Awesome

BAP Tutorial

Introduction

In this tutorial, we will develop a non-trivial plugin that will verify that in a program a certain sequence of calls doesn't happen. We will develop the analysis in both OCaml and Python. You can choose either path or even both paths.

Takeaways

After finishing this tutorial you will learn how to use BAP basic capabilities and how to extend BAP using our plugin system. You will learn how to examine programs, by looking into their intermediate representations (IR) or disassembly. You will also learn how to run binaries in the emulated environment.

OCaml Path Takeaways

how to build and install plugins
how to use command-line arguments in a plugin
how to work with IR and different graph representations.

Python Path Takeaways

how to interact with bap from the Python universe
how to use IR visitors
how to create graphs

Preparation

To complete the OCaml path of the tutorial you will need an OCaml development environment, bap, and all the necessary libraries. The easiest option is to install BAP from opam. But if you don't want to spend time on that, don't worry. We have prepared a Vagrantfile, that will set up your virtual machine using Vagrant. If you have vagrant installed on your machine the following will set up and prepare the necessary development environment.

# if you don't want to install bap from opam yourself
wget http://tiny.cc/Vagrantfile
vagrant up
vagrant ssh

The last command will ssh you into the installed virtual machine, that has BAP and all the necessary libraries readily available. It also has Emacs preinstalled and configured to work with OCaml.

If you're going to take only the Python part of the tutorial and do not want to install any virtual machines, then you can use BAP's binary distribution. If you're using some Debian derivative, then the following will install BAP on your system.

wget https://github.com/BinaryAnalysisPlatform/bap/releases/download/v2.3.0/{bap,libbap,libbap-dev}_2.3.0.deb
sudo dpkg -i {bap,libbap,libbap-dev}_2.3.0.deb
sudo install pip
sudo pip install bap networkx

First Steps

For OCaml, create a new folder somewhere in your home folder, and enter into it, e.g.,

mkdir check_path
cd check_path

Then use Emacs to create an empty file

emacs check_path.ml

Note: it is mandatory to develop each BAP plugin in a separate folder.

After Emacs has started, use M-x merlin-use <RET> bap command to tell Merlin that you're going to use the bap library. See our wiki for more instructions about Merlin.

If you're developing in Python then do the same but create a file with the '.py extension, e.g.,

emacs check_path.py

First steps with bap

We will need some binary to play with. Especially for this tutorial, we created a binary that is small, but still non-trivial. You can get it using wget

wget tiny.cc/bap-echo -O exe

If you're running on an x86 machine you can run the binary by

chmod a+x exe
./exe

The binary will parse and print some verses from Shakespeare. It will also echo all its arguments.

You can also run the binary using bap. The binary will not be run on an actual CPU, but emulated using the Primus Framework.

To output the executable in the default assembler use -dasm option, e.g.,

bap ./exe -dasm

ProTip: if you prefer the intel syntax use --llvm-x86-syntax=intel.

To get the IR just use the -d option, e.g.,

bap ./exe -d

You can also use -dcfg to get the control-flow graphs, and -dcallgraph to dump the program call graph. To get the full list of supported formats for the project data structure use the bap list formats -f project command, e.g.,

$ bap list formats -f project
  Bap.Std.Project:
    symbols (1.0.0)        print symbol table
    ogre (1.0.0)           print the file specification in the OGRE format
    marshal (2.0.0)        OCaml standard marshaling format
    knowledge (1.0.0)      dumps the knowledge base
    graph (1.0.0)          print unlabeled CFG for each procedure
    cfg (1.0.0)            print rich CFG for each procedure
    callgraph (1.0.0)      print program callgraph in DOT format
    bir (1.0.0)            print program in IR
    bil.sexp (1.0.0)       print BIL instructions in Sexp format
    bil.adt (1.0.0)        print BIL instructions in ADT format
    bil (1.0.0)            print BIL instructions
    asm.sexp (1.0.0)       print assembly instructions as it was decoded
    asm.decoded (1.0.0)    print assembly instructions as it was decoded
    asm.adt (1.0.0)        print assembly instruction endcoded in ADT format
    asm (1.0.0)            print assembly instructions
    adt (1.0.0)            print program IR in ADT format

For more options see bap --help or man bap. To get information about a plugin named <PLUGIN use the bap --<PLUGIN>-help> option. e.g., for the optimization plugin, bap --optimization-help.

ProTip: if you have a GUI setup, then you can use xdot for interactive visualization of control flow graphs:

bap exe -dcfg --print-symbol=main | xdot -

Theoretical Background

The purpose of our analysis is to verify that for all execution paths in a program, a call to the function f is not followed by a call to the function g, where f and g are specified by a user. We can express this property concisely using Linear Temporal Logic:

f -> G not(g)

where f and g are propositions that are true if a call to a the corresponding function occurred, and G x is the LTL operator that requires the predicate x to be true on the entire subsequent path.

We will write a policy checker, that will verify that this property holds for all execution paths, or will provide a counter-example, proving the opposite. The checker will be static and sound (wrt to the soundness of the CFG).

Now, let's figure out how we will prove this property:

If there are no calls to f or to g or to both, then the verification property is trivially true, i.e., not f \/ not g This property has a strong "must" modality. Since the call indeed can't happen, so the property must hold for all paths.
If f calls g, either directly or by calling a function that calls g, then f occurs before g. The call to g still may not occur, as we ignore path constraints, so this gives us the may modality. We will express this as calls(f,g).
If there is a subroutine that has two calls that lead to f and g correspondingly, and these calls are reachable in the control flow graph of the subroutine, then g may be called after f with the may modality. (Again, we do not consider the feasibility of the path constraint). More formally, this property can be expressed as:
```
 sites(p,q) := calls(p,f) /\ calls(q,g) /\ reaches(p,q)
```
where reaches(x,y) is true if there exists a control flow graph with a path between x and y.

To summarize, to prove that f -> G not(g) holds for the static model of our program, we need to prove that the following doesn't hold

f /\ g /\ (calls(f,g) \/ ∃p∃q, sites(p,q))

OCaml implementation

Boilerplate Code

We will start with several open and include statements that will prepare our development context. Add the following to the top of the check_path.ml file.

open Core_kernel
open Poly
open Bap.Std
open Graphlib.Std
open Format
include Self()

After you saved the file with C-x C-s, the (No errors) message should appear in the mini-buffer. If it's not true, then make sure that you have set up Merlin and didn't make any errors in your code.

Defining the Command Line Interface

We will create a Cmdline module with the definition of our command line interface. It is not required to put the code in a separate module, but it looks cleaner and makes it easier for a reader to identify that this particular section of code deals with the command line interface

let main src dst proj = ()

module Cmdline = struct
  open Config
  let src = param string "src" ~doc:"Name of the source function"
  let dst = param string "dst" ~doc:"Name of the sink function"

  let () = when_ready (fun {get=(!!)} ->
      Project.register_pass' (main !!src !!dst))

  let () = manpage [
      `S "DESCRIPTION";
      `P
        "Checks whether there is an execution path that contains a call
    to the $(v,src) function before a call to the $(v,dst)
    parameter. Outputs a possible (shortest) path that can
    lead to the policy violation."
    ]
end

The code above will specify that our plugin has two command line parameters, one is named src and another dst, both having type string. It will also register the main function as a pass in the analysis framework. The following stanza probably needs some explanations:

  let () = when_ready (fun {get=(!!)} ->
      Project.register_pass' (main !!src !!dst))

The when_ready function takes one argument, that is a function that will be called as soon as system configuration is processed and command-line arguments are parsed. The function will take a record, that has only one field get, that is a function on itself, that will take a value of type 'a param and will extract a value of type 'a from it. The {get=(!!)} syntax, binds the get function to the (!!) unary operator that can be used nicely to extract values from parameters.

To summarize, a user specifies the command-line interface using the param function. Then the Config.when_ready function is used to bind the get function, which is a key that will unlock the parameters, and extract the values that are written there by the command-line parser. The interface is described thoroughly in the reference documentation.

Building, Installing, and Running

The following three commands will build, install, and run our analysis:

bapbuild check_path.plugin
bapbundle install check_path.plugin
bap ./exe --check-path-src=malloc --check-path-dst=free --pass=check-path

The bapbuild check_path.plugin command will compile the check_path.plugin from the check_path.ml. The bapbundle install check_path.plugin command will install the plugin into the place, where BAP will automatically load (but not run) it.

The last command will run bap on a binary and instruct it to process the binary with a pass named check-path. By default, a pass name has the same name as the plugin name, except that all underscores are substituted with dashes.

The --check-path-src=malloc will pass malloc as an argument to the src parameter. Note, that to pass an argument to a plugin parameter you need to prefix the parameter with the plugin name.

When developing, it is convenient to run all three commands at once, with one keystroke. Fortunately, Emacs allows us to do this. To compile for the first time, type

M-x compile <RET> bapbuild check_path.plugin && bapbundle install check_path.plugin && bap ./exe --check-path-src=malloc --check-path-dst=free --pass=check-path

This will run all three commands. You can also create a Makefile, or store this command in some build script.

Whenever you need to recompile, just hit C-c C-c and Emacs will recompile using the last compilation command.

See the Emacs manual for more information about compilation.

Trivial Proof

We will first check the trivial proof, i.e., when one of the functions, or both, are not present. For that we will write a function, that will translate a name of a function, as specified by a user, to the term identifier of this function.

let find_sub prog name =
  Term.enum sub_t prog |>
  Seq.find_map ~f:(fun s ->
      Option.some_if (Sub.name s = name) (Term.tid s))

Now we can extend the main function:

let main src dst proj =
  let prog = Project.program proj in
  match Option.both (find_sub prog src) (find_sub prog dst) with
  | None -> printf "satisfied (trivially)\n"
  | Some (src,dst) -> printf "not implemented\n"

The Skeleton Of Our Proof

We will now start writing a verify function, that will try to prove that the relation doesn't hold by providing a concrete counter-example. The verify function will take the program in the Intermediate Representation (IR) and two term identifiers, one for the source function and another for the destination function.

The function will try to find a path in the program call graph, that leads from the source function to the destination function. If there is no such path then for each control flow graph we will try to find a path between a callsite that reaches the source function and a callsite that reaches the destination function.

We will work with two different graph representations, one for the call graph, and another for the control flow graph, let's introduce the following abbreviations:

module CG = Graphs.Callgraph
module CFG = Graphs.Tid

Our verify function return type will be proof option, i.e., it will return Some proof if the proof was found, otherwise an absence of a proof is the proof that the relation will hold unless there is a not yet discovered indirect call.

Let's define the proof type as a variant type with two branches that correspond to the two cases in our proof strategy:

type proof =
  | Calls of CG.edge path
  | Sites of CFG.edge path

We will use the path data structure from the Graphlib library, as a particular representation of the proof term. Since we would like to output these paths, we will start with the implementation of two pretty-printing functions for both variants. Since they share the same polymorphic data structure, we will first write a polymorphic function that will print all types of paths, and then we will specialize this function to our two concrete types:

let pp_path get_src get_dst ppf path =
  let print_node n = fprintf ppf "%s" (Tid.name n) in
  print_node (get_src (Path.start path));
  Seq.iter (Path.edges path) ~f:(fun e ->
      pp_print_string ppf " -> ";
      print_node (get_dst e))

let pp_calls = pp_path CG.Edge.src CG.Edge.dst
let pp_sites = pp_path CFG.Edge.src CFG.Edge.dst

The pp_path function is parametrized with the get_src and get_dst functions that will extract, correspondingly, the source node and the destination node of an edge. Then it will print the source of the first edge, and iterate over all edges and print their destinations. Recall, that an edge consists of two endpoints and a label (we will ignore the label for now). Thus a path is a sequence of m edges, connecting m+1 nodes, e.g.,

(s, n0); (n0, n1); ...; (sm, d)
  /\       /\              /\
  ||       ||              ||
  e0       e1              em

Since we want our output to be concise, we will print only m+1 nodes, i.e., in our case s,n0,n1,...,d.

Now, let's update the main function:

let main src dst proj =
  let prog = Project.program proj in
  match Option.both (find_sub prog src) (find_sub prog dst) with
  | None -> printf "satisfied (trivially)@\n"
  | Some (src,dst) -> match verify src dst prog with
    | None -> printf "satisfied (no counter-example was found)@\n"
    | Some p ->
      printf "unsatisfied by ";
      match p with
      | Calls p -> printf "calls via %a@\n" pp_calls p
      | Sites p -> printf "callsites via %a@\n" pp_sites p

We will also provide a stub for the verify function, so that we can compile and run our program (and keep Merlin happy):

let verify src dst prog : proof option = None

Implementing `verify`

First of all, let's try to prove, that there is a path in the program call graph from the source function to the destination function. It would be even easier than the trivial proof:

let verify src dst prog : proof option =
  let cg = Program.to_graph prog in
  match Graphlib.shortest_path (module CG) cg src dst with
  | Some path -> Some (Calls path)
  | None -> None

Proving that for all control flow graphs there is no callsite to the destination function that is reachable from a callsite to the source, would be a little bit harder. We need to enumerate all subroutines, obtain their control flow graphs, and then check the connectivity of each pair of calls of the form: call to the destination, call to the source function.

We will address these problems using the bottom-up approach. We will start with the supporting code, that will enumerate all interesting callsites from a subroutine term. For that, we need to identify whether a callsite, that is a call term, may lead to an invocation of a target function. That would be the reaches predicate:

let reaches cg callee target =
  Graphlib.is_reachable (module CG) cg callee target

Now we are ready to write the callsites function, which will take the call graph cg, the target function, and the subroutine term sub, and return a sequence of calls that has a destination function, that reaches target in the call graph.

let callsites cg target sub =
  Term.enum blk_t sub |>
  Seq.concat_map ~f:(fun blk ->
      Term.enum jmp_t blk |> Seq.filter_map ~f:(fun j ->
          match Jmp.kind j with
          | Goto _ | Ret _ | Int (_,_) -> None
          | Call dst -> match Call.target dst with
            | Direct tid when reaches cg tid target ->
              Some (Term.tid blk)
            | _ -> None))

In our implementation, we are ignoring local jumps, as well as return statements and CPU interrupts. We also ignore indirect jumps and calls, expecting that it is the task of another analysis to make all possible control flow explicit in the graph.

Now we have all the necessary tools to finish our analysis, we will update the verify function:

let verify src dst prog : proof option =
  let cg = Program.to_graph prog in
  match Graphlib.shortest_path (module CG) cg src dst with
  | Some path -> Some (Calls path)
  | None ->
    Term.enum sub_t prog |> Seq.find_map ~f:(fun sub ->
        let g = Sub.to_graph sub in
        Seq.find_map (callsites cg src sub) ~f:(fun sc ->
            Seq.find_map (callsites cg dst sub) ~f:(fun dc ->
                if Tid.equal sc dc then None
                else Graphlib.shortest_path (module CFG) g sc dc))) |>
    Option.map ~f:(fun p -> Sites p)

That's it, for each subroutine in the program, we generate the control flow graph, then we iterate over all callsites that reach the source function, and over all callsites that reach the destination function, and check that if these two callsites are not equal, then there should not be a path between them, and if there is, then it is the proof of the policy violation. We added the equality test, because if the same callsite (and we denote a callsite by the basic block that hosts the call) reaches both the destination and the source, then it means that we have two distinct calls in the basic block that are mutually exclusive by the invariant of a control flow graph, or that we have one call where the destination function is invoked strictly before the source function. The only other case is ruled out by the first phase of our analysis, that proved that there is no path between the source function and the destination function in the call graph.

Python Implementation

The Python implementation will follow closely the OCaml implementation, so it is recommended to read the latter, to get some insights. The main difference is that in Python we need to build graphs manually. It is quite possible, that we will introduce the code for graph building in our next release of Python bindings, but so far (as of version 1.3) we need to build them manually. Anyway, building a graph is an excellent showcase of how to work with the complex IR structures. But we will start from the easy stuff first.

Running bap from Python

Unfortunately, currently, we do not provide an interface for writing real plugins in Python and so far, the only way to use BAP from Python is to run the bap executable and to parse its output. Fortunately, we provide some supporting code for this, so running bap is as easy as:

import bap
proj = bap.run('./echo')

We recommend running bap in some interactive REPL, such as IPython, that makes it easy to explore the data structures that are returned from the bap module functions.

Setting up the user interface

We will use the standard argparse module for argument parsing, and the Networkx library to work with graphs. So let's start with writing some boilerplate code:

import bap
import networkx as nx
import argparse

def verify(prog, src_name, dst_name):
    pass

description = """
Verifies the safety property that the DST function is never called
after the SRC function is called
"""

def main():
    parser = argparse.ArgumentParser(description=description)
    parser.add_argument('filename', help='target filename')
    parser.add_argument('-s', '--src', required=True, help='the source function')
    parser.add_argument('-d', '--dst', required=True, help='the sink function')
    args = parser.parse_args()
    proj = bap.run(args.filename)
    result = verify(proj.program, args.src, args.dst)
    if result:
        print('unsatisfied ({} are reachable via {})'.format(str(result[0]), str(result[1])))
    else:
        print('satisfied')

if __name__ == '__main__':
    main()

We specify one positional argument for the target binary, as well as two mandatory arguments, the names of the source and destination functions. Our verification function is expected to return None, if no counter-example was found, or a tuple if we found a proof. The first element of the tuple represents the kind of the proof, i.e., whether we found a relation in a call graph or a control flow graph, and the second element is the path itself.

Building Graphs

BAP represents graphs using [Algebraic Data Types][6]. It is easier to work with ADT using the Visitor pattern. The idea is simple. A complex ADT, like the IR in our case, is a tree with many different kinds of branches and leaves. The base visitor class will traverse the tree, and every time an element of type Xxx is entered, a method named enter_Xxx will be called (if it exists). For example, when a basic block is entered, the enter_Blk method will be invoked with the particular instance of the block.

We will implement GraphsBuilder that will descent through the program IR and build two kinds of graphs simultaneously - the program callgraph, and a list of control flow graphs (for each subroutine). Our builder state (other than graphs themselves) will include the current subroutine and the current basic block:

class GraphsBuilder(bap.adt.Visitor):
    def __init__(self):
        self.callgraph = nx.DiGraph()
        self.cfgs = []
        self.sub = None
        self.blk = None

Every time we enter into a new subroutine term we create an empty digraph and append it to the list of control flow graphs. We also add a node to the call graph and label this node with the reference to the newly created control flow graph - this will make it easier for us to get the CFG of a function, the functionality that we will need later. We also update the currently processed subroutine (we do not really store the subroutine number, as we only interested in the term identifier of the subroutine).

    def enter_Sub(self, sub):
        cfg = nx.DiGraph()
        self.callgraph.add_node(sub.id.number, name=sub.name, sub=sub, cfg=cfg)
        self.sub = sub.id.number
        self.cfgs.append(cfg)

When we enter a basic block, we push its term identifier as a node to the currently built control flow graph and update the current block number:

    def enter_Blk(self, blk):
        self.cfgs[-1].add_node(blk.id.number, blk=blk)
        self.blk = blk.id.number

When we see a call we need to add an edge to the call graph and an edge to the current control flow graph to represent the fallthrough edge (if the call is not marked as a non-return call).

    def enter_Call(self, jmp):
        callee = direct(jmp.target[0])
        if callee:

        fall = direct(jmp.target[1]) if len(jmp.target) == 2 else None
        if fall:
            self.cfgs[-1].add_edge(self.blk, fall.number, jmp=jmp)

Note, that we are ignoring indirect calls, and using the following helper function to extract the destination of a direct jump:

def direct(jmp):
    return jmp.arg if jmp is not None and jmp.constr == 'Direct' else None

Whenever we enter the Goto statement, that represents a transfer of the control flow between two basic blocks in a function, we just add a new edge, if this transfer is explicit (i.e., direct)

    def enter_Goto(self, jmp):
        dst = direct(jmp.target)
        if dst:
            self.cfgs[-1].add_edge(self.blk, dst.number, jmp=jmp)

Finally, when a CPU exception occurs, we just add a fallthrough edge to the currently built control flow graph

    def enter_Exn(self, exn):
        fall = exn.target[1]
        if fall:
            self.cfgs[-1].add_edge(self.blk, fall.number, exn=exn)

Implementing the Callsites Collector

As in the OCaml version, we will collect all callsites that reach targets in which we are interested. The implementation is rather straight-forward, whenever we see a call, we look at the call destination, and if it is known and reaches either the source function, or the destination function, we add it to the corresponding list.

class CallsitesCollector(bap.adt.Visitor):
    def __init__(self, callgraph, src, dst):
        self.callgraph = callgraph
        self.caller = None
        self.src = src
        self.dst = dst
        self.srcs = []
        self.dsts = []

    def clear(self):
        self.srcs = []
        self.dsts = []

    def enter_Blk(self, blk):
        self.caller = blk.id.number

    def enter_Call(self,jmp):
        callee = direct(jmp.target[0])
        if callee:
            if nx.has_path(self.callgraph, callee.number, self.src):
                self.srcs.append(self.caller)
            if nx.has_path(self.callgraph, callee.number, self.dst):
                self.dsts.append(self.caller)

We also added the clear method that will drop both lists, as we are planning to reuse the same collector for all subroutines.

Implementing The Verification Procedure

Now we have a full arsenal of tools to implement the verification procedure. It is rather straightforward, we will start by finding target functions, using the find method that is provided by all instances of the Seq class. If one of the targets is not present, then the verification condition is satisfied trivially and we bail out of the function. Otherwise, we instantiate GraphsBuilder and run it to build the graphs. We then look at the call graph to see if there is a path between the source and destination functions. If there is no such path, then we are looking at each CFG for a pair of callsites to the target functions that have a path. And if there is one, then it is returned.

def verify(prog, src_name, dst_name):
    src = prog.subs.find(src_name)
    dst = prog.subs.find(dst_name)
    if src is None or dst is None:
        return None

    graphs = GraphsBuilder()
    graphs.run(prog)
    cg = graphs.callgraph

    if nx.has_path(cg, src.id.number, dst.id.number):
        return ('calls', nx.shortest_path(cg, src.id.number, dst.id.number))

    calls = CallsitesCollector(graphs.callgraph, src.id.number, dst.id.number)

    for sub in prog.subs:
        calls.run(sub)
        cfg = graphs.callgraph.nodes[sub.id.number]['cfg']
        for src in calls.srcs:
            for dst in calls.dsts:
                if src != dst and nx.has_path(cfg, src, dst):
                    return ('sites', nx.shortest_path(cfg, src, dst))
        calls.clear()

    return None