Execution Engines

When you ask for a step of execution to happen in angr, something has to actually perform the step. angr uses a series of engines (subclasses of the SimEngine class) to emulate the effects that of a given section of code has on an input state. The execution core of angr simply tries all the available engines in sequence, taking the first one that is able to handle the step. The following is the default list of engines, in order:
The failure engine kicks in when the previous step took us to some uncontinuable state
The syscall engine kicks in when the previous step ended in a syscall
The hook engine kicks in when the current address is hooked
The unicorn engine kicks in when the UNICORN state option is enabled and there is no symbolic data in the state
The VEX engine kicks in as the final fallback.
SimSuccessors
The code that actually tries all the engines in turn is project.factory.successors(state, **kwargs), which passes its arguments onto each of the engines. This function is at the heart of state.step() and simulation_manager.step(). It returns a SimSuccessors object, which we discussed briefly before. The purpose of SimSuccessors is to perform a simple categorization of the successor states, stored in various list attributes. They are:
Attribute
Guard Condition
Instruction Pointer
Description
successors
True (can be symbolic, but constrained to True)
Can be symbolic (but 256 solutions or less; see unconstrained_successors).
A normal, satisfiable successor state to the state processed by the engine. The instruction pointer of this state may be symbolic (i.e., a computed jump based on user input), so the state might actually represent several potential continuations of execution going forward.
unsat_successors
False (can be symbolic, but constrained to False).
Can be symbolic.
Unsatisfiable successors. These are successors whose guard conditions can only be false (i.e., jumps that cannot be taken, or the default branch of jumps that must be taken).
flat_successors
True (can be symbolic, but constrained to True).
Concrete value.
As noted above, states in the successors list can have symbolic instruction pointers. This is rather confusing, as elsewhere in the code (i.e., in SimEngineVEX.process, when it's time to step that state forward), we make assumptions that a single program state only represents the execution of a single spot in the code. To alleviate this, when we encounter states in successors with symbolic instruction pointers, we compute all possible concrete solutions (up to an arbitrary threshold of 256) for them, and make a copy of the state for each such solution. We call this process "flattening". These flat_successors are states, each of which has a different, concrete instruction pointer. For example, if the instruction pointer of a state in successors was X+5, where X had constraints of X > 0x800000 and X <= 0x800010, we would flatten it into 16 different flat_successors states, one with an instruction pointer of 0x800006, one with 0x800007, and so on until 0x800015.
unconstrained_successors
True (can be symbolic, but constrained to True).
Symbolic (with more than 256 solutions).
During the flattening procedure described above, if it turns out that there are more than 256 possible solutions for the instruction pointer, we assume that the instruction pointer has been overwritten with unconstrained data (i.e., a stack overflow with user data). This assumption is not sound in general. Such states are placed in unconstrained_successors and not in successors.
all_successors
Anything
Can be symbolic.
This is successors + unsat_successors + unconstrained_successors.
Breakpoints
TODO: rewrite this to fix the narrative
Like any decent execution engine, angr supports breakpoints. This is pretty cool! A point is set as follows:
1
>>> import angr
2
>>> b = angr.Project('examples/fauxware/fauxware')
3
​
4
# get our state
5
>>> s = b.factory.entry_state()
6
​
7
# add a breakpoint. This breakpoint will drop into ipdb right before a memory write happens.
8
>>> s.inspect.b('mem_write')
9
​
10
# on the other hand, we can have a breakpoint trigger right *after* a memory write happens. 
11
# we can also have a callback function run instead of opening ipdb.
12
>>> def debug_func(state):
13
...     print("State %s is about to do a memory write!")
14
​
15
>>> s.inspect.b('mem_write', when=angr.BP_AFTER, action=debug_func)
16
​
17
# or, you can have it drop you in an embedded IPython!
18
>>> s.inspect.b('mem_write', when=angr.BP_AFTER, action=angr.BP_IPYTHON)
Copied!
There are many other places to break than a memory write. Here is the list. You can break at BP_BEFORE or BP_AFTER for each of these events.
Event type
Event meaning
mem_read
Memory is being read.
mem_write
Memory is being written.
address_concretization
A symbolic memory access is being resolved.
reg_read
A register is being read.
reg_write
A register is being written.
tmp_read
A temp is being read.
tmp_write
A temp is being written.
expr
An expression is being created (i.e., a result of an arithmetic operation or a constant in the IR).
statement
An IR statement is being translated.
instruction
A new (native) instruction is being translated.
irsb
A new basic block is being translated.
constraints
New constraints are being added to the state.
exit
A successor is being generated from execution.
fork
A symbolic execution state has forked into multiple states.
symbolic_variable
A new symbolic variable is being created.
call
A call instruction is hit.
return
A ret instruction is hit.
simprocedure
A simprocedure (or syscall) is executed.
dirty
A dirty IR callback is executed.
These events expose different attributes:
Event type
Attribute name
Attribute availability
Attribute meaning
mem_read
mem_read_address
BP_BEFORE or BP_AFTER
The address at which memory is being read.
mem_read
mem_read_expr
BP_AFTER
The expression at that address.
mem_read
mem_read_length
BP_BEFORE or BP_AFTER
The length of the memory read.
mem_read
mem_read_condition
BP_BEFORE or BP_AFTER
The condition of the memory read.
mem_write
mem_write_address
BP_BEFORE or BP_AFTER
The address at which memory is being written.
mem_write
mem_write_length
BP_BEFORE or BP_AFTER
The length of the memory write.
mem_write
mem_write_expr
BP_BEFORE or BP_AFTER
The expression that is being written.
mem_write
mem_write_condition
BP_BEFORE or BP_AFTER
The condition of the memory write.
reg_read
reg_read_offset
BP_BEFORE or BP_AFTER
The offset of the register being read.
reg_read
reg_read_length
BP_BEFORE or BP_AFTER
The length of the register read.
reg_read
reg_read_expr
BP_AFTER
The expression in the register.
reg_read
reg_read_condition
BP_BEFORE or BP_AFTER
The condition of the register read.
reg_write
reg_write_offset
BP_BEFORE or BP_AFTER
The offset of the register being written.
reg_write
reg_write_length
BP_BEFORE or BP_AFTER
The length of the register write.
reg_write
reg_write_expr
BP_BEFORE or BP_AFTER
The expression that is being written.
reg_write
reg_write_condition
BP_BEFORE or BP_AFTER
The condition of the register write.
tmp_read
tmp_read_num
BP_BEFORE or BP_AFTER
The number of the temp being read.
tmp_read
tmp_read_expr
BP_AFTER
The expression of the temp.
tmp_write
tmp_write_num
BP_BEFORE or BP_AFTER
The number of the temp written.
These attributes can be accessed as members of state.inspect during the appropriate breakpoint callback to access the appropriate values. You can even modify these value to modify further uses of the values!
1
>>> def track_reads(state):
2
...     print('Read', state.inspect.mem_read_expr, 'from', state.inspect.mem_read_address)
3
...
4
>>> s.inspect.b('mem_read', when=angr.BP_AFTER, action=track_reads)
Copied!
Additionally, each of these properties can be used as a keyword argument to inspect.b to make the breakpoint conditional:
1
# This will break before a memory write if 0x1000 is a possible value of its target expression
2
>>> s.inspect.b('mem_write', mem_write_address=0x1000)
3
​
4
# This will break before a memory write if 0x1000 is the *only* value of its target expression
5
>>> s.inspect.b('mem_write', mem_write_address=0x1000, mem_write_address_unique=True)
6
​
7
# This will break after instruction 0x8000, but only 0x1000 is a possible value of the last expression that was read from memory
8
>>> s.inspect.b('instruction', when=angr.BP_AFTER, instruction=0x8000, mem_read_expr=0x1000)
Copied!
Cool stuff! In fact, we can even specify a function as a condition:
1
# this is a complex condition that could do anything! In this case, it makes sure that RAX is 0x41414141 and
2
# that the basic block starting at 0x8004 was executed sometime in this path's history
3
>>> def cond(state):
4
...     return state.eval(state.regs.rax, cast_to=str) == 'AAAA' and 0x8004 in state.inspect.backtrace
5
​
6
>>> s.inspect.b('mem_write', condition=cond)
Copied!
That is some cool stuff!
Caution about mem_read breakpoint
The mem_read breakpoint gets triggered anytime there are memory reads by either the executing program or the binary analysis. If you are using breakpoint on mem_read and also using state.mem to load data from memory addresses, then know that the breakpoint will be fired as you are technically reading memory.
So if you want to load data from memory and not trigger any mem_read breakpoint you have had set up, then use state.memory.load with the keyword arguments disable_actions=True and inspect=False.
This is also true for state.find and you can use the same keyword arguments to prevent mem_read breakpoints from firing.
Simulation Managers
Analyses
Last modified 2yr ago
Export as PDF
Copy link
Contents
SimSuccessors
Breakpoints
Caution about mem_read breakpoint