Merge pull request #14272 from taosdata/feature/3.0_mhli

refactor(sync): optimized only one replica
This commit is contained in:
Li Minghao 2022-06-27 11:05:39 +08:00 committed by GitHub
commit 486f268353
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
20 changed files with 173 additions and 106 deletions

View File

@ -524,7 +524,7 @@ void syncReconfigFinishLog2(char* s, const SyncReconfigFinish* pMsg);
int32_t syncNodeOnPingCb(SSyncNode* ths, SyncPing* pMsg);
int32_t syncNodeOnPingReplyCb(SSyncNode* ths, SyncPingReply* pMsg);
int32_t syncNodeOnTimeoutCb(SSyncNode* ths, SyncTimeout* pMsg);
int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg);
int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg, SyncIndex* pRetIndex);
int32_t syncNodeOnRequestVoteCb(SSyncNode* ths, SyncRequestVote* pMsg);
int32_t syncNodeOnRequestVoteReplyCb(SSyncNode* ths, SyncRequestVoteReply* pMsg);
int32_t syncNodeOnAppendEntriesCb(SSyncNode* ths, SyncAppendEntries* pMsg);
@ -541,7 +541,7 @@ int32_t syncNodeOnSnapshotRspCb(SSyncNode* ths, SyncSnapshotRsp* pMsg);
// -----------------------------------------
typedef int32_t (*FpOnPingCb)(SSyncNode* ths, SyncPing* pMsg);
typedef int32_t (*FpOnPingReplyCb)(SSyncNode* ths, SyncPingReply* pMsg);
typedef int32_t (*FpOnClientRequestCb)(SSyncNode* ths, SyncClientRequest* pMsg);
typedef int32_t (*FpOnClientRequestCb)(SSyncNode* ths, SyncClientRequest* pMsg, SyncIndex* pRetIndex);
typedef int32_t (*FpOnRequestVoteCb)(SSyncNode* ths, SyncRequestVote* pMsg);
typedef int32_t (*FpOnRequestVoteReplyCb)(SSyncNode* ths, SyncRequestVoteReply* pMsg);
typedef int32_t (*FpOnAppendEntriesCb)(SSyncNode* ths, SyncAppendEntries* pMsg);

View File

@ -442,7 +442,7 @@ int32_t mndProcessSyncMsg(SRpcMsg *pMsg) {
syncPingReplyDestroy(pSyncMsg);
} else if (pMsg->msgType == TDMT_SYNC_CLIENT_REQUEST) {
SyncClientRequest *pSyncMsg = syncClientRequestFromRpcMsg2(pMsg);
code = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg);
code = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg, NULL);
syncClientRequestDestroy(pSyncMsg);
} else if (pMsg->msgType == TDMT_SYNC_REQUEST_VOTE) {
SyncRequestVote *pSyncMsg = syncRequestVoteFromRpcMsg2(pMsg);
@ -491,7 +491,7 @@ int32_t mndProcessSyncMsg(SRpcMsg *pMsg) {
syncPingReplyDestroy(pSyncMsg);
} else if (pMsg->msgType == TDMT_SYNC_CLIENT_REQUEST) {
SyncClientRequest *pSyncMsg = syncClientRequestFromRpcMsg2(pMsg);
code = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg);
code = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg, NULL);
syncClientRequestDestroy(pSyncMsg);
} else if (pMsg->msgType == TDMT_SYNC_REQUEST_VOTE) {
SyncRequestVote *pSyncMsg = syncRequestVoteFromRpcMsg2(pMsg);

View File

@ -137,6 +137,25 @@ void vnodeProposeMsg(SQueueInfo *pInfo, STaosQall *qall, int32_t numOfMsgs) {
vError("vgId:%d, failed to pre-process msg:%p since %s", vgId, pMsg, terrstr());
} else {
code = syncPropose(pVnode->sync, pMsg, vnodeIsMsgWeak(pMsg->msgType));
if (code == 1) {
do {
static int32_t cnt = 0;
if (cnt++ % 1000 == 1) {
vInfo("vgId:%d, msg:%p apply right now, apply index:%ld, msgtype:%s,%d", vgId, pMsg,
pMsg->info.conn.applyIndex, TMSG_INFO(pMsg->msgType), pMsg->msgType);
}
} while (0);
SRpcMsg rsp = {.code = pMsg->code, .info = pMsg->info};
if (vnodeProcessWriteReq(pVnode, pMsg, pMsg->info.conn.applyIndex, &rsp) < 0) {
rsp.code = terrno;
vInfo("vgId:%d, msg:%p failed to apply right now since %s", vgId, pMsg, terrstr());
}
if (rsp.info.handle != NULL) {
tmsgSendRsp(&rsp);
}
}
}
}
@ -163,11 +182,13 @@ void vnodeProposeMsg(SQueueInfo *pInfo, STaosQall *qall, int32_t numOfMsgs) {
SRpcMsg rsp = {.code = TSDB_CODE_RPC_REDIRECT, .info = pMsg->info};
tmsgSendRedirectRsp(&rsp, &newEpSet);
} else {
if (code != 1) {
if (terrno != 0) code = terrno;
vError("vgId:%d, msg:%p failed to propose since %s, code:0x%x", vgId, pMsg, tstrerror(code), code);
SRpcMsg rsp = {.code = code, .info = pMsg->info};
tmsgSendRsp(&rsp);
}
}
vGTrace("vgId:%d, msg:%p is freed, code:0x%x", vgId, pMsg, code);
rpcFreeCont(pMsg->pCont);
@ -260,7 +281,7 @@ int32_t vnodeProcessSyncReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
SyncClientRequest *pSyncMsg = syncClientRequestFromRpcMsg2(pRpcMsg);
assert(pSyncMsg != NULL);
ret = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg);
ret = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg, NULL);
syncClientRequestDestroy(pSyncMsg);
} else if (pRpcMsg->msgType == TDMT_SYNC_REQUEST_VOTE) {
@ -359,16 +380,10 @@ static void vnodeSyncCommitMsg(SSyncFSM *pFsm, const SRpcMsg *pMsg, SFsmCbMeta c
SyncIndex beginIndex = SYNC_INDEX_INVALID;
char logBuf[256] = {0};
if (pFsm->FpGetSnapshotInfo != NULL) {
(*pFsm->FpGetSnapshotInfo)(pFsm, &snapshot);
beginIndex = snapshot.lastApplyIndex;
}
if (cbMeta.index > beginIndex) {
snprintf(
logBuf, sizeof(logBuf),
snprintf(logBuf, sizeof(logBuf),
"==callback== ==CommitCb== execute, pFsm:%p, index:%ld, isWeak:%d, code:%d, state:%d %s, beginIndex :%ld\n",
pFsm, cbMeta.index, cbMeta.isWeak, cbMeta.code, cbMeta.state, syncUtilState2String(cbMeta.state), beginIndex);
pFsm, cbMeta.index, cbMeta.isWeak, cbMeta.code, cbMeta.state, syncUtilState2String(cbMeta.state),
beginIndex);
syncRpcMsgLog2(logBuf, (SRpcMsg *)pMsg);
SRpcMsg rpcMsg = {.msgType = pMsg->msgType, .contLen = pMsg->contLen};
@ -377,16 +392,6 @@ static void vnodeSyncCommitMsg(SSyncFSM *pFsm, const SRpcMsg *pMsg, SFsmCbMeta c
syncGetAndDelRespRpc(pVnode->sync, cbMeta.seqNum, &rpcMsg.info);
rpcMsg.info.conn.applyIndex = cbMeta.index;
tmsgPutToQueue(&pVnode->msgCb, APPLY_QUEUE, &rpcMsg);
} else {
char logBuf[256] = {0};
snprintf(logBuf, sizeof(logBuf),
"==callback== ==CommitCb== do not execute, pFsm:%p, index:%ld, isWeak:%d, code:%d, state:%d %s, "
"beginIndex :%ld\n",
pFsm, cbMeta.index, cbMeta.isWeak, cbMeta.code, cbMeta.state, syncUtilState2String(cbMeta.state),
beginIndex);
syncRpcMsgLog2(logBuf, (SRpcMsg *)pMsg);
}
}
static void vnodeSyncPreCommitMsg(SSyncFSM *pFsm, const SRpcMsg *pMsg, SFsmCbMeta cbMeta) {

View File

@ -50,7 +50,7 @@ typedef struct SSyncIO {
void *pSyncNode;
int32_t (*FpOnSyncPing)(SSyncNode *pSyncNode, SyncPing *pMsg);
int32_t (*FpOnSyncPingReply)(SSyncNode *pSyncNode, SyncPingReply *pMsg);
int32_t (*FpOnSyncClientRequest)(SSyncNode *pSyncNode, SyncClientRequest *pMsg);
int32_t (*FpOnSyncClientRequest)(SSyncNode *pSyncNode, SyncClientRequest *pMsg, SyncIndex *pRetIndex);
int32_t (*FpOnSyncRequestVote)(SSyncNode *pSyncNode, SyncRequestVote *pMsg);
int32_t (*FpOnSyncRequestVoteReply)(SSyncNode *pSyncNode, SyncRequestVoteReply *pMsg);
int32_t (*FpOnSyncAppendEntries)(SSyncNode *pSyncNode, SyncAppendEntries *pMsg);

View File

@ -169,7 +169,7 @@ SSyncNode* syncNodeOpen(const SSyncInfo* pSyncInfo);
void syncNodeStart(SSyncNode* pSyncNode);
void syncNodeStartStandBy(SSyncNode* pSyncNode);
void syncNodeClose(SSyncNode* pSyncNode);
int32_t syncNodePropose(SSyncNode* pSyncNode, const SRpcMsg* pMsg, bool isWeak);
int32_t syncNodePropose(SSyncNode* pSyncNode, SRpcMsg* pMsg, bool isWeak);
// option
bool syncNodeSnapshotEnable(SSyncNode* pSyncNode);
@ -233,6 +233,7 @@ SyncIndex syncNodeGetPreIndex(SSyncNode* pSyncNode, SyncIndex index);
SyncTerm syncNodeGetPreTerm(SSyncNode* pSyncNode, SyncIndex index);
int32_t syncNodeGetPreIndexTerm(SSyncNode* pSyncNode, SyncIndex index, SyncIndex* pPreIndex, SyncTerm* pPreTerm);
bool syncNodeIsOptimizedOneReplica(SSyncNode* ths, SRpcMsg* pMsg);
int32_t syncNodeCommit(SSyncNode* ths, SyncIndex beginIndex, SyncIndex endIndex, uint64_t flag);
int32_t syncNodeUpdateNewConfigIndex(SSyncNode* ths, SSyncCfg* pNewCfg);

View File

@ -31,7 +31,10 @@ extern "C" {
typedef struct SSyncLogStoreData {
SSyncNode* pSyncNode;
SWal* pWal;
TdThreadMutex mutex;
SWalReadHandle* pWalHandle;
// SyncIndex beginIndex; // valid begin index, default 0, may be set beginIndex > 0
} SSyncLogStoreData;

View File

@ -102,6 +102,7 @@ void syncMaybeAdvanceCommitIndex(SSyncNode* pSyncNode) {
}
}
// maybe execute fsm
if (newCommitIndex > pSyncNode->commitIndex) {
SyncIndex beginIndex = pSyncNode->commitIndex + 1;
SyncIndex endIndex = newCommitIndex;

View File

@ -281,7 +281,7 @@ static void *syncIOConsumerFunc(void *param) {
if (io->FpOnSyncClientRequest != NULL) {
SyncClientRequest *pSyncMsg = syncClientRequestFromRpcMsg2(pRpcMsg);
ASSERT(pSyncMsg != NULL);
io->FpOnSyncClientRequest(io->pSyncNode, pSyncMsg);
io->FpOnSyncClientRequest(io->pSyncNode, pSyncMsg, NULL);
syncClientRequestDestroy(pSyncMsg);
}

View File

@ -50,7 +50,7 @@ static int32_t syncNodeAppendNoop(SSyncNode* ths);
// process message ----
int32_t syncNodeOnPingCb(SSyncNode* ths, SyncPing* pMsg);
int32_t syncNodeOnPingReplyCb(SSyncNode* ths, SyncPingReply* pMsg);
int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg);
int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg, SyncIndex* pRetIndex);
// life cycle
static void syncFreeNode(void* param);
@ -627,7 +627,7 @@ int32_t syncPropose(int64_t rid, SRpcMsg* pMsg, bool isWeak) {
return ret;
}
int32_t syncNodePropose(SSyncNode* pSyncNode, const SRpcMsg* pMsg, bool isWeak) {
int32_t syncNodePropose(SSyncNode* pSyncNode, SRpcMsg* pMsg, bool isWeak) {
int32_t ret = 0;
char eventLog[128];
@ -664,13 +664,34 @@ int32_t syncNodePropose(SSyncNode* pSyncNode, const SRpcMsg* pMsg, bool isWeak)
SRpcMsg rpcMsg;
syncClientRequest2RpcMsg(pSyncMsg, &rpcMsg);
// optimized one replica
if (syncNodeIsOptimizedOneReplica(pSyncNode, pMsg)) {
SyncIndex retIndex;
int32_t code = syncNodeOnClientRequestCb(pSyncNode, pSyncMsg, &retIndex);
if (code == 0) {
pMsg->info.conn.applyIndex = retIndex;
rpcFreeCont(rpcMsg.pCont);
syncRespMgrDel(pSyncNode->pSyncRespMgr, seqNum);
ret = 1;
sDebug("vgId:%d optimized index:%ld success, msgtype:%s,%d", pSyncNode->vgId, retIndex,
TMSG_INFO(pMsg->msgType), pMsg->msgType);
} else {
ret = -1;
terrno = TSDB_CODE_SYN_INTERNAL_ERROR;
sError("vgId:%d optimized index:%ld error, msgtype:%s,%d", pSyncNode->vgId, retIndex, TMSG_INFO(pMsg->msgType),
pMsg->msgType);
}
} else {
if (pSyncNode->FpEqMsg != NULL && (*pSyncNode->FpEqMsg)(pSyncNode->msgcb, &rpcMsg) == 0) {
ret = 0;
} else {
ret = -1;
terrno = TSDB_CODE_SYN_INTERNAL_ERROR;
sError("syncPropose pSyncNode->FpEqMsg is NULL");
sError("enqueue msg error, FpEqMsg is NULL");
}
}
syncClientRequestDestroy(pSyncMsg);
goto _END;
@ -2377,7 +2398,7 @@ int32_t syncNodeOnPingReplyCb(SSyncNode* ths, SyncPingReply* pMsg) {
// /\ UNCHANGED <<messages, serverVars, candidateVars,
// leaderVars, commitIndex>>
//
int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg) {
int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg, SyncIndex* pRetIndex) {
int32_t ret = 0;
syncClientRequestLog2("==syncNodeOnClientRequestCb==", pMsg);
@ -2436,6 +2457,14 @@ int32_t syncNodeOnClientRequestCb(SSyncNode* ths, SyncClientRequest* pMsg) {
rpcFreeCont(rpcMsg.pCont);
}
if (pRetIndex != NULL) {
if (ret == 0 && pEntry != NULL) {
*pRetIndex = pEntry->index;
} else {
*pRetIndex = SYNC_INDEX_INVALID;
}
}
syncEntryDestory(pEntry);
return ret;
}
@ -2600,6 +2629,10 @@ static int32_t syncNodeProposeConfigChangeFinish(SSyncNode* ths, SyncReconfigFin
return 0;
}
bool syncNodeIsOptimizedOneReplica(SSyncNode* ths, SRpcMsg* pMsg) {
return (ths->replicaNum == 1 && syncUtilUserCommit(pMsg->msgType) && ths->vgId != 1);
}
int32_t syncNodeCommit(SSyncNode* ths, SyncIndex beginIndex, SyncIndex endIndex, uint64_t flag) {
int32_t code = 0;
ESyncState state = flag;
@ -2621,7 +2654,20 @@ int32_t syncNodeCommit(SSyncNode* ths, SyncIndex beginIndex, SyncIndex endIndex,
syncEntry2OriginalRpc(pEntry, &rpcMsg);
// user commit
if (ths->pFsm->FpCommitCb != NULL && syncUtilUserCommit(pEntry->originalRpcType)) {
if ((ths->pFsm->FpCommitCb != NULL) && syncUtilUserCommit(pEntry->originalRpcType)) {
bool internalExecute = true;
if ((ths->replicaNum == 1) && ths->restoreFinish && (ths->vgId != 1)) {
internalExecute = false;
}
do {
char logBuf[128];
snprintf(logBuf, sizeof(logBuf), "index:%ld, internalExecute:%d", i, internalExecute);
syncNodeEventLog(ths, logBuf);
} while (0);
// execute fsm in apply thread, or execute outside syncPropose
if (internalExecute) {
SFsmCbMeta cbMeta = {0};
cbMeta.index = pEntry->index;
cbMeta.lastConfigIndex = syncNodeGetSnapshotConfigIndex(ths, cbMeta.index);
@ -2635,6 +2681,7 @@ int32_t syncNodeCommit(SSyncNode* ths, SyncIndex beginIndex, SyncIndex endIndex,
ths->pFsm->FpCommitCb(ths->pFsm, &rpcMsg, cbMeta);
}
}
// config change
if (pEntry->originalRpcType == TDMT_SYNC_CONFIG_CHANGE) {

View File

@ -257,6 +257,8 @@ static int32_t raftLogGetEntry(struct SSyncLogStore* pLogStore, SyncIndex index,
return -1;
}
taosThreadMutexLock(&(pData->mutex));
code = walReadWithHandle(pWalHandle, index);
if (code != 0) {
int32_t err = terrno;
@ -281,6 +283,7 @@ static int32_t raftLogGetEntry(struct SSyncLogStore* pLogStore, SyncIndex index,
terrno = saveErr;
*/
taosThreadMutexUnlock(&(pData->mutex));
return code;
}
@ -301,6 +304,7 @@ static int32_t raftLogGetEntry(struct SSyncLogStore* pLogStore, SyncIndex index,
terrno = saveErr;
*/
taosThreadMutexUnlock(&(pData->mutex));
return code;
}
@ -364,6 +368,7 @@ SSyncLogStore* logStoreCreate(SSyncNode* pSyncNode) {
pData->pWal = pSyncNode->pWal;
ASSERT(pData->pWal != NULL);
taosThreadMutexInit(&(pData->mutex), NULL);
pData->pWalHandle = walOpenReadHandle(pData->pWal);
ASSERT(pData->pWalHandle != NULL);
@ -408,9 +413,14 @@ SSyncLogStore* logStoreCreate(SSyncNode* pSyncNode) {
void logStoreDestory(SSyncLogStore* pLogStore) {
if (pLogStore != NULL) {
SSyncLogStoreData* pData = pLogStore->data;
taosThreadMutexLock(&(pData->mutex));
if (pData->pWalHandle != NULL) {
walCloseReadHandle(pData->pWalHandle);
pData->pWalHandle = NULL;
}
taosThreadMutexUnlock(&(pData->mutex));
taosThreadMutexDestroy(&(pData->mutex));
taosMemoryFree(pLogStore->data);
taosMemoryFree(pLogStore);
@ -460,6 +470,8 @@ SSyncRaftEntry* logStoreGetEntry(SSyncLogStore* pLogStore, SyncIndex index) {
SWal* pWal = pData->pWal;
if (index >= SYNC_INDEX_BEGIN && index <= logStoreLastIndex(pLogStore)) {
taosThreadMutexLock(&(pData->mutex));
// SWalReadHandle* pWalHandle = walOpenReadHandle(pWal);
SWalReadHandle* pWalHandle = pData->pWalHandle;
ASSERT(pWalHandle != NULL);
@ -503,6 +515,7 @@ SSyncRaftEntry* logStoreGetEntry(SSyncLogStore* pLogStore, SyncIndex index) {
terrno = saveErr;
*/
taosThreadMutexUnlock(&(pData->mutex));
return pEntry;
} else {

View File

@ -419,15 +419,12 @@ void test6() {
logStoreDestory(pLogStore);
cleanup();
// restart
init();
pLogStore = logStoreCreate(pSyncNode);
assert(pLogStore);
pSyncNode->pLogStore = pLogStore;
do {
SyncIndex firstVer = walGetFirstVer(pWal);
SyncIndex lastVer = walGetLastVer(pWal);