bool run(OperationContext* txn, const string&, BSONObj& cmdObj, int, string& errmsg,
BSONObjBuilder& result, bool) {
Timer t;
ConnectionString hostConn(cmdObj["host"].String(),
ConnectionString::SYNC);
unsigned seed = (unsigned) number_field(cmdObj, "seed", 0);
int numThreads = (int) number_field(cmdObj, "numThreads", 4);
int wait = (int) number_field(cmdObj, "wait", 10000);
log() << "Starting " << this->name << " with -" << endl
<< " seed: " << seed << endl
<< " numThreads: " << numThreads << endl
<< " total wait: " << wait << endl << endl;
// Skew host clocks if needed
try {
skewClocks( hostConn, cmdObj );
}
catch( DBException e ) {
errmsg = str::stream() << "Clocks could not be skewed." << causedBy( e );
return false;
}
count.store(0);
keepGoing = true;
vector<shared_ptr<boost::thread> > threads;
vector<shared_ptr<BSONObjBuilder> > results;
for (int i = 0; i < numThreads; i++) {
results.push_back(shared_ptr<BSONObjBuilder> (new BSONObjBuilder()));
threads.push_back(shared_ptr<boost::thread> (new boost::thread(
stdx::bind(&TestDistLockWithSkew::runThread, this,
hostConn, (unsigned) i, seed + i, boost::ref(cmdObj),
boost::ref(*(results[i].get()))))));
}
sleepsecs(wait / 1000);
keepGoing = false;
bool errors = false;
for (unsigned i = 0; i < threads.size(); i++) {
threads[i]->join();
errors = errors || results[i].get()->obj()["errors"].Bool();
}
result.append("count", count.loadRelaxed());
result.append("errors", errors);
result.append("timeMS", t.millis());
return !errors;
}
bool run(OperationContext* txn, const string&, BSONObj& cmdObj, int, string& errmsg,
BSONObjBuilder& result, bool) {
long long skew = (long long) number_field(cmdObj, "skew", 0);
log() << "Adjusting jsTime() clock skew to " << skew << endl;
jsTimeVirtualSkew( skew );
log() << "JSTime adjusted, now is " << jsTime() << endl;
return true;
}
void runThread(ConnectionString& hostConn, unsigned threadId, unsigned seed,
BSONObj& cmdObj, BSONObjBuilder& result) {
stringstream ss;
ss << "thread-" << threadId;
setThreadName(ss.str().c_str());
// Lock name
string lockName = string_field(cmdObj, "lockName", this->name + "_lock");
// Range of clock skew in diff threads
int skewRange = (int) number_field(cmdObj, "skewRange", 1);
// How long to wait with the lock
int threadWait = (int) number_field(cmdObj, "threadWait", 30);
if(threadWait <= 0) threadWait = 1;
// Max amount of time (ms) a thread waits before checking the lock again
int threadSleep = (int) number_field(cmdObj, "threadSleep", 30);
if(threadSleep <= 0) threadSleep = 1;
// How long until the lock is forced in ms, only compared locally
unsigned long long takeoverMS = (unsigned long long) number_field(cmdObj, "takeoverMS", 0);
// Whether or not we should hang some threads
int hangThreads = (int) number_field(cmdObj, "hangThreads", 0);
boost::mt19937 gen((boost::mt19937::result_type) seed);
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomSkew(gen, boost::uniform_int<>(0, skewRange));
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomWait(gen, boost::uniform_int<>(1, threadWait));
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomSleep(gen, boost::uniform_int<>(1, threadSleep));
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomNewLock(gen, boost::uniform_int<>(0, 3));
int skew = 0;
if (!lock.get()) {
// Pick a skew, but the first two threads skew the whole range
if(threadId == 0)
skew = -skewRange / 2;
else if(threadId == 1)
skew = skewRange / 2;
else skew = randomSkew() - (skewRange / 2);
// Skew this thread
jsTimeVirtualThreadSkew( skew );
log() << "Initializing lock with skew of " << skew << " for thread " << threadId << endl;
lock.reset(new DistributedLock(hostConn, lockName, takeoverMS, true ));
log() << "Skewed time " << jsTime() << " for thread " << threadId << endl
<< " max wait (with lock: " << threadWait << ", after lock: " << threadSleep << ")" << endl
<< " takeover in " << takeoverMS << "(ms remote)" << endl;
}
DistributedLock* myLock = lock.get();
bool errors = false;
BSONObj lockObj;
while (keepGoing) {
try {
if (myLock->lock_try("Testing distributed lock with skew.", false, &lockObj )) {
log() << "**** Locked for thread " << threadId << " with ts " << lockObj["ts"] << endl;
if( count.loadRelaxed() % 2 == 1 && ! myLock->lock_try( "Testing lock re-entry.", true ) ) {
errors = true;
log() << "**** !Could not re-enter lock already held" << endl;
break;
}
if( count.loadRelaxed() % 3 == 1 && myLock->lock_try( "Testing lock non-re-entry.", false ) ) {
errors = true;
log() << "**** !Invalid lock re-entry" << endl;
break;
}
int before = count.addAndFetch(1);
int sleep = randomWait();
sleepmillis(sleep);
int after = count.loadRelaxed();
if(after != before) {
errors = true;
log() << "**** !Bad increment while sleeping with lock for: " << sleep << "ms" << endl;
break;
}
// Unlock only half the time...
if(hangThreads == 0 || threadId % hangThreads != 0) {
log() << "**** Unlocking for thread " << threadId << " with ts " << lockObj["ts"] << endl;
myLock->unlock( &lockObj );
}
else {
log() << "**** Not unlocking for thread " << threadId << endl;
verify( DistributedLock::killPinger( *myLock ) );
// We're simulating a crashed process...
break;
}
}
}
catch( const DBException& ex ) {
log() << "*** !Could not try distributed lock." << causedBy( ex ) << endl;
break;
}
// Create a new lock 1/3 of the time
if( randomNewLock() > 1 ){
lock.reset(new DistributedLock( hostConn, lockName, takeoverMS, true ));
myLock = lock.get();
}
sleepmillis(randomSleep());
}
result << "errors" << errors
<< "skew" << skew
<< "takeover" << (long long) takeoverMS
<< "localTimeout" << (takeoverMS > 0);
}
void runThread(ConnectionString& hostConn, unsigned threadId, unsigned seed,
BSONObj& cmdObj, BSONObjBuilder& result) {
stringstream ss;
ss << "thread-" << threadId;
setThreadName(ss.str().c_str());
// Lock name
string lockName = string_field(cmdObj, "lockName", this->name + "_lock");
// Range of clock skew in diff threads
int skewRange = (int) number_field(cmdObj, "skewRange", 1);
// How long to wait with the lock
int threadWait = (int) number_field(cmdObj, "threadWait", 30);
if(threadWait <= 0) threadWait = 1;
// Max amount of time (ms) a thread waits before checking the lock again
int threadSleep = (int) number_field(cmdObj, "threadSleep", 30);
if(threadSleep <= 0) threadSleep = 1;
// (Legacy) how long until the lock is forced in mins, measured locally
int takeoverMins = (int) number_field(cmdObj, "takeoverMins", 0);
// How long until the lock is forced in ms, only compared locally
unsigned long long takeoverMS = (unsigned long long) number_field(cmdObj, "takeoverMS", 0);
// Whether or not we should hang some threads
int hangThreads = (int) number_field(cmdObj, "hangThreads", 0);
boost::mt19937 gen((boost::mt19937::result_type) seed);
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomSkew(gen, boost::uniform_int<>(0, skewRange));
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomWait(gen, boost::uniform_int<>(1, threadWait));
boost::variate_generator<boost::mt19937&, boost::uniform_int<> > randomSleep(gen, boost::uniform_int<>(1, threadSleep));
int skew = 0;
bool legacy = (takeoverMins > 0);
if (!lock.get()) {
// Pick a skew, but the first two threads skew the whole range
if(threadId == 0)
skew = -skewRange / 2;
else if(threadId == 1)
skew = skewRange / 2;
else skew = randomSkew() - (skewRange / 2);
// Skew this thread
jsTimeVirtualThreadSkew( skew );
log() << "Initializing lock with skew of " << skew << " for thread " << threadId << endl;
lock.reset(new DistributedLock(hostConn, lockName, legacy ? (unsigned long long)takeoverMins : takeoverMS, true, legacy));
log() << "Skewed time " << jsTime() << " for thread " << threadId << endl
<< " max wait (with lock: " << threadWait << ", after lock: " << threadSleep << ")" << endl
<< " takeover in " << (legacy ? (unsigned long long)takeoverMins : takeoverMS) << (legacy ? " (mins local)" : "(ms remote)") << endl;
}
DistributedLock* myLock = lock.get();
bool errors = false;
while (keepGoing) {
try {
if (myLock->lock_try("Testing distributed lock with skew.")) {
log() << "**** Locked for thread " << threadId << endl;
count++;
int before = count;
int sleep = randomWait();
sleepmillis(sleep);
int after = count;
if(after != before) {
errors = true;
log() << "**** !Bad increment while sleeping with lock for: " << sleep << "ms" << endl;
break;
}
// Unlock only half the time...
if(hangThreads == 0 || threadId % hangThreads != 0) {
log() << "**** Unlocking for thread " << threadId << endl;
myLock->unlock();
}
else {
log() << "**** Not unlocking for thread " << threadId << endl;
DistributedLock::killPinger( *myLock );
// We're simulating a crashed process...
break;
}
}
}
catch( LockException& e ) {
log() << "*** !Could not try distributed lock." << m_caused_by(e) << endl;
break;
}
sleepmillis(randomSleep());
}
result << "errors" << errors
<< "skew" << skew
<< "takeover" << (long long) (legacy ? takeoverMS : (unsigned long long)takeoverMins)
<< "localTimeout" << (takeoverMS > 0);
}