/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.cassandra.concurrent;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicReference;
import java.util.concurrent.locks.LockSupport;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.apache.cassandra.utils.JVMStabilityInspector;
final class SEPWorker extends AtomicReference<SEPWorker.Work> implements Runnable
{
private static final Logger logger = LoggerFactory.getLogger(SEPWorker.class);
final Long workerId;
final Thread thread;
final SharedExecutorPool pool;
// prevStopCheck stores the value of pool.stopCheck after we last incremented it; if it hasn't changed,
// we know nobody else was spinning in the interval, so we increment our soleSpinnerSpinTime accordingly,
// and otherwise we set it to zero; this is then used to terminate the final spinning thread, as the coordinated
// strategy can only work when there are multiple threads spinning (as more sleep time must elapse than real time)
long prevStopCheck = 0;
long soleSpinnerSpinTime = 0;
SEPWorker(Long workerId, Work initialState, SharedExecutorPool pool)
{
this.pool = pool;
this.workerId = workerId;
thread = new Thread(this, pool.poolName + "-Worker-" + workerId);
thread.setDaemon(true);
set(initialState);
thread.start();
}
public void run()
{
/**
* we maintain two important invariants:
* 1) after exiting spinning phase, we ensure at least one more task on _each_ queue will be processed
* promptly after we begin, assuming any are outstanding on any pools. this is to permit producers to
* avoid signalling if there are _any_ spinning threads. we achieve this by simply calling maybeSchedule()
* on each queue if on decrementing the spin counter we hit zero.
* 2) before processing a task on a given queue, we attempt to assign another worker to the _same queue only_;
* this allows a producer to skip signalling work if the task queue is currently non-empty, and in conjunction
* with invariant (1) ensures that if any thread was spinning when a task was added to any executor, that
* task will be processed immediately if work permits are available
*/
SEPExecutor assigned = null;
Runnable task = null;
try
{
while (true)
{
if (isSpinning() && !selfAssign())
{
doWaitSpin();
continue;
}
// if stop was signalled, go to sleep (don't try self-assign; being put to sleep is rare, so let's obey it
// whenever we receive it - though we don't apply this constraint to producers, who may reschedule us before
// we go to sleep)
if (stop())
while (isStopped())
LockSupport.park();
// we can be assigned any state from STOPPED, so loop if we don't actually have any tasks assigned
assigned = get().assigned;
if (assigned == null)
continue;
task = assigned.tasks.poll();
// if we do have tasks assigned, nobody will change our state so we can simply set it to WORKING
// (which is also a state that will never be interrupted externally)
set(Work.WORKING);
boolean shutdown;
while (true)
{
// before we process any task, we maybe schedule a new worker _to our executor only_; this
// ensures that even once all spinning threads have found work, if more work is left to be serviced
// and permits are available, it will be dealt with immediately.
assigned.maybeSchedule();
// we know there is work waiting, as we have a work permit, so poll() will always succeed
task.run();
task = null;
// if we're shutting down, or we fail to take a permit, we don't perform any more work
if ((shutdown = assigned.shuttingDown) || !assigned.takeTaskPermit())
break;
task = assigned.tasks.poll();
}
// return our work permit, and maybe signal shutdown
assigned.returnWorkPermit();
if (shutdown && assigned.getActiveCount() == 0)
assigned.shutdown.signalAll();
assigned = null;
// try to immediately reassign ourselves some work; if we fail, start spinning
if (!selfAssign())
startSpinning();
}
}
catch (Throwable t)
{
JVMStabilityInspector.inspectThrowable(t);
while (true)
{
if (get().assigned != null)
{
assigned = get().assigned;
set(Work.WORKING);
}
if (assign(Work.STOPPED, true))
break;
}
if (assigned != null)
assigned.returnWorkPermit();
if (task != null)
logger.error("Failed to execute task, unexpected exception killed worker: {}", t);
else
logger.error("Unexpected exception killed worker: {}", t);
}
}
// try to assign this worker the provided work
// valid states to assign are SPINNING, STOP_SIGNALLED, (ASSIGNED);
// restores invariants of the various states (e.g. spinningCount, descheduled collection and thread park status)
boolean assign(Work work, boolean self)
{
Work state = get();
while (state.canAssign(self))
{
if (!compareAndSet(state, work))
{
state = get();
continue;
}
// if we were spinning, exit the state (decrement the count); this is valid even if we are already spinning,
// as the assigning thread will have incremented the spinningCount
if (state.isSpinning())
stopSpinning();
// if we're being descheduled, place ourselves in the descheduled collection
if (work.isStop())
pool.descheduled.put(workerId, this);
// if we're currently stopped, and the new state is not a stop signal
// (which we can immediately convert to stopped), unpark the worker
if (state.isStopped() && (!work.isStop() || !stop()))
LockSupport.unpark(thread);
return true;
}
return false;
}
// try to assign ourselves an executor with work available
private boolean selfAssign()
{
// if we aren't permitted to assign in this state, fail
if (!get().canAssign(true))
return false;
for (SEPExecutor exec : pool.executors)
{
if (exec.takeWorkPermit(true))
{
Work work = new Work(exec);
// we successfully started work on this executor, so we must either assign it to ourselves or ...
if (assign(work, true))
return true;
// ... if we fail, schedule it to another worker
pool.schedule(work);
// and return success as we must have already been assigned a task
assert get().assigned != null;
return true;
}
}
return false;
}
// we can only call this when our state is WORKING, and no other thread may change our state in this case;
// so in this case only we do not need to CAS. We increment the spinningCount and add ourselves to the spinning
// collection at the same time
private void startSpinning()
{
assert get() == Work.WORKING;
pool.spinningCount.incrementAndGet();
set(Work.SPINNING);
}
// exit the spinning state; if there are no remaining spinners, we immediately try and schedule work for all executors
// so that any producer is safe to not spin up a worker when they see a spinning thread (invariant (1) above)
private void stopSpinning()
{
if (pool.spinningCount.decrementAndGet() == 0)
for (SEPExecutor executor : pool.executors)
executor.maybeSchedule();
prevStopCheck = soleSpinnerSpinTime = 0;
}
// perform a sleep-spin, incrementing pool.stopCheck accordingly
private void doWaitSpin()
{
// pick a random sleep interval based on the number of threads spinning, so that
// we should always have a thread about to wake up, but most threads are sleeping
long sleep = 10000L * pool.spinningCount.get();
sleep = Math.min(1000000, sleep);
sleep *= Math.random();
sleep = Math.max(10000, sleep);
long start = System.nanoTime();
// place ourselves in the spinning collection; if we clash with another thread just exit
Long target = start + sleep;
if (pool.spinning.putIfAbsent(target, this) != null)
return;
LockSupport.parkNanos(sleep);
// remove ourselves (if haven't been already) - we should be at or near the front, so should be cheap-ish
pool.spinning.remove(target, this);
// finish timing and grab spinningTime (before we finish timing so it is under rather than overestimated)
long end = System.nanoTime();
long spin = end - start;
long stopCheck = pool.stopCheck.addAndGet(spin);
maybeStop(stopCheck, end);
if (prevStopCheck + spin == stopCheck)
soleSpinnerSpinTime += spin;
else
soleSpinnerSpinTime = 0;
prevStopCheck = stopCheck;
}
private static final long stopCheckInterval = TimeUnit.MILLISECONDS.toNanos(10L);
// stops a worker if elapsed real time is less than elapsed spin time, as this implies the equivalent of
// at least one worker achieved nothing in the interval. we achieve this by maintaining a stopCheck which
// is initialised to a negative offset from realtime; as we spin we add to this value, and if we ever exceed
// realtime we have spun too much and deschedule; if we get too far behind realtime, we reset to our initial offset
private void maybeStop(long stopCheck, long now)
{
long delta = now - stopCheck;
if (delta <= 0)
{
// if stopCheck has caught up with present, we've been spinning too much, so if we can atomically
// set it to the past again, we should stop a worker
if (pool.stopCheck.compareAndSet(stopCheck, now - stopCheckInterval))
{
// try and stop ourselves;
// if we've already been assigned work stop another worker
if (!assign(Work.STOP_SIGNALLED, true))
pool.schedule(Work.STOP_SIGNALLED);
}
}
else if (soleSpinnerSpinTime > stopCheckInterval && pool.spinningCount.get() == 1)
{
// permit self-stopping
assign(Work.STOP_SIGNALLED, true);
}
else
{
// if stop check has gotten too far behind present, update it so new spins can affect it
while (delta > stopCheckInterval * 2 && !pool.stopCheck.compareAndSet(stopCheck, now - stopCheckInterval))
{
stopCheck = pool.stopCheck.get();
delta = now - stopCheck;
}
}
}
private boolean isSpinning()
{
return get().isSpinning();
}
private boolean stop()
{
return get().isStop() && compareAndSet(Work.STOP_SIGNALLED, Work.STOPPED);
}
private boolean isStopped()
{
return get().isStopped();
}
/**
* Represents, and communicates changes to, a worker's work state - there are three non-actively-working
* states (STOP_SIGNALLED, STOPPED, AND SPINNING) and two working states: WORKING, and (ASSIGNED), the last
* being represented by a non-static instance with its "assigned" executor set.
*
* STOPPED: indicates the worker is descheduled, and whilst accepts work in this state (causing it to
* be rescheduled) it will generally not be considered for work until all other worker threads are busy.
* In this state we should be present in the pool.descheduled collection, and should be parked
* -> (ASSIGNED)|SPINNING
* STOP_SIGNALLED: the worker has been asked to deschedule itself, but has not yet done so; only entered from a SPINNING
* state, and generally communicated to itself, but maybe set from any worker. this state may be preempted
* and replaced with (ASSIGNED) or SPINNING
* In this state we should be present in the pool.descheduled collection
* -> (ASSIGNED)|STOPPED|SPINNING
* SPINNING: indicates the worker has no work to perform, so is performing a friendly wait-based-spinning
* until it either is (ASSIGNED) some work (by itself or another thread), or sent STOP_SIGNALLED
* In this state we _may_ be in the pool.spinning collection (but only if we are in the middle of a sleep)
* -> (ASSIGNED)|STOP_SIGNALLED|SPINNING
* (ASSIGNED): asks the worker to perform some work against the specified executor, and preassigns a task permit
* from that executor so that in this state there is always work to perform.
* In general a worker assigns itself this state, but sometimes it may assign another worker the state
* either if there is work outstanding and no-spinning threads, or there is a race to self-assign
* -> WORKING
* WORKING: indicates the worker is actively processing an executor's task queue; in this state it accepts
* no state changes/communications, except from itself; it usually exits this mode into SPINNING,
* but if work is immediately available on another executor it self-triggers (ASSIGNED)
* -> SPINNING|(ASSIGNED)
*/
static final class Work
{
static final Work STOP_SIGNALLED = new Work();
static final Work STOPPED = new Work();
static final Work SPINNING = new Work();
static final Work WORKING = new Work();
final SEPExecutor assigned;
Work(SEPExecutor executor)
{
this.assigned = executor;
}
private Work()
{
this.assigned = null;
}
boolean canAssign(boolean self)
{
// we can assign work if there isn't new work already assigned and either
// 1) we are assigning to ourselves
// 2) the worker we are assigning to is not already in the middle of WORKING
return assigned == null && (self || !isWorking());
}
boolean isSpinning()
{
return this == Work.SPINNING;
}
boolean isWorking()
{
return this == Work.WORKING;
}
boolean isStop()
{
return this == Work.STOP_SIGNALLED;
}
boolean isStopped()
{
return this == Work.STOPPED;
}
boolean isAssigned()
{
return assigned != null;
}
}
}