/** Main driving function for clone detection. This is the class that chooses inputs, runs each function, and looks at the
* outputs to decide how to partition the functions. It does this repeatedly in order to build a PartitionForest. The
* analyze() method is the main entry point. */
class CloneDetector {
protected:
static const char *name; /**< For --debug output. */
RSIM_Thread *thread; /**< Thread where analysis is running. */
PartitionForest partition; /**< Partitioning of functions into similarity sets. */
enum { MAX_ITERATIONS = 10 }; /**< Maximum number of times we run the functions; max number of input sets. */
enum { MAX_SIMSET_SIZE = 3 }; /**< Any similarity set containing more than this many functions will be partitioned. */
public:
CloneDetector(RSIM_Thread *thread): thread(thread) {}
// Allocate a page of memory in the process address space.
rose_addr_t allocate_page(rose_addr_t hint=0) {
RSIM_Process *proc = thread->get_process();
rose_addr_t addr = proc->mem_map(hint, 4096, MemoryMap::MM_PROT_RW, MAP_ANONYMOUS, 0, -1);
assert((int64_t)addr>=0 || (int64_t)addr<-256); // disallow error numbers
return addr;
}
// Obtain a memory map for disassembly
MemoryMap *disassembly_map(RSIM_Process *proc) {
MemoryMap *map = new MemoryMap(proc->get_memory(), MemoryMap::COPY_SHALLOW);
map->prune(MemoryMap::MM_PROT_READ); // don't let the disassembler read unreadable memory, else it will segfault
// Removes execute permission for any segment whose debug name does not contain the name of the executable. When
// comparing two different executables for clones, we probably don't need to compare code that came from dynamically
// linked libraries since they will be identical in both executables.
struct Pruner: MemoryMap::Visitor {
std::string exename;
Pruner(const std::string &exename): exename(exename) {}
virtual bool operator()(const MemoryMap*, const Extent&, const MemoryMap::Segment &segment_) {
MemoryMap::Segment *segment = const_cast<MemoryMap::Segment*>(&segment_);
if (segment->get_name().find(exename)==std::string::npos) {
unsigned p = segment->get_mapperms();
p &= ~MemoryMap::MM_PROT_EXEC;
segment->set_mapperms(p);
}
return true;
}
} pruner(proc->get_exename());
map->traverse(pruner);
return map;
}
// Get all the functions defined for this process image. We do this by disassembling the entire process executable memory
// and using CFG analysis to figure out where the functions are located.
Functions find_functions(RTS_Message *m, RSIM_Process *proc) {
m->mesg("%s triggered; disassembling entire specimen image...\n", name);
MemoryMap *map = disassembly_map(proc);
std::ostringstream ss;
map->dump(ss, " ");
m->mesg("%s: using this memory map for disassembly:\n%s", name, ss.str().c_str());
SgAsmBlock *gblk = proc->disassemble(false/*take no shortcuts*/, map);
delete map; map=NULL;
std::vector<SgAsmFunction*> functions = SageInterface::querySubTree<SgAsmFunction>(gblk);
#if 0 /*DEBUGGING [Robb P. Matzke 2013-02-12]*/
// Prune the function list to contain only what we want.
for (std::vector<SgAsmFunction*>::iterator fi=functions.begin(); fi!=functions.end(); ++fi) {
if ((*fi)->get_name().compare("_Z1fRi")!=0)
*fi = NULL;
}
functions.erase(std::remove(functions.begin(), functions.end(), (SgAsmFunction*)NULL), functions.end());
#endif
return Functions(functions.begin(), functions.end());
}
// Perform a pointer-detection analysis on each function. We'll need the results in order to determine whether a function
// input should consume a pointer or a non-pointer from the input value set.
typedef std::map<SgAsmFunction*, CloneDetection::PointerDetector> PointerDetectors;
PointerDetectors detect_pointers(RTS_Message *m, RSIM_Thread *thread, const Functions &functions) {
// Choose an SMT solver. This is completely optional. Pointer detection still seems to work fairly well (and much,
// much faster) without an SMT solver.
SMTSolver *solver = NULL;
#if 0 // optional code
if (YicesSolver::available_linkage())
solver = new YicesSolver;
#endif
PointerDetectors retval;
CloneDetection::InstructionProvidor *insn_providor = new CloneDetection::InstructionProvidor(thread->get_process());
for (Functions::iterator fi=functions.begin(); fi!=functions.end(); ++fi) {
m->mesg("%s: performing pointer detection analysis for \"%s\" at 0x%08"PRIx64,
name, (*fi)->get_name().c_str(), (*fi)->get_entry_va());
CloneDetection::PointerDetector pd(insn_providor, solver);
pd.initial_state().registers.gpr[x86_gpr_sp] = SYMBOLIC_VALUE<32>(thread->policy.INITIAL_STACK);
pd.initial_state().registers.gpr[x86_gpr_bp] = SYMBOLIC_VALUE<32>(thread->policy.INITIAL_STACK);
//pd.set_debug(stderr);
pd.analyze(*fi);
retval.insert(std::make_pair(*fi, pd));
#if 1 /*DEBUGGING [Robb P. Matzke 2013-01-24]*/
if (m->get_file()) {
const CloneDetection::PointerDetector::Pointers plist = pd.get_pointers();
for (CloneDetection::PointerDetector::Pointers::const_iterator pi=plist.begin(); pi!=plist.end(); ++pi) {
std::ostringstream ss;
if (pi->type & BinaryAnalysis::PointerAnalysis::DATA_PTR)
ss <<"data ";
if (pi->type & BinaryAnalysis::PointerAnalysis::CODE_PTR)
ss <<"code ";
ss <<"pointer at " <<pi->address;
m->mesg(" %s", ss.str().c_str());
}
}
#endif
}
return retval;
}
// Randomly choose a set of input values. The set will consist of the specified number of non-pointers and pointers. The
// non-pointer values are chosen randomly, but limited to a certain range. The pointers are chosen randomly to be null or
// non-null and the non-null values each have one page allocated via simulated mmap() (i.e., the non-null values themselves
// are not actually random).
InputValues choose_inputs(size_t nintegers, size_t npointers) {
static unsigned integer_modulus = 256; // arbitrary;
static unsigned nonnull_denom = 3; // probability of a non-null pointer is 1/N
CloneDetection::InputValues inputs;
for (size_t i=0; i<nintegers; ++i)
inputs.add_integer(rand() % integer_modulus);
for (size_t i=0; i<npointers; ++i)
inputs.add_pointer(rand()%nonnull_denom ? 0 : allocate_page());
return inputs;
}
// Run a single function, look at its outputs, and insert it into the correct place in the PartitionForest
void insert_function(SgAsmFunction *func, InputValues &inputs, CloneDetection::PointerDetector &pointers,
PartitionForest &partition, PartitionForest::Vertex *parent) {
CloneDetection::Outputs<RSIM_SEMANTICS_VTYPE> *outputs = fuzz_test(func, inputs, pointers);
OutputValues concrete_outputs = outputs->get_values();
partition.insert(func, concrete_outputs, parent);
}
// Analyze a single function by running it with the specified inputs and collecting its outputs. */
CloneDetection::Outputs<RSIM_SEMANTICS_VTYPE> *fuzz_test(SgAsmFunction *function, CloneDetection::InputValues &inputs,
const CloneDetection::PointerDetector &pointers) {
RSIM_Process *proc = thread->get_process();
RTS_Message *m = thread->tracing(TRACE_MISC);
m->mesg("==========================================================================================");
m->mesg("%s: fuzz testing function \"%s\" at 0x%08"PRIx64, name, function->get_name().c_str(), function->get_entry_va());
// Not sure if saving/restoring memory state is necessary. I don't thing machine memory is adjusted by the semantic
// policy's writeMemory() or readMemory() operations after the policy is triggered to enable our analysis. But it
// shouldn't hurt to save/restore anyway, and it's fast. [Robb Matzke 2013-01-14]
proc->mem_transaction_start(name);
pt_regs_32 saved_regs = thread->get_regs();
// Trigger the analysis, resetting it to start executing the specified function using the input values and pointer
// variable addresses we selected previously.
thread->policy.trigger(function->get_entry_va(), &inputs, &pointers);
// "Run" the function using our semantic policy. The function will not "run" in the normal sense since: since our
// policy has been triggered, memory access, function calls, system calls, etc. will all operate differently. See
// CloneDetectionSemantics.h and CloneDetectionTpl.h for details.
try {
thread->main();
} catch (const Disassembler::Exception &e) {
// Probably due to the analyzed function's RET instruction, but could be from other things as well. In any case, we
// stop analyzing the function when this happens.
m->mesg("%s: function disassembly failed at 0x%08"PRIx64": %s", name, e.ip, e.mesg.c_str());
} catch (const CloneDetection::InsnLimitException &e) {
// The analysis might be in an infinite loop, such as when analyzing "void f() { while(1); }"
m->mesg("%s: %s", name, e.mesg.c_str());
} catch (const RSIM_Semantics::InnerPolicy<>::Halt &e) {
// The x86 HLT instruction appears in some functions (like _start) as a failsafe to terminate a process. We need
// to intercept it and terminate only the function analysis.
m->mesg("%s: function executed HLT instruction at 0x%08"PRIx64, name, e.ip);
}
// Gather the function's outputs before restoring machine state.
bool verbose = true;
CloneDetection::Outputs<RSIM_SEMANTICS_VTYPE> *outputs = thread->policy.get_outputs(verbose);
thread->init_regs(saved_regs);
proc->mem_transaction_rollback(name);
return outputs;
}
