant address, we add an edge to the CFG of the correct type, and add the destination block to the set of blocks to be analyzed.
In the event of a function call, the destination block is also considered the start of a new function. If the target function is known to return, the block after the call is also analyzed.
In the event of a return, the current function is marked as returning, and the appropriate edges in the callgraph and CFG are updated.
For all indirect jumps (block exits with a non-constant destination) Indirect Jump Resolution is performed.

Finding function starts¶

CFGFast supports multiple ways of deciding where a function starts and ends.

First the binary’s main entry point will be analyzed. For binaries with symbols (e.g., non-stripped ELF and PE binaries) all function symbols will be used as possible starting points. For binaries without symbols, such as stripped binaries, or binaries loaded using the blob loader backend, CFG will scan the binary for a set of function prologues defined for the binary’s architecture. Finally, by default, the binary’s entire code section will be scanned for executable contents, regardless of prologues or symbols.

In addition to these, as with CFGEmulated, function starts will also be considered when they are the target of a “call” instruction on the given architecture.

All of these options can be disabled

FakeRets and function returns¶

When a function call is observed, we first assume that the callee function eventually returns, and treat the block after it as part of the caller function. This inferred control-flow edge is known as a “FakeRet”. If, in analyzing the callee, we find this not to be true, we update the CFG, removing this “FakeRet”, and updating the callgraph and function blocks accordingly. As such, the CFG is recovered twice. In doing this, the set of blocks in each function, and whether the function returns, can be recovered and propagated directly.

Indirect Jump Resolution¶

Options¶

These are the most useful options when working with CFGFast:

Option	Description
force_complete_scan	(Default: True) Treat the entire binary as code for the purposes of function detection. If you have a blob (e.g., mixed code and data) you want to turn this off.
function_starts	A list of addresses, to use as entry points into the analysis.
normalize	(Default: False) Normalize the resulting functions (e.g., each basic block belongs to at most one function, back-edges point to the start of basic blocks)
resolve_indirect_jumps	(Default: True) Perform additional analysis to attempt to find targets for every indirect jump found during CFG creation.
more!	Examine the docstring on p.analyses.CFGFast for more up-to-date options

CFGEmulated details¶

Options¶

The most common options for CFGEmulated include:

Option	Description
context_sensitivity_level	This sets the context sensitivity level of the analysis. See the context sensitivity level section below for more information. This is 1 by default.
starts	A list of addresses, to use as entry points into the analysis.
avoid_runs	A list of addresses to ignore in the analysis.
call_depth	Limit the depth of the analysis to some number calls. This is useful for checking which functions a specific function can directly jump to (by setting `call_depth` to 1).
initial_state	An initial state can be provided to the CFG, which it will use throughout its analysis.
keep_state	To save memory, the state at each basic block is discarded by default. If `keep_state` is True, the state is saved in the CFGNode.
enable_symbolic_back_traversal	Whether to enable an intensive technique for resolving indirect jumps
enable_advanced_backward_slicing	Whether to enable another intensive technique for resolving direct jumps
more!	Examine the docstring on p.analyses.CFGEmulated for more up-to-date options

Context Sensitivity Level¶

angr constructs a CFG by executing every basic block and seeing where it goes. This introduces some challenges: a basic block can act differently in different contexts. For example, if a block ends in a function return, the target of that return will be different, depending on different callers of the function containing that basic block.

The context sensitivity level is, conceptually, the number of such callers to keep on the callstack. To explain this concept, let’s look at the following code:

void error(char *error)
{
    puts(error);
}

void alpha()
{
    puts("alpha");
    error("alpha!");
}

void beta()
{
    puts("beta");
    error("beta!");
}

void main()
{
    alpha();
    beta();
}

The above sample has four call chains: main>alpha>puts, main>alpha>error>puts and main>beta>puts, and main>beta>error>puts. While, in this case, angr can probably execute both call chains, this becomes unfeasible for larger binaries. Thus, angr executes the blocks with states limited by the context sensitivity level. That is, each function is re-analyzed for each unique context that it is called in.

For example, the puts() function above will be analyzed with the following contexts, given different context sensitivity levels:

Level	Meaning	Contexts
0	Callee-only	`puts`
1	One caller, plus callee	`alpha>puts` `beta>puts` `error>puts`
2	Two callers, plus callee	`alpha>error>puts` `main>alpha>puts` `beta>error>puts` `main>beta>puts`
3	Three callers, plus callee	`main>alpha>error>puts` `main>alpha>puts` `main>beta>error>puts` `main>beta>puts`

The upside of increasing the context sensitivity level is that more information can be gleaned from the CFG. For example, with context sensitivity of 1, the CFG will show that, when called from alpha, puts returns to alpha, when called from error, puts returns to error, and so forth. With context sensitivity of 0, the CFG simply shows that puts returns to alpha, beta, and error. This, specifically, is the context sensitivity level used in IDA. The downside of increasing the context sensitivity level is that it exponentially increases the analysis time.