Unicorn反混淆：恢复被OLLVM保护的程序(一)

一、目标

现在很多程序利用ollvm的控制流平坦化来增加逆向分析的难度。 控制流平坦化 (control flow flattening)的基本思想主要是通过一个主分发器来控制程序基本块的执行流程，例如下图是正常的执行流程

1:show1

经过控制流平坦化后的执行流程就如下图:

1:show2

这样可以模糊基本块之间的前后关系，增加程序分析的难度。

二、分析

这里我们以 check_passwd_arm_flat 为例来尝试恢复被ollvm混淆后的程序。先拖进ida，从流程图上可以看到典型的 控制流平台化 之后的结果：

1:cfg

恢复的流程是 分块→找出真实块→确定真实块之间的调用关系→Patch二进制程序

分块

filename = "./check_passwd_arm_flat"
start_addr = 0x83B0
end_addr = 0x87D4

project = angr.Project(filename, load_options={'auto_load_libs': False})
print(hex(project.entry))
cfg = project.analyses.CFGFast(regions=[(start_addr,end_addr)],normalize='True',force_complete_scan=False)

target_function = cfg.functions.get(start_addr)
#将angr的cfg转化为转化为类似ida的cfg
supergraph = am_graph.to_supergraph(target_function.transition_graph) 

找出真实块、序言、retn块和无用块

# get prologue_node and retn_node
prologue_node = None
for node in supergraph.nodes():if supergraph.in_degree(node) == 0:prologue_node = nodeif supergraph.out_degree(node) == 0 and len(node.out_branches) == 0:retn_node = node

print("序言块={},retn块={}".format(hex(prologue_node.addr),hex(retn_node.addr))) 

Tip:

在实战中，需要具体分析出 主分发器, 不一定序言块之后的就一定是主分发器，也不一定存在预处理器。

def get_relevant_nop_nodes(supergraph, main_dispatcher_node, prologue_node, retn_node):# relevant_nodes = list(supergraph.predecessors(pre_dispatcher_node))relevant_nodes = []nop_nodes = []for node in supergraph.nodes():# print(hex(node.addr))# 和主分发器有联系，并且size大于8的，认为是真实块if supergraph.has_edge(node, main_dispatcher_node) and node.size > 8:# XXX: use node.size is faster than to create a blockrelevant_nodes.append(node)# print(hex(node.addr))continueif node.addr in (prologue_node.addr, retn_node.addr, main_dispatcher_node.addr):continue# 非真实块的默认要干掉nop_nodes.append(node)return relevant_nodes, nop_nodes 

输出结果:

******************* 真实块 ************************
序言块: 0x83b0
主分发器: 0x87d0
retn块: 0x87c0
真实块: ['0x875c', '0x86f4', '0x8794', '0x8658', '0x8628', '0x8714', '0x86d4', '0x87ac', '0x8770', '0x8694', '0x864c', '0x8734', '0x8788', '0x86b0', '0x867c'] 

确定真实块之间的调用关系

网上确定真实块之间的调用关系大多都是用 angr 的符号执行来实现，不过angr是个强大的工程，强练容易走火入魔，需要从基础练起。 无名侠 提供了一个Unicorn模拟执行的思路来寻找两个真实块的关系。所以本文使用Unicorn来确定真实块之前的调用关系。

我们只想得到ollvm路径，而不是真实代码块的运行结果，因此要尽可能屏蔽非ollvm的内存操作。具体屏蔽方法稍后介绍。下面这段代码初始化Unicorn的虚拟CPU，并映射程序代码内存以及栈空间，最后调用hook_add设置UC_HOOK_CODE和UC_HOOK_MEM_UNMAPPED的事件回调。UC_HOOK_CODE回调会在每条指令执行前被调用，UC_HOOK_MEM_UNMAPPED会在内存异常的时候调用。

# 初始化
load_base = 0
emu = Uc(UC_ARCH_ARM, UC_MODE_ARM | UC_MODE_LITTLE_ENDIAN)

# 映射代码段 0x8000是 check_passwd_arm_flat 代码段的基址
emu.mem_map(0x8000, 4 * 1024 * 1024)
emu.mem_write(0x8000,binByte)

STACK_ADDR = 0x7F000000
STACK_SIZE = 1024 * 1024
start_addr = None

emu.mem_map(STACK_ADDR, STACK_SIZE)
emu.hook_add(UC_HOOK_CODE, hook_code)
emu.hook_add(UC_HOOK_MEM_UNMAPPED,hook_mem_access) 

真实块之间的关系有两种：1、顺序 2、分支，针对本文的例子，真实块里面的分支指令有 moveq movne movlt movgt，是时候祭出这个表了：

条件字段表

Tip:

分支指令需要具体情况具体分析，没有通用一劳永逸的解决。除非多费点功夫把所有的分支指令都处理一遍。我怀疑 angr 的符号执行就是把这个活给干了。

# 分支处理
if ins.mnemonic != 'mov' and ins.mnemonic.startswith('mov'):print(">>> branch 0x%x:\t%s\t%s" %(ins.address, ins.mnemonic, ins.op_str))if branch_control == 1 :vZ = (uc.reg_read(UC_ARM_REG_CPSR) & 0x40000000) >> 30vN = (uc.reg_read(UC_ARM_REG_CPSR) & 0x80000000) >> 31vV = (uc.reg_read(UC_ARM_REG_CPSR) & 0x10000000) >> 28if ins.mnemonic == 'moveq' or ins.mnemonic == 'movne' :if vZ == 0:uc.reg_write(UC_ARM_REG_CPSR,uc.reg_read(UC_ARM_REG_CPSR) | 0x40000000)print("Z 0->1 change cpsr = 0x%x" % uc.reg_read(UC_ARM_REG_CPSR))else:uc.reg_write(UC_ARM_REG_CPSR,uc.reg_read(UC_ARM_REG_CPSR) & 0xBFFFFFFF)print("Z 1->0 change cpsr = 0x%x" % uc.reg_read(UC_ARM_REG_CPSR))elif ins.mnemonic == 'movgt':if vZ == 0 and vN == vV:uc.reg_write(UC_ARM_REG_CPSR,uc.reg_read(UC_ARM_REG_CPSR) | 0x40000000)print("GT 0->1 change cpsr = 0x%x" % uc.reg_read(UC_ARM_REG_CPSR))else:uc.reg_write(UC_ARM_REG_CPSR,uc.reg_read(UC_ARM_REG_CPSR) & 0x2FFFFFFF)print("GT 1->0 change cpsr = 0x%x" % uc.reg_read(UC_ARM_REG_CPSR))elif ins.mnemonic == 'movlt':if vN != vV :uc.reg_write(UC_ARM_REG_CPSR,uc.reg_read(UC_ARM_REG_CPSR) & 0x6FFFFFFF)print("lt != -> = change cpsr = 0x%x" % uc.reg_read(UC_ARM_REG_CPSR))else:uc.reg_write(UC_ARM_REG_CPSR,uc.reg_read(UC_ARM_REG_CPSR) & 0xEFFFFFFF)print("lt = -> != change cpsr = 0x%x" % uc.reg_read(UC_ARM_REG_CPSR))else:print(">>> None " + ins.mnemonic) 

启动虚拟机的函数叫find_path，用于寻找真实块的下一个代码块。branch为分支控制。 如果branch = 1，则虚拟机在遇到movxx指令的时候会走movxx条件分支。

def find_path(uc,start_addr,branch = None):global block_startaddrglobal distAddrglobal isSucessglobal branch_controltry:block_startaddr = start_addrisSucess = FalsedistAddr = 0branch_control = branchuc.emu_start(start_addr,0x10000)print("emu end..")except UcError as e:pc = uc.reg_read(UC_ARM_REG_PC)if pc != 0:print("find_path UcError: %spc:%x" % (e,pc))return Noneelse:print("find_path ERROR: %spc:%x" % (e,pc))if isSucess:return distAddrreturn None 

控制流集成 使用队列的方式来路径搜索，起始搜索从函数入口开始。函数入口根据offset变量指定。 queue中的元素是一个二元组，第一项为执行地址，第二项为寄存器环境。每次搜索开始的时候从queue中获取一个将要搜索的真实块，设置寄存器，调用find_path搜索下一个真实块，将搜索到的真实块与新寄存器放入队列（保证上下文完整）使用这样做的好处就是可以搜索任意队列中的代码块，并且寄存器环境一定是和该代码块一致的。

queue = [(push_entry,None)]
flow = defaultdict(list)
patch_instrs = {}

while len(queue) != 0:env = queue.pop()pc = env[0]set_context(emu,env[1])if pc in flow:#print "???"continueflow[pc] = []print('------------------- run %#x---------------------' % pc)block = project.factory.block(pc)has_branches = False# 寻找有分支的代码块for ins in block.capstone.insns:if ins.insn.mnemonic != 'mov' and ins.insn.mnemonic.startswith('mov'):if pc not in patch_instrs:patch_instrs[pc] = inshas_branches = Trueif has_branches:# 有分支的代码块跑两次，一次正常，一次分支ctx = get_context(emu)p1 = find_path(emu,pc,0)if p1 != None:queue.append((p1,get_context(emu)))flow[pc].append(p1)set_context(emu,ctx)p2 = find_path(emu,pc,1)if p1 == p2:p2 = Noneif p2 != None:queue.append((p2,get_context(emu)))flow[pc].append(p2)else:p = find_path(emu,pc)if p != None:queue.append((p,get_context(emu)))flow[pc].append(p)

print("Emulation arm code done") 

路径探索，需要禁用掉一切函数调用、非栈空间内存访问，当虚拟机指令有内存操作需求时，判断目标内存地址范围是否在栈中，如果不在栈中则跳过该指令, 在本例中有一些内存访问代码段的固定值，这部分指令需要支持。 禁用的指令有bl、blx，只要识别bl前缀即可。

flag_pass = False

for b in ban_ins:if ins.mnemonic.find(b) != -1:flag_pass = Truebreak

if ins.op_str.find('[') != -1:if ins.op_str.find('[sp') == -1:flag_pass = Truefor op in ins.operands:# print(op.type)if op.type == ARM_OP_MEM:addr = 0if op.value.mem.base != 0:addr += uc.reg_read(reg_ctou(ins.reg_name(op.value.mem.base)))elif op.value.index != 0:addr += uc.reg_read(reg_ctou(ins.reg_name(op.value.mem.index)))elif op.value.disp != 0:addr += op.value.disp# 内存操作在栈区域if addr >= 0x7F000000 and addr < 0x7F000000 + 1024 * 1024 :flag_pass = False# 内存操作在代码区域if addr >= 0x8000 and addr < 0x9000: flag_pass = False

if flag_pass:print("will pass 0x%x:\t%s\t%s" %(ins.address, ins.mnemonic, ins.op_str))uc.reg_write(UC_ARM_REG_PC, address + size)return 

最后打印出找到的真实块之前的调用关系

************************flow******************************
0x83b0:['0x8628']
0x8628:['0x864c', '0x8658']
0x8658:['0x867c']
0x867c:['0x8628']
0x864c:['0x8694']
0x8694:['0x8794', '0x86b0']
0x86b0:['0x8788', '0x86d4']
0x86d4:['0x8788', '0x86f4']
0x86f4:['0x8714', '0x8788']
0x8788:['0x87ac']
0x87ac:['0x87c0']
0x87c0:[]
0x8714:['0x8788', '0x8734']
0x8734:['0x8770', '0x875c']
0x875c:['0x87c0']
0x8770:['0x8788']
0x8794:['0x87ac'] 

Patch二进制程序

首先把无用块都改成nop指令

for nop_node in nop_nodes:fill_nop(origin_data, nop_node.addr-base_addr,nop_node.size, project.arch) 

然后针对没有产生分支的真实块把最后一条指令改成jmp指令跳转到下一真实块

print("{} jmp {}".format(hex(parent),hex(childs[0])) )
# 把最后一条指令改成jmp指令跳转到下一真实块
parent_block = project.factory.block(parent)
last_instr = parent_block.capstone.insns[-1]
file_offset = last_instr.address - base_addr

patch_value = ins_b_jmp_hex_arm(last_instr.address, childs[0], 'b')
if project.arch.memory_endness == "Iend_BE": patch_value = patch_value[::-1]
patch_instruction(origin_data, file_offset, patch_value) 

针对产生分支的真实块把MOVX指令改成相应的条件跳转指令跳向符合条件的分支，例如moveq 改成beq ，再在这条之后添加b 指令跳向另一分支

instr = patch_instrs[parent]
# print("0x%x: %s \t%s\t%s" % (instr.insn.address,getByteStr(instr.insn.bytes), instr.insn.mnemonic, instr.insn.op_str))
file_offset = instr.insn.address - base_addr

parent_block = project.factory.block(parent)
fill_nop(origin_data, file_offset, parent_block.addr + parent_block.size - base_addr - file_offset, project.arch)

# patch the movx instruction to bx instruction
bx_cond = 'b' + instr.insn.mnemonic[len('mov'):]
patch_value = ins_b_jmp_hex_arm(instr.insn.address, childs[0], bx_cond)
if project.arch.memory_endness == 'Iend_BE':patch_value = patch_value[::-1]
patch_instruction(origin_data, file_offset, patch_value)

file_offset += 4
# patch the next instruction to b instrcution
patch_value = ins_b_jmp_hex_arm(instr.insn.address+4, childs[1], 'b')
if project.arch.memory_endness == 'Iend_BE':patch_value = patch_value[::-1]
patch_instruction(origin_data, file_offset, patch_value) 

最后用ida查看修复之后的cfg

1:cfgex

可以看到CFG跟原来的大致一样，然后反编译恢复出原始代码

bool __fastcall check_password(_BYTE *a1)
{int v2; // [sp+18h] [bp-10h]int i; // [sp+1Ch] [bp-Ch]v2 = 0;for ( i = 0; a1[i]; ++i )v2 += (unsigned __int8)a1[i];return i == 4&& v2 == 0x1A1&& (signed int)(unsigned __int8)a1[3] > 'c'&& (signed int)(unsigned __int8)a1[3] >= 'e'&& *a1 == 'b'&& ((unsigned __int8)a1[3] ^ 0xD) == (unsigned __int8)a1[1];
}