belliottsmith commented on code in PR #4324: URL: https://github.com/apache/cassandra/pull/4324#discussion_r2355157988
########## src/java/org/apache/cassandra/metrics/ThreadLocalMetrics.java: ########## @@ -0,0 +1,377 @@ +/* + * 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.metrics; + +import java.lang.ref.PhantomReference; +import java.lang.ref.ReferenceQueue; +import java.util.BitSet; +import java.util.Collections; +import java.util.List; +import java.util.Set; +import java.util.concurrent.ConcurrentHashMap; +import java.util.concurrent.CopyOnWriteArrayList; +import java.util.concurrent.TimeUnit; +import java.util.concurrent.TimeoutException; +import java.util.concurrent.atomic.AtomicInteger; +import java.util.concurrent.atomic.AtomicLongArray; +import java.util.concurrent.locks.Lock; +import java.util.concurrent.locks.ReadWriteLock; +import java.util.concurrent.locks.ReentrantReadWriteLock; + +import com.google.common.annotations.VisibleForTesting; + +import io.netty.util.concurrent.FastThreadLocal; +import org.apache.cassandra.concurrent.Shutdownable; + +import static com.google.common.collect.ImmutableList.of; +import static org.apache.cassandra.concurrent.ExecutorFactory.Global.executorFactory; +import static org.apache.cassandra.concurrent.InfiniteLoopExecutor.SimulatorSafe.UNSAFE; +import static org.apache.cassandra.utils.ExecutorUtils.shutdownAndWait; + +/** + * A thread-local counter implementation designed to use in metrics as an alternative to LongAdder used by Dropwizard metrics. + * This implementation has reduced write (increment) CPU usage costs in exchange for a higher read cost. + * We keep and increment parts of a counter locally for each thread. + * To reduce memory footprint per counter they are grouped together to a long[] array for each thread. + * A position of a counter value is the same for every thread for the same counter id. + * Piggyback volatile visibility is expected for readers who execute getCount method to see recent writes to thread local arrays. + * If a metric is not used anymore the position in the array is reused. Phantom references are used to track aliveness of metric users. + * When a thread died the counter values accumulated by it are transfered to a shared summaryValues collection. + * Threads death is tracked using 2 approaches: FastThreadLocal.onRemoval callback and phantom references to Thread objects. + */ +public class ThreadLocalMetrics +{ + private static final int INITIAL_COUNTERS_CAPACITY = 16; + + static final AtomicInteger idGenerator = new AtomicInteger(); + + private static final Object freeMetricIdSetGuard = new Object(); + + @VisibleForTesting + static final BitSet freeMetricIdSet = new BitSet(); + + private static final List<ThreadLocalMetrics> allThreadLocalMetrics = new CopyOnWriteArrayList<>(); + + /* the lock is used to coordinate the threads which: + * 1) transfer values from a dead thread to summaryValues + * 2) calculate a getCount value. + * Using this lock we want to avoid + * a value lost while moving it in getCount + * as well as a double-counting + */ + private static final ReadWriteLock summaryLock = new ReentrantReadWriteLock(); + + private static final FastThreadLocal<ThreadLocalMetrics> threadLocalMetricsCurrent = new FastThreadLocal<>() + { + @Override + protected ThreadLocalMetrics initialValue() + { + ThreadLocalMetrics result = new ThreadLocalMetrics(); + allThreadLocalMetrics.add(result); + destroyWhenUnreachable(Thread.currentThread(), result::release); + return result; + } + + // this method is invoked when a thread is going to finish, but it works only for FastThreadLocalThread + // so, we use phantom references for other cases + @Override + protected void onRemoval(ThreadLocalMetrics value) + { + value.release(); + } + }; + + private static volatile AtomicLongArray summaryValues = new AtomicLongArray(INITIAL_COUNTERS_CAPACITY); + + private static final Shutdownable cleaner; + private static final Set<PhantomReference<Object>> phantomReferences = Collections.newSetFromMap(new ConcurrentHashMap<>()); + private static final ReferenceQueue<Object> referenceQueue = new ReferenceQueue<>(); + + static + { + cleaner = executorFactory().infiniteLoop("ThreadLocalMetrics-Cleaner", ThreadLocalMetrics::cleanupOneReference, UNSAFE); + } + + // we assume that counterValues can be only extended + private long[] counterValues = new long[INITIAL_COUNTERS_CAPACITY]; + + private static void cleanupOneReference() throws InterruptedException + { + Object obj = referenceQueue.remove(100); + if (obj instanceof MetricIdReference) + { + ((MetricIdReference) obj).release(); + phantomReferences.remove(obj); + } + else if (obj instanceof MetricCleanerReference) + { + ((MetricCleanerReference) obj).release(); + phantomReferences.remove(obj); + } + } + + private static class MetricIdReference extends PhantomReference<Object> + { + private final int metricId; + + public MetricIdReference(Object referent, ReferenceQueue<? super Object> q, int metricId) + { + super(referent, q); + this.metricId = metricId; + } + + public void release() + { + recycleMetricId(metricId); + } + } + + private static class MetricCleanerReference extends PhantomReference<Object> + { + private final MetricCleaner metricCleaner; + + public MetricCleanerReference(Object referent, ReferenceQueue<? super Object> q, MetricCleaner metricCleaner) + { + super(referent, q); + this.metricCleaner = metricCleaner; + } + + public void release() + { + metricCleaner.clean(); + } + } + + interface MetricCleaner + { + void clean(); + } + + static void destroyWhenUnreachable(Object referent, int metricId) + { + phantomReferences.add(new MetricIdReference(referent, referenceQueue, metricId)); + } + + static void destroyWhenUnreachable(Object referent, MetricCleaner metricCleaner) + { + phantomReferences.add(new MetricCleanerReference(referent, referenceQueue, metricCleaner)); + } + + @VisibleForTesting + public static void shutdownCleaner(long timeout, TimeUnit unit) throws InterruptedException, TimeoutException + { + shutdownAndWait(timeout, unit, of(cleaner)); + } + + private void release() + { + // Using this lock while moving we want to avoid races with readers in getCount + // such races can cause a transfered value lost or its double-counting by a reader + Lock lock = summaryLock.writeLock(); + lock.lock(); + try + { + // we may try to release ThreadLocalMetrics 2 times: onRemoval and by PhantomReference + // so this if check is needed to avoid a potential double release + if (allThreadLocalMetrics.remove(this)) + { + for (int metricId = 0; metricId < counterValues.length; metricId++) + { + long value = counterValues[metricId]; + if (value != 0) + updateSummary(metricId, value); + } + } + } + finally + { + lock.unlock(); + } + } + + /** + * If we already have ThreadLocalMetrics instance looked up for the current thread + * we can use this method to avoid thread local lookup costs. + * It can be used if you need to update several counters at the same time. + * @param metricId metric to add a value + * @param n valuen to add, can be negative number as well + */ + public void addNonStatic(int metricId, long n) + { + getNonStatic(metricId)[metricId] += n; + } + + public static void add(int metricId, long n) + { + get(metricId)[metricId] += n; + } + + private static long getCount(int metricId, boolean resetToZero) + { + long result; + Lock readLock = summaryLock.readLock(); + readLock.lock(); + try + { + result = getSummaryValue(metricId); + for (ThreadLocalMetrics threadLocalMetrics : allThreadLocalMetrics) + { + long count = 0; + long[] currentCounterValues = threadLocalMetrics.counterValues; + // currentCounterValues is extended for a thread when a value for metricId is reported in the thread + if (metricId < currentCounterValues.length) + count = currentCounterValues[metricId]; + result += count; + } + if (resetToZero) + updateSummary(metricId, -result); // compensative reset without writing to thread local values + } + finally + { + readLock.unlock(); + } + return result; + } + + // must be executed under summaryLock + private static long getSummaryValue(int metricId) + { + return summaryValues.get(metricId); + } + + // must be executed under summaryLock + private static void updateSummary(int metricId, long value) + { + summaryValues.getAndAdd(metricId, value); + } + + public static long getCount(int metricId) + { + return getCount(metricId, false); + } + + public static long getCountAndReset(int metricId) + { + return getCount(metricId, true); + } + + public static ThreadLocalMetrics get() { + return threadLocalMetricsCurrent.get(); + } + + private static long[] get(int metricId) + { + ThreadLocalMetrics threadLocalMetrics = ThreadLocalMetrics.get(); + return threadLocalMetrics.getNonStatic(metricId); + } + + private long[] getNonStatic(int metricId) + { + long[] currentCounterValues = counterValues; + if (metricId < currentCounterValues.length) + return currentCounterValues; + + long[] newCounterValues = new long[calculateNewCapacity(metricId)]; + System.arraycopy(currentCounterValues, 0, newCounterValues, 0, currentCounterValues.length); + counterValues = newCounterValues; + return newCounterValues; + } + + private static int calculateNewCapacity(int metricId) + { + return Math.max(metricId + 1, (int)(metricId * 1.1) ); + } + + static int allocateMetricId() + { + int metricId; + synchronized (freeMetricIdSetGuard) + { + metricId = freeMetricIdSet.nextSetBit(0); + if (metricId >= 0) + freeMetricIdSet.clear(metricId); + } + if (metricId < 0) + metricId = idGenerator.getAndIncrement(); + + if (metricId >= summaryValues.length()) // double-checked locking + { + Lock lock = summaryLock.writeLock(); + lock.lock(); + try + { + if (metricId >= summaryValues.length()) + { + AtomicLongArray newSummaryValues = new AtomicLongArray(calculateNewCapacity(metricId)); + for (int i = 0; i < summaryValues.length(); i++) + newSummaryValues.set(i, summaryValues.get(i)); + summaryValues = newSummaryValues; + } + } + finally + { + lock.unlock(); + } + } + return metricId; + } + + static void recycleMetricId(int metricId) + { + // we use lock here to avoid potential issues when a metric is releasing and a thread is detected as dead at the same time + // in this case we may clean a summary value and later the thread removal logic may re-add a non-zero summary value + Lock lock = summaryLock.writeLock(); + lock.lock(); + try + { + for (ThreadLocalMetrics threadLocalMetrics : allThreadLocalMetrics) + { + long[] currentCounterValues = threadLocalMetrics.counterValues; + if (metricId < currentCounterValues.length) + currentCounterValues[metricId] = 0; Review Comment: Yes. I don't know whether we really want to solve it or not, as it isn't trivial. The simplest approach is to introduce some volatile variable we read on each operation. This would ideally be `counterValues` (which is free on x86, but not weaker memory models) but we would need to come up with a scheme to allow us to write a no-op that isn't a no-op (since we cannot actually change `counterValues` under the owning thread). So we might need a second `counterValues0` value we write on the owning thread whenever we resize, and the zeroing thread reads this and writes it to `counterValues` to make the zero visible. This would definitely solve the problem, but it would come with some performance penalty on non-TSO systems. Another option would be to periodically checkpoint each `ThreadLocalMetrics` object and only recirculate the id after each thread has checkpointed. But this is quite complicated for something that is unlikely to be a bug that could ever manifest in a real system. Another option would be to zero on release of the id, then simply wait some number of minutes before we mark the id reusable, which more than puts this into "physically impossible even if not technically guaranteed by the JMM" territory. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
