hcq/openGauss-server
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

issue#I5UDM6 LogAccessExclusiveLock() when primary

Browse Source 
node acquire AccessExclusiveLock for systable

数据库中表（例如系统表pg_database在大量drop database情况下）因为vacuum时会触发
truncate操作，备机replay truncate操作的同时，有监控程序查询该表（
xlog_block_smgr_redo_truncate采用LockRelFileNode给该表的relfilenode加锁了，但
scan操作是通过LockRelationOid给relation的oid加锁，导致replay truncate和select操
作并行了），使被truncate操作invalid的buffer再次被加载到缓冲池中，在之后新的数据
插入到此原本应该已经invalid的数据页时产生错误。query在initscan时通过smgrnblocks
拿到nblocks，如果此时replay trucate log先做了InvalidateBuffer，再做
CacheInvalidateSmgr、smgr_truncate，会导致scan时再次讲invalid的buffer加载到
bufferpool中，产生错误。
回看pg代码，发现在lock relation时候，主机加AccessExclusiveLock时候会记录一条
XLOG_STANDBY_LOCK的日志，当备机回放到该日志时会加上AccessExclusiveLock。但是在
og中，因为原本分布式代码的残留，只在用户表时记了该日志，所以导致系统表会出现上
述问题。
This commit is contained in:
Chunling Wang
2022-09-30 16:34:36 +08:00
parent 6f87123c7a
commit ecfc320dd6
1 changed files with 2 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/gausskernel/storage/lmgr/lock.cpp
View File

@ -754,30 +754,8 @@ static LockAcquireResult LockAcquireExtendedXC(const LOCKTAG *locktag, LOCKMODE
*/
if (lockmode >= AccessExclusiveLock && locktag->locktag_type == LOCKTAG_RELATION && !RecoveryInProgress() &&
XLogStandbyInfoActive()) {
/*
* In a scenario like:
*
* 1, openGauss run vacuum full or autovacuum pg_class, insert AccessExclusiveLock xlog.
* 2, datanode crash, vacuum full abort.
* 3, datanode restart in pending mode, start recovery.
* 4, startup thread acquire pg_class's AccessExclusiveLock.
* 5, startup thread complete recovery and wait for notify.
* 6, cm agent connect datanode, need to init relcache file.
* 7, cm agent connect want to acquire pg_class's AccessShareLock,
* but the AccessExclusiveLock lock is hold by startup thread.
* 8, dead lock, datanode hang.
*
* Other system tables like pg_attribute/pg_type.. also have such problem.
*
* To solve this problem, we don't insert AccessExclusiveLock xlog for system tables.
* This change may cause exception when primary run vacuum full system table while
* standby access the system table at the same time.
*
*/
if (locktag->locktag_field2 > FirstNormalObjectId) {
LogAccessExclusiveLockPrepare();
log_lock = true;
}
LogAccessExclusiveLockPrepare();
log_lock = true;
}
/*