/**
Copyright (C) SYSTAP, LLC DBA Blazegraph 2006-2016. All rights reserved.
Contact:
SYSTAP, LLC DBA Blazegraph
2501 Calvert ST NW #106
Washington, DC 20008
licenses@blazegraph.com
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
package com.bigdata.io.writecache;
import java.io.IOException;
import java.nio.ByteBuffer;
import java.nio.channels.FileChannel;
import java.util.concurrent.atomic.AtomicReference;
import com.bigdata.counters.CAT;
import com.bigdata.counters.CounterSet;
import com.bigdata.counters.Instrument;
import com.bigdata.io.DirectBufferPool;
import com.bigdata.io.FileChannelUtility;
import com.bigdata.io.IBufferAccess;
import com.bigdata.io.IReopenChannel;
import com.bigdata.rwstore.RWStore;
/**
* The BufferedWrite merges/elides sorted scattered writes to minimize IO
* requests and maximize IO rates. This has a net positive effect on SAS, SATA,
* and SSD.
*
* @author Martyn Cutcher
*
* @todo unit tests (this is used by RWStore and so is in general tested as part
* of that class, but it does not have its own test suite and it should not
* be all that difficult to write one, especially if we factor out an API
* for reporting the slotSize and then use a mock object in place of the
* RWStore).
*
* @todo harmonize with {@link CounterSet} for reporting purposes.
*/
public class BufferedWrite {
/**
* Used to determine the size of the allocation slot onto which a record is
* being written. This is used to pad the size of the IO out to the size of
* the slot. This can improve the IO efficiency When the slots are sized so
* as to fall on multiples of sector boundaries.
*/
private final IBufferedWriter m_store;
/**
* The direct {@link ByteBuffer} used to combine writes which are contiguous
* into a single IO.
*/
// private final ByteBuffer m_data;
private final AtomicReference<IBufferAccess> m_data = new AtomicReference<IBufferAccess>();
/**
* The offset on the backing channel at which the data in {@link #m_data}
* will be written when it is flushed to the backing channel. This is
* <code>-1</code> initially (and when reset) as a flag indicating that
* there is no data in {@link #m_data} and that the next record written by
* the caller on the buffer will assign the {@link #m_startAddr starting
* offset} of the data in the buffer.
* <p>
* Guarded by synchronized(this) (paranoia)
*/
private long m_startAddr = -1;
/**
* The offset of the backing channel at which the next byte would be written
* if it were appended to the data already present in {@link #m_data}.
* <p>
* Guarded by synchronized(this) (paranoia)
*/
private long m_endAddr = 0;
private final RWStore.StoreCounters<?> m_storeCounters;
public BufferedWrite(final IBufferedWriter store) throws InterruptedException {
if (store == null)
throw new IllegalArgumentException();
m_store = store;
m_storeCounters = m_store.getStoreCounters();
m_data.set( DirectBufferPool.INSTANCE.acquire() );
}
/**
* Release the direct buffer associated with this object.
*
* @throws InterruptedException
*/
// /*
// * Note: Consider adding synchronized(this) here to guard against the
// * possibility that the buffer could be released (and hence recycled) while
// * a write operation was occurring concurrently, However, this raises the
// * specter that a lock ordering problem could cause a deadlock.
// */
// synchronized
public void release() throws InterruptedException {
final IBufferAccess tmp = m_data.get();
if (tmp == null) {
// Already closed.
return;
}
if (m_data.compareAndSet(tmp/* expected */, null/* update */)) {
tmp.release();
}
}
/**
* Used to zero pad slots in buffered writes.
*
* @see <a href="http://trac.blazegraph.com/ticket/721#comment:10"> HA1 </a>
*/
static private final byte[] s_zeros = new byte[256];
/**
* Buffer a write.
*
* @param offset
* The offset on the backing channel at which the data should be
* written.
* @param data
* The data.
* @param opener
* The object which knows how to re-open the backing channel.
* @return The #of write IOs performed during this method call.
*
* @throws IOException
*/
synchronized
public int write(final long offset, final ByteBuffer data,
final IReopenChannel<FileChannel> opener) throws IOException {
m_storeCounters.bufferDataWrites++;
final int data_len = data.remaining();
final int slot_len = m_store.getSlotSize(data_len);
int nwrites = 0;
final ByteBuffer m_data = this.m_data.get().buffer();
if (slot_len > m_data.remaining()) {
/*
* There is not enough room in [m_data] to absorb the caller's data
* record, so we have to flush first.
*/
nwrites += flush(opener);
}
if (m_startAddr == -1) {
/*
* The buffer will begin to absorb data destined for the [offset]
* into the backing channel specified for the caller's data record.
*/
m_startAddr = m_endAddr = offset;
} else if (m_endAddr != offset) {
/*
* If this is NOT a contiguous write then flush existing content.
* After the flush, the buffer will begin to absorb data destined
* for the [offset] into the backing channel specified for the
* caller's data record.
*/
nwrites += flush(opener);
m_startAddr = m_endAddr = offset;
}
// copy the caller's record into the buffer.
m_data.put(data);
// if data_len < slot_len then clear remainder of buffer
int padding = slot_len - data_len;
while (padding > 0) {
if (padding > s_zeros.length) {
m_data.put(s_zeros);
padding -= s_zeros.length;
} else {
m_data.put(s_zeros, 0, padding);
break;
}
}
// update the file offset by the size of the allocation slot
m_endAddr += slot_len;
// update the buffer position by the size of the allocation slot.
final long pos = m_endAddr - m_startAddr;
m_data.position((int) pos);
return nwrites;
}
/**
* Flush buffered data to the backing channel.
*
* @param opener
* The object which knows how to re-open the backing channel.
*
* @return The #of write IOs performed during this method call.
*
* @throws IOException
*/
synchronized
public int flush(final IReopenChannel<FileChannel> opener)
throws IOException {
final ByteBuffer m_data = this.m_data.get().buffer();
if (m_data.position() == 0) {
// NOP.
return 0;
}
// increment by the amount of data currently in the buffer.
m_storeCounters.bufferDataBytes += m_data.position();
// write out the data in the buffer onto the backing channel.
m_data.flip();
final int nwrites = FileChannelUtility.writeAll(opener, m_data, m_startAddr);
m_storeCounters.bufferFileWrites += nwrites;
reset();
return nwrites;
}
/**
* Reset the buffer position and limit and clear the starting offset on the
* file to <code>-1</code>.
*/
synchronized
public void reset() {
final IBufferAccess tmp = m_data.get();
if (tmp == null) {
// Already closed.
return;
}
final ByteBuffer m_data = tmp.buffer();
// reset the buffer state.
//m_data.position(0);
//m_data.limit(m_data.capacity());
m_data.clear();
m_startAddr = -1;
m_endAddr = 0;
}
public String getStats(final StringBuffer buf, final boolean reset) {
final String ret = "BufferedWrites, data: " + m_storeCounters.bufferDataWrites + ", file: " + m_storeCounters.bufferFileWrites + ", bytes: " + m_storeCounters.bufferDataBytes;
if (buf != null) {
buf.append(ret + "\n");
}
if (reset) {
m_storeCounters.bufferFileWrites = 0;
m_storeCounters.bufferDataWrites = 0;
m_storeCounters.bufferDataBytes = 0;
}
return ret;
}
}