NanaChiyo0721/Shiroha
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Rewrite buffered linear master sync logic & sync locks

Browse Source 
Generated by ai(uhm idk which model(x) but the code is ok)
This commit is contained in:
NanaChiyo0721
2026-07-18 06:44:47 +08:00
parent 0cf1ddec5c
commit d5733936ac
1 changed files with 484 additions and 323 deletions
Download Patch File Download Diff File Expand all files Collapse all files
shiroha-server/src/main/java/io/nanachiyo0721/shiroha/data/BufferedLinearRegionFile.java
+429 -268
View File
@@ -29,15 +29,30 @@ import java.nio.file.Files;
import java.nio.file.Path; import java.nio.file.Path;
import java.nio.file.StandardCopyOption; import java.nio.file.StandardCopyOption;
import java.nio.file.StandardOpenOption; import java.nio.file.StandardOpenOption;
import java.util.Arrays;
import java.util.concurrent.atomic.AtomicLong; import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.locks.ReadWriteLock; import java.util.concurrent.locks.ReadWriteLock;
import java.util.concurrent.locks.ReentrantReadWriteLock; import java.util.concurrent.locks.ReentrantReadWriteLock;
/**
* Lock hierarchy (always acquire top to bottom, never the reverse):
* <ol>
* <li>{@code syncLock} — serializes master file syncs against close</li>
* <li>{@code Bucket.lock} — per-bucket lazy-load guard</li>
* <li>{@code masterFileLock} — master file read / append / replace</li>
* <li>{@code regionObjectLock} — in-memory sector table + swap file channel</li>
* </ol>
* The atomic flags (closed / synced / beingSynced / lastWritten) and the bucket
* epochs are lock-free and may be touched while holding any (or no) lock.
*/
public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.RegionFile { public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.RegionFile {
private static final double SWAP_FILE_AUTO_COMPACT_PERCENT = 3.0 / 5.0; // 60 % private static final double SWAP_FILE_AUTO_COMPACT_PERCENT = 3.0 / 5.0; // 60 %
private static final long SWAP_FILE_AUTO_COMPACT_SIZE = 1024 * 1024; // 1 MiB private static final long SWAP_FILE_AUTO_COMPACT_SIZE = 1024 * 1024; // 1 MiB
// master file WAL appends leave the replaced bucket records behind as garbage; once
// it piles up past this threshold the next sync compacts via a full tmp-file rewrite
private static final double MASTER_FILE_AUTO_COMPACT_PERCENT = SWAP_FILE_AUTO_COMPACT_PERCENT;
private static final long MASTER_FILE_AUTO_COMPACT_SIZE = SWAP_FILE_AUTO_COMPACT_SIZE;
private static final long SWAP_FILE_SUPER_BLOCK = 0x1145141919810L; private static final long SWAP_FILE_SUPER_BLOCK = 0x1145141919810L;
private static final int SWAP_FILE_HASH_SEED = 0x0721; // ～(∠・ω< )⌒★ private static final int SWAP_FILE_HASH_SEED = 0x0721; // ～(∠・ω< )⌒★
private static final byte SWAP_FILE_VERSION = 0x02; // ver 2.0 private static final byte SWAP_FILE_VERSION = 0x02; // ver 2.0
@@ -74,6 +89,10 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
private final Path masterFilePath; private final Path masterFilePath;
private final Path swapFilePath; private final Path swapFilePath;
// outermost lock: serializes syncToMasterFile() against closeInternal(), so the
// swap channel can never be torn down while a sync is still reading from it
private final Object syncLock = new Object();
private final ReadWriteLock regionObjectLock = new ReentrantReadWriteLock(); private final ReadWriteLock regionObjectLock = new ReentrantReadWriteLock();
private final XXHash32 xxHash32 = XXHashFactory.fastestInstance().hash32(); private final XXHash32 xxHash32 = XXHashFactory.fastestInstance().hash32();
private Sector[] sectors = new Sector[1024]; private Sector[] sectors = new Sector[1024];
@@ -134,6 +153,34 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
} }
private static void transferFully(FileChannel source, long sourceOffset, long count, FileChannel target, long targetOffset) throws IOException {
target.position(targetOffset);
long transferred = 0;
while (transferred < count) {
transferred += source.transferTo(sourceOffset + transferred, count - transferred, target);
}
}
// replaces target with source, deleting source if both attempts fail
private static void atomicReplace(Path source, Path target) throws IOException {
try {
Files.move(source, target, StandardCopyOption.REPLACE_EXISTING, StandardCopyOption.ATOMIC_MOVE);
} catch (Throwable e) {
// atomic move might be unsupported on some file systems, so give it an attempt to retry without atomic move
try {
Files.move(source, target, StandardCopyOption.REPLACE_EXISTING);
} catch (Throwable ex) {
e.addSuppressed(ex);
// delete file that failed to replace
Files.deleteIfExists(source);
throw new IOException("Failed to replace " + target + "!", e);
}
}
}
private void cleanUpSwapFile() throws IOException { private void cleanUpSwapFile() throws IOException {
Files.deleteIfExists(this.swapFilePath); Files.deleteIfExists(this.swapFilePath);
} }
@@ -142,6 +189,10 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
final int bucketIndex = chunkIndex >> BUCKET_SHIFT; final int bucketIndex = chunkIndex >> BUCKET_SHIFT;
final Bucket bucket = this.buckets[bucketIndex]; final Bucket bucket = this.buckets[bucketIndex];
if (bucket.loaded) { // volatile fast path
return;
}
// bucket lock -> master read lock -> swap write lock // bucket lock -> master read lock -> swap write lock
synchronized (bucket.lock) { synchronized (bucket.lock) {
if (bucket.loaded) { if (bucket.loaded) {
@@ -153,32 +204,26 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
} }
private long markBucketDirty(int chunkIndex) { // used by the legacy parsers: their data goes through the write path directly,
return this.markBucketDirtyByIndex(chunkIndex >> BUCKET_SHIFT); // so the bucket must be flagged loaded first to avoid a recursive lazy-load
private void markBucketLoaded(int chunkIndex) {
final Bucket bucket = this.buckets[chunkIndex >> BUCKET_SHIFT];
synchronized (bucket.lock) {
bucket.loaded = true;
}
} }
private long markBucketDirtyByIndex(int bucketIndex) { private void markBucketDirty(int chunkIndex) {
final Bucket bucket = this.buckets[bucketIndex]; this.buckets[chunkIndex >> BUCKET_SHIFT].writeEpoch.incrementAndGet();
return bucket.writeEpoch.incrementAndGet();
} }
private long getBucketWriteEpoch(int bucketIndex) { private long getBucketWriteEpoch(int bucketIndex) {
final Bucket bucket = this.buckets[bucketIndex]; return this.buckets[bucketIndex].writeEpoch.get();
return bucket.writeEpoch.get();
}
private long getBucketSyncedEpoch(int bucketIndex) {
final Bucket bucket = this.buckets[bucketIndex];
return bucket.syncedEpoch.get();
} }
private void markBucketSynced(int bucketIndex, long syncedEpoch) { private void markBucketSynced(int bucketIndex, long syncedEpoch) {
final Bucket bucket = this.buckets[bucketIndex]; this.buckets[bucketIndex].syncedEpoch.accumulateAndGet(syncedEpoch, Math::max);
bucket.syncedEpoch.accumulateAndGet(syncedEpoch, Math::max);
} }
private boolean isBucketDirty(int bucketIndex) { private boolean isBucketDirty(int bucketIndex) {
@@ -223,14 +268,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
public void syncIfNeeded() throws IOException { public void syncIfNeeded() throws IOException {
// the sync operation is just coping the data from swap file to the master file
// so we could acquire read lock simply so that we won't block any other read operations
try { try {
// skip if closed already
if (this.isClosed()) {
return;
}
this.syncToMasterFile(); this.syncToMasterFile();
} finally { } finally {
BEING_SYNCED_HANDLE.setVolatile(this, false); // mark as not being synced BEING_SYNCED_HANDLE.setVolatile(this, false); // mark as not being synced
@@ -238,14 +276,21 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
private void syncToMasterFile() throws IOException { private void syncToMasterFile() throws IOException {
// prevent multiple syncs in the same time // serialized against close: the swap channel cannot go away under a running sync
synchronized (this.syncLock) {
// skip if closed already
if (this.isClosedRaw()) {
return;
}
// fast skip when there is nothing to sync; writers flip the flag back
// via markAsToSync() which triggers the next round
if (!SYNCED_HANDLE.compareAndSet(this, false, true)) { if (!SYNCED_HANDLE.compareAndSet(this, false, true)) {
return; return;
} }
try { try {
// this.masterFileParser.writeMainFile(this.masterFilePath); this.masterFileParser.sync(this.masterFilePath);
this.masterFileParser.writeMainFileBucketed(this.masterFilePath);
} catch (Throwable e) { } catch (Throwable e) {
// set back // set back
SYNCED_HANDLE.setVolatile(this, false); SYNCED_HANDLE.setVolatile(this, false);
@@ -253,6 +298,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
throw new IOException("Failed to sync to master file!", e); throw new IOException("Failed to sync to master file!", e);
} }
} }
}
private void tryLoadOldBlinearMasterFileData() throws IOException { private void tryLoadOldBlinearMasterFileData() throws IOException {
this.masterFileParser.tryParseMainFileOld(this.masterFilePath); this.masterFileParser.tryParseMainFileOld(this.masterFilePath);
@@ -329,29 +375,20 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
return; return;
} }
long spareSize = this.currentAcquiredIndex; long liveSize = 0;
spareSize -= this.headerSize();
for (Sector sector : this.sectors) { for (Sector sector : this.sectors) {
// skip no data sectors // skip no data sectors
if (!sector.hasData()) { if (!sector.hasData()) {
continue; continue;
} }
spareSize -= sector.length; liveSize += sector.length;
} }
long sectorSize = 0; // everything acquired but not covered by a live sector is garbage
for (Sector sector : this.sectors) { final long spareSize = this.currentAcquiredIndex - this.headerSize() - liveSize;
// skip no data sectors
if (!sector.hasData()) {
continue;
}
sectorSize += sector.length; final boolean compactRequested = spareSize > SWAP_FILE_AUTO_COMPACT_SIZE && (double) spareSize > ((double) liveSize) * SWAP_FILE_AUTO_COMPACT_PERCENT;
}
final boolean compactRequested = spareSize > SWAP_FILE_AUTO_COMPACT_SIZE && (double) spareSize > ((double) sectorSize) * SWAP_FILE_AUTO_COMPACT_PERCENT;
// try auto compact to clean the garbage area // try auto compact to clean the garbage area
if (compactRequested) { if (compactRequested) {
@@ -371,25 +408,59 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
private void closeInternal() throws IOException { private void closeInternal() throws IOException {
this.syncIfNeeded(); synchronized (this.syncLock) {
if (this.isClosedRaw()) {
// already closed (possibly by a compact disaster path): just make sure
// both channels are really gone — close is idempotent
this.regionObjectLock.writeLock().lock();
try {
this.swapFileChannel.close();
} finally {
this.regionObjectLock.writeLock().unlock();
}
this.masterFileParser.close();
return;
}
// final sync so no buffered data is lost; holding syncLock also guarantees no
// concurrent flusher sync is still running when we tear down below.
// if this throws we deliberately stay open: the flusher can retry the sync
// later, and the not-yet-synced swap data is not dropped on the floor
this.syncToMasterFile();
IOException failure = null;
this.regionObjectLock.writeLock().lock(); this.regionObjectLock.writeLock().lock();
try { try {
this.markClosed(); this.markClosed();
this.swapFileChannel.close(); this.swapFileChannel.close();
} catch (IOException e) {
failure = e;
} finally { } finally {
this.regionObjectLock.writeLock().unlock(); this.regionObjectLock.writeLock().unlock();
} }
try {
// acquired after the region lock is fully released, never inside it (lock hierarchy)
this.masterFileParser.close();
} catch (IOException e) {
if (failure == null) failure = e; else failure.addSuppressed(e);
} }
private void markClosed() throws IOException { if (failure != null) {
if (!CLOSED_HANDLE.compareAndSet(this, false, true)) { throw failure;
throw new IOException("Already closed!"); }
}
} }
private void markClosed() {
// lenient CAS: the disaster path of compactSwapFile() may have closed us already
if (CLOSED_HANDLE.compareAndSet(this, false, true)) {
this.flusher.removeFile(this); this.flusher.removeFile(this);
} }
}
private void compactSwapFile() throws IOException { private void compactSwapFile() throws IOException {
this.writeSwapFileHeaders(true, true); // save headers for compact this.writeSwapFileHeaders(true, true); // save headers for compact
@@ -413,7 +484,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
long newAcquiredIndex; long newAcquiredIndex;
final Path targetTemp = new File(this.swapFilePath.toString() + ".tmp").toPath(); final Path targetTemp = Path.of(this.swapFilePath + ".tmp");
try (FileChannel tempChannel = FileChannel.open( try (FileChannel tempChannel = FileChannel.open(
targetTemp, targetTemp,
@@ -423,7 +494,6 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
StandardOpenOption.TRUNCATE_EXISTING StandardOpenOption.TRUNCATE_EXISTING
)) { )) {
long offsetPointer = this.headerSize(); long offsetPointer = this.headerSize();
tempChannel.position(offsetPointer);
for (Sector sector : newSectorsToBeReplaced) { for (Sector sector : newSectorsToBeReplaced) {
// skip cleared or no data-contained sectors // skip cleared or no data-contained sectors
@@ -432,7 +502,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
// transfer to target // transfer to target
sector.transferTo(this.swapFileChannel, tempChannel); transferFully(this.swapFileChannel, sector.offset, sector.length, tempChannel, offsetPointer);
// recalculate the offset and length // recalculate the offset and length
final Sector newRecalculated = new Sector(sector.index, offsetPointer, sector.length); final Sector newRecalculated = new Sector(sector.index, offsetPointer, sector.length);
@@ -459,26 +529,8 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
// replace swap file // replace swap file
try { try {
Files.move( atomicReplace(targetTemp, this.swapFilePath);
targetTemp,
this.swapFilePath,
StandardCopyOption.REPLACE_EXISTING,
StandardCopyOption.ATOMIC_MOVE
);
} catch (Throwable e) { } catch (Throwable e) {
// atomic move might be unsupported on some file systems, so give it an attempt to retry without atomic move
try {
Files.move(
targetTemp,
this.swapFilePath,
StandardCopyOption.REPLACE_EXISTING
);
} catch (Throwable ex) {
// now we are totally failed
e.addSuppressed(ex);
// delete file that failed to replace
Files.deleteIfExists(targetTemp);
// recalculate acquired index // recalculate acquired index
this.recalculateAcquiredIndex(); this.recalculateAcquiredIndex();
// reopen closed channel // reopen closed channel
@@ -487,8 +539,6 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
this.markClosed(); // prevent new writing & sync operations this.markClosed(); // prevent new writing & sync operations
throw new IOException("Failed to replace original swap file!", e); throw new IOException("Failed to replace original swap file!", e);
} }
}
try { try {
// reopen file channel // reopen file channel
@@ -545,10 +595,6 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
private @Nullable ByteBuffer readChunkDataRaw(int index) throws IOException { private @Nullable ByteBuffer readChunkDataRaw(int index) throws IOException {
return this.readChunkDataRaw(index, true);
}
private @Nullable ByteBuffer readChunkDataRaw(int index, boolean acquireLock) throws IOException {
final ByteBuffer raw; final ByteBuffer raw;
this.regionObjectLock.readLock().lock(); this.regionObjectLock.readLock().lock();
@@ -606,6 +652,8 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
private void writeChunk(int x, int z, @NotNull ByteBuffer data) throws IOException { private void writeChunk(int x, int z, @NotNull ByteBuffer data) throws IOException {
final int chunkIndex = getChunkIndex(x, z); final int chunkIndex = getChunkIndex(x, z);
this.ensureBucketLoaded(chunkIndex);
if (data.remaining() > MAX_SIZE_PER_CHUNK) { if (data.remaining() > MAX_SIZE_PER_CHUNK) {
throw new RegionFileStorage.RegionFileSizeException("Writing too large chunk, limit : " + MAX_SIZE_PER_CHUNK + " but got : " + data.remaining()); throw new RegionFileStorage.RegionFileSizeException("Writing too large chunk, limit : " + MAX_SIZE_PER_CHUNK + " but got : " + data.remaining());
} }
@@ -703,11 +751,6 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
@Override @Override
public void write(@NotNull ChunkPos pos, ByteBuffer buf) throws IOException { public void write(@NotNull ChunkPos pos, ByteBuffer buf) throws IOException {
final int chunkIndex = getChunkIndex(pos.x(), pos.z());
this.ensureBucketLoaded(chunkIndex);
this.writeChunk(pos.x(), pos.z(), buf); this.writeChunk(pos.x(), pos.z(), buf);
} }
@@ -819,16 +862,6 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
this.length = length; this.length = length;
} }
public void transferTo(@NotNull FileChannel source, @NotNull FileChannel target) throws IOException {
long transferred = 0;
while (transferred < this.length) {
transferred += source.transferTo(
this.offset + transferred,
this.length - transferred,
target);
}
}
public @NotNull ByteBuffer read(@NotNull FileChannel channel) throws IOException { public @NotNull ByteBuffer read(@NotNull FileChannel channel) throws IOException {
final ByteBuffer result = ByteBuffer.allocate((int) this.length); final ByteBuffer result = ByteBuffer.allocate((int) this.length);
@@ -906,9 +939,6 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
public void close() throws IOException { public void close() throws IOException {
ByteBuffer bytebuffer = ByteBuffer.wrap(this.buf, 0, this.count); ByteBuffer bytebuffer = ByteBuffer.wrap(this.buf, 0, this.count);
final int chunkIndex = getChunkIndex(this.pos.x(), this.pos.z());
BufferedLinearRegionFile.this.ensureBucketLoaded(chunkIndex);
BufferedLinearRegionFile.this.writeChunk(this.pos.x(), this.pos.z(), bytebuffer); BufferedLinearRegionFile.this.writeChunk(this.pos.x(), this.pos.z(), bytebuffer);
BufferedLinearRegionFile.this.flushInternal(); BufferedLinearRegionFile.this.flushInternal();
@@ -916,197 +946,341 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
} }
private class LinearMasterFileParser { private class LinearMasterFileParser {
// V3 new format layout: // V3 bucketed format layout:
// [0, 14): header — superblock(8) + version(1) + compressionLevel(1) + xxHash32Seed(4) // [0, 14): header — superblock(8) + version(1) + compressionLevel(1) + xxHash32Seed(4)
// [14, 142): position table — BUCKET_COUNT(16) × long(8) each; 0 = no data for that bucket // [14, 142): position table — BUCKET_COUNT(16) × long(8) each; 0 = no data for that bucket
// [526, EOF): bucket data — originalLen(int) + compressedLen(int) + compressedData // [142, EOF): bucket records — originalLen(int) + compressedLen(int) + compressedData
private static final long V3_POS_TABLE_OFFSET = 14L; private static final int V3_HEADER_SIZE = 14;
private static final long V3_POS_TABLE_OFFSET = V3_HEADER_SIZE;
private static final int V3_POS_TABLE_SIZE = BUCKET_COUNT * Long.BYTES; // 128 private static final int V3_POS_TABLE_SIZE = BUCKET_COUNT * Long.BYTES; // 128
private static final long V3_DATA_AREA_OFFSET = V3_POS_TABLE_OFFSET + V3_POS_TABLE_SIZE; // 142 private static final long V3_DATA_AREA_OFFSET = V3_POS_TABLE_OFFSET + V3_POS_TABLE_SIZE; // 142
private static final int V3_RECORD_HEADER_SIZE = Integer.BYTES * 2; // originalLen + compressedLen
private final ReadWriteLock masterFileLock = new ReentrantReadWriteLock(); private final ReadWriteLock masterFileLock = new ReentrantReadWriteLock();
public void writeMainFileBucketed(@NotNull Path mainFile) throws IOException { // WAL(append) state, guarded by masterFileLock: null until the first sync after
// open has fully rewritten the master file; afterwards syncs only append changed
// buckets to the tail and update the position table in place.
// recordSizes mirrors positionTable (size of each live record) so the garbage
// ratio can be computed without touching the disk
private @Nullable FileChannel appendChannel;
private long[] positionTable;
private long[] recordSizes;
private long appendOffset;
// a consistent snapshot of one bucket taken from the swap file;
// payload == null means the bucket holds no chunks at all
private record BucketRecord(long epoch, byte @Nullable [] payload) {
}
// must be called under syncLock (see syncToMasterFile)
public void sync(@NotNull Path mainFile) throws IOException {
this.masterFileLock.writeLock().lock();
try {
// full rewrite on the first sync after open, and afterwards whenever the
// appended garbage passed the auto-compact threshold: writes a tmp file,
// then atomically replaces the master file with it
if (this.appendChannel == null || this.shouldCompactMasterFile()) {
this.rewriteFully(mainFile);
} else {
// WAL-style otherwise: only append the dirty buckets
this.appendDirtyBuckets();
}
} finally {
this.masterFileLock.writeLock().unlock();
}
}
// only valid in WAL mode (appendChannel != null); mirrors the swap file heuristic
private boolean shouldCompactMasterFile() {
long liveSize = 0;
for (final long size : this.recordSizes) {
liveSize += size;
}
final long spareSize = this.appendOffset - V3_DATA_AREA_OFFSET - liveSize;
return spareSize > MASTER_FILE_AUTO_COMPACT_SIZE && (double) spareSize > ((double) liveSize) * MASTER_FILE_AUTO_COMPACT_PERCENT;
}
private void rewriteFully(@NotNull Path mainFile) throws IOException {
// a compacting rewrite replaces the whole file, so drop the old append channel
// first; if anything below fails, the next sync just retries via this path
if (this.appendChannel != null) {
this.appendChannel.close();
this.appendChannel = null;
}
final Path tmpFilePath = Path.of(mainFile + ".tmp"); final Path tmpFilePath = Path.of(mainFile + ".tmp");
final long[] syncedBucketEpochs = new long[BUCKET_COUNT]; final long[] syncedBucketEpochs = new long[BUCKET_COUNT];
final long[] newPositionTable = new long[BUCKET_COUNT]; final long[] newPositionTable = new long[BUCKET_COUNT];
final long[] newRecordSizes = new long[BUCKET_COUNT];
final long newAppendOffset;
// note: there is no necessary hold the write lock for this stuff // open the old file to copy the non-dirty buckets over
// as the truly write operations only happens on replacing the master file (see the file move call below this hunk) try (FileChannel oldChannel = this.openV3MasterFile(mainFile)) {
// and we had CAS flags to prevent multiple synchronization happening at the same time final long[] oldPositionTable = oldChannel == null ? null : this.parseOffsetTable(oldChannel);
// Open old file to copy non-dirty buckets
long[] oldPositionTable = null;
FileChannel oldChannel = null;
this.masterFileLock.writeLock().lock();
try {
if (Files.exists(mainFile)) {
try {
oldChannel = FileChannel.open(mainFile, StandardOpenOption.READ);
if (oldChannel.size() >= V3_DATA_AREA_OFFSET) {
final ByteBuffer hdr = ByteBuffer.allocate(14);
readFullyAt(oldChannel, hdr, 0);
hdr.flip();
if (hdr.getLong() == MASTER_FILE_SUPER_BLOCK && hdr.get() == MASTER_FILE_VERSION_BUCKET) {
oldPositionTable = this.parseOffsetTable(oldChannel);
} else {
oldChannel.close();
oldChannel = null;
}
} else {
oldChannel.close();
oldChannel = null;
}
} catch (Throwable e) {
if (oldChannel != null) {
try {
oldChannel.close();
} catch (IOException e2) {
e.addSuppressed(e2);
}
}
throw new RuntimeException(e);
}
}
BufferedLinearRegionFile.this.regionObjectLock.readLock().lock();
try (FileChannel outChannel = FileChannel.open(tmpFilePath, try (FileChannel outChannel = FileChannel.open(tmpFilePath,
StandardOpenOption.CREATE, StandardOpenOption.WRITE, StandardOpenOption.TRUNCATE_EXISTING)) { StandardOpenOption.CREATE, StandardOpenOption.WRITE, StandardOpenOption.TRUNCATE_EXISTING)) {
this.writeV3Header(outChannel);
// Write header (14 bytes) // position table placeholder (all zeros, filled in at the end)
final ByteBuffer header = ByteBuffer.allocate(14);
header.putLong(MASTER_FILE_SUPER_BLOCK);
header.put(MASTER_FILE_VERSION_BUCKET);
header.put(BufferedLinearRegionFile.this.compressionLevel);
header.putInt(BufferedLinearRegionFile.this.xxHash32Seed);
header.flip();
writeFullyAt(outChannel, header, 0);
// Write position table placeholder (all zeros, filled in at the end)
writeFullyAt(outChannel, ByteBuffer.allocate(V3_POS_TABLE_SIZE), V3_POS_TABLE_OFFSET); writeFullyAt(outChannel, ByteBuffer.allocate(V3_POS_TABLE_SIZE), V3_POS_TABLE_OFFSET);
long dataOffset = V3_DATA_AREA_OFFSET; long dataOffset = V3_DATA_AREA_OFFSET;
for (int bucketIndex = 0; bucketIndex < BUCKET_COUNT; bucketIndex++) { for (int bucketIndex = 0; bucketIndex < BUCKET_COUNT; bucketIndex++) {
final long bucketWriteEpoch = BufferedLinearRegionFile.this.getBucketWriteEpoch(bucketIndex); if (BufferedLinearRegionFile.this.isBucketDirty(bucketIndex)) {
final boolean isBucketDirty = bucketWriteEpoch != BufferedLinearRegionFile.this.getBucketSyncedEpoch(bucketIndex); final BucketRecord record = this.buildBucketRecord(bucketIndex);
if (record.payload() != null) {
writeFullyAt(outChannel, ByteBuffer.wrap(record.payload()), dataOffset);
newPositionTable[bucketIndex] = dataOffset;
newRecordSizes[bucketIndex] = record.payload().length;
dataOffset += record.payload().length;
}
// else: the bucket is empty now, its table entry stays 0
syncedBucketEpochs[bucketIndex] = record.epoch();
} else if (oldPositionTable != null && oldPositionTable[bucketIndex] != 0) {
// not dirty: copy the record bytes straight from the old file
final long oldOffset = oldPositionTable[bucketIndex];
final ByteBuffer lens = this.readRecordLengths(oldChannel, oldOffset);
lens.getInt(); // skip originalLen
final long recordSize = V3_RECORD_HEADER_SIZE + (long) lens.getInt();
transferFully(oldChannel, oldOffset, recordSize, outChannel, dataOffset);
newPositionTable[bucketIndex] = dataOffset;
newRecordSizes[bucketIndex] = recordSize;
dataOffset += recordSize;
}
}
// write the finalized position table
writeFullyAt(outChannel, this.encodePositionTable(newPositionTable), V3_POS_TABLE_OFFSET);
outChannel.force(true);
newAppendOffset = dataOffset;
}
}
atomicReplace(tmpFilePath, mainFile);
// enter WAL mode: keep the freshly written master file open for appending syncs
this.appendChannel = FileChannel.open(mainFile, StandardOpenOption.READ, StandardOpenOption.WRITE);
this.positionTable = newPositionTable;
this.recordSizes = newRecordSizes;
this.appendOffset = newAppendOffset;
this.markBucketsSynced(syncedBucketEpochs);
}
private void appendDirtyBuckets() throws IOException {
final FileChannel channel = this.appendChannel;
final long[] syncedBucketEpochs = new long[BUCKET_COUNT];
final long[] newPositionTable = this.positionTable.clone();
final long[] newRecordSizes = this.recordSizes.clone();
long dataOffset = this.appendOffset;
boolean anyDirty = false;
for (int bucketIndex = 0; bucketIndex < BUCKET_COUNT; bucketIndex++) {
if (!BufferedLinearRegionFile.this.isBucketDirty(bucketIndex)) {
continue;
}
final BucketRecord record = this.buildBucketRecord(bucketIndex);
if (record.payload() != null) {
writeFullyAt(channel, ByteBuffer.wrap(record.payload()), dataOffset);
newPositionTable[bucketIndex] = dataOffset;
newRecordSizes[bucketIndex] = record.payload().length;
dataOffset += record.payload().length;
} else {
// the bucket is empty now
newPositionTable[bucketIndex] = 0;
newRecordSizes[bucketIndex] = 0;
}
syncedBucketEpochs[bucketIndex] = record.epoch();
anyDirty = true;
}
if (!anyDirty) {
return;
}
// make the appended records durable before the position table may point at them
channel.force(false);
// commit the new tail first: even a torn position table write can then never
// cause a later append to overwrite records the on-disk table already references
this.appendOffset = dataOffset;
writeFullyAt(channel, this.encodePositionTable(newPositionTable), V3_POS_TABLE_OFFSET);
channel.force(true);
this.positionTable = newPositionTable;
this.recordSizes = newRecordSizes;
this.markBucketsSynced(syncedBucketEpochs);
}
// snapshots one bucket under a short read lock (raw sector bytes only);
// LZ4 decompression and zstd compression both run outside any lock so
// writers are only blocked while the raw bytes are copied
private @NotNull BucketRecord buildBucketRecord(int bucketIndex) throws IOException {
final int baseChunkIndex = bucketIndex << BUCKET_SHIFT;
final ByteBuffer[] rawSectors = new ByteBuffer[BUCKET_SIZE];
final long epoch;
BufferedLinearRegionFile.this.regionObjectLock.readLock().lock();
try {
// the epoch is taken before the data: writes completing afterwards bump
// it further, so they simply get picked up by the next sync round
epoch = BufferedLinearRegionFile.this.getBucketWriteEpoch(bucketIndex);
for (int i = 0; i < BUCKET_SIZE; i++) {
final Sector sector = BufferedLinearRegionFile.this.sectors[baseChunkIndex + i];
rawSectors[i] = sector.hasData() ? sector.read(BufferedLinearRegionFile.this.swapFileChannel) : null;
}
} finally {
BufferedLinearRegionFile.this.regionObjectLock.readLock().unlock();
}
if (isBucketDirty) {
final int baseChunk = bucketIndex << BUCKET_SHIFT;
final ByteArrayOutputStream rawBuf = new ByteArrayOutputStream(); final ByteArrayOutputStream rawBuf = new ByteArrayOutputStream();
final DataOutputStream rawOut = new DataOutputStream(rawBuf); final DataOutputStream rawOut = new DataOutputStream(rawBuf);
boolean hasAny = false; boolean hasAny = false;
for (int i = 0; i < BUCKET_SIZE; i++) { for (int i = 0; i < BUCKET_SIZE; i++) {
// swap read lock final ByteBuffer rawSector = rawSectors[i];
final ByteBuffer data = BufferedLinearRegionFile.this.readChunkDataRaw(baseChunk + i, false);
// note: null -> no data contained // note: null -> no data contained
if (data == null) { if (rawSector == null) {
rawOut.writeInt(0); rawOut.writeInt(0);
} else { continue;
final byte[] arr = new byte[data.remaining()]; }
data.get(arr);
final ByteBuffer chunkData = BufferedLinearRegionFile.this.compressingOps.decompress(rawSector);
final byte[] arr = new byte[chunkData.remaining()];
chunkData.get(arr);
rawOut.writeInt(arr.length); rawOut.writeInt(arr.length);
rawOut.write(arr); rawOut.write(arr);
hasAny = true; hasAny = true;
} }
}
rawOut.flush(); rawOut.flush();
if (hasAny) { if (!hasAny) {
return new BucketRecord(epoch, null);
}
final byte[] raw = rawBuf.toByteArray(); final byte[] raw = rawBuf.toByteArray();
final byte[] compressed = Zstd.compress(raw, BufferedLinearRegionFile.this.compressionLevel); final byte[] compressed = Zstd.compress(raw, BufferedLinearRegionFile.this.compressionLevel);
newPositionTable[bucketIndex] = dataOffset; final ByteBuffer payload = ByteBuffer.allocate(V3_RECORD_HEADER_SIZE + compressed.length);
payload.putInt(raw.length); // original (uncompressed) length
payload.putInt(compressed.length); // compressed length
payload.put(compressed);
final ByteBuffer bucketBuf = ByteBuffer.allocate(8 + compressed.length); return new BucketRecord(epoch, payload.array());
bucketBuf.putInt(raw.length); // original (uncompressed) length
bucketBuf.putInt(compressed.length); // compressed length
bucketBuf.put(compressed);
bucketBuf.flip();
writeFullyAt(outChannel, bucketBuf, dataOffset);
dataOffset += bucketBuf.limit();
} }
// else: newPositionTable[bucketIndex] stays 0
syncedBucketEpochs[bucketIndex] = bucketWriteEpoch; private void markBucketsSynced(long[] syncedBucketEpochs) {
} else { for (int i = 0; i < syncedBucketEpochs.length; i++) {
// Not dirty: copy bytes from old file if available // note: a dirty bucket always has a write epoch >= 1, so 0 = untouched
if (oldPositionTable != null && oldPositionTable[bucketIndex] != 0) { if (syncedBucketEpochs[i] != 0L) {
final long oldOffset = oldPositionTable[bucketIndex]; BufferedLinearRegionFile.this.markBucketSynced(i, syncedBucketEpochs[i]);
final ByteBuffer lensBuf = ByteBuffer.allocate(8);
readFullyAt(oldChannel, lensBuf, oldOffset);
lensBuf.flip();
lensBuf.getInt(); // skip originalLen
final int compressedLen = lensBuf.getInt();
final long bucketTotalSize = 8L + compressedLen;
newPositionTable[bucketIndex] = dataOffset;
outChannel.position(dataOffset);
long remaining = bucketTotalSize;
long srcPos = oldOffset;
while (remaining > 0) {
final long transferred = oldChannel.transferTo(srcPos, remaining, outChannel);
srcPos += transferred;
remaining -= transferred;
}
dataOffset += bucketTotalSize;
} }
} }
} }
// Write the finalized position table // opens the master file for reading if it exists and is a valid V3 bucketed file, else null
final ByteBuffer posTableBuf = ByteBuffer.allocate(V3_POS_TABLE_SIZE); private @Nullable FileChannel openV3MasterFile(@NotNull Path mainFile) throws IOException {
for (final long pos : newPositionTable) { if (!Files.exists(mainFile)) {
posTableBuf.putLong(pos); return null;
}
posTableBuf.flip();
writeFullyAt(outChannel, posTableBuf, V3_POS_TABLE_OFFSET);
outChannel.force(true);
} finally {
BufferedLinearRegionFile.this.regionObjectLock.readLock().unlock();
if (oldChannel != null) {
oldChannel.close();
}
} }
final FileChannel channel = FileChannel.open(mainFile, StandardOpenOption.READ);
try { try {
Files.move(tmpFilePath, mainFile, StandardCopyOption.REPLACE_EXISTING, StandardCopyOption.ATOMIC_MOVE); if (channel.size() >= V3_DATA_AREA_OFFSET) {
final ByteBuffer header = ByteBuffer.allocate(V3_HEADER_SIZE);
readFullyAt(channel, header, 0);
header.flip();
if (header.getLong() == MASTER_FILE_SUPER_BLOCK && header.get() == MASTER_FILE_VERSION_BUCKET) {
return channel;
}
}
} catch (Throwable e) { } catch (Throwable e) {
try { try {
Files.move(tmpFilePath, mainFile, StandardCopyOption.REPLACE_EXISTING); channel.close();
} catch (Throwable ex) { } catch (IOException e2) {
Files.deleteIfExists(tmpFilePath); e.addSuppressed(e2);
e.addSuppressed(ex);
throw new IOException("Failed to replace master file!", e);
} }
throw e;
}
channel.close();
return null;
}
private void writeV3Header(@NotNull FileChannel channel) throws IOException {
final ByteBuffer header = ByteBuffer.allocate(V3_HEADER_SIZE);
header.putLong(MASTER_FILE_SUPER_BLOCK);
header.put(MASTER_FILE_VERSION_BUCKET);
header.put(BufferedLinearRegionFile.this.compressionLevel);
header.putInt(BufferedLinearRegionFile.this.xxHash32Seed);
header.flip();
writeFullyAt(channel, header, 0);
}
private @NotNull ByteBuffer encodePositionTable(long[] table) {
final ByteBuffer buf = ByteBuffer.allocate(V3_POS_TABLE_SIZE);
for (final long pos : table) {
buf.putLong(pos);
}
return buf.flip();
}
private @NotNull ByteBuffer readRecordLengths(@NotNull FileChannel channel, long recordOffset) throws IOException {
final ByteBuffer lens = ByteBuffer.allocate(V3_RECORD_HEADER_SIZE);
readFullyAt(channel, lens, recordOffset);
return lens.flip();
}
public void close() throws IOException {
this.masterFileLock.writeLock().lock();
try {
if (this.appendChannel != null) {
this.appendChannel.close();
this.appendChannel = null;
} }
} finally { } finally {
this.masterFileLock.writeLock().unlock(); this.masterFileLock.writeLock().unlock();
} }
for (int i = 0; i < syncedBucketEpochs.length; i++) {
final long syncedEpoch = syncedBucketEpochs[i];
if (syncedEpoch != 0L) {
BufferedLinearRegionFile.this.markBucketSynced(i, syncedEpoch);
}
}
} }
private void loadBucketsFor(Path file, int bucketIndex) throws IOException { private void loadBucketsFor(@NotNull Path file, int bucketIndex) throws IOException {
final int beginChunkIndex = bucketIndex << BUCKET_SHIFT; final int beginChunkIndex = bucketIndex << BUCKET_SHIFT;
this.masterFileLock.readLock().lock(); this.masterFileLock.readLock().lock();
try { try {
final ByteBuffer decompressed;
if (this.appendChannel != null) {
// WAL mode: reuse the always-open channel and the cached position table
decompressed = this.readBucketData(this.appendChannel, this.positionTable[bucketIndex]);
} else {
if (!Files.exists(file)) { if (!Files.exists(file)) {
return; return;
} }
@@ -1116,7 +1290,22 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
return; return;
} }
final ByteBuffer headerBuf = ByteBuffer.allocate(14); this.checkV3Header(channel);
decompressed = this.readBucketData(channel, this.parseOffsetTable(channel)[bucketIndex]);
}
}
if (decompressed != null) {
this.loadChunksFromBucketData(decompressed, beginChunkIndex);
}
} finally {
this.masterFileLock.readLock().unlock();
}
}
private void checkV3Header(@NotNull FileChannel channel) throws IOException {
final ByteBuffer headerBuf = ByteBuffer.allocate(V3_HEADER_SIZE);
readFullyAt(channel, headerBuf, 0); readFullyAt(channel, headerBuf, 0);
headerBuf.flip(); headerBuf.flip();
@@ -1128,31 +1317,23 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
if (version != MASTER_FILE_VERSION_BUCKET) if (version != MASTER_FILE_VERSION_BUCKET)
throw new IOException("Unknown version: " + version); throw new IOException("Unknown version: " + version);
// compressionLevel and hashSeed consumed but not used here // compressionLevel and hashSeed are not used here
headerBuf.get(); }
headerBuf.getInt();
final long[] posTable = this.parseOffsetTable(channel); // reads and decompresses one bucket record; null when the table entry is empty
private @Nullable ByteBuffer readBucketData(@NotNull FileChannel channel, long recordOffset) throws IOException {
if (recordOffset == 0) {
return null;
}
// New format: jump directly to bucket data final ByteBuffer lens = this.readRecordLengths(channel, recordOffset);
final long bucketDataOffset = posTable[bucketIndex]; final int originalLen = lens.getInt();
if (bucketDataOffset == 0) return; final int compressedLen = lens.getInt();
final ByteBuffer lensBuf = ByteBuffer.allocate(8);
readFullyAt(channel, lensBuf, bucketDataOffset);
lensBuf.flip();
final int originalLen = lensBuf.getInt();
final int compressedLen = lensBuf.getInt();
final byte[] compressedData = new byte[compressedLen]; final byte[] compressedData = new byte[compressedLen];
readFullyAt(channel, ByteBuffer.wrap(compressedData), bucketDataOffset + 8); readFullyAt(channel, ByteBuffer.wrap(compressedData), recordOffset + V3_RECORD_HEADER_SIZE);
final ByteBuffer decompressed = ByteBuffer.wrap(Zstd.decompress(compressedData, originalLen)); return ByteBuffer.wrap(Zstd.decompress(compressedData, originalLen));
this.loadChunksFromBucketData(decompressed, beginChunkIndex);
}
} finally {
this.masterFileLock.readLock().unlock();
}
} }
private long @NonNull [] parseOffsetTable(FileChannel channel) throws IOException { private long @NonNull [] parseOffsetTable(FileChannel channel) throws IOException {
@@ -1162,10 +1343,8 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
final long[] table = new long[BUCKET_COUNT]; final long[] table = new long[BUCKET_COUNT];
Arrays.fill(table, 0L);
for (int i = 0; i < BUCKET_COUNT; i++) { for (int i = 0; i < BUCKET_COUNT; i++) {
final long pos = buf.getLong(); table[i] = buf.getLong();
table[i] = pos;
} }
return table; return table;
@@ -1253,12 +1432,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
bucketBuffer.get(chunkData); bucketBuffer.get(chunkData);
// Mark bucket as loaded. writeChunk() bumps the bucket epoch so it gets synced to the new master format. // Mark bucket as loaded. writeChunk() bumps the bucket epoch so it gets synced to the new master format.
final int blinearBucketIndex = chunkIndex >> BUCKET_SHIFT; BufferedLinearRegionFile.this.markBucketLoaded(chunkIndex);
final Bucket bucket = BufferedLinearRegionFile.this.buckets[blinearBucketIndex];
synchronized (bucket.lock) {
bucket.loaded = true;
}
// Use writeChunk to go through the full path (adds length + timestamp + xxhash header) // Use writeChunk to go through the full path (adds length + timestamp + xxhash header)
BufferedLinearRegionFile.this.writeChunk(chunkX, chunkZ, ByteBuffer.wrap(chunkData)); BufferedLinearRegionFile.this.writeChunk(chunkX, chunkZ, ByteBuffer.wrap(chunkData));
@@ -1305,13 +1479,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
final ByteBuffer sectorDataNioBuffer = ByteBuffer.wrap(sectorData); final ByteBuffer sectorDataNioBuffer = ByteBuffer.wrap(sectorData);
final int bucketIndex = index >> BUCKET_SHIFT; BufferedLinearRegionFile.this.markBucketLoaded(index);
final Bucket bucket = BufferedLinearRegionFile.this.buckets[bucketIndex];
synchronized (bucket.lock) {
bucket.loaded = true;
}
BufferedLinearRegionFile.this.writeChunkDataRaw(index, sectorDataNioBuffer, false); BufferedLinearRegionFile.this.writeChunkDataRaw(index, sectorDataNioBuffer, false);
} }
} }
@@ -1359,14 +1527,7 @@ public class BufferedLinearRegionFile implements io.nanachiyo0721.shiroha.data.R
final int x = posByAxis[0]; final int x = posByAxis[0];
final int z = posByAxis[1]; final int z = posByAxis[1];
BufferedLinearRegionFile.this.markBucketLoaded(i);
final int bucketIndex = i >> BUCKET_SHIFT;
final Bucket bucket = BufferedLinearRegionFile.this.buckets[bucketIndex];
synchronized (bucket.lock) {
bucket.loaded = true;
}
BufferedLinearRegionFile.this.writeChunk(x, z, chunkDataNioBuffer); BufferedLinearRegionFile.this.writeChunk(x, z, chunkDataNioBuffer);
} }
} }