PostgreSQL啟動恢復讀取checkpoint記錄失敗的條件
阿新 • • 發佈：2018-08-05
oca gets 文件 expected ogre ESS available logs ani
1、首先讀取ControlFile->checkPoint指向的checkpoint
2、如果讀取失敗，slave直接abort退出，master再次讀取ControlFile->prevCheckPoint指向的checkpoint
StartupXLOG->
|--checkPointLoc = ControlFile->checkPoint;
|--record = ReadCheckpointRecord(xlogreader, checkPointLoc, 1, true):
|-- if (record != NULL){
...
}else if (StandbyMode){
ereport(PANIC,(errmsg("could not locate a valid checkpoint record")));
}else{
checkPointLoc = ControlFile->prevCheckPoint;
record = ReadCheckpointRecord(xlogreader, checkPointLoc, 2, true);
if (record != NULL){
InRecovery = true;//標記下面進入recovery
}else{
ereport(PANIC,(errmsg("could not locate a valid checkpoint record")));
}
}
一、那麽什麽條件下讀取的checkpoint記錄record==NULL?
1、ControlFile->checkPoint % XLOG_BLCKSZ < SizeOfXLogShortPHD 2、ReadRecord(xlogreader, ControlFile->checkPoint, LOG, true)返回NULL 3、ReadRecord讀到的record!=NULL && record->xl_rmid != RM_XLOG_ID 4、ReadRecord讀到的record!=NULL && info != XLOG_CHECKPOINT_SHUTDOWN && info != XLOG_CHECKPOINT_ONLINE 5、ReadRecord讀到的record!=NULL && record->xl_tot_len != SizeOfXLogRecord + SizeOfXLogRecordDataHeaderShort + sizeof(CheckPoint)
二、ReadRecord函數返回NULL的條件
ReadRecord(xlogreader, ControlFile->checkPoint, LOG, true) |--record = XLogReadRecord(xlogreader, ControlFile->checkPoint, &errormsg); |-- 2.1 record==NULL && ！StandbyMode |-- 2.2 record!=NULL && !tliInHistory(xlogreader->latestPageTLI, expectedTLEs) /*----- note:只要讀取了一頁xlog,就會賦值為該頁第一個記錄的時間線 XLogReaderValidatePageHeader -->xlogreader->latestPageTLI=hdr->xlp_tli; ------*/
三、XlogReadRecord讀取checkpoint返回NULL的條件？
XLogReadRecord(xlogreader, ControlFile->checkPoint, &errormsg)
targetPagePtr = ControlFile->checkPoint - (ControlFile->checkPoint % XLOG_BLCKSZ);
targetRecOff = ControlFile->checkPoint % XLOG_BLCKSZ;
readOff = ReadPageInternal(state,targetPagePtr, Min(targetRecOff + SizeOfXLogRecord, XLOG_BLCKSZ));
pageHeaderSize = XLogPageHeaderSize((XLogPageHeader) state->readBuf);
record = (XLogRecord *) (state->readBuf + RecPtr % XLOG_BLCKSZ);
total_len = record->xl_tot_len;
-------------
1、readOff < 0
2、0< targetRecOff < pageHeaderSize
3、(((XLogPageHeader) state->readBuf)->xlp_info & XLP_FIRST_IS_CONTRECORD) && targetRecOff == pageHeaderSize
page頭有跨頁的record並且checkpoint定位的偏移正好在頁頭尾部
4、targetRecOff <= XLOG_BLCKSZ - SizeOfXLogRecord &&
!ValidXLogRecordHeader(state, ControlFile->checkPoint, state->ReadRecPtr, record,randAccess)
---（record->xl_tot_len < SizeOfXLogRecord || record->xl_rmid > RM_MAX_ID || record->xl_prev != state->ReadRecPtr）
5、targetRecOff > XLOG_BLCKSZ - SizeOfXLogRecord && total_len < SizeOfXLogRecord
6、total_len > state->readRecordBufSize && !allocate_recordbuf(state, total_len)
一旦該記錄損壞，total_len的長度非常大的話，就需要allocate_recordbuf擴展state->readbuf，可能因此分配失敗abort
記錄的checksum需要等待全部讀取完整記錄後才校驗
-------------三、ReadPageInternal返回的readOff返回小於0的條件
ReadPageInternal(state,targetPagePtr, Min(targetRecOff + SizeOfXLogRecord, XLOG_BLCKSZ))
1、第一次read wal文件，readLen = state->read_page：讀取第一頁。readLen < 0
2、readLen>0 && !XLogReaderValidatePageHeader(state, targetSegmentPtr, state->readBuf)
--
3、讀取checkpoint所在頁readLen = state->read_page: readLen < 0
4、readLen > 0 && readLen <= SizeOfXLogShortPHD
5、!XLogReaderValidatePageHeader(state, pageptr, (char *) hdr)四、XLogPageRead何時返回值<0 ?
/*
1、WaitForWALToBecomeAvailable open失敗
2、lseek 失敗 && ！StandbyMode
3、read失敗 && ！StandbyMode
4、校驗page頭失敗 && ！StandbyMode
如果是StandbyMode，則會重新retry->WaitForWALToBecomeAvailable,切換日誌源進行open
*/
!WaitForWALToBecomeAvailable(targetPagePtr + reqLen,private->randAccess,1,targetRecPtr)//open
|-- return -1
readOff = targetPageOff;
if (lseek(readFile, (off_t) readOff, SEEK_SET) < 0){
！StandbyMode:: return -1
}
if (read(readFile, readBuf, XLOG_BLCKSZ) != XLOG_BLCKSZ){
！StandbyMode:: return -1
}
XLogReaderValidatePageHeader(xlogreader, targetPagePtr, readBuf)
！StandbyMode:: return -1五、WaitForWALToBecomeAvailable何時返回false？
--XLOG_FROM_ARCHIVE | XLOG_FROM_PG_WAL
1、先XLogFileReadAnyTLI open日誌：
1、遍歷時間線列表裏的每一個時間線，從最新的開始
2、當讀取checkpoint的時候，source是XLOG_FROM_ANY
3、先找歸檔的日誌進行open；如果open失敗再找WAL日誌進行open
4、如果都沒有open成功，則向前找時間線，open前一個時間線segno和文件號相同的文件進行open
5、open成功後expectedTLEs被賦值為當前時間線列表的所有值
2、如果open失敗，則切換日誌源：XLOG_FROM_ARCHIVE | XLOG_FROM_PG_WAL -> XLOG_FROM_STREAM
3、切換日誌源後，XLOG_FROM_ARCHIVE | XLOG_FROM_PG_WAL 則：
slave && promote ：return false
!StandbyMode：return false
--XLOG_FROM_STREAM
1、!WalRcvStreaming()即receiver進程掛了，切換日誌源
2、CheckForStandbyTrigger()切換日誌源
3、XLOG_FROM_STREAM->XLOG_FROM_ARCHIVEPostgreSQL啟動恢復讀取checkpoint記錄失敗的條件