/**
* Prepare a transaction.
*
* @param session the session
* @param transaction the name of the transaction
*/
void prepareCommit(Session session, String transaction) {
if (trace.isDebugEnabled()) {
trace.debug("log prepare commit s: " + session.getId() + ", " + transaction);
}
if (store.getDatabase().getPageStore() == null) {
// database already closed
return;
}
// store it on a separate log page
int pageSize = store.getPageSize();
pageOut.flush();
pageOut.fillPage();
Data buffer = getBuffer();
buffer.writeByte((byte) PREPARE_COMMIT);
buffer.writeVarInt(session.getId());
buffer.writeString(transaction);
if (buffer.length() >= PageStreamData.getCapacity(pageSize)) {
throw DbException.getInvalidValueException("transaction name (too long)", transaction);
}
write(buffer);
// store it on a separate log page
flushOut();
pageOut.fillPage();
if (store.getDatabase().getFlushOnEachCommit()) {
flush();
}
}
/**
* A table is truncated.
*
* @param session the session
* @param tableId the table id
*/
void logTruncate(Session session, int tableId) {
if (trace.isDebugEnabled()) {
trace.debug("log truncate s: " + session.getId() + " table: " + tableId);
}
session.addLogPos(logSectionId, logPos);
logPos++;
Data buffer = getBuffer();
buffer.writeByte((byte) TRUNCATE);
buffer.writeVarInt(session.getId());
buffer.writeVarInt(tableId);
write(buffer);
}
private void freeLogPages(IntArray pages) {
if (trace.isDebugEnabled()) {
trace.debug("log frees " + pages.get(0) + ".." + pages.get(pages.size() - 1));
}
Data buffer = getBuffer();
buffer.writeByte((byte) FREE_LOG);
int size = pages.size();
buffer.writeVarInt(size);
for (int i = 0; i < size; i++) {
buffer.writeVarInt(pages.get(i));
}
write(buffer);
}
/**
* Mark a transaction as committed.
*
* @param sessionId the session
*/
void commit(int sessionId) {
if (trace.isDebugEnabled()) {
trace.debug("log commit s: " + sessionId);
}
if (store.getDatabase().getPageStore() == null) {
// database already closed
return;
}
Data buffer = getBuffer();
buffer.writeByte((byte) COMMIT);
buffer.writeVarInt(sessionId);
write(buffer);
if (store.getDatabase().getFlushOnEachCommit()) {
flush();
}
}
/**
* A record is added to a table, or removed from a table.
*
* @param session the session
* @param tableId the table id
* @param row the row to add
* @param add true if the row is added, false if it is removed
*/
void logAddOrRemoveRow(Session session, int tableId, Row row, boolean add) {
if (trace.isDebugEnabled()) {
trace.debug(
"log "
+ (add ? "+" : "-")
+ " s: "
+ session.getId()
+ " table: "
+ tableId
+ " row: "
+ row);
}
session.addLogPos(logSectionId, logPos);
logPos++;
Data data = dataBuffer;
data.reset();
int columns = row.getColumnCount();
data.writeVarInt(columns);
data.checkCapacity(row.getByteCount(data));
if (session.isRedoLogBinaryEnabled()) {
for (int i = 0; i < columns; i++) {
data.writeValue(row.getValue(i));
}
} else {
for (int i = 0; i < columns; i++) {
Value v = row.getValue(i);
if (v.getType() == Value.BYTES) {
data.writeValue(ValueNull.INSTANCE);
} else {
data.writeValue(v);
}
}
}
Data buffer = getBuffer();
buffer.writeByte((byte) (add ? ADD : REMOVE));
buffer.writeVarInt(session.getId());
buffer.writeVarInt(tableId);
buffer.writeVarLong(row.getKey());
if (add) {
buffer.writeVarInt(data.length());
buffer.checkCapacity(data.length());
buffer.write(data.getBytes(), 0, data.length());
}
write(buffer);
}
/**
* Add an undo entry to the log. The page data is only written once until the next checkpoint.
*
* @param pageId the page id
* @param page the old page data
*/
void addUndo(int pageId, Data page) {
if (undo.get(pageId) || freeing) {
return;
}
if (trace.isDebugEnabled()) {
trace.debug("log undo " + pageId);
}
if (SysProperties.CHECK) {
if (page == null) {
DbException.throwInternalError("Undo entry not written");
}
}
undo.set(pageId);
undoAll.set(pageId);
Data buffer = getBuffer();
buffer.writeByte((byte) UNDO);
buffer.writeVarInt(pageId);
if (page.getBytes()[0] == 0) {
buffer.writeVarInt(1);
} else {
int pageSize = store.getPageSize();
if (COMPRESS_UNDO) {
int size = compress.compress(page.getBytes(), pageSize, compressBuffer, 0);
if (size < pageSize) {
buffer.writeVarInt(size);
buffer.checkCapacity(size);
buffer.write(compressBuffer, 0, size);
} else {
buffer.writeVarInt(0);
buffer.checkCapacity(pageSize);
buffer.write(page.getBytes(), 0, pageSize);
}
} else {
buffer.writeVarInt(0);
buffer.checkCapacity(pageSize);
buffer.write(page.getBytes(), 0, pageSize);
}
}
write(buffer);
}
/** Free up all pages allocated by the log. */
void free() {
if (trace.isDebugEnabled()) {
trace.debug("log free");
}
int currentDataPage = 0;
if (pageOut != null) {
currentDataPage = pageOut.getCurrentDataPageId();
pageOut.freeReserved();
}
try {
freeing = true;
int first = 0;
int loopDetect = 1024, loopCount = 0;
PageStreamTrunk.Iterator it = new PageStreamTrunk.Iterator(store, firstTrunkPage);
while (firstTrunkPage != 0 && firstTrunkPage < store.getPageCount()) {
PageStreamTrunk t = it.next();
if (t == null) {
if (it.canDelete()) {
store.free(firstTrunkPage, false);
}
break;
}
if (loopCount++ >= loopDetect) {
first = t.getPos();
loopCount = 0;
loopDetect *= 2;
} else if (first != 0 && first == t.getPos()) {
throw DbException.throwInternalError("endless loop at " + t);
}
t.free(currentDataPage);
firstTrunkPage = t.getNextTrunk();
}
} finally {
freeing = false;
}
}
/**
* Run one recovery stage. There are three recovery stages: 0: only the undo steps are run
* (restoring the state before the last checkpoint). 1: the pages that are used by the transaction
* log are allocated. 2: the committed operations are re-applied.
*
* @param stage the recovery stage
* @return whether the transaction log was empty
*/
boolean recover(int stage) {
if (trace.isDebugEnabled()) {
trace.debug("log recover stage: " + stage);
}
if (stage == RECOVERY_STAGE_ALLOCATE) {
PageInputStream in = new PageInputStream(store, logKey, firstTrunkPage, firstDataPage);
usedLogPages = in.allocateAllPages();
in.close();
return true;
}
PageInputStream pageIn = new PageInputStream(store, logKey, firstTrunkPage, firstDataPage);
DataReader in = new DataReader(pageIn);
int logId = 0;
Data data = store.createData();
boolean isEmpty = true;
try {
int pos = 0;
while (true) {
int x = in.readByte();
if (x < 0) {
break;
}
pos++;
isEmpty = false;
if (x == UNDO) {
int pageId = in.readVarInt();
int size = in.readVarInt();
if (size == 0) {
in.readFully(data.getBytes(), store.getPageSize());
} else if (size == 1) {
// empty
Arrays.fill(data.getBytes(), 0, store.getPageSize(), (byte) 0);
} else {
in.readFully(compressBuffer, size);
try {
compress.expand(compressBuffer, 0, size, data.getBytes(), 0, store.getPageSize());
} catch (ArrayIndexOutOfBoundsException e) {
DbException.convertToIOException(e);
}
}
if (stage == RECOVERY_STAGE_UNDO) {
if (!undo.get(pageId)) {
if (trace.isDebugEnabled()) {
trace.debug("log undo {0}", pageId);
}
store.writePage(pageId, data);
undo.set(pageId);
undoAll.set(pageId);
} else {
if (trace.isDebugEnabled()) {
trace.debug("log undo skip {0}", pageId);
}
}
}
} else if (x == ADD) {
int sessionId = in.readVarInt();
int tableId = in.readVarInt();
Row row = readRow(in, data);
if (stage == RECOVERY_STAGE_UNDO) {
store.allocateIfIndexRoot(pos, tableId, row);
} else if (stage == RECOVERY_STAGE_REDO) {
if (isSessionCommitted(sessionId, logId, pos)) {
if (trace.isDebugEnabled()) {
trace.debug("log redo + table: " + tableId + " s: " + sessionId + " " + row);
}
store.redo(tableId, row, true);
} else {
if (trace.isDebugEnabled()) {
trace.debug("log ignore s: " + sessionId + " + table: " + tableId + " " + row);
}
}
}
} else if (x == REMOVE) {
int sessionId = in.readVarInt();
int tableId = in.readVarInt();
long key = in.readVarLong();
if (stage == RECOVERY_STAGE_REDO) {
if (isSessionCommitted(sessionId, logId, pos)) {
if (trace.isDebugEnabled()) {
trace.debug("log redo - table: " + tableId + " s:" + sessionId + " key: " + key);
}
store.redoDelete(tableId, key);
} else {
if (trace.isDebugEnabled()) {
trace.debug("log ignore s: " + sessionId + " - table: " + tableId + " " + key);
}
}
}
} else if (x == TRUNCATE) {
int sessionId = in.readVarInt();
int tableId = in.readVarInt();
if (stage == RECOVERY_STAGE_REDO) {
if (isSessionCommitted(sessionId, logId, pos)) {
if (trace.isDebugEnabled()) {
trace.debug("log redo truncate table: " + tableId);
}
store.redoTruncate(tableId);
} else {
if (trace.isDebugEnabled()) {
trace.debug("log ignore s: " + sessionId + " truncate table: " + tableId);
}
}
}
} else if (x == PREPARE_COMMIT) {
int sessionId = in.readVarInt();
String transaction = in.readString();
if (trace.isDebugEnabled()) {
trace.debug("log prepare commit " + sessionId + " " + transaction + " pos: " + pos);
}
if (stage == RECOVERY_STAGE_UNDO) {
int page = pageIn.getDataPage();
setPrepareCommit(sessionId, page, transaction);
}
} else if (x == ROLLBACK) {
int sessionId = in.readVarInt();
if (trace.isDebugEnabled()) {
trace.debug("log rollback " + sessionId + " pos: " + pos);
}
// ignore - this entry is just informational
} else if (x == COMMIT) {
int sessionId = in.readVarInt();
if (trace.isDebugEnabled()) {
trace.debug("log commit " + sessionId + " pos: " + pos);
}
if (stage == RECOVERY_STAGE_UNDO) {
setLastCommitForSession(sessionId, logId, pos);
}
} else if (x == NOOP) {
// nothing to do
} else if (x == CHECKPOINT) {
logId++;
} else if (x == FREE_LOG) {
int count = in.readVarInt();
for (int i = 0; i < count; i++) {
int pageId = in.readVarInt();
if (stage == RECOVERY_STAGE_REDO) {
if (!usedLogPages.get(pageId)) {
store.free(pageId, false);
}
}
}
} else {
if (trace.isDebugEnabled()) {
trace.debug("log end");
break;
}
}
}
} catch (DbException e) {
if (e.getErrorCode() == ErrorCode.FILE_CORRUPTED_1) {
trace.debug("log recovery stopped");
} else {
throw e;
}
} catch (IOException e) {
trace.debug("log recovery completed");
}
undo = new BitField();
if (stage == RECOVERY_STAGE_REDO) {
usedLogPages = null;
}
return isEmpty;
}
/**
* Write the operation to the trace system if debug trace is enabled.
*
* @param operation the operation performed
* @param id the id of the operation
*/
public void traceOperation(String operation, int id) {
if (trace.isDebugEnabled()) {
trace.debug("{0} {1}", operation, id);
}
}