/*
* Copyright 2022 The Netty Project
*
* The Netty Project 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:
*
* https://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 io.netty.buffer;
import io.netty.util.ByteProcessor;
import io.netty.util.IllegalReferenceCountException;
import io.netty.util.NettyRuntime;
import io.netty.util.Recycler.EnhancedHandle;
import io.netty.util.ReferenceCounted;
import io.netty.util.concurrent.FastThreadLocal;
import io.netty.util.concurrent.FastThreadLocalThread;
import io.netty.util.internal.ObjectPool;
import io.netty.util.internal.ObjectUtil;
import io.netty.util.internal.PlatformDependent;
import io.netty.util.internal.ReferenceCountUpdater;
import io.netty.util.internal.SuppressJava6Requirement;
import io.netty.util.internal.SystemPropertyUtil;
import io.netty.util.internal.ThreadExecutorMap;
import io.netty.util.internal.UnstableApi;
import java.io.IOException;
import java.io.InputStream;
import java.io.OutputStream;
import java.nio.ByteBuffer;
import java.nio.ByteOrder;
import java.nio.channels.ClosedChannelException;
import java.nio.channels.FileChannel;
import java.nio.channels.GatheringByteChannel;
import java.nio.channels.ScatteringByteChannel;
import java.util.Arrays;
import java.util.Queue;
import java.util.Set;
import java.util.concurrent.ConcurrentLinkedQueue;
import java.util.concurrent.CopyOnWriteArraySet;
import java.util.concurrent.atomic.AtomicIntegerFieldUpdater;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.atomic.AtomicReferenceFieldUpdater;
import java.util.concurrent.locks.StampedLock;
/**
* An auto-tuning pooling allocator, that follows an anti-generational hypothesis.
*
* The allocator is organized into a list of Magazines, and each magazine has a chunk-buffer that they allocate buffers
* from.
*
* The magazines hold the mutexes that ensure the thread-safety of the allocator, and each thread picks a magazine
* based on the id of the thread. This spreads the contention of multi-threaded access across the magazines.
* If contention is detected above a certain threshold, the number of magazines are increased in response to the
* contention.
*
* The magazines maintain histograms of the sizes of the allocations they do. The histograms are used to compute the
* preferred chunk size. The preferred chunk size is one that is big enough to service 10 allocations of the
* 99-percentile size. This way, the chunk size is adapted to the allocation patterns.
*
* Computing the preferred chunk size is a somewhat expensive operation. Therefore, the frequency with which this is
* done, is also adapted to the allocation pattern. If a newly computed preferred chunk is the same as the previous
* preferred chunk size, then the frequency is reduced. Otherwise, the frequency is increased.
*
* This allows the allocator to quickly respond to changes in the application workload,
* without suffering undue overhead from maintaining its statistics.
*
* Since magazines are "relatively thread-local", the allocator has a central queue that allow excess chunks from any
* magazine, to be shared with other magazines.
* The {@link #createSharedChunkQueue()} method can be overridden to customize this queue.
*/
@SuppressJava6Requirement(reason = "Guarded by version check")
@UnstableApi
final class AdaptivePoolingAllocator implements AdaptiveByteBufAllocator.AdaptiveAllocatorApi {
enum MagazineCaching {
EventLoopThreads,
FastThreadLocalThreads,
None
}
private static final int EXPANSION_ATTEMPTS = 3;
private static final int INITIAL_MAGAZINES = 4;
private static final int RETIRE_CAPACITY = 4 * 1024;
private static final int MIN_CHUNK_SIZE = 128 * 1024;
private static final int MAX_STRIPES = NettyRuntime.availableProcessors() * 2;
private static final int BUFS_PER_CHUNK = 10; // For large buffers, aim to have about this many buffers per chunk.
/**
* The maximum size of a pooled chunk, in bytes. Allocations bigger than this will never be pooled.
*
* This number is 10 MiB, and is derived from the limitations of internal histograms.
*/
private static final int MAX_CHUNK_SIZE =
BUFS_PER_CHUNK * (1 << AllocationStatistics.HISTO_MAX_BUCKET_SHIFT); // 10 MiB.
/**
* The capacity if the central queue that allow chunks to be shared across magazines.
* The default size is {@link NettyRuntime#availableProcessors()},
* and the maximum number of magazines is twice this.
*
* This means the maximum amount of memory that we can have allocated-but-not-in-use is
* 5 * {@link NettyRuntime#availableProcessors()} * {@link #MAX_CHUNK_SIZE} bytes.
*/
private static final int CENTRAL_QUEUE_CAPACITY = SystemPropertyUtil.getInt(
"io.netty.allocator.centralQueueCapacity", NettyRuntime.availableProcessors());
/**
* The capacity if the magazine local buffer queue. This queue just pools the outer ByteBuf instance and not
* the actual memory and so helps to reduce GC pressure.
*/
private static final int MAGAZINE_BUFFER_QUEUE_CAPACITY = SystemPropertyUtil.getInt(
"io.netty.allocator.magazineBufferQueueCapacity", 1024);
private static final Object NO_MAGAZINE = Boolean.TRUE;
private final ChunkAllocator chunkAllocator;
private final Queue centralQueue;
private final StampedLock magazineExpandLock;
private volatile Magazine[] magazines;
private final FastThreadLocal threadLocalMagazine;
private final Set liveCachedMagazines;
private volatile boolean freed;
AdaptivePoolingAllocator(ChunkAllocator chunkAllocator, MagazineCaching magazineCaching) {
ObjectUtil.checkNotNull(chunkAllocator, "chunkAllocator");
ObjectUtil.checkNotNull(magazineCaching, "magazineCaching");
this.chunkAllocator = chunkAllocator;
centralQueue = ObjectUtil.checkNotNull(createSharedChunkQueue(), "centralQueue");
magazineExpandLock = new StampedLock();
if (magazineCaching != MagazineCaching.None) {
assert magazineCaching == MagazineCaching.EventLoopThreads ||
magazineCaching == MagazineCaching.FastThreadLocalThreads;
final boolean cachedMagazinesNonEventLoopThreads =
magazineCaching == MagazineCaching.FastThreadLocalThreads;
final Set liveMagazines = new CopyOnWriteArraySet();
threadLocalMagazine = new FastThreadLocal() {
@Override
protected Object initialValue() {
if (cachedMagazinesNonEventLoopThreads || ThreadExecutorMap.currentExecutor() != null) {
Magazine mag = new Magazine(AdaptivePoolingAllocator.this, false);
if (FastThreadLocalThread.willCleanupFastThreadLocals(Thread.currentThread())) {
// Only add it to the liveMagazines if we can guarantee that onRemoval(...) is called,
// as otherwise we might end up holding the reference forever.
liveMagazines.add(mag);
}
return mag;
}
return NO_MAGAZINE;
}
@Override
protected void onRemoval(final Object value) throws Exception {
if (value != NO_MAGAZINE) {
liveMagazines.remove(value);
}
}
};
liveCachedMagazines = liveMagazines;
} else {
threadLocalMagazine = null;
liveCachedMagazines = null;
}
Magazine[] mags = new Magazine[INITIAL_MAGAZINES];
for (int i = 0; i < mags.length; i++) {
mags[i] = new Magazine(this);
}
magazines = mags;
}
/**
* Create a thread-safe multi-producer, multi-consumer queue to hold chunks that spill over from the
* internal Magazines.
*
* Each Magazine can only hold two chunks at any one time: the chunk it currently allocates from,
* and the next-in-line chunk which will be used for allocation once the current one has been used up.
* This queue will be used by magazines to share any excess chunks they allocate, so that they don't need to
* allocate new chunks when their current and next-in-line chunks have both been used up.
*
* The simplest implementation of this method is to return a new {@link ConcurrentLinkedQueue}.
* However, the {@code CLQ} is unbounded, and this means there's no limit to how many chunks can be cached in this
* queue.
*
* Each chunk in this queue can be up to {@link #MAX_CHUNK_SIZE} in size, so it is recommended to use a bounded
* queue to limit the maximum memory usage.
*
* The default implementation will create a bounded queue with a capacity of {@link #CENTRAL_QUEUE_CAPACITY}.
*
* @return A new multi-producer, multi-consumer queue.
*/
private static Queue createSharedChunkQueue() {
return PlatformDependent.newFixedMpmcQueue(CENTRAL_QUEUE_CAPACITY);
}
@Override
public ByteBuf allocate(int size, int maxCapacity) {
return allocate(size, maxCapacity, Thread.currentThread(), null);
}
private AdaptiveByteBuf allocate(int size, int maxCapacity, Thread currentThread, AdaptiveByteBuf buf) {
if (size <= MAX_CHUNK_SIZE) {
int sizeBucket = AllocationStatistics.sizeBucket(size); // Compute outside of Magazine lock for better ILP.
FastThreadLocal threadLocalMagazine = this.threadLocalMagazine;
if (threadLocalMagazine != null && currentThread instanceof FastThreadLocalThread) {
Object mag = threadLocalMagazine.get();
if (mag != NO_MAGAZINE) {
Magazine magazine = (Magazine) mag;
if (buf == null) {
buf = magazine.newBuffer();
}
boolean allocated = magazine.tryAllocate(size, sizeBucket, maxCapacity, buf);
assert allocated : "Allocation of threadLocalMagazine must always succeed";
return buf;
}
}
long threadId = currentThread.getId();
Magazine[] mags;
int expansions = 0;
do {
mags = magazines;
int mask = mags.length - 1;
int index = (int) (threadId & mask);
for (int i = 0, m = Integer.numberOfTrailingZeros(~mask); i < m; i++) {
Magazine mag = mags[index + i & mask];
if (buf == null) {
buf = mag.newBuffer();
}
if (mag.tryAllocate(size, sizeBucket, maxCapacity, buf)) {
// Was able to allocate.
return buf;
}
}
expansions++;
} while (expansions <= EXPANSION_ATTEMPTS && tryExpandMagazines(mags.length));
}
// The magazines failed us, or the buffer is too big to be pooled.
return allocateFallback(size, maxCapacity, currentThread, buf);
}
private AdaptiveByteBuf allocateFallback(int size, int maxCapacity, Thread currentThread, AdaptiveByteBuf buf) {
// If we don't already have a buffer, obtain one from the most conveniently available magazine.
Magazine magazine;
if (buf != null) {
Chunk chunk = buf.chunk;
magazine = chunk != null && chunk != Magazine.MAGAZINE_FREED?
chunk.magazine : getFallbackMagazine(currentThread);
} else {
magazine = getFallbackMagazine(currentThread);
buf = magazine.newBuffer();
}
// Create a one-off chunk for this allocation.
AbstractByteBuf innerChunk = chunkAllocator.allocate(size, maxCapacity);
Chunk chunk = new Chunk(innerChunk, magazine, false);
try {
chunk.readInitInto(buf, size, maxCapacity);
} finally {
// As the chunk is an one-off we need to always call release explicitly as readInitInto(...)
// will take care of retain once when successful. Once The AdaptiveByteBuf is released it will
// completely release the Chunk and so the contained innerChunk.
chunk.release();
}
return buf;
}
private Magazine getFallbackMagazine(Thread currentThread) {
Object tlMag;
FastThreadLocal threadLocalMagazine = this.threadLocalMagazine;
if (threadLocalMagazine != null &&
currentThread instanceof FastThreadLocalThread &&
(tlMag = threadLocalMagazine.get()) != NO_MAGAZINE) {
return (Magazine) tlMag;
}
Magazine[] mags = magazines;
return mags[(int) currentThread.getId() & mags.length - 1];
}
/**
* Allocate into the given buffer. Used by {@link AdaptiveByteBuf#capacity(int)}.
*/
void allocate(int size, int maxCapacity, AdaptiveByteBuf into) {
AdaptiveByteBuf result = allocate(size, maxCapacity, Thread.currentThread(), into);
assert result == into: "Re-allocation created separate buffer instance";
}
@Override
public long usedMemory() {
long sum = 0;
for (Chunk chunk : centralQueue) {
sum += chunk.capacity();
}
for (Magazine magazine : magazines) {
sum += magazine.usedMemory.get();
}
if (liveCachedMagazines != null) {
for (Magazine magazine : liveCachedMagazines) {
sum += magazine.usedMemory.get();
}
}
return sum;
}
private boolean tryExpandMagazines(int currentLength) {
if (currentLength >= MAX_STRIPES) {
return true;
}
long writeLock = magazineExpandLock.tryWriteLock();
if (writeLock != 0) {
try {
Magazine[] mags = magazines;
if (mags.length >= MAX_STRIPES || mags.length > currentLength || freed) {
return true;
}
int preferredChunkSize = mags[0].sharedPrefChunkSize;
Magazine[] expanded = new Magazine[mags.length * 2];
for (int i = 0, l = expanded.length; i < l; i++) {
Magazine m = new Magazine(this);
m.localPrefChunkSize = preferredChunkSize;
m.sharedPrefChunkSize = preferredChunkSize;
expanded[i] = m;
}
magazines = expanded;
for (Magazine magazine : mags) {
magazine.free();
}
} finally {
magazineExpandLock.unlockWrite(writeLock);
}
}
return true;
}
private boolean offerToQueue(Chunk buffer) {
if (freed) {
return false;
}
return centralQueue.offer(buffer);
}
// Ensure that we release all previous pooled resources when this object is finalized. This is needed as otherwise
// we might end up with leaks. While these leaks are usually harmless in reality it would still at least be
// very confusing for users.
@Override
protected void finalize() throws Throwable {
try {
super.finalize();
} finally {
free();
}
}
private void free() {
freed = true;
long stamp = magazineExpandLock.writeLock();
try {
Magazine[] mags = magazines;
for (Magazine magazine : mags) {
magazine.free();
}
} finally {
magazineExpandLock.unlockWrite(stamp);
}
for (;;) {
Chunk chunk = centralQueue.poll();
if (chunk == null) {
break;
}
chunk.release();
}
}
static int sizeBucket(int size) {
return AllocationStatistics.sizeBucket(size);
}
@SuppressWarnings("checkstyle:finalclass") // Checkstyle mistakenly believes this class should be final.
private static class AllocationStatistics {
private static final int MIN_DATUM_TARGET = 1024;
private static final int MAX_DATUM_TARGET = 65534;
private static final int INIT_DATUM_TARGET = 9;
private static final int HISTO_MIN_BUCKET_SHIFT = 13; // Smallest bucket is 1 << 13 = 8192 bytes in size.
private static final int HISTO_MAX_BUCKET_SHIFT = 20; // Biggest bucket is 1 << 20 = 1 MiB bytes in size.
private static final int HISTO_BUCKET_COUNT = 1 + HISTO_MAX_BUCKET_SHIFT - HISTO_MIN_BUCKET_SHIFT; // 8 buckets.
private static final int HISTO_MAX_BUCKET_MASK = HISTO_BUCKET_COUNT - 1;
private static final int SIZE_MAX_MASK = MAX_CHUNK_SIZE - 1;
protected final AdaptivePoolingAllocator parent;
private final boolean shareable;
private final short[][] histos = {
new short[HISTO_BUCKET_COUNT], new short[HISTO_BUCKET_COUNT],
new short[HISTO_BUCKET_COUNT], new short[HISTO_BUCKET_COUNT],
};
private short[] histo = histos[0];
private final int[] sums = new int[HISTO_BUCKET_COUNT];
private int histoIndex;
private int datumCount;
private int datumTarget = INIT_DATUM_TARGET;
protected volatile int sharedPrefChunkSize = MIN_CHUNK_SIZE;
protected volatile int localPrefChunkSize = MIN_CHUNK_SIZE;
private AllocationStatistics(AdaptivePoolingAllocator parent, boolean shareable) {
this.parent = parent;
this.shareable = shareable;
}
protected void recordAllocationSize(int bucket) {
histo[bucket]++;
if (datumCount++ == datumTarget) {
rotateHistograms();
}
}
static int sizeBucket(int size) {
if (size == 0) {
return 0;
}
// Minimum chunk size is 128 KiB. We'll only make bigger chunks if the 99-percentile is 16 KiB or greater,
// so we truncate and roll up the bottom part of the histogram to 8 KiB.
// The upper size band is 1 MiB, and that gives us exactly 8 size buckets,
// which is a magical number for JIT optimisations.
int normalizedSize = size - 1 >> HISTO_MIN_BUCKET_SHIFT & SIZE_MAX_MASK;
return Math.min(Integer.SIZE - Integer.numberOfLeadingZeros(normalizedSize), HISTO_MAX_BUCKET_MASK);
}
private void rotateHistograms() {
short[][] hs = histos;
for (int i = 0; i < HISTO_BUCKET_COUNT; i++) {
sums[i] = (hs[0][i] & 0xFFFF) + (hs[1][i] & 0xFFFF) + (hs[2][i] & 0xFFFF) + (hs[3][i] & 0xFFFF);
}
int sum = 0;
for (int count : sums) {
sum += count;
}
int targetPercentile = (int) (sum * 0.99);
int sizeBucket = 0;
for (; sizeBucket < sums.length; sizeBucket++) {
if (sums[sizeBucket] > targetPercentile) {
break;
}
targetPercentile -= sums[sizeBucket];
}
int percentileSize = 1 << sizeBucket + HISTO_MIN_BUCKET_SHIFT;
int prefChunkSize = Math.max(percentileSize * BUFS_PER_CHUNK, MIN_CHUNK_SIZE);
localPrefChunkSize = prefChunkSize;
if (shareable) {
for (Magazine mag : parent.magazines) {
prefChunkSize = Math.max(prefChunkSize, mag.localPrefChunkSize);
}
}
if (sharedPrefChunkSize != prefChunkSize) {
// Preferred chunk size changed. Increase check frequency.
datumTarget = Math.max(datumTarget >> 1, MIN_DATUM_TARGET);
sharedPrefChunkSize = prefChunkSize;
} else {
// Preferred chunk size did not change. Check less often.
datumTarget = Math.min(datumTarget << 1, MAX_DATUM_TARGET);
}
histoIndex = histoIndex + 1 & 3;
histo = histos[histoIndex];
datumCount = 0;
Arrays.fill(histo, (short) 0);
}
/**
* Get the preferred chunk size, based on statistics from the {@linkplain #recordAllocationSize(int) recorded}
* allocation sizes.
*
* This method must be thread-safe.
*
* @return The currently preferred chunk allocation size.
*/
protected int preferredChunkSize() {
return sharedPrefChunkSize;
}
}
@SuppressJava6Requirement(reason = "Guarded by version check")
private static final class Magazine extends AllocationStatistics {
private static final AtomicReferenceFieldUpdater NEXT_IN_LINE;
static {
NEXT_IN_LINE = AtomicReferenceFieldUpdater.newUpdater(Magazine.class, Chunk.class, "nextInLine");
}
private static final Chunk MAGAZINE_FREED = new Chunk();
private static final ObjectPool EVENT_LOOP_LOCAL_BUFFER_POOL = ObjectPool.newPool(
new ObjectPool.ObjectCreator() {
@Override
public AdaptiveByteBuf newObject(ObjectPool.Handle handle) {
return new AdaptiveByteBuf(handle);
}
});
private Chunk current;
@SuppressWarnings("unused") // updated via NEXT_IN_LINE
private volatile Chunk nextInLine;
private final AtomicLong usedMemory;
private final StampedLock allocationLock;
private final Queue bufferQueue;
private final ObjectPool.Handle handle;
Magazine(AdaptivePoolingAllocator parent) {
this(parent, true);
}
Magazine(AdaptivePoolingAllocator parent, boolean shareable) {
super(parent, shareable);
if (shareable) {
// We only need the StampedLock if this Magazine will be shared across threads.
allocationLock = new StampedLock();
bufferQueue = PlatformDependent.newFixedMpmcQueue(MAGAZINE_BUFFER_QUEUE_CAPACITY);
handle = new ObjectPool.Handle() {
@Override
public void recycle(AdaptiveByteBuf self) {
bufferQueue.offer(self);
}
};
} else {
allocationLock = null;
bufferQueue = null;
handle = null;
}
usedMemory = new AtomicLong();
}
public boolean tryAllocate(int size, int sizeBucket, int maxCapacity, AdaptiveByteBuf buf) {
if (allocationLock == null) {
// This magazine is not shared across threads, just allocate directly.
return allocate(size, sizeBucket, maxCapacity, buf);
}
// Try to retrieve the lock and if successful allocate.
long writeLock = allocationLock.tryWriteLock();
if (writeLock != 0) {
try {
return allocate(size, sizeBucket, maxCapacity, buf);
} finally {
allocationLock.unlockWrite(writeLock);
}
}
return false;
}
private boolean allocate(int size, int sizeBucket, int maxCapacity, AdaptiveByteBuf buf) {
recordAllocationSize(sizeBucket);
Chunk curr = current;
if (curr != null) {
if (curr.remainingCapacity() > size) {
curr.readInitInto(buf, size, maxCapacity);
// We still have some bytes left that we can use for the next allocation, just early return.
return true;
}
// At this point we know that this will be the last time current will be used, so directly set it to
// null and release it once we are done.
current = null;
if (curr.remainingCapacity() == size) {
try {
curr.readInitInto(buf, size, maxCapacity);
return true;
} finally {
curr.release();
}
}
}
Chunk last = curr;
assert current == null;
// The fast-path for allocations did not work.
//
// Try to fetch the next "Magazine local" Chunk first, if this this fails as we don't have
// one setup we will poll our centralQueue. If this fails as well we will just allocate a new Chunk.
//
// In any case we will store the Chunk as the current so it will be used again for the next allocation and
// so be "reserved" by this Magazine for exclusive usage.
if (nextInLine != null) {
curr = NEXT_IN_LINE.getAndSet(this, null);
if (curr == MAGAZINE_FREED) {
// Allocation raced with a stripe-resize that freed this magazine.
restoreMagazineFreed();
return false;
}
} else {
curr = parent.centralQueue.poll();
if (curr == null) {
curr = newChunkAllocation(size);
}
}
current = curr;
assert current != null;
if (last != null) {
if (last.remainingCapacity() < RETIRE_CAPACITY) {
last.release();
} else {
transferChunk(last);
}
}
if (curr.remainingCapacity() > size) {
curr.readInitInto(buf, size, maxCapacity);
} else if (curr.remainingCapacity() == size) {
try {
curr.readInitInto(buf, size, maxCapacity);
} finally {
// Release in a finally block so even if readInitInto(...) would throw we would still correctly
// release the current chunk before null it out.
curr.release();
current = null;
}
} else {
Chunk newChunk = newChunkAllocation(size);
try {
newChunk.readInitInto(buf, size, maxCapacity);
if (curr.remainingCapacity() < RETIRE_CAPACITY) {
curr.release();
current = newChunk;
} else {
transferChunk(newChunk);
}
newChunk = null;
} finally {
if (newChunk != null) {
assert current == null;
// Something went wrong, let's ensure we not leak the newChunk.
newChunk.release();
}
}
}
return true;
}
private void restoreMagazineFreed() {
Chunk next = NEXT_IN_LINE.getAndSet(this, MAGAZINE_FREED);
if (next != null && next != MAGAZINE_FREED) {
next.release(); // A chunk snuck in through a race. Release it after restoring MAGAZINE_FREED state.
}
}
private void transferChunk(Chunk current) {
if (NEXT_IN_LINE.compareAndSet(this, null, current)
|| parent.offerToQueue(current)) {
return;
}
Chunk nextChunk = NEXT_IN_LINE.get(this);
if (nextChunk != null && current.remainingCapacity() > nextChunk.remainingCapacity()) {
if (NEXT_IN_LINE.compareAndSet(this, nextChunk, current)) {
if (nextChunk != MAGAZINE_FREED) {
nextChunk.release();
}
return;
}
}
// Next-in-line is occupied AND the central queue is full.
// Rare that we should get here, but we'll only do one allocation out of this chunk, then.
current.release();
}
private Chunk newChunkAllocation(int promptingSize) {
int size = Math.max(promptingSize * BUFS_PER_CHUNK, preferredChunkSize());
ChunkAllocator chunkAllocator = parent.chunkAllocator;
return new Chunk(chunkAllocator.allocate(size, size), this, true);
}
boolean trySetNextInLine(Chunk chunk) {
return NEXT_IN_LINE.compareAndSet(this, null, chunk);
}
void free() {
// Release the current Chunk and the next that was stored for later usage.
restoreMagazineFreed();
long stamp = allocationLock.writeLock();
try {
if (current != null) {
current.release();
current = null;
}
} finally {
allocationLock.unlockWrite(stamp);
}
}
public AdaptiveByteBuf newBuffer() {
AdaptiveByteBuf buf;
if (handle == null) {
buf = EVENT_LOOP_LOCAL_BUFFER_POOL.get();
} else {
buf = bufferQueue.poll();
if (buf == null) {
buf = new AdaptiveByteBuf(handle);
}
}
buf.resetRefCnt();
buf.discardMarks();
return buf;
}
}
private static final class Chunk implements ReferenceCounted {
private final AbstractByteBuf delegate;
final Magazine magazine;
private final int capacity;
private final boolean pooled;
private int allocatedBytes;
private static final long REFCNT_FIELD_OFFSET =
ReferenceCountUpdater.getUnsafeOffset(Chunk.class, "refCnt");
private static final AtomicIntegerFieldUpdater AIF_UPDATER =
AtomicIntegerFieldUpdater.newUpdater(Chunk.class, "refCnt");
private static final ReferenceCountUpdater updater =
new ReferenceCountUpdater() {
@Override
protected AtomicIntegerFieldUpdater updater() {
return AIF_UPDATER;
}
@Override
protected long unsafeOffset() {
return REFCNT_FIELD_OFFSET;
}
};
// Value might not equal "real" reference count, all access should be via the updater
@SuppressWarnings({"unused", "FieldMayBeFinal"})
private volatile int refCnt;
Chunk() {
// Constructor only used by the MAGAZINE_FREED sentinel.
delegate = null;
magazine = null;
capacity = 0;
pooled = false;
}
Chunk(AbstractByteBuf delegate, Magazine magazine, boolean pooled) {
this.delegate = delegate;
this.magazine = magazine;
this.pooled = pooled;
capacity = delegate.capacity();
magazine.usedMemory.getAndAdd(capacity);
updater.setInitialValue(this);
}
@Override
public Chunk touch(Object hint) {
return this;
}
@Override
public int refCnt() {
return updater.refCnt(this);
}
@Override
public Chunk retain() {
return updater.retain(this);
}
@Override
public Chunk retain(int increment) {
return updater.retain(this, increment);
}
@Override
public Chunk touch() {
return this;
}
@Override
public boolean release() {
if (updater.release(this)) {
deallocate();
return true;
}
return false;
}
@Override
public boolean release(int decrement) {
if (updater.release(this, decrement)) {
deallocate();
return true;
}
return false;
}
private void deallocate() {
Magazine mag = magazine;
AdaptivePoolingAllocator parent = mag.parent;
int chunkSize = mag.preferredChunkSize();
int memSize = delegate.capacity();
if (!pooled || memSize < chunkSize || memSize > chunkSize + (chunkSize >> 1)) {
// Drop the chunk if the parent allocator is closed, or if the chunk is smaller than the
// preferred chunk size, or over 50% larger than the preferred chunk size.
mag.usedMemory.getAndAdd(-capacity());
delegate.release();
} else {
updater.resetRefCnt(this);
delegate.setIndex(0, 0);
allocatedBytes = 0;
if (!mag.trySetNextInLine(this)) {
if (!parent.offerToQueue(this)) {
// The central queue is full. Ensure we release again as we previously did use resetRefCnt()
// which did increase the reference count by 1.
boolean released = updater.release(this);
delegate.release();
assert released;
}
}
}
}
public void readInitInto(AdaptiveByteBuf buf, int size, int maxCapacity) {
int startIndex = allocatedBytes;
allocatedBytes = startIndex + size;
Chunk chunk = this;
chunk.retain();
try {
buf.init(delegate, chunk, 0, 0, startIndex, size, maxCapacity);
chunk = null;
} finally {
if (chunk != null) {
// If chunk is not null we know that buf.init(...) failed and so we need to manually release
// the chunk again as we retained it before calling buf.init(...).
chunk.release();
}
}
}
public int remainingCapacity() {
return capacity - allocatedBytes;
}
public int capacity() {
return capacity;
}
}
static final class AdaptiveByteBuf extends AbstractReferenceCountedByteBuf {
private final ObjectPool.Handle handle;
private int adjustment;
private AbstractByteBuf rootParent;
Chunk chunk;
private int length;
private ByteBuffer tmpNioBuf;
private boolean hasArray;
private boolean hasMemoryAddress;
AdaptiveByteBuf(ObjectPool.Handle recyclerHandle) {
super(0);
handle = ObjectUtil.checkNotNull(recyclerHandle, "recyclerHandle");
}
void init(AbstractByteBuf unwrapped, Chunk wrapped, int readerIndex, int writerIndex,
int adjustment, int capacity, int maxCapacity) {
this.adjustment = adjustment;
chunk = wrapped;
length = capacity;
maxCapacity(maxCapacity);
setIndex0(readerIndex, writerIndex);
hasArray = unwrapped.hasArray();
hasMemoryAddress = unwrapped.hasMemoryAddress();
rootParent = unwrapped;
tmpNioBuf = unwrapped.internalNioBuffer(adjustment, capacity).slice();
}
private AbstractByteBuf rootParent() {
final AbstractByteBuf rootParent = this.rootParent;
if (rootParent != null) {
return rootParent;
}
throw new IllegalReferenceCountException();
}
@Override
public int capacity() {
return length;
}
@Override
public ByteBuf capacity(int newCapacity) {
if (newCapacity == capacity()) {
ensureAccessible();
return this;
}
checkNewCapacity(newCapacity);
if (newCapacity < capacity()) {
length = newCapacity;
setIndex0(Math.min(readerIndex(), newCapacity), Math.min(writerIndex(), newCapacity));
return this;
}
// Reallocation required.
ByteBuffer data = tmpNioBuf;
data.clear();
tmpNioBuf = null;
Chunk chunk = this.chunk;
Magazine magazine = chunk.magazine;
AdaptivePoolingAllocator allocator = magazine.parent;
int readerIndex = this.readerIndex;
int writerIndex = this.writerIndex;
allocator.allocate(newCapacity, maxCapacity(), this);
tmpNioBuf.put(data);
tmpNioBuf.clear();
chunk.release();
this.readerIndex = readerIndex;
this.writerIndex = writerIndex;
return this;
}
@Override
public ByteBufAllocator alloc() {
return rootParent().alloc();
}
@Override
public ByteOrder order() {
return rootParent().order();
}
@Override
public ByteBuf unwrap() {
return null;
}
@Override
public boolean isDirect() {
return rootParent().isDirect();
}
@Override
public int arrayOffset() {
return idx(rootParent().arrayOffset());
}
@Override
public boolean hasMemoryAddress() {
return hasMemoryAddress;
}
@Override
public long memoryAddress() {
ensureAccessible();
return rootParent().memoryAddress() + adjustment;
}
@Override
public ByteBuffer nioBuffer(int index, int length) {
checkIndex(index, length);
return rootParent().nioBuffer(idx(index), length);
}
@Override
public ByteBuffer internalNioBuffer(int index, int length) {
checkIndex(index, length);
return (ByteBuffer) internalNioBuffer().position(index).limit(index + length);
}
private ByteBuffer internalNioBuffer() {
return (ByteBuffer) tmpNioBuf.clear();
}
@Override
public ByteBuffer[] nioBuffers(int index, int length) {
checkIndex(index, length);
return rootParent().nioBuffers(idx(index), length);
}
@Override
public boolean hasArray() {
return hasArray;
}
@Override
public byte[] array() {
ensureAccessible();
return rootParent().array();
}
@Override
public ByteBuf copy(int index, int length) {
checkIndex(index, length);
return rootParent().copy(idx(index), length);
}
@Override
public int nioBufferCount() {
return rootParent().nioBufferCount();
}
@Override
protected byte _getByte(int index) {
return rootParent()._getByte(idx(index));
}
@Override
protected short _getShort(int index) {
return rootParent()._getShort(idx(index));
}
@Override
protected short _getShortLE(int index) {
return rootParent()._getShortLE(idx(index));
}
@Override
protected int _getUnsignedMedium(int index) {
return rootParent()._getUnsignedMedium(idx(index));
}
@Override
protected int _getUnsignedMediumLE(int index) {
return rootParent()._getUnsignedMediumLE(idx(index));
}
@Override
protected int _getInt(int index) {
return rootParent()._getInt(idx(index));
}
@Override
protected int _getIntLE(int index) {
return rootParent()._getIntLE(idx(index));
}
@Override
protected long _getLong(int index) {
return rootParent()._getLong(idx(index));
}
@Override
protected long _getLongLE(int index) {
return rootParent()._getLongLE(idx(index));
}
@Override
public ByteBuf getBytes(int index, ByteBuf dst, int dstIndex, int length) {
checkIndex(index, length);
rootParent().getBytes(idx(index), dst, dstIndex, length);
return this;
}
@Override
public ByteBuf getBytes(int index, byte[] dst, int dstIndex, int length) {
checkIndex(index, length);
rootParent().getBytes(idx(index), dst, dstIndex, length);
return this;
}
@Override
public ByteBuf getBytes(int index, ByteBuffer dst) {
checkIndex(index, dst.remaining());
rootParent().getBytes(idx(index), dst);
return this;
}
@Override
protected void _setByte(int index, int value) {
rootParent()._setByte(idx(index), value);
}
@Override
protected void _setShort(int index, int value) {
rootParent()._setShort(idx(index), value);
}
@Override
protected void _setShortLE(int index, int value) {
rootParent()._setShortLE(idx(index), value);
}
@Override
protected void _setMedium(int index, int value) {
rootParent()._setMedium(idx(index), value);
}
@Override
protected void _setMediumLE(int index, int value) {
rootParent()._setMediumLE(idx(index), value);
}
@Override
protected void _setInt(int index, int value) {
rootParent()._setInt(idx(index), value);
}
@Override
protected void _setIntLE(int index, int value) {
rootParent()._setIntLE(idx(index), value);
}
@Override
protected void _setLong(int index, long value) {
rootParent()._setLong(idx(index), value);
}
@Override
protected void _setLongLE(int index, long value) {
rootParent().setLongLE(idx(index), value);
}
@Override
public ByteBuf setBytes(int index, byte[] src, int srcIndex, int length) {
checkIndex(index, length);
rootParent().setBytes(idx(index), src, srcIndex, length);
return this;
}
@Override
public ByteBuf setBytes(int index, ByteBuf src, int srcIndex, int length) {
checkIndex(index, length);
rootParent().setBytes(idx(index), src, srcIndex, length);
return this;
}
@Override
public ByteBuf setBytes(int index, ByteBuffer src) {
checkIndex(index, src.remaining());
rootParent().setBytes(idx(index), src);
return this;
}
@Override
public ByteBuf getBytes(int index, OutputStream out, int length)
throws IOException {
checkIndex(index, length);
if (length != 0) {
ByteBufUtil.readBytes(alloc(), internalNioBuffer().duplicate(), index, length, out);
}
return this;
}
@Override
public int getBytes(int index, GatheringByteChannel out, int length)
throws IOException {
return out.write(internalNioBuffer(index, length).duplicate());
}
@Override
public int getBytes(int index, FileChannel out, long position, int length)
throws IOException {
return out.write(internalNioBuffer(index, length).duplicate(), position);
}
@Override
public int setBytes(int index, InputStream in, int length)
throws IOException {
checkIndex(index, length);
final AbstractByteBuf rootParent = rootParent();
if (rootParent.hasArray()) {
return rootParent.setBytes(idx(index), in, length);
}
byte[] tmp = ByteBufUtil.threadLocalTempArray(length);
int readBytes = in.read(tmp, 0, length);
if (readBytes <= 0) {
return readBytes;
}
setBytes(index, tmp, 0, readBytes);
return readBytes;
}
@Override
public int setBytes(int index, ScatteringByteChannel in, int length)
throws IOException {
try {
return in.read(internalNioBuffer(index, length).duplicate());
} catch (ClosedChannelException ignored) {
return -1;
}
}
@Override
public int setBytes(int index, FileChannel in, long position, int length)
throws IOException {
try {
return in.read(internalNioBuffer(index, length).duplicate(), position);
} catch (ClosedChannelException ignored) {
return -1;
}
}
@Override
public int forEachByte(int index, int length, ByteProcessor processor) {
checkIndex(index, length);
int ret = rootParent().forEachByte(idx(index), length, processor);
return forEachResult(ret);
}
@Override
public int forEachByteDesc(int index, int length, ByteProcessor processor) {
checkIndex(index, length);
int ret = rootParent().forEachByteDesc(idx(index), length, processor);
return forEachResult(ret);
}
private int forEachResult(int ret) {
if (ret < adjustment) {
return -1;
}
return ret - adjustment;
}
@Override
public boolean isContiguous() {
return rootParent().isContiguous();
}
private int idx(int index) {
return index + adjustment;
}
@Override
protected void deallocate() {
if (chunk != null) {
chunk.release();
}
tmpNioBuf = null;
chunk = null;
rootParent = null;
if (handle instanceof EnhancedHandle) {
EnhancedHandle enhancedHandle = (EnhancedHandle) handle;
enhancedHandle.unguardedRecycle(this);
} else {
handle.recycle(this);
}
}
}
/**
* The strategy for how {@link AdaptivePoolingAllocator} should allocate chunk buffers.
*/
interface ChunkAllocator {
/**
* Allocate a buffer for a chunk. This can be any kind of {@link AbstractByteBuf} implementation.
* @param initialCapacity The initial capacity of the returned {@link AbstractByteBuf}.
* @param maxCapacity The maximum capacity of the returned {@link AbstractByteBuf}.
* @return The buffer that represents the chunk memory.
*/
AbstractByteBuf allocate(int initialCapacity, int maxCapacity);
}
}
