PooledByteBufAllocator

newDirectBuffer

protected ByteBuf newDirectBuffer(int initialCapacity, int maxCapacity) {
    // Obtain each thread's own cache from ThreadLocal.
    PoolThreadCache cache = threadCache.get();
    PoolArena<ByteBuffer> directArena = cache.directArena;

    final ByteBuf buf;
    if (directArena != null) {
        // Allocate memory.
        buf = directArena.allocate(cache, initialCapacity, maxCapacity);
    } else {
        buf = PlatformDependent.hasUnsafe() ?
                UnsafeByteBufUtil.newUnsafeDirectByteBuf(this, initialCapacity, maxCapacity) :
                new UnpooledDirectByteBuf(this, initialCapacity, maxCapacity);
    }

    return toLeakAwareBuffer(buf);
}

There are mainly two steps. First obtain PoolThreadCache, then obtain PoolArena from it. Second, allocate memory through PoolArena.

Obtain PoolThreadCache Through PoolThreadLocalCache

PoolThreadLocalCache is actually a FastThreadLocal. It inherits Netty’s own FastThreadLocal implementation, which is a kind of ThreadLocal.

Netty’s own ThreadLocal class is faster than Java’s.

final class PoolThreadLocalCache extends FastThreadLocal<PoolThreadCache> {
    private final boolean useCacheForAllThreads;

    PoolThreadLocalCache(boolean useCacheForAllThreads) {
        this.useCacheForAllThreads = useCacheForAllThreads;
    }

    private <T> PoolArena<T> leastUsedArena(PoolArena<T>[] arenas) {
        if (arenas == null || arenas.length == 0) {
            return null;
        }

        PoolArena<T> minArena = arenas[0];
        for (int i = 1; i < arenas.length; i++) {
            PoolArena<T> arena = arenas[i];
            if (arena.numThreadCaches.get() < minArena.numThreadCaches.get()) {
                minArena = arena;
            }
        }

        return minArena;
    }
}

initialValue

@Override
protected synchronized PoolThreadCache initialValue() {
    final PoolArena<byte[]> heapArena = leastUsedArena(heapArenas);
    final PoolArena<ByteBuffer> directArena = leastUsedArena(directArenas);

    Thread current = Thread.currentThread();
    if (useCacheForAllThreads || current instanceof FastThreadLocalThread) {
        return new PoolThreadCache(
                heapArena, directArena, tinyCacheSize, smallCacheSize, normalCacheSize,
                DEFAULT_MAX_CACHED_BUFFER_CAPACITY, DEFAULT_CACHE_TRIM_INTERVAL);
    }
    // No caching so just use 0 as sizes.
    return new PoolThreadCache(heapArena, directArena, 0, 0, 0, 0, 0);
}

You can see that there are two PoolArena instances: one for heap memory and one for direct memory. Finally a PoolThreadCache is constructed.

We can look at the sizes of heapArenas and directArenas.

// Twice the number of CPU cores.
final int defaultMinNumArena = NettyRuntime.availableProcessors() * 2;
final int defaultChunkSize = DEFAULT_PAGE_SIZE << DEFAULT_MAX_ORDER;
DEFAULT_NUM_HEAP_ARENA = Math.max(0,
        SystemPropertyUtil.getInt(
                "io.netty.allocator.numHeapArenas",
                (int) Math.min(
                        defaultMinNumArena,
                        runtime.maxMemory() / defaultChunkSize / 2 / 3)));
DEFAULT_NUM_DIRECT_ARENA = Math.max(0,
        SystemPropertyUtil.getInt(
                "io.netty.allocator.numDirectArenas",
                (int) Math.min(
                        defaultMinNumArena,
                        PlatformDependent.maxDirectMemory() / defaultChunkSize / 2 / 3)));

Why is it twice the CPU count? Because the number of NIO threads created earlier is also twice the CPU count, so each thread can have its own arena.

Create PoolThreadCache

Allocate Memory in PoolThreadLocalCache