util/concurrent/Striped.java

/*
 * Copyright (C) 2011 The Guava Authors
 *
 * Licensed 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 com.google.common.util.concurrent;

import com.google.common.annotations.Beta;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.MoreObjects;
import com.google.common.base.Preconditions;
import com.google.common.base.Supplier;
import com.google.common.collect.ImmutableList;
import com.google.common.collect.Iterables;
import com.google.common.collect.MapMaker;
import com.google.common.math.IntMath;
import com.google.common.primitives.Ints;

import java.lang.ref.Reference;
import java.lang.ref.ReferenceQueue;
import java.lang.ref.WeakReference;
import java.math.RoundingMode;
import java.util.Arrays;
import java.util.Collections;
import java.util.List;
import java.util.concurrent.ConcurrentMap;
import java.util.concurrent.Semaphore;
import java.util.concurrent.atomic.AtomicReferenceArray;
import java.util.concurrent.locks.Lock;
import java.util.concurrent.locks.ReadWriteLock;
import java.util.concurrent.locks.ReentrantLock;
import java.util.concurrent.locks.ReentrantReadWriteLock;

/**
 * A striped {@code Lock/Semaphore/ReadWriteLock}. This offers the underlying lock striping
 * similar to that of {@code ConcurrentHashMap} in a reusable form, and extends it for
 * semaphores and read-write locks. Conceptually, lock striping is the technique of dividing a lock
 * into many <i>stripes</i>, increasing the granularity of a single lock and allowing independent
 * operations to lock different stripes and proceed concurrently, instead of creating contention
 * for a single lock.
 *
 * <p>The guarantee provided by this class is that equal keys lead to the same lock (or semaphore),
 * i.e. {@code if (key1.equals(key2))} then {@code striped.get(key1) == striped.get(key2)}
 * (assuming {@link Object#hashCode()} is correctly implemented for the keys). Note
 * that if {@code key1} is <strong>not</strong> equal to {@code key2}, it is <strong>not</strong>
 * guaranteed that {@code striped.get(key1) != striped.get(key2)}; the elements might nevertheless
 * be mapped to the same lock. The lower the number of stripes, the higher the probability of this
 * happening.
 *
 * <p>There are three flavors of this class: {@code Striped<Lock>}, {@code Striped<Semaphore>},
 * and {@code Striped<ReadWriteLock>}. For each type, two implementations are offered:
 * {@linkplain #lock(int) strong} and {@linkplain #lazyWeakLock(int) weak}
 * {@code Striped<Lock>}, {@linkplain #semaphore(int, int) strong} and {@linkplain
 * #lazyWeakSemaphore(int, int) weak} {@code Striped<Semaphore>}, and {@linkplain
 * #readWriteLock(int) strong} and {@linkplain #lazyWeakReadWriteLock(int) weak}
 * {@code Striped<ReadWriteLock>}. <i>Strong</i> means that all stripes (locks/semaphores) are
 * initialized eagerly, and are not reclaimed unless {@code Striped} itself is reclaimable.
 * <i>Weak</i> means that locks/semaphores are created lazily, and they are allowed to be reclaimed
 * if nobody is holding on to them. This is useful, for example, if one wants to create a {@code
 * Striped<Lock>} of many locks, but worries that in most cases only a small portion of these
 * would be in use.
 *
 * <p>Prior to this class, one might be tempted to use {@code Map<K, Lock>}, where {@code K}
 * represents the task. This maximizes concurrency by having each unique key mapped to a unique
 * lock, but also maximizes memory footprint. On the other extreme, one could use a single lock
 * for all tasks, which minimizes memory footprint but also minimizes concurrency. Instead of
 * choosing either of these extremes, {@code Striped} allows the user to trade between required
 * concurrency and memory footprint. For example, if a set of tasks are CPU-bound, one could easily
 * create a very compact {@code Striped<Lock>} of {@code availableProcessors() * 4} stripes,
 * instead of possibly thousands of locks which could be created in a {@code Map<K, Lock>}
 * structure.
 *
 * @author Dimitris Andreou
 * @since 13.0
 */
@Beta
public abstract class Striped<L> {
  /**
   * If there are at least this many stripes, we assume the memory usage of a ConcurrentMap will be
   * smaller than a large array.  (This assumes that in the lazy case, most stripes are unused. As
   * always, if many stripes are in use, a non-lazy striped makes more sense.)
   */
  private static final int LARGE_LAZY_CUTOFF = 1024;

  private Striped() {}

  /**
   * Returns the stripe that corresponds to the passed key. It is always guaranteed that if
   * {@code key1.equals(key2)}, then {@code get(key1) == get(key2)}.
   *
   * @param key an arbitrary, non-null key
   * @return the stripe that the passed key corresponds to
   */
  public abstract L get(Object key);

  /**
   * Returns the stripe at the specified index. Valid indexes are 0, inclusively, to
   * {@code size()}, exclusively.
   *
   * @param index the index of the stripe to return; must be in {@code [0...size())}
   * @return the stripe at the specified index
   */
  public abstract L getAt(int index);

  /**
   * Returns the index to which the given key is mapped, so that getAt(indexFor(key)) == get(key).
   */
  abstract int indexFor(Object key);

  /**
   * Returns the total number of stripes in this instance.
   */
  public abstract int size();

  /**
   * Returns the stripes that correspond to the passed objects, in ascending (as per
   * {@link #getAt(int)}) order. Thus, threads that use the stripes in the order returned
   * by this method are guaranteed to not deadlock each other.
   *
   * <p>It should be noted that using a {@code Striped<L>} with relatively few stripes, and
   * {@code bulkGet(keys)} with a relative large number of keys can cause an excessive number
   * of shared stripes (much like the birthday paradox, where much fewer than anticipated birthdays
   * are needed for a pair of them to match). Please consider carefully the implications of the
   * number of stripes, the intended concurrency level, and the typical number of keys used in a
   * {@code bulkGet(keys)} operation. See <a href="http://www.mathpages.com/home/kmath199.htm">Balls
   * in Bins model</a> for mathematical formulas that can be used to estimate the probability of
   * collisions.
   *
   * @param keys arbitrary non-null keys
   * @return the stripes corresponding to the objects (one per each object, derived by delegating
   *         to {@link #get(Object)}; may contain duplicates), in an increasing index order.
   */
  public Iterable<L> bulkGet(Iterable<?> keys) {
    // Initially using the array to store the keys, then reusing it to store the respective L's
    final Object[] array = Iterables.toArray(keys, Object.class);
    if (array.length == 0) {
      return ImmutableList.of();
    }
    int[] stripes = new int[array.length];
    for (int i = 0; i < array.length; i++) {
      stripes[i] = indexFor(array[i]);
    }
    Arrays.sort(stripes);
    // optimize for runs of identical stripes
    int previousStripe = stripes[0];
    array[0] = getAt(previousStripe);
    for (int i = 1; i < array.length; i++) {
      int currentStripe = stripes[i];
      if (currentStripe == previousStripe) {
        array[i] = array[i - 1];
      } else {
        array[i] = getAt(currentStripe);
        previousStripe = currentStripe;
      }
    }
    /*
     * Note that the returned Iterable holds references to the returned stripes, to avoid
     * error-prone code like:
     *
     * Striped<Lock> stripedLock = Striped.lazyWeakXXX(...)'
     * Iterable<Lock> locks = stripedLock.bulkGet(keys);
     * for (Lock lock : locks) {
     *   lock.lock();
     * }
     * operation();
     * for (Lock lock : locks) {
     *   lock.unlock();
     * }
     *
     * If we only held the int[] stripes, translating it on the fly to L's, the original locks
     * might be garbage collected after locking them, ending up in a huge mess.
     */
    @SuppressWarnings("unchecked") // we carefully replaced all keys with their respective L's
    List<L> asList = (List<L>) Arrays.asList(array);
    return Collections.unmodifiableList(asList);
  }

  // Static factories

  /**
   * Creates a {@code Striped<Lock>} with eagerly initialized, strongly referenced locks.
   * Every lock is reentrant.
   *
   * @param stripes the minimum number of stripes (locks) required
   * @return a new {@code Striped<Lock>}
   */
  public static Striped<Lock> lock(int stripes) {
    return new CompactStriped<Lock>(stripes, new Supplier<Lock>() {
      @Override public Lock get() {
        return new PaddedLock();
      }
    });
  }

  /**
   * Creates a {@code Striped<Lock>} with lazily initialized, weakly referenced locks.
   * Every lock is reentrant.
   *
   * @param stripes the minimum number of stripes (locks) required
   * @return a new {@code Striped<Lock>}
   */
  public static Striped<Lock> lazyWeakLock(int stripes) {
    return lazy(stripes, new Supplier<Lock>() {
      @Override public Lock get() {
        return new ReentrantLock(false);
      }
    });
  }

  private static <L> Striped<L> lazy(int stripes, Supplier<L> supplier) {
    return stripes < LARGE_LAZY_CUTOFF
        ? new SmallLazyStriped<L>(stripes, supplier)
        : new LargeLazyStriped<L>(stripes, supplier);
  }

  /**
   * Creates a {@code Striped<Semaphore>} with eagerly initialized, strongly referenced semaphores,
   * with the specified number of permits.
   *
   * @param stripes the minimum number of stripes (semaphores) required
   * @param permits the number of permits in each semaphore
   * @return a new {@code Striped<Semaphore>}
   */
  public static Striped<Semaphore> semaphore(int stripes, final int permits) {
    return new CompactStriped<Semaphore>(stripes, new Supplier<Semaphore>() {
      @Override public Semaphore get() {
        return new PaddedSemaphore(permits);
      }
    });
  }

  /**
   * Creates a {@code Striped<Semaphore>} with lazily initialized, weakly referenced semaphores,
   * with the specified number of permits.
   *
   * @param stripes the minimum number of stripes (semaphores) required
   * @param permits the number of permits in each semaphore
   * @return a new {@code Striped<Semaphore>}
   */
  public static Striped<Semaphore> lazyWeakSemaphore(int stripes, final int permits) {
    return lazy(stripes, new Supplier<Semaphore>() {
      @Override public Semaphore get() {
        return new Semaphore(permits, false);
      }
    });
  }

  /**
   * Creates a {@code Striped<ReadWriteLock>} with eagerly initialized, strongly referenced
   * read-write locks. Every lock is reentrant.
   *
   * @param stripes the minimum number of stripes (locks) required
   * @return a new {@code Striped<ReadWriteLock>}
   */
  public static Striped<ReadWriteLock> readWriteLock(int stripes) {
    return new CompactStriped<ReadWriteLock>(stripes, READ_WRITE_LOCK_SUPPLIER);
  }

  /**
   * Creates a {@code Striped<ReadWriteLock>} with lazily initialized, weakly referenced
   * read-write locks. Every lock is reentrant.
   *
   * @param stripes the minimum number of stripes (locks) required
   * @return a new {@code Striped<ReadWriteLock>}
   */
  public static Striped<ReadWriteLock> lazyWeakReadWriteLock(int stripes) {
    return lazy(stripes, READ_WRITE_LOCK_SUPPLIER);
  }

  // ReentrantReadWriteLock is large enough to make padding probably unnecessary
  private static final Supplier<ReadWriteLock> READ_WRITE_LOCK_SUPPLIER =
      new Supplier<ReadWriteLock>() {
    @Override public ReadWriteLock get() {
      return new ReentrantReadWriteLock();
    }
  };

  private abstract static class PowerOfTwoStriped<L> extends Striped<L> {
    /** Capacity (power of two) minus one, for fast mod evaluation */
    final int mask;

    PowerOfTwoStriped(int stripes) {
      Preconditions.checkArgument(stripes > 0, "Stripes must be positive");
      this.mask = stripes > Ints.MAX_POWER_OF_TWO ? ALL_SET : ceilToPowerOfTwo(stripes) - 1;
    }

    @Override final int indexFor(Object key) {
      int hash = smear(key.hashCode());
      return hash & mask;
    }

    @Override public final L get(Object key) {
      return getAt(indexFor(key));
    }
  }

  /**
   * Implementation of Striped where 2^k stripes are represented as an array of the same length,
   * eagerly initialized.
   */
  private static class CompactStriped<L> extends PowerOfTwoStriped<L> {
    /** Size is a power of two. */
    private final Object[] array;

    private CompactStriped(int stripes, Supplier<L> supplier) {
      super(stripes);
      Preconditions.checkArgument(stripes <= Ints.MAX_POWER_OF_TWO, "Stripes must be <= 2^30)");

      this.array = new Object[mask + 1];
      for (int i = 0; i < array.length; i++) {
        array[i] = supplier.get();
      }
    }

    @SuppressWarnings("unchecked") // we only put L's in the array
    @Override public L getAt(int index) {
      return (L) array[index];
    }

    @Override public int size() {
      return array.length;
    }
  }

  /**
   * Implementation of Striped where up to 2^k stripes can be represented, using an
   * AtomicReferenceArray of size 2^k. To map a user key into a stripe, we take a k-bit slice of the
   * user key's (smeared) hashCode(). The stripes are lazily initialized and are weakly referenced.
   */
  @VisibleForTesting static class SmallLazyStriped<L> extends PowerOfTwoStriped<L> {
    final AtomicReferenceArray<ArrayReference<? extends L>> locks;
    final Supplier<L> supplier;
    final int size;
    final ReferenceQueue<L> queue = new ReferenceQueue<L>();

    SmallLazyStriped(int stripes, Supplier<L> supplier) {
      super(stripes);
      this.size = (mask == ALL_SET) ? Integer.MAX_VALUE : mask + 1;
      this.locks = new AtomicReferenceArray<ArrayReference<? extends L>>(size);
      this.supplier = supplier;
    }

    @Override public L getAt(int index) {
      if (size != Integer.MAX_VALUE) {
        Preconditions.checkElementIndex(index, size());
      } // else no check necessary, all index values are valid
      ArrayReference<? extends L> existingRef = locks.get(index);
      L existing = existingRef == null ? null : existingRef.get();
      if (existing != null) {
        return existing;
      }
      L created = supplier.get();
      ArrayReference<L> newRef = new ArrayReference<L>(created, index, queue);
      while (!locks.compareAndSet(index, existingRef, newRef)) {
        // we raced, we need to re-read and try again
        existingRef = locks.get(index);
        existing = existingRef == null ? null : existingRef.get();
        if (existing != null) {
          return existing;
        }
      }
      drainQueue();
      return created;
    }

    // N.B. Draining the queue is only necessary to ensure that we don't accumulate empty references
    // in the array.  We could skip this if we decide we don't care about holding on to Reference
    // objects indefinitely.
    private void drainQueue() {
      Reference<? extends L> ref;
      while ((ref = queue.poll()) != null) {
        // We only ever register ArrayReferences with the queue so this is always safe.
        ArrayReference<? extends L> arrayRef = (ArrayReference<? extends L>) ref;
        // Try to clear out the array slot, n.b. if we fail that is fine, in either case the
        // arrayRef will be out of the array after this step.
        locks.compareAndSet(arrayRef.index, arrayRef, null);
      }
    }

    @Override public int size() {
      return size;
    }

    private static final class ArrayReference<L> extends WeakReference<L> {
      final int index;

      ArrayReference(L referent, int index, ReferenceQueue<L> queue) {
        super(referent, queue);
        this.index = index;
      }
    }
  }

  /**
   * Implementation of Striped where up to 2^k stripes can be represented, using a ConcurrentMap
   * where the key domain is [0..2^k). To map a user key into a stripe, we take a k-bit slice of the
   * user key's (smeared) hashCode(). The stripes are lazily initialized and are weakly referenced.
   */
  @VisibleForTesting static class LargeLazyStriped<L> extends PowerOfTwoStriped<L> {
    final ConcurrentMap<Integer, L> locks;
    final Supplier<L> supplier;
    final int size;

    LargeLazyStriped(int stripes, Supplier<L> supplier) {
      super(stripes);
      this.size = (mask == ALL_SET) ? Integer.MAX_VALUE : mask + 1;
      this.supplier = supplier;
      this.locks = new MapMaker().weakValues().makeMap();
    }

    @Override public L getAt(int index) {
      if (size != Integer.MAX_VALUE) {
        Preconditions.checkElementIndex(index, size());
      } // else no check necessary, all index values are valid
      L existing = locks.get(index);
      if (existing != null) {
        return existing;
      }
      L created = supplier.get();
      existing = locks.putIfAbsent(index, created);
      return MoreObjects.firstNonNull(existing, created);
    }

    @Override public int size() {
      return size;
    }
  }

  /**
   * A bit mask were all bits are set.
   */
  private static final int ALL_SET = ~0;

  private static int ceilToPowerOfTwo(int x) {
    return 1 << IntMath.log2(x, RoundingMode.CEILING);
  }

  /*
   * This method was written by Doug Lea with assistance from members of JCP
   * JSR-166 Expert Group and released to the public domain, as explained at
   * http://creativecommons.org/licenses/publicdomain
   *
   * As of 2010/06/11, this method is identical to the (package private) hash
   * method in OpenJDK 7's java.util.HashMap class.
   */
  // Copied from java/com/google/common/collect/Hashing.java
  private static int smear(int hashCode) {
    hashCode ^= (hashCode >>> 20) ^ (hashCode >>> 12);
    return hashCode ^ (hashCode >>> 7) ^ (hashCode >>> 4);
  }

  private static class PaddedLock extends ReentrantLock {
    /*
     * Padding from 40 into 64 bytes, same size as cache line. Might be beneficial to add
     * a fourth long here, to minimize chance of interference between consecutive locks,
     * but I couldn't observe any benefit from that.
     */
    @SuppressWarnings("unused")
    long q1, q2, q3;

    PaddedLock() {
      super(false);
    }
  }

  private static class PaddedSemaphore extends Semaphore {
    // See PaddedReentrantLock comment
    @SuppressWarnings("unused")
    long q1, q2, q3;

    PaddedSemaphore(int permits) {
      super(permits, false);
    }
  }
}