未验证 提交 5b300b58 编写于 作者: S Shengliang Guan 提交者: GitHub

Merge pull request #16044 from taosdata/fix/mnode

fix: deadlock of mnode if its state changed
......@@ -87,10 +87,11 @@ typedef struct {
typedef struct {
tsem_t syncSem;
int64_t sync;
bool standby;
SReplica replica;
int32_t errCode;
int32_t transId;
SRWLatch lock;
int8_t standby;
int8_t leaderTransferFinish;
} SSyncMgmt;
......
......@@ -60,15 +60,19 @@ void mndSyncCommitMsg(struct SSyncFSM *pFsm, const SRpcMsg *pMsg, SFsmCbMeta cbM
sdbSetApplyInfo(pMnode->pSdb, cbMeta.index, cbMeta.term, cbMeta.lastConfigIndex);
}
taosRLockLatch(&pMgmt->lock);
if (transId <= 0) {
taosRUnLockLatch(&pMgmt->lock);
mError("trans:%d, invalid commit msg", transId);
} else if (transId == pMgmt->transId) {
taosRUnLockLatch(&pMgmt->lock);
if (pMgmt->errCode != 0) {
mError("trans:%d, failed to propose since %s", transId, tstrerror(pMgmt->errCode));
}
pMgmt->transId = 0;
tsem_post(&pMgmt->syncSem);
} else {
taosRUnLockLatch(&pMgmt->lock);
STrans *pTrans = mndAcquireTrans(pMnode, transId);
if (pTrans != NULL) {
mDebug("trans:%d, execute in mnode which not leader", transId);
......@@ -115,6 +119,7 @@ void mndReConfig(struct SSyncFSM *pFsm, const SRpcMsg *pMsg, SReConfigCbMeta cbM
mInfo("trans:-1, sync reconfig is proposed, saved:%d code:0x%x, index:%" PRId64 " term:%" PRId64, pMgmt->transId,
cbMeta.code, cbMeta.index, cbMeta.term);
taosWLockLatch(&pMgmt->lock);
if (pMgmt->transId == -1) {
if (pMgmt->errCode != 0) {
mError("trans:-1, failed to propose sync reconfig since %s", tstrerror(pMgmt->errCode));
......@@ -122,6 +127,7 @@ void mndReConfig(struct SSyncFSM *pFsm, const SRpcMsg *pMsg, SReConfigCbMeta cbM
pMgmt->transId = 0;
tsem_post(&pMgmt->syncSem);
}
taosWUnLockLatch(&pMgmt->lock);
}
int32_t mndSnapshotStartRead(struct SSyncFSM *pFsm, void *pParam, void **ppReader) {
......@@ -170,12 +176,24 @@ static void mndBecomeFollower(struct SSyncFSM *pFsm) {
SMnode *pMnode = pFsm->data;
mDebug("vgId:1, become follower");
// clear old leader resource
taosWLockLatch(&pMnode->syncMgmt.lock);
if (pMnode->syncMgmt.transId != 0) {
pMnode->syncMgmt.transId = 0;
tsem_post(&pMnode->syncMgmt.syncSem);
}
taosWUnLockLatch(&pMnode->syncMgmt.lock);
}
static void mndBecomeLeader(struct SSyncFSM *pFsm) {
SMnode *pMnode = pFsm->data;
mDebug("vgId:1, become leader");
SMnode *pMnode = pFsm->data;
taosWLockLatch(&pMnode->syncMgmt.lock);
if (pMnode->syncMgmt.transId != 0) {
pMnode->syncMgmt.transId = 0;
tsem_post(&pMnode->syncMgmt.syncSem);
}
taosWUnLockLatch(&pMnode->syncMgmt.lock);
}
SSyncFSM *mndSyncMakeFsm(SMnode *pMnode) {
......@@ -202,6 +220,8 @@ SSyncFSM *mndSyncMakeFsm(SMnode *pMnode) {
int32_t mndInitSync(SMnode *pMnode) {
SSyncMgmt *pMgmt = &pMnode->syncMgmt;
taosInitRWLatch(&pMgmt->lock);
pMgmt->transId = 0;
SSyncInfo syncInfo = {.vgId = 1, .FpSendMsg = mndSyncSendMsg, .FpEqMsg = mndSyncEqMsg};
snprintf(syncInfo.path, sizeof(syncInfo.path), "%s%ssync", pMnode->path, TD_DIRSEP);
......@@ -254,11 +274,14 @@ int32_t mndSyncPropose(SMnode *pMnode, SSdbRaw *pRaw, int32_t transId) {
memcpy(req.pCont, pRaw, req.contLen);
pMgmt->errCode = 0;
taosWLockLatch(&pMgmt->lock);
pMgmt->transId = transId;
taosWUnLockLatch(&pMgmt->lock);
mTrace("trans:%d, will be proposed", pMgmt->transId);
const bool isWeak = false;
int32_t code = syncPropose(pMgmt->sync, &req, isWeak);
if (code == 0) {
tsem_wait(&pMgmt->syncSem);
} else if (code == -1 && terrno == TSDB_CODE_SYN_NOT_LEADER) {
......@@ -286,10 +309,12 @@ void mndSyncStart(SMnode *pMnode) {
}
void mndSyncStop(SMnode *pMnode) {
taosWLockLatch(&pMnode->syncMgmt.lock);
if (pMnode->syncMgmt.transId != 0) {
pMnode->syncMgmt.transId = 0;
tsem_post(&pMnode->syncMgmt.syncSem);
}
taosWUnLockLatch(&pMnode->syncMgmt.lock);
}
bool mndIsMaster(SMnode *pMnode) {
......
......@@ -308,7 +308,8 @@ struct SVnode {
SSink* pSink;
tsem_t canCommit;
int64_t sync;
int32_t blockCount;
SRWLatch lock;
bool blocked;
bool restored;
tsem_t syncSem;
SQHandle* pQuery;
......
......@@ -85,7 +85,8 @@ SVnode *vnodeOpen(const char *path, STfs *pTfs, SMsgCb msgCb) {
pVnode->state.commitTerm = info.state.commitTerm;
pVnode->pTfs = pTfs;
pVnode->msgCb = msgCb;
pVnode->blockCount = 0;
taosInitRWLatch(&pVnode->lock);
pVnode->blocked = false;
tsem_init(&pVnode->syncSem, 0, 0);
tsem_init(&(pVnode->canCommit), 0, 1);
......
......@@ -28,20 +28,28 @@ static inline bool vnodeIsMsgWeak(tmsg_t type) { return false; }
static inline void vnodeWaitBlockMsg(SVnode *pVnode, const SRpcMsg *pMsg) {
if (vnodeIsMsgBlock(pMsg->msgType)) {
const STraceId *trace = &pMsg->info.traceId;
vGTrace("vgId:%d, msg:%p wait block, type:%s", pVnode->config.vgId, pMsg, TMSG_INFO(pMsg->msgType));
pVnode->blockCount = 1;
tsem_wait(&pVnode->syncSem);
taosWLockLatch(&pVnode->lock);
if (!pVnode->blocked) {
vGTrace("vgId:%d, msg:%p wait block, type:%s", pVnode->config.vgId, pMsg, TMSG_INFO(pMsg->msgType));
pVnode->blocked = true;
taosWUnLockLatch(&pVnode->lock);
tsem_wait(&pVnode->syncSem);
} else {
taosWUnLockLatch(&pVnode->lock);
}
}
}
static inline void vnodePostBlockMsg(SVnode *pVnode, const SRpcMsg *pMsg) {
if (vnodeIsMsgBlock(pMsg->msgType)) {
const STraceId *trace = &pMsg->info.traceId;
if (pVnode->blockCount) {
taosWLockLatch(&pVnode->lock);
if (pVnode->blocked) {
vGTrace("vgId:%d, msg:%p post block, type:%s", pVnode->config.vgId, pMsg, TMSG_INFO(pMsg->msgType));
pVnode->blockCount = 0;
pVnode->blocked = false;
tsem_post(&pVnode->syncSem);
}
taosWUnLockLatch(&pVnode->lock);
}
}
......@@ -677,6 +685,12 @@ static void vnodeBecomeFollower(struct SSyncFSM *pFsm) {
vDebug("vgId:%d, become follower", pVnode->config.vgId);
// clear old leader resource
taosWLockLatch(&pVnode->lock);
if (pVnode->blocked) {
pVnode->blocked = false;
tsem_post(&pVnode->syncSem);
}
taosWUnLockLatch(&pVnode->lock);
}
static void vnodeBecomeLeader(struct SSyncFSM *pFsm) {
......
......@@ -210,6 +210,8 @@ if $rows != 3 then
return -1
endi
sql_error select * from performance_schema.PERF_OFFSETS;
sql show create stable stb;
if $rows != 1 then
return -1
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册