源码解析angr的模拟执行
https://p6.toutiaoimg.com/large/pgc-image/955af34524c64a20a1ceaa2a5df0d889前言
angr是很有名的二进制符号实行工具，网上有很多关于angr的源码解析的文章。但是好像还没有关于angr模仿实行模块的解析。而模仿实行部分也是angr中相当重要的一个部分。因此，本文将解析angr模仿实行部分的源码，来帮助大家了解angr模仿实行的基本原理。
概述
当我们用angr去符号实行的时候，最基本的几个操作如下面代码所示：导入代码（第1行）、导入二进制（第2行）、确定初始状态（第3行）、构建simulation_manager对象（第4行）、模仿实行（第5行）。而到底angr是怎么符号实行的呢？因此就需要深入simulation_manager的源码（sim_manager.py）去一探究竟了。
import angrp = angr.Project("xxxx")entry_state = p.factory.entry_state()simgr = p.factory.simgr(entry_state)#simgr是simulation_manager的别名simgr.explore(find=xxxx)simulation_manager这个类位于angr/sim_manager.py文件里。
simulation_manager是angr中模仿实行管理器。主要的操作对象是步伐的状态对象（sim_state）。状态都被放在stash里，可以往前实行、过滤、合并或者移到别的stash里。stash里可以明白为是放状态的一个列表，stash有这么几种，分别表示状态的状态：
(1) active：生存接下来要实行的状态
(2) deadended：由于某些原因不能再继续实行下去，好比没有合法的指令、下个节点的状态不可解，或者有一个非法的指令指针。
(3) pruned：当使用lazy_sovles的策略时，只有在必要的时候才去检查状态是否可解。当发现一个不可求解的节点后，将其后面的节点都优化掉，放在pruned里。
(4) unconstrained：好比PC被用户数据或者其他类型的符号变量所控制，导致不知道实行哪个指令。
(5) unsat：不可求解的状态。好比，输入同时为AAAA和BBBB。
接下来看看源码，源码中提示我们看simulation_manager的三个重要方法：step、explore、use_technique。
use_technique
angr里有自带很多启发式的路径探索方法。这个函数就是让simulation_manager能够调用外部写好的启发式路径搜刮方法。官方给出的几个样例里，除了经典的深度优先搜刮、也有检测内存使用情况、CMU论文里的Veritest（合并循环的状态）等等策略。
代码首先先判定tech是否属于ExplorationTechnique这个类。然后setup方法开始初始化。然后把tech防到techniques列表中去，这也意味着可以使用多种策略。这里的hookset暂时没有看懂。
def use_technique(self, tech): """ Use an exploration technique with this SimulationManager. Techniques can be found in :mod:`angr.exploration_techniques`. :param tech: An ExplorationTechnique object that contains code to modify this SimulationManager's behavior. :type tech: ExplorationTechnique :return: The technique that was added, for convenience """ if not isinstance(tech, ExplorationTechnique): raise SimulationManagerError # XXX: as promised tech.project = self._project tech.setup(self) HookSet.install_hooks(self, **tech._get_hooks()) self._techniques.append(tech) return techexplore
先来看看看explore函数的参数，有stash，n，find，avoid等参数。explore函数的功能是从某个类型的stash，好比active，开始寻找满足find条件的，需要避免avoid条件的状态，直到找了n次，或者找到了num_find个状态。然后找到的状态都会塞到find_stash里，筛选的状态都会放在avoid_stash里。
其中find和avoid参数可以是一个地址，或者一堆地址的集合或者列表，甚至可以是一个函数，以状态为输入，输出True 或者False，来表示该状态是否是要寻找的状态。如果angr的CFG作为cfg的参数并且find是一个地址或者一个列表或者集合，那么到达不了目标状态的状态就会先把提前筛选掉。
def explore(self, stash='active', n=None, find=None, avoid=None, find_stash='found', avoid_stash='avoid', cfg=None, num_find=1, **kwargs): """ Tick stash "stash" forward (up to "n" times or until "num_find" states are found), looking for condition "find", avoiding condition "avoid". Stores found states into "find_stash' and avoided states into "avoid_stash". The "find" and "avoid" parameters may be any of: - An address to find - A set or list of addresses to find - A function that takes a state and returns whether or not it matches. If an angr CFG is passed in as the "cfg" parameter and "find" is either a number or a list or a set, then any states which cannot possibly reach a success state without going through a failure state will be preemptively avoided. """ num_find += len(self._stashes) if find_stash in self._stashes else 0 tech = self.use_technique(Explorer(find, avoid, find_stash, avoid_stash, cfg, num_find)) # Modify first Veritesting so that they can work together. deviation_filter_saved = None for t in self._techniques: if isinstance(t,Veritesting): deviation_filter_saved = t.options.get("deviation_filter",None) if deviation_filter_saved is not None: t.options["deviation_filter"] = lambda s: tech.find(s) or tech.avoid(s) or deviation_filter_saved(s) else: t.options["deviation_filter"] = lambda s: tech.find(s) or tech.avoid(s) break try: self.run(stash=stash, n=n, **kwargs) finally: self.remove_technique(tech) for t in self._techniques: if isinstance(t,Veritesting): if deviation_filter_saved is None: del t.options["deviation_filter"] else: t.options["deviation_filter"] = deviation_filter_saved break return self宏观来看explore函数分为三部分：初始化，兼容veritest策略，探索（run）。兼容veritest策略的代码占了很多，对于明白veritest策略与其他策略的关系很有帮助，但是对我们明白符号实行帮助较小，这里就不赘述了。
首先，更新num_find的参数为设定的num_find参数加上找到的状态。接着，用传入的参数find，avoid等生成Explorer对象，然后再用use_technique方法生成一个tech对象。这里为什么要生成Explore对象，然后再用use_technique方法？
Explorer对象继续了ExplorationTechnique类，以是他也是一种探索策略，并且是一种最基础的策略。
而符号实行过程中，可以使用多种策略，那么如何综合这些策略呢？angr是把他们都放在了simulationmanager里的.techniques列表里，而use_technique方法的作用正是把策略对象放进这个techniques列表里。
num_find += len(self._stashes) if find_stash in self._stashes else 0 tech = self.use_technique(Explorer(find, avoid, find_stash, avoid_stash, cfg, num_find))初始化后，接下来就是去探索状态部分。简朴的一个try，finally语句。不论run的结果如何，最后都把基础探索策略移出_techniques列表里。
try: self.run(stash=stash, n=n, **kwargs) finally: self.remove_technique(tech)run函数的代码如下，思路很简朴，根据当前的探索策略，一直探索，直到到达一个完整的状态。如果策略里没定义完整的策略，那就把stash里的状态都跑完。run里涉及到了后面会讲的step函数，这里可以先简朴明白为单步符号实行。
def run(self, stash='active', n=None, until=None, **kwargs): """ Run until the SimulationManager has reached a completed state, according to the current exploration techniques. If no exploration techniques that define a completion state are being used, run until there is nothing left to run. :param stash: Operate on this stash :param n: Step at most this many times :param until: If provided, should be a function that takes a SimulationManager and returns True or False. Stepping will terminate when it is True. :return: The simulation manager, for chaining. :rtype: SimulationManager """ for _ in (itertools.count() if n is None else range(0, n)): if not self.complete() and self._stashes: self.step(stash=stash, **kwargs) if not (until and until(self)): continue break return selfstep
最后就是这个比较复杂的step函数了，可以看作是符号实行的基本单元了。相比explore函数的参数多了selector_func，step_func，successor_func，filter_func，until。这些参数的意思代码解释写得比较清楚了，就简朴翻译一下。这些参数都是一个以状态为输入，返回各种东西（好比bool值，后继节点等）的一个函数，类似下面的代码。
def fun(state): if state.addr == xxxx: return True else: return False
[*]selector_func：如果为True，将会继续步进，反之会被生存。
[*]successor_func：返回的是后继节点，后面将会使用这些后继节点去符号实行。反之，则是使用project.factory.successors的后继节点。
[*]filter_func：返回的是stash的名字。filter_func的主要作用是给状态分类，分到各个stash里去。
[*]step_func：与前面参数不同，输入是为simulation_manger对象，并返回simulation_manager对象。这个函数会在simulation_manager对象每次step的时候被调用。
def step(self, stash='active', n=None, selector_func=None, step_func=None, successor_func=None, until=None, filter_func=None, **run_args): """ Step a stash of states forward and categorize the successors appropriately. The parameters to this function allow you to control everything about the stepping and categorization process. :param stash: The name of the stash to step (default: 'active') :param selector_func: If provided, should be a function that takes a state and returns a boolean. If True, the state will be stepped. Otherwise, it will be kept as-is. :param step_func: If provided, should be a function that takes a SimulationManager and returns a SimulationManager. Will be called with the SimulationManager at every step. Note that this function should not actually perform any stepping - it is meant to be a maintenance function called after each step. :param successor_func:If provided, should be a function that takes a state and return its successors. Otherwise, project.factory.successors will be used. :param filter_func: If provided, should be a function that takes a state and return the name of the stash, to which the state should be moved. :param until: (DEPRECATED) If provided, should be a function that takes a SimulationManager and returns True or False. Stepping will terminate when it is True. :param n: (DEPRECATED) The number of times to step (default: 1 if "until" is not provided) Additionally, you can pass in any of the following keyword args for project.factory.successors: :param jumpkind: The jumpkind of the previous exit :param addr: An address to execute at instead of the state's ip. :param stmt_whitelist:A list of stmt indexes to which to confine execution. :param last_stmt: A statement index at which to stop execution. :param thumb: Whether the block should be lifted in ARM's THUMB mode. :param backup_state: A state to read bytes from instead of using project memory. :param opt_level: The VEX optimization level to use. :param insn_bytes: A string of bytes to use for the block instead of the project. :param size: The maximum size of the block, in bytes. :param num_inst: The maximum number of instructions. :param traceflags: traceflags to be passed to VEX. Default: 0 :returns: The simulation manager, for chaining. :rtype: SimulationManager """ l.info("Stepping %s of %s", stash, self) # 88 bucket = defaultdict(list) for state in self._fetch_states(stash=stash): goto = self.filter(state, filter_func=filter_func) if isinstance(goto, tuple): goto, state = goto if goto not in (None, stash): bucket.append(state) continue if not self.selector(state, selector_func=selector_func): bucket.append(state) continue pre_errored = len(self._errored) successors = self.step_state(state, successor_func=successor_func, **run_args) # handle degenerate stepping cases here. desired behavior: # if a step produced only unsat states, always add them to the unsat stash since this usually indicates a bug # if a step produced sat states and save_unsat is False, drop the unsats # if a step produced no successors, period, add the original state to deadended # first check if anything happened besides unsat. that gates all this behavior if not any(v for k, v in successors.items() if k != 'unsat') and len(self._errored) == pre_errored: # then check if there were some unsats if successors.get('unsat', []): # only unsats. current setup is acceptable. pass else: # no unsats. we've deadended. bucket['deadended'].append(state) continue else: # there were sat states. it's okay to drop the unsat ones if the user said so. if not self._save_unsat: successors.pop('unsat', None) for to_stash, successor_states in successors.items(): bucket.extend(successor_states) self._clear_states(stash=stash) for to_stash, states in bucket.items(): self._store_states(to_stash or stash, states) if step_func is not None: return step_func(self) return self首先，从stash里取出一个状态，调用filter函数看下该状态最后要去哪个stash里，如果不是当前的stash，则把该状态塞到应该放的stash的地方，然后取下一个状态。调用selector函数，选择要生存的状态。
bucket = defaultdict(list)# 依次从stash里取出状态for state in self._fetch_states(stash=stash): goto = self.filter(state, filter_func=filter_func) # 返回的是个元组，（状态该去的stash，状态） if isinstance(goto, tuple): goto, state = goto #如果要去的stash不是当前的stash，也不是None， if goto not in (None, stash): # 那么把他放进该去的stash里，就不管他了。也就筛选掉了。 bucket.append(state) # continue # 如果selector函数返回False，则需要生存该状态到当前的stash if not self.selector(state, selector_func=selector_func): # 生存状态 bucket.append(state) continue如果没有触发selector或者filter，就去找后继节点。这里调用了step_state函数。
for state in self._fetch_states(stash=stash): ... successors = self.step_state(state, successor_func=successor_func, **run_args)step_state函数如下所示，这个函数主要是处理后继节点的状态。将不可解的状态，无约束的状态都放在相应的stash里。
def step_state(self, state, successor_func=None, **run_args): """ Don't use this function manually - it is meant to interface with exploration techniques. """ try: successors = self.successors(state, successor_func=successor_func, **run_args) stashes = {None: successors.flat_successors, 'unsat': successors.unsat_successors, 'unconstrained': successors.unconstrained_successors} except: ... return stashes由于step_state函数大概会发生很多错误，因此后续的代码是去做后继节点错误状态的处理。
for state in self._fetch_states(stash=stash): ... #如果有后继节点有任何一个unsat状态或者发生了新的错误 if not any(v for k, v in successors.items() if k != 'unsat') and len(self._errored) == pre_errored: #对于unsat状态，就先不管他 if successors.get('unsat', []): # only unsats. current setup is acceptable. pass else: #如果不是unsat，那分析遇到某些原因终止了，把该状态加到deadended的stash里去。 bucket['deadended'].append(state) continue else: # 如果没有设置生存unsat状态，就把后继节点的unsat状态丢出去。 if not self._save_unsat: successors.pop('unsat', None)接下来就把后继节点加到bucket的to_stash或者stash里去。自此，这个for循环就竣事了。
for state in self._fetch_states(stash=stash): ... for to_stash, successor_states in successors.items(): bucket.extend(successor_states)剩下就是一些收尾工作，清空当前stash里的状态，然后再把bucket的内容存到simulation_manager对象的stash里去。
self._clear_states(stash=stash)for to_stash, states in bucket.items(): self._store_states(to_stash or stash, states)如果有设置step_func，就去调用step_func。由此也能看到step_func是在step函数最后调用的。
if step_func is not None: return step_func(self)总结
angr模仿实行部分的主要代码就解析到这里了，希望大家能够对angr的模仿实行有更深的明白。在明白了angr这部分的内容之后，应该能够比较轻易地扩展angr的探索策略。
本文只涉及了sim_manager.py中的几个重要地方，如果要纯熟使用simulation_manager的各种功能的话，好比split，merge等等，还需要再看看源码。
参考资料： https://github.com/angr/angr/tree/master/angr
页:
[1]