eval.py

#!/usr/bin/env python

# Copyright (c) 2012, Vasilis Pappas <vpappas@cs.columbia.edu>
# This file is part of Orp http://nsl.cs.columbia.edu/projects/orp

import time
import pickle
import optparse
import itertools
import sys
import inp
import func
import gadget
import swap
import preserv
import equiv
import reorder


def _dump(filename, data):
    print "+ dumping", filename
    with open(filename, "wb") as f:
        pickle.dump(data, f)


def get_changed_bytes(functions):
    """Applies all the randomizations on each function and returns the
    superset of the changed bytes for each technique (coverage)."""

    swap_b, preserv_b, equiv_b, reorder_b = set(), set(), set(), set()

    for f in filter(lambda x: x.level != -1, functions.itervalues()):
        # swap
        swap.liveness_analysis(f.code)
        live_regs = swap.get_reg_live_subsets(f.instrs, f.code, f.igraph)
        swap.split_reg_live_subsets(live_regs, f.code)
        swaps = swap.get_reg_swaps(live_regs)
        swap_b |= swap.do_single_swaps(swaps, gen_patched=False)

        # preserv
        preservs, avail_regs = preserv.get_reg_preservations(f)
        preserv_b |= preserv.do_reg_preservs(f.instrs, f.blocks, preservs,
                                             avail_regs, gen_patched=False)

        # equiv
        equiv_b |= equiv.do_equiv_instrs(f.instrs, gen_patched=False)

        # reorder
        reorder_b |= reorder.do_reordering(f.blocks, gen_patched=False)

    return swap_b, preserv_b, equiv_b, reorder_b


def eval_coverage(input_file, dump_eval_data=False):
    """Evaluates the effectiveness of the randomization techniques in
    breaking/eliminating gadgets (coverage)."""

    def _check_gadget(g, changed, ins_hit):
        """Checks if the given gadget was broken by these changed bytes. In addition,
        it keeps track of which of the gadget instructions where changed. Returns
        True, if the gadget was broken, False otherwise."""

        broke = False

        for i_start, i_end in zip(g.addrs, g.addrs[1:] + [g.end]):
            if any(b in changed for b in xrange(i_start, i_end)):
                ins_hit.add(i_start)
                broke = True

        return broke

    start_time = time.time()

    # get the changed byte sets
    functions = inp.get_functions(input_file)
    levels = func.classify_functions(functions)
    func.analyze_functions(functions, levels)
    swap_b, preserv_b, equiv_b, reorder_b = get_changed_bytes(functions)

    # get the gadgets
    gadgets = gadget.get_simple_gadgets(input_file)

    # initialize output vars
    unchanged = set()
    eliminated = set()
    swapped = set()
    preserved = set()
    equived = set()
    reordered = set()
    ins_hit = set()
    ins_hit_data = {2: [0] * 3, 3: [0] * 4, 4: [0] * 5, 5: [0] * 6}

    # check each gadget one by one to see which were broken/eliminated etc
    for g in gadgets:

        ins_hit.clear()
        broke_this_g = False

        # check if the gadget is broken
        if _check_gadget(g, swap_b, ins_hit):
            swapped.add(g)
            broke_this_g = True
        if _check_gadget(g, preserv_b, ins_hit):
            preserved.add(g)
            broke_this_g = True
        if g.overlap and _check_gadget(g, equiv_b, ins_hit):
            equived.add(g)
            broke_this_g = True
        if g.overlap and _check_gadget(g, reorder_b, ins_hit):
            reordered.add(g)
            broke_this_g = True
        elif not g.overlap and g.func_ea and g.func_ea in functions:
            # XXX: it's extremely difficult to find cases where gadgets will be broke
            # this way because the gadget extraction algorithm usualy stops on jumps,
            # which are the baseic block boarders. So, most non-overlapping gadgets
            # are whole BBs.
            f = functions[g.func_ea]
            for ins in (f.code[a] for a in xrange(g.start, g.end) if a in f.code):
                if ins.addr != ins.raddr and ins.raddr < g.start or ins.raddr > g.end:
                    reordered.add(g)
                    broke_this_g = True
                    break

        # one of the few :)
        if not broke_this_g:
            unchanged.add(g)

        # if any of the gadget's instruction was changed
        if len(ins_hit) > 0:
            # update the gadget isntruction hit histogram
            for i in ins_hit:
                ins_hit_data[len(g.addrs)][g.addrs.index(i)] += 1
            ins_hit_data[len(g.addrs)][-1] += 1
            # check if the gadget was elliminated!
            if g.addrs[-1] in ins_hit and ((g.addrs[-1] in swap_b or
                                                    g.addrs[-1] in preserv_b) or (g.overlap and (
                            g.addrs[-1] in equiv_b or g.addrs[-1] in reorder_b))):
                eliminated.add(g)

    # output the evaluation data..
    print "--------------------------- done -------------------------------"

    c_len = float(len(swap_b | preserv_b | equiv_b | reorder_b))
    print "total changed bytes: %d" % c_len
    print "  swap: %d (%.1f%%)" % (len(swap_b), 100 * len(swap_b) / c_len)
    print "  preserv: %d (%.1f%%)" % (len(preserv_b), 100 * len(preserv_b) / c_len)
    print "  equiv: %d (%.1f%%)" % (len(equiv_b), 100 * len(equiv_b) / c_len)
    print "  reorder: %d (%.1f%%)" % (len(reorder_b), 100 * len(reorder_b) / c_len)

    print "total gadgets: %d" % len(gadgets)

    if len(gadgets) > 0:

        print "  instructions hit within gadgets (last column is the total):"
        for k, v in ins_hit_data.iteritems():
            print "    glen%d:" % k, ' '.join([str(c) for c in v])

        reds = set((g for g in gadgets if g.red))

        r_len = len(reds)
        g_len = float(len(gadgets))
        print "  red: %d (%.1f%%)" % (r_len, 100 * r_len / g_len)

        gr_len = float(len(gadgets - reds))
        u_len = len(unchanged)
        ur_len = float(len(unchanged - reds))
        print "  unchanged: %d (%.1f%%) - no red: %d (%.1f%%)" % (
            u_len, 100 * u_len / g_len, ur_len, 100 * ur_len / gr_len)

        e_len = len(eliminated)
        print "  eliminated: %d (%.1f%%) - no red: %.1f%%" % (
            e_len, 100 * e_len / g_len, 100 * e_len / gr_len)

        b_len = len(gadgets - unchanged)
        print "  broken: %d (%.1f%%) - no red: %.1f%%" % (
            b_len, 100 * b_len / g_len, 100 * b_len / gr_len)

        s_len = len(swapped)
        print "    swap: %d (%.1f%%) - no red: %.1f%%" % (
            s_len, 100 * s_len / g_len, 100 * s_len / gr_len)

        p_len = len(preserved)
        print "    preserv: %d (%.1f%%) - no red: %.1f%%" % (
            p_len, 100 * p_len / g_len, 100 * p_len / gr_len)

        q_len = len(equived)
        print "    equiv: %d (%.1f%%) - no red: %.1f%%" % (
            q_len, 100 * q_len / g_len, 100 * q_len / gr_len)

        o_len = len(reordered)
        print "    reorder: %d (%.1f%%) - no red: %.1f%%" % (
            o_len, 100 * o_len / g_len, 100 * o_len / gr_len)

        avg_g_blen = sum((g.end - g.start for g in gadgets)) / g_len
        avg_ur_blen = sum((g.end - g.start for g in unchanged - reds)) / ur_len
        print "  avg size (bytes): %.2f - unchanged (no red): %.2f" % (
            avg_g_blen, avg_ur_blen)

        avg_g_ilen = sum((g.ins_num for g in gadgets)) / g_len
        avg_ur_ilen = sum((g.ins_num for g in unchanged - reds)) / ur_len
        print "  avg size (insns): %.2f - unchanged (no red): %.2f" % (
            avg_g_ilen, avg_ur_ilen)

    print "evaluation took", int(time.time() - start_time), "seconds"

    if dump_eval_data:
        _dump(input_file + ".changed_bytes_swap", swap_b)
        _dump(input_file + ".changed_bytes_preserv", preserv_b)
        _dump(input_file + ".changed_bytes_equiv", equiv_b)
        _dump(input_file + ".changed_bytes_reorder", reorder_b)

        _dump(input_file + ".gadgets_unchanged", unchanged)
        _dump(input_file + ".gadgets_swapped", swapped)
        _dump(input_file + ".gadgets_preserved", preserved)
        _dump(input_file + ".gadgets_equived", equived)
        _dump(input_file + ".gadgets_reordered", reordered)

    # sanity checks
    if swapped | preserved | equived | reordered != gadgets - unchanged:
        print "BUG: gadget sets check failed!!"


def get_entropy(f, g, live_regs, swaps, preservs, avail_regs):
    """Given a gadget, it returns the number of different ways that it can be
    broken. A second return value indicates whether this gadget can be
    eliminated (entropy)."""

    def _add_if_hit(g, diff, diffs):
        gdiff = tuple(d for d in diff if g.start <= d[0] < g.end)  # d[0] is ea
        old_diffs_len = len(diffs)
        if len(gdiff) > 0:
            diffs.add(gdiff)
        return old_diffs_len != len(diffs)

    entropy = 1
    eliminated = False

    # 1. swaps
    swap_diffs = set()
    gswaps = []  # filter swaps that actually hit the gadget
    gbytes = range(g.start, g.end)

    for s in swaps:
        # sbytes = [xrange(i.addr, i.addr+i.inst_len) for i in swap.get_instrs()]
        # if set(itertools.chain(*sbytes)) & set(xrange(g.start, g.end)):
        #  gswaps.append(swap)
        for ins in s.get_instrs():
            if any(ins.addr <= b < ins.addr + len(ins.bytes) for b in gbytes):
                gswaps.append(s)
                break

    # XXX use only at most 10 swaps here .. otherwise it takes forever
    for i, swap_comb in enumerate(swap.gen_swap_combinations(gswaps[:10])):
        success, changed = swap.apply_swap_comb(swap_comb)
        if not success:
            continue
        diff = inp.get_diff(changed)
        _add_if_hit(g, diff, swap_diffs)
        for ins in changed:
            ins.reset_changed()
        if i > 1000:
            print "killing after 1000 swap combinations"
            break

    entropy *= len(swap_diffs) + 1
    eliminated = any(g.addrs[-1] == d[1] for d in itertools.chain(*swap_diffs))

    print "\tswaps:", len(swap_diffs)

    # 2. preservs
    preserv_diffs = set()

    # 2.1 global substitution
    implicit = reduce(lambda x, y: x | y, [i.implicit for i in f.instrs if not i.f_exit], set())
    preserv_regs = [reg for reg, pushes, pops in preservs if reg not in implicit]
    preserv_combs = []

    if len(preserv_regs) >= 2:
        preserv_combs.extend(itertools.combinations(preserv_regs, 2))

    preserv_combs.extend(itertools.product(preserv_regs, avail_regs))

    for r1, r2 in preserv_combs:
        for ins in f.instrs:
            if not ins.swap_registers(r1, r2):
                break
        else:
            diff = inp.get_diff(f.instrs)
            _add_if_hit(g, diff, preserv_diffs)
        for ins in f.instrs:
            ins.reset_changed()

    # 2.2 reorder pushes/pops
    # XXX: a preserved register is used only after it's pushed!
    #      and only before it's poped!
    # augment each ins in the preservs with blocks and find first/last
    preserv_blocks = set()

    for reg, pushes, pops in preservs:
        for ins in itertools.chain(pushes, pops):
            ins.block = next((b for b in f.blocks if ins in b.instrs), None)
            if not hasattr(ins.block, 'first'):
                ins.block.first = ins.block.instrs.index(ins)
                ins.block.last = ins.block.instrs.index(ins)
            else:
                ins.block.first = min(ins.block.first, ins.block.instrs.index(ins))
                ins.block.last = max(ins.block.last, ins.block.instrs.index(ins))
            preserv_blocks.add(ins.block)

    # group preservs in the same block and reorder them
    preservs_groups = []

    for reg, pushes, pops in preservs:
        if len(preservs_groups) == 0:
            preservs_groups.append([(reg, pushes, pops)])
            continue
        for group in preservs_groups:
            if (all(p1.block == p2.block for p1, p2 in zip(group[0][1], pushes)) and
                    all(p1.block == p2.block for p1, p2 in zip(group[0][2], pops))):
                group.append((reg, pushes, pops))
                break
            else:
                preservs_groups.append([(reg, pushes, pops)])

    # do reorder of pushes/pops
    for group in (g for g in preservs_groups if len(g) > 1):
        for group_perm in itertools.permutations(group):
            for block in preserv_blocks:
                block.rinstrs = block.instrs[:]
                block.curr_first = block.first
                block.curr_last = block.last
            for reg, pushes, pops in group:
                for push in pushes:
                    push.block.rinstrs.remove(push)
                    push.block.rinstrs.insert(push.block.curr_first, push)
                    push.block.curr_first += 1
                for pop in reversed(pops):
                    pop.block.rinstrs.remove(pop)
                    pop.block.rinstrs.insert(pop.block.curr_last - 1, pop)
                    pop.block.curr_last -= 1
            diff = set()
            for block in preserv_blocks:
                diff.update(inp.get_block_diff(block))
            _add_if_hit(g, diff, preserv_diffs)

    entropy *= len(preserv_diffs) + 1
    eliminated = eliminated or any(g.addrs[-1] == d[1] for d in itertools.chain(*preserv_diffs))

    print "\tpreservs:", len(preserv_diffs)

    # overlaping gadgets!
    if g.overlap:

        # 3. equiv
        equiv_instrs = []

        for ins in f.instrs:
            if equiv.check_equiv(ins):
                diff = inp.get_diff([ins])
                if any(g.start <= ea < g.end for ea, orig, curr in diff):
                    equiv_instrs.append(ins)
                eliminated = eliminated or any(g.addrs[-1] == d[1] for d in diff)
                ins.reset_changed()

        sr_equiv = [i for i in equiv_instrs if i.bytes[i.opc_off] in objs.same_regs]
        entropy *= 2 ** (len(equiv_instrs) - len(sr_equiv))
        entropy *= 5 ** len(sr_equiv)

        print "\tequiv (instrs):", len(equiv_instrs), "(%d sr_equiv)" % len(sr_equiv)

        # 4. reorder
        reorder_diffs = set()

        # find the block(s) of the gadget
        for block in f.blocks:
            if block.begin <= g.start < block.end or block.begin <= g.end < block.end or (
                            g.start <= block.begin and g.end >= block.end):
                # dag = reorder.BuildBBDependenceDAG(block)
                # for order in itertools.permutations(block.instrs):
                #  # all should be forward edges, from left to right
                #  if any(order.index(u) < order.index(v) for u, v in dag.edges()):
                #    continue
                for order in reorder.gen_topological_sortings(block):
                    block.rinstrs = order
                    diff = inp.get_block_diff(block)
                    _add_if_hit(g, diff, reorder_diffs)

        eliminated = any(g.addrs[-1] == d[1] for d in itertools.chain(*reorder_diffs))
        entropy *= len(reorder_diffs) + 1

        print "\treorder:", len(reorder_diffs)

    if not g.overlap:
        # 4. reorder in non-overlapping
        reorder_breakes = 0

        # find the block(s) of the gadget
        for block in f.blocks:
            if block.begin <= g.start < block.end or block.begin <= g.end < block.end or (
                            g.start <= block.begin and g.end >= block.end):
                # print "will work on block:", block
                for order in reorder.gen_topological_sortings(block):
                    # print "test this order:", order
                    block.rinstrs = order
                    # update the address of reordered instrs
                    for i, rins in enumerate(block.rinstrs):
                        if i == 0:
                            rins.raddr = block.begin
                        else:
                            rins.raddr = block.rinstrs[i - 1].raddr + len(block.rinstrs[i - 1].bytes)
                    for ins in (f.code[a] for a in g.addrs if a in f.code):
                        if ins.addr != ins.raddr and ins.raddr < g.start or ins.raddr >= g.end:
                            reorder_breakes += 1
                            break

        entropy *= reorder_breakes + 1

        print "\treorder (non-overlapping):", reorder_breakes

    print "\tgadget entropy:", entropy, "eliminated:", eliminated

    return entropy, eliminated


def eval_exploit(input_file, dump_eval_data=False):
    """Evaluates a single exploit and checks whether it is prevented. If so,
    it also calulated the number of different permutations that break it
    (entropy)."""

    start_time = time.time()

    # load and analyze the input file
    functions = inp.get_functions(input_file)
    levels = func.classify_functions(functions)
    func.analyze_functions(functions, levels)

    # get the gadgets
    try:
        gadgets = gadget.get_payload_gadgets(input_file)
    except Exception, e:
        print "ERROR: Could not open payload file"
        print e
        sys.exit(1)

    # initialize output vars
    gadgets_hit = set()
    func_gads = {}
    expl_entropy = 1
    elim_gadgets = brk_gadgets = 0
    color = {True: "red", False: "blue"}
    overlap = {True: ",overlap", False: ""}

    # group the gadgets based on the function they belong
    for g in gadgets:
        if g.func_ea not in func_gads:
            func_gads[g.func_ea] = []
        func_gads[g.func_ea].append(g)

    for func_ea, gadgets in func_gads.iteritems():
        # there is nothing we can do ..
        if func_ea == None:
            continue

        f = functions[func_ea]

        # do analysis!
        swap.liveness_analysis(f.code)
        live_regs = swap.get_reg_live_subsets(f.instrs, f.code, f.igraph)
        swap.split_reg_live_subsets(live_regs, f.code)
        swaps = swap.get_reg_swaps(live_regs)
        preservs, avail_regs = preserv.get_reg_preservations(f)

        for g in gadgets:

            print "0x%08x (%s%s) # %s" % (g.start,
                                          color[g.red], overlap[g.overlap], g.string)

            if g.red:
                continue

            if g.func_ea != g.end_func_ea:
                print "gadget spans in two or more functions!"

            entropy, eliminated = get_entropy(f, g, live_regs, swaps, preservs, avail_regs)

            expl_entropy *= entropy

            if entropy > 1:
                brk_gadgets += 1
                gadgets_hit.add(g)

            if eliminated:
                elim_gadgets += 1

    print "all:", len(gadgets), "eliminated:", elim_gadgets, "broken:", brk_gadgets
    print "payload entropy:", expl_entropy

    print "evaluation took", int(time.time() - start_time), "seconds"

    if dump_eval_data:
        _dump(input_file + ".gadgets_hit", gadgets_hit)


if __name__ == "__main__":

    parser = optparse.OptionParser("usage: %prog -ec input_file")

    parser.add_option("-c", "--coverage", dest="coverage",
                      action="store_true", default=False,
                      help="evaluate the coverage")

    parser.add_option("-e", "--exploit", dest="exploit",
                      action="store_true", default=False,
                      help="evaluate the exploit")

    (options, args) = parser.parse_args()

    if len(args) == 0:
        parser.error("no input file")
    elif len(args) > 1:
        parser.error("more than one input files")

    if options.coverage:
        # import cProfile
        # cProfile.run('eval_coverage(args[0])', 'evalprof')
        eval_coverage(args[0])
    elif options.exploit:
        eval_exploit(args[0])
    else:
        parser.error("must specify -e or -c")