void
Group::onSessionInitiateFailure(const ProcessPtr &process, Session *session) {
vector<Callback> actions;
TRACE_POINT();
// Standard resource management boilerplate stuff...
PoolPtr pool = getPool();
boost::unique_lock<boost::mutex> lock(pool->syncher);
assert(process->isAlive());
assert(isAlive() || getLifeStatus() == SHUTTING_DOWN);
UPDATE_TRACE_POINT();
P_DEBUG("Could not initiate a session with process " <<
process->inspect() << ", detaching from pool if possible");
if (!pool->detachProcessUnlocked(process, actions)) {
P_DEBUG("Process was already detached");
}
pool->fullVerifyInvariants();
lock.unlock();
runAllActions(actions);
}
// The 'self' parameter is for keeping the current Group object alive while this thread is running.
void
Group::finalizeRestart(GroupPtr self, Options options, RestartMethod method,
SpawnerFactoryPtr spawnerFactory, unsigned int restartsInitiated,
vector<Callback> postLockActions)
{
TRACE_POINT();
Pool::runAllActions(postLockActions);
postLockActions.clear();
this_thread::disable_interruption di;
this_thread::disable_syscall_interruption dsi;
// Create a new spawner.
SpawnerPtr newSpawner = spawnerFactory->create(options);
SpawnerPtr oldSpawner;
UPDATE_TRACE_POINT();
PoolPtr pool = getPool();
Pool::DebugSupportPtr debug = pool->debugSupport;
if (debug != NULL && debug->restarting) {
this_thread::restore_interruption ri(di);
this_thread::restore_syscall_interruption rsi(dsi);
this_thread::interruption_point();
debug->debugger->send("About to end restarting");
debug->messages->recv("Finish restarting");
}
ScopedLock l(pool->syncher);
if (!isAlive()) {
P_DEBUG("Group " << name << " is shutting down, so aborting restart");
return;
}
if (restartsInitiated != this->restartsInitiated) {
// Before this restart could be finalized, another restart command was given.
// The spawner we just created might be out of date now so we abort.
P_DEBUG("Restart of group " << name << " aborted because a new restart was initiated concurrently");
if (debug != NULL && debug->restarting) {
debug->debugger->send("Restarting aborted");
}
return;
}
// Run some sanity checks.
pool->fullVerifyInvariants();
assert(m_restarting);
UPDATE_TRACE_POINT();
// Atomically swap the new spawner with the old one.
resetOptions(options);
oldSpawner = spawner;
spawner = newSpawner;
m_restarting = false;
if (shouldSpawn()) {
spawn();
} else if (isWaitingForCapacity()) {
P_INFO("Group " << name << " is waiting for capacity to become available. "
"Trying to shutdown another idle process to free capacity...");
if (pool->forceFreeCapacity(this, postLockActions) != NULL) {
spawn();
} else {
P_INFO("There are no processes right now that are eligible "
"for shutdown. Will try again later.");
}
}
verifyInvariants();
l.unlock();
oldSpawner.reset();
Pool::runAllActions(postLockActions);
P_DEBUG("Restart of group " << name << " done");
if (debug != NULL && debug->restarting) {
debug->debugger->send("Restarting done");
}
}
void
Group::spawnThreadRealMain(const SpawnerPtr &spawner, const Options &options, unsigned int restartsInitiated) {
TRACE_POINT();
this_thread::disable_interruption di;
this_thread::disable_syscall_interruption dsi;
PoolPtr pool = getPool();
Pool::DebugSupportPtr debug = pool->debugSupport;
bool done = false;
while (!done) {
bool shouldFail = false;
if (debug != NULL && debug->spawning) {
UPDATE_TRACE_POINT();
this_thread::restore_interruption ri(di);
this_thread::restore_syscall_interruption rsi(dsi);
this_thread::interruption_point();
string iteration;
{
LockGuard g(debug->syncher);
debug->spawnLoopIteration++;
iteration = toString(debug->spawnLoopIteration);
}
P_DEBUG("Begin spawn loop iteration " << iteration);
debug->debugger->send("Begin spawn loop iteration " +
iteration);
vector<string> cases;
cases.push_back("Proceed with spawn loop iteration " + iteration);
cases.push_back("Fail spawn loop iteration " + iteration);
MessagePtr message = debug->messages->recvAny(cases);
shouldFail = message->name == "Fail spawn loop iteration " + iteration;
}
ProcessPtr process;
ExceptionPtr exception;
try {
UPDATE_TRACE_POINT();
this_thread::restore_interruption ri(di);
this_thread::restore_syscall_interruption rsi(dsi);
if (shouldFail) {
throw SpawnException("Simulated failure");
} else {
process = spawner->spawn(options);
process->setGroup(shared_from_this());
}
} catch (const thread_interrupted &) {
break;
} catch (const tracable_exception &e) {
exception = copyException(e);
// Let other (unexpected) exceptions crash the program so
// gdb can generate a backtrace.
}
UPDATE_TRACE_POINT();
ScopeGuard guard(boost::bind(Process::forceTriggerShutdownAndCleanup, process));
boost::unique_lock<boost::mutex> lock(pool->syncher);
if (!isAlive()) {
if (process != NULL) {
P_DEBUG("Group is being shut down so dropping process " <<
process->inspect() << " which we just spawned and exiting spawn loop");
} else {
P_DEBUG("The group is being shut down. A process failed "
"to be spawned anyway, so ignoring this error and exiting "
"spawn loop");
}
// We stop immediately because any previously assumed invariants
// may have been violated.
break;
} else if (restartsInitiated != this->restartsInitiated) {
if (process != NULL) {
P_DEBUG("A restart was issued for the group, so dropping process " <<
process->inspect() << " which we just spawned and exiting spawn loop");
} else {
P_DEBUG("A restart was issued for the group. A process failed "
"to be spawned anyway, so ignoring this error and exiting "
"spawn loop");
}
// We stop immediately because any previously assumed invariants
// may have been violated.
break;
}
verifyInvariants();
assert(m_spawning);
assert(processesBeingSpawned > 0);
processesBeingSpawned--;
assert(processesBeingSpawned == 0);
UPDATE_TRACE_POINT();
vector<Callback> actions;
if (process != NULL) {
AttachResult result = attach(process, actions);
if (result == AR_OK) {
guard.clear();
if (getWaitlist.empty()) {
pool->assignSessionsToGetWaiters(actions);
} else {
assignSessionsToGetWaiters(actions);
}
P_DEBUG("New process count = " << enabledCount <<
", remaining get waiters = " << getWaitlist.size());
} else {
done = true;
P_DEBUG("Unable to attach spawned process " << process->inspect());
if (result == AR_ANOTHER_GROUP_IS_WAITING_FOR_CAPACITY) {
pool->possiblySpawnMoreProcessesForExistingGroups();
}
}
} else {
// TODO: sure this is the best thing? if there are
// processes currently alive we should just use them.
P_ERROR("Could not spawn process for group " << name <<
": " << exception->what() << "\n" <<
exception->backtrace());
if (enabledCount == 0) {
enableAllDisablingProcesses(actions);
}
Pool::assignExceptionToGetWaiters(getWaitlist, exception, actions);
pool->assignSessionsToGetWaiters(actions);
done = true;
}
done = done
|| (processLowerLimitsSatisfied() && getWaitlist.empty())
|| processUpperLimitsReached()
|| pool->atFullCapacity(false);
m_spawning = !done;
if (done) {
P_DEBUG("Spawn loop done");
} else {
processesBeingSpawned++;
P_DEBUG("Continue spawning");
}
UPDATE_TRACE_POINT();
pool->fullVerifyInvariants();
lock.unlock();
UPDATE_TRACE_POINT();
runAllActions(actions);
UPDATE_TRACE_POINT();
}
if (debug != NULL && debug->spawning) {
debug->debugger->send("Spawn loop done");
}
}
// The 'self' parameter is for keeping the current Group object alive while this thread is running.
void
Group::spawnThreadOOBWRequest(GroupPtr self, ProcessPtr process) {
TRACE_POINT();
this_thread::disable_interruption di;
this_thread::disable_syscall_interruption dsi;
Socket *socket;
Connection connection;
PoolPtr pool = getPool();
Pool::DebugSupportPtr debug = pool->debugSupport;
UPDATE_TRACE_POINT();
P_DEBUG("Performing OOBW request for process " << process->inspect());
if (debug != NULL && debug->oobw) {
debug->debugger->send("OOBW request about to start");
debug->messages->recv("Proceed with OOBW request");
}
UPDATE_TRACE_POINT();
{
// Standard resource management boilerplate stuff...
boost::unique_lock<boost::mutex> lock(pool->syncher);
if (OXT_UNLIKELY(!process->isAlive()
|| process->enabled == Process::DETACHED
|| !isAlive()))
{
return;
}
if (process->enabled != Process::DISABLED) {
UPDATE_TRACE_POINT();
P_INFO("Out-of-Band Work canceled: process " << process->inspect() <<
" was concurrently re-enabled.");
if (debug != NULL && debug->oobw) {
debug->debugger->send("OOBW request canceled");
}
return;
}
assert(process->oobwStatus == Process::OOBW_IN_PROGRESS);
assert(process->sessions == 0);
socket = process->sessionSockets.top();
assert(socket != NULL);
}
UPDATE_TRACE_POINT();
unsigned long long timeout = 1000 * 1000 * 60; // 1 min
try {
this_thread::restore_interruption ri(di);
this_thread::restore_syscall_interruption rsi(dsi);
// Grab a connection. The connection is marked as fail in order to
// ensure it is closed / recycled after this request (otherwise we'd
// need to completely read the response).
connection = socket->checkoutConnection();
connection.fail = true;
ScopeGuard guard(boost::bind(&Socket::checkinConnection, socket, connection));
// This is copied from RequestHandler when it is sending data using the
// "session" protocol.
char sizeField[sizeof(uint32_t)];
SmallVector<StaticString, 10> data;
data.push_back(StaticString(sizeField, sizeof(uint32_t)));
data.push_back(makeStaticStringWithNull("REQUEST_METHOD"));
data.push_back(makeStaticStringWithNull("OOBW"));
data.push_back(makeStaticStringWithNull("PASSENGER_CONNECT_PASSWORD"));
data.push_back(makeStaticStringWithNull(process->connectPassword));
uint32_t dataSize = 0;
for (unsigned int i = 1; i < data.size(); i++) {
dataSize += (uint32_t) data[i].size();
}
Uint32Message::generate(sizeField, dataSize);
gatheredWrite(connection.fd, &data[0], data.size(), &timeout);
// We do not care what the actual response is ... just wait for it.
UPDATE_TRACE_POINT();
waitUntilReadable(connection.fd, &timeout);
} catch (const SystemException &e) {
P_ERROR("*** ERROR: " << e.what() << "\n" << e.backtrace());
} catch (const TimeoutException &e) {
P_ERROR("*** ERROR: " << e.what() << "\n" << e.backtrace());
}
UPDATE_TRACE_POINT();
vector<Callback> actions;
{
// Standard resource management boilerplate stuff...
PoolPtr pool = getPool();
boost::unique_lock<boost::mutex> lock(pool->syncher);
if (OXT_UNLIKELY(!process->isAlive() || !isAlive())) {
return;
}
process->oobwStatus = Process::OOBW_NOT_ACTIVE;
if (process->enabled == Process::DISABLED) {
enable(process, actions);
assignSessionsToGetWaiters(actions);
}
pool->fullVerifyInvariants();
initiateNextOobwRequest();
}
UPDATE_TRACE_POINT();
runAllActions(actions);
actions.clear();
UPDATE_TRACE_POINT();
P_DEBUG("Finished OOBW request for process " << process->inspect());
if (debug != NULL && debug->oobw) {
debug->debugger->send("OOBW request finished");
}
}
void
Group::onSessionClose(const ProcessPtr &process, Session *session) {
TRACE_POINT();
// Standard resource management boilerplate stuff...
PoolPtr pool = getPool();
boost::unique_lock<boost::mutex> lock(pool->syncher);
assert(process->isAlive());
assert(isAlive() || getLifeStatus() == SHUTTING_DOWN);
P_TRACE(2, "Session closed for process " << process->inspect());
verifyInvariants();
UPDATE_TRACE_POINT();
/* Update statistics. */
process->sessionClosed(session);
assert(process->getLifeStatus() == Process::ALIVE);
assert(process->enabled == Process::ENABLED
|| process->enabled == Process::DISABLING
|| process->enabled == Process::DETACHED);
if (process->enabled == Process::ENABLED) {
pqueue.decrease(process->pqHandle, process->busyness());
}
/* This group now has a process that's guaranteed to be not
* totally busy.
*/
assert(!process->isTotallyBusy());
bool detachingBecauseOfMaxRequests = false;
bool detachingBecauseCapacityNeeded = false;
bool shouldDetach =
( detachingBecauseOfMaxRequests = (
options.maxRequests > 0
&& process->processed >= options.maxRequests
)) || (
detachingBecauseCapacityNeeded = (
process->sessions == 0
&& getWaitlist.empty()
&& (
!pool->getWaitlist.empty()
|| anotherGroupIsWaitingForCapacity()
)
)
);
bool shouldDisable =
process->enabled == Process::DISABLING
&& process->sessions == 0
&& enabledCount > 0;
if (shouldDetach || shouldDisable) {
vector<Callback> actions;
if (shouldDetach) {
if (detachingBecauseCapacityNeeded) {
/* Someone might be trying to get() a session for a different
* group that couldn't be spawned because of lack of pool capacity.
* If this group isn't under sufficiently load (as apparent by the
* checked conditions) then now's a good time to detach
* this process or group in order to free capacity.
*/
P_DEBUG("Process " << process->inspect() << " is no longer totally "
"busy; detaching it in order to make room in the pool");
} else {
/* This process has processed its maximum number of requests,
* so we detach it.
*/
P_DEBUG("Process " << process->inspect() <<
" has reached its maximum number of requests (" <<
options.maxRequests << "); detaching it");
}
pool->detachProcessUnlocked(process, actions);
} else {
removeProcessFromList(process, disablingProcesses);
addProcessToList(process, disabledProcesses);
removeFromDisableWaitlist(process, DR_SUCCESS, actions);
maybeInitiateOobw(process);
}
pool->fullVerifyInvariants();
lock.unlock();
runAllActions(actions);
} else {
// This could change process->enabled.
maybeInitiateOobw(process);
if (!getWaitlist.empty() && process->enabled == Process::ENABLED) {
/* If there are clients on this group waiting for a process to
* become available then call them now.
*/
UPDATE_TRACE_POINT();
// Already calls verifyInvariants().
assignSessionsToGetWaitersQuickly(lock);
}
}
}