/* * Copyright (c) 2019 TAOS Data, Inc. * * This program is free software: you can use, redistribute, and/or modify * it under the terms of the GNU Affero General Public License, version 3 * or later ("AGPL"), as published by the Free Software Foundation. * * This program is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. * * You should have received a copy of the GNU Affero General Public License * along with this program. If not, see . */ #define _DEFAULT_SOURCE #include "syncPipeline.h" #include "syncCommit.h" #include "syncIndexMgr.h" #include "syncInt.h" #include "syncRaftEntry.h" #include "syncRaftStore.h" #include "syncReplication.h" #include "syncRespMgr.h" #include "syncSnapshot.h" #include "syncUtil.h" int64_t syncLogBufferGetEndIndex(SSyncLogBuffer* pBuf) { taosThreadMutexLock(&pBuf->mutex); int64_t index = pBuf->endIndex; taosThreadMutexUnlock(&pBuf->mutex); return index; } int32_t syncLogBufferAppend(SSyncLogBuffer* pBuf, SSyncNode* pNode, SSyncRaftEntry* pEntry) { taosThreadMutexLock(&pBuf->mutex); syncLogBufferValidate(pBuf); SyncIndex index = pEntry->index; if (index - pBuf->startIndex >= pBuf->size) { sError("vgId:%d, failed to append due to sync log buffer full. index:%" PRId64 "", pNode->vgId, index); goto _out; } ASSERT(index == pBuf->endIndex); SSyncRaftEntry* pExist = pBuf->entries[index % pBuf->size].pItem; ASSERT(pExist == NULL); // initial log buffer with at least one item, e.g. commitIndex SSyncRaftEntry* pMatch = pBuf->entries[(index - 1 + pBuf->size) % pBuf->size].pItem; ASSERTS(pMatch != NULL, "no matched log entry"); ASSERT(pMatch->index + 1 == index); SSyncLogBufEntry tmp = {.pItem = pEntry, .prevLogIndex = pMatch->index, .prevLogTerm = pMatch->term}; pBuf->entries[index % pBuf->size] = tmp; pBuf->endIndex = index + 1; syncLogBufferValidate(pBuf); taosThreadMutexUnlock(&pBuf->mutex); return 0; _out: syncLogBufferValidate(pBuf); syncEntryDestroy(pEntry); taosThreadMutexUnlock(&pBuf->mutex); return -1; } SyncTerm syncLogReplMgrGetPrevLogTerm(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncIndex index) { SSyncLogBuffer* pBuf = pNode->pLogBuf; SSyncRaftEntry* pEntry = NULL; SyncIndex prevIndex = index - 1; SyncTerm prevLogTerm = -1; terrno = TSDB_CODE_SUCCESS; if (prevIndex == -1 && pNode->pLogStore->syncLogBeginIndex(pNode->pLogStore) == 0) return 0; if (prevIndex > pBuf->matchIndex) { terrno = TSDB_CODE_WAL_LOG_NOT_EXIST; return -1; } ASSERT(index - 1 == prevIndex); if (prevIndex >= pBuf->startIndex) { pEntry = pBuf->entries[(prevIndex + pBuf->size) % pBuf->size].pItem; ASSERTS(pEntry != NULL, "no log entry found"); prevLogTerm = pEntry->term; return prevLogTerm; } if (pMgr && pMgr->startIndex <= prevIndex && prevIndex < pMgr->endIndex) { int64_t timeMs = pMgr->states[(prevIndex + pMgr->size) % pMgr->size].timeMs; ASSERTS(timeMs != 0, "no log entry found"); prevLogTerm = pMgr->states[(prevIndex + pMgr->size) % pMgr->size].term; ASSERT(prevIndex == 0 || prevLogTerm != 0); return prevLogTerm; } SSnapshot snapshot; if (pNode->pFsm->FpGetSnapshotInfo(pNode->pFsm, &snapshot) == 0 && prevIndex == snapshot.lastApplyIndex) { return snapshot.lastApplyTerm; } if (pNode->pLogStore->syncLogGetEntry(pNode->pLogStore, prevIndex, &pEntry) == 0) { prevLogTerm = pEntry->term; syncEntryDestroy(pEntry); pEntry = NULL; return prevLogTerm; } sError("vgId:%d, failed to get log term since %s. index: %" PRId64 "", pNode->vgId, terrstr(), prevIndex); terrno = TSDB_CODE_WAL_LOG_NOT_EXIST; return -1; } SSyncRaftEntry* syncEntryBuildDummy(SyncTerm term, SyncIndex index, int32_t vgId) { return syncEntryBuildNoop(term, index, vgId); } int32_t syncLogValidateAlignmentOfCommit(SSyncNode* pNode, SyncIndex commitIndex) { SyncIndex firstVer = pNode->pLogStore->syncLogBeginIndex(pNode->pLogStore); if (firstVer > commitIndex + 1) { sError("vgId:%d, firstVer of WAL log greater than tsdb commit version + 1. firstVer: %" PRId64 ", tsdb commit version: %" PRId64 "", pNode->vgId, firstVer, commitIndex); return -1; } SyncIndex lastVer = pNode->pLogStore->syncLogLastIndex(pNode->pLogStore); if (lastVer < commitIndex) { sError("vgId:%d, lastVer of WAL log less than tsdb commit version. lastVer: %" PRId64 ", tsdb commit version: %" PRId64 "", pNode->vgId, lastVer, commitIndex); return -1; } return 0; } int32_t syncLogBufferInitWithoutLock(SSyncLogBuffer* pBuf, SSyncNode* pNode) { ASSERTS(pNode->pLogStore != NULL, "log store not created"); ASSERTS(pNode->pFsm != NULL, "pFsm not registered"); ASSERTS(pNode->pFsm->FpGetSnapshotInfo != NULL, "FpGetSnapshotInfo not registered"); SSnapshot snapshot; if (pNode->pFsm->FpGetSnapshotInfo(pNode->pFsm, &snapshot) < 0) { sError("vgId:%d, failed to get snapshot info since %s", pNode->vgId, terrstr()); goto _err; } SyncIndex commitIndex = snapshot.lastApplyIndex; SyncTerm commitTerm = TMAX(snapshot.lastApplyTerm, 0); if (syncLogValidateAlignmentOfCommit(pNode, commitIndex)) { terrno = TSDB_CODE_WAL_LOG_INCOMPLETE; goto _err; } SyncIndex lastVer = pNode->pLogStore->syncLogLastIndex(pNode->pLogStore); ASSERT(lastVer >= commitIndex); SyncIndex toIndex = lastVer; // update match index pBuf->commitIndex = commitIndex; pBuf->matchIndex = toIndex; pBuf->endIndex = toIndex + 1; // load log entries in reverse order SSyncLogStore* pLogStore = pNode->pLogStore; SyncIndex index = toIndex; SSyncRaftEntry* pEntry = NULL; bool takeDummy = false; while (true) { if (index <= pBuf->commitIndex) { takeDummy = true; break; } if (pLogStore->syncLogGetEntry(pLogStore, index, &pEntry) < 0) { sError("vgId:%d, failed to get log entry since %s. index:%" PRId64 "", pNode->vgId, terrstr(), index); break; } bool taken = false; int emptySize = 5; if (toIndex - index + 1 <= pBuf->size - emptySize) { SSyncLogBufEntry tmp = {.pItem = pEntry, .prevLogIndex = -1, .prevLogTerm = -1}; pBuf->entries[index % pBuf->size] = tmp; taken = true; } if (index < toIndex) { pBuf->entries[(index + 1) % pBuf->size].prevLogIndex = pEntry->index; pBuf->entries[(index + 1) % pBuf->size].prevLogTerm = pEntry->term; } if (!taken) { syncEntryDestroy(pEntry); pEntry = NULL; break; } index--; } // put a dummy record at commitIndex if present in log buffer if (takeDummy) { ASSERT(index == pBuf->commitIndex); SSyncRaftEntry* pDummy = syncEntryBuildDummy(commitTerm, commitIndex, pNode->vgId); if (pDummy == NULL) { terrno = TSDB_CODE_OUT_OF_MEMORY; goto _err; } SSyncLogBufEntry tmp = {.pItem = pDummy, .prevLogIndex = commitIndex - 1, .prevLogTerm = commitTerm}; pBuf->entries[(commitIndex + pBuf->size) % pBuf->size] = tmp; if (index < toIndex) { pBuf->entries[(index + 1) % pBuf->size].prevLogIndex = commitIndex; pBuf->entries[(index + 1) % pBuf->size].prevLogTerm = commitTerm; } } // update startIndex pBuf->startIndex = takeDummy ? index : index + 1; sInfo("vgId:%d, init sync log buffer. buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); // validate syncLogBufferValidate(pBuf); return 0; _err: return -1; } int32_t syncLogBufferInit(SSyncLogBuffer* pBuf, SSyncNode* pNode) { taosThreadMutexLock(&pBuf->mutex); int32_t ret = syncLogBufferInitWithoutLock(pBuf, pNode); taosThreadMutexUnlock(&pBuf->mutex); return ret; } int32_t syncLogBufferReInit(SSyncLogBuffer* pBuf, SSyncNode* pNode) { taosThreadMutexLock(&pBuf->mutex); for (SyncIndex index = pBuf->startIndex; index < pBuf->endIndex; index++) { SSyncRaftEntry* pEntry = pBuf->entries[(index + pBuf->size) % pBuf->size].pItem; if (pEntry == NULL) continue; syncEntryDestroy(pEntry); pEntry = NULL; memset(&pBuf->entries[(index + pBuf->size) % pBuf->size], 0, sizeof(pBuf->entries[0])); } pBuf->startIndex = pBuf->commitIndex = pBuf->matchIndex = pBuf->endIndex = 0; int32_t ret = syncLogBufferInitWithoutLock(pBuf, pNode); if (ret < 0) { sError("vgId:%d, failed to re-initialize sync log buffer since %s.", pNode->vgId, terrstr()); } taosThreadMutexUnlock(&pBuf->mutex); return ret; } FORCE_INLINE SyncTerm syncLogBufferGetLastMatchTerm(SSyncLogBuffer* pBuf) { SyncIndex index = pBuf->matchIndex; SSyncRaftEntry* pEntry = pBuf->entries[(index + pBuf->size) % pBuf->size].pItem; ASSERT(pEntry != NULL); return pEntry->term; } int32_t syncLogBufferAccept(SSyncLogBuffer* pBuf, SSyncNode* pNode, SSyncRaftEntry* pEntry, SyncTerm prevTerm) { taosThreadMutexLock(&pBuf->mutex); syncLogBufferValidate(pBuf); int32_t ret = -1; SyncIndex index = pEntry->index; SyncIndex prevIndex = pEntry->index - 1; SyncTerm lastMatchTerm = syncLogBufferGetLastMatchTerm(pBuf); SSyncRaftEntry* pExist = NULL; bool inBuf = true; if (index <= pBuf->commitIndex) { sTrace("vgId:%d, already committed. index: %" PRId64 ", term: %" PRId64 ". log buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pEntry->index, pEntry->term, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); SyncTerm term = syncLogReplMgrGetPrevLogTerm(NULL, pNode, index + 1); ASSERT(pEntry->term >= 0); if (term == pEntry->term) { ret = 0; } goto _out; } if (index - pBuf->startIndex >= pBuf->size) { sWarn("vgId:%d, out of buffer range. index: %" PRId64 ", term: %" PRId64 ". log buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pEntry->index, pEntry->term, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); goto _out; } if (index > pBuf->matchIndex && lastMatchTerm != prevTerm) { sWarn("vgId:%d, not ready to accept. index: %" PRId64 ", term: %" PRId64 ": prevterm: %" PRId64 " != lastmatch: %" PRId64 ". log buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pEntry->index, pEntry->term, prevTerm, lastMatchTerm, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); goto _out; } // check current in buffer pExist = syncLogBufferGetOneEntry(pBuf, pNode, index, &inBuf); if (pExist != NULL) { ASSERT(pEntry->index == pExist->index); if (pEntry->term != pExist->term) { (void)syncLogBufferRollback(pBuf, pNode, index); } else { sTrace("vgId:%d, duplicate log entry received. index: %" PRId64 ", term: %" PRId64 ". log buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pEntry->index, pEntry->term, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); SyncTerm existPrevTerm = syncLogReplMgrGetPrevLogTerm(NULL, pNode, index); ASSERT(pEntry->term == pExist->term && (pEntry->index > pBuf->matchIndex || prevTerm == existPrevTerm)); ret = 0; goto _out; } } // update ASSERT(pBuf->entries[index % pBuf->size].pItem == NULL); SSyncLogBufEntry tmp = {.pItem = pEntry, .prevLogIndex = prevIndex, .prevLogTerm = prevTerm}; pEntry = NULL; pBuf->entries[index % pBuf->size] = tmp; // update end index pBuf->endIndex = TMAX(index + 1, pBuf->endIndex); // success ret = 0; _out: syncEntryDestroy(pEntry); if (!inBuf) { syncEntryDestroy(pExist); pExist = NULL; } syncLogBufferValidate(pBuf); taosThreadMutexUnlock(&pBuf->mutex); return ret; } int32_t syncLogStorePersist(SSyncLogStore* pLogStore, SSyncRaftEntry* pEntry) { ASSERT(pEntry->index >= 0); SyncIndex lastVer = pLogStore->syncLogLastIndex(pLogStore); if (lastVer >= pEntry->index && pLogStore->syncLogTruncate(pLogStore, pEntry->index) < 0) { sError("failed to truncate log store since %s. from index:%" PRId64 "", terrstr(), pEntry->index); return -1; } lastVer = pLogStore->syncLogLastIndex(pLogStore); ASSERT(pEntry->index == lastVer + 1); if (pLogStore->syncLogAppendEntry(pLogStore, pEntry) < 0) { sError("failed to append sync log entry since %s. index:%" PRId64 ", term:%" PRId64 "", terrstr(), pEntry->index, pEntry->term); return -1; } lastVer = pLogStore->syncLogLastIndex(pLogStore); ASSERT(pEntry->index == lastVer); return 0; } int64_t syncLogBufferProceed(SSyncLogBuffer* pBuf, SSyncNode* pNode, SyncTerm* pMatchTerm) { taosThreadMutexLock(&pBuf->mutex); syncLogBufferValidate(pBuf); SSyncLogStore* pLogStore = pNode->pLogStore; int64_t matchIndex = pBuf->matchIndex; while (pBuf->matchIndex + 1 < pBuf->endIndex) { int64_t index = pBuf->matchIndex + 1; ASSERT(index >= 0); // try to proceed SSyncLogBufEntry* pBufEntry = &pBuf->entries[index % pBuf->size]; SyncIndex prevLogIndex = pBufEntry->prevLogIndex; SyncTerm prevLogTerm = pBufEntry->prevLogTerm; SSyncRaftEntry* pEntry = pBufEntry->pItem; if (pEntry == NULL) { sTrace("vgId:%d, cannot proceed match index in log buffer. no raft entry at next pos of matchIndex:%" PRId64, pNode->vgId, pBuf->matchIndex); goto _out; } ASSERT(index == pEntry->index); // match SSyncRaftEntry* pMatch = pBuf->entries[(pBuf->matchIndex + pBuf->size) % pBuf->size].pItem; ASSERT(pMatch != NULL); ASSERT(pMatch->index == pBuf->matchIndex); ASSERT(pMatch->index + 1 == pEntry->index); ASSERT(prevLogIndex == pMatch->index); if (pMatch->term != prevLogTerm) { sInfo( "vgId:%d, mismatching sync log entries encountered. " "{ index:%" PRId64 ", term:%" PRId64 " } " "{ index:%" PRId64 ", term:%" PRId64 ", prevLogIndex:%" PRId64 ", prevLogTerm:%" PRId64 " } ", pNode->vgId, pMatch->index, pMatch->term, pEntry->index, pEntry->term, prevLogIndex, prevLogTerm); goto _out; } // increase match index pBuf->matchIndex = index; sTrace("vgId:%d, log buffer proceed. start index: %" PRId64 ", match index: %" PRId64 ", end index: %" PRId64, pNode->vgId, pBuf->startIndex, pBuf->matchIndex, pBuf->endIndex); // replicate on demand (void)syncNodeReplicateWithoutLock(pNode); // persist if (syncLogStorePersist(pLogStore, pEntry) < 0) { sError("vgId:%d, failed to persist sync log entry from buffer since %s. index:%" PRId64, pNode->vgId, terrstr(), pEntry->index); goto _out; } ASSERT(pEntry->index == pBuf->matchIndex); // update my match index matchIndex = pBuf->matchIndex; syncIndexMgrSetIndex(pNode->pMatchIndex, &pNode->myRaftId, pBuf->matchIndex); } // end of while _out: pBuf->matchIndex = matchIndex; if (pMatchTerm) { *pMatchTerm = pBuf->entries[(matchIndex + pBuf->size) % pBuf->size].pItem->term; } syncLogBufferValidate(pBuf); taosThreadMutexUnlock(&pBuf->mutex); return matchIndex; } int32_t syncLogFsmExecute(SSyncNode* pNode, SSyncFSM* pFsm, ESyncState role, SyncTerm term, SSyncRaftEntry* pEntry) { ASSERTS(pFsm->FpCommitCb != NULL, "No commit cb registered for the FSM"); if ((pNode->replicaNum == 1) && pNode->restoreFinish && pNode->vgId != 1) { return 0; } if (pNode->vgId != 1 && vnodeIsMsgBlock(pEntry->originalRpcType)) { sTrace("vgId:%d, blocking msg ready to execute. index:%" PRId64 ", term: %" PRId64 ", type: %s", pNode->vgId, pEntry->index, pEntry->term, TMSG_INFO(pEntry->originalRpcType)); } SRpcMsg rpcMsg = {0}; syncEntry2OriginalRpc(pEntry, &rpcMsg); SFsmCbMeta cbMeta = {0}; cbMeta.index = pEntry->index; cbMeta.lastConfigIndex = syncNodeGetSnapshotConfigIndex(pNode, pEntry->index); cbMeta.isWeak = pEntry->isWeak; cbMeta.code = 0; cbMeta.state = role; cbMeta.seqNum = pEntry->seqNum; cbMeta.term = pEntry->term; cbMeta.currentTerm = term; cbMeta.flag = -1; (void)syncRespMgrGetAndDel(pNode->pSyncRespMgr, cbMeta.seqNum, &rpcMsg.info); int32_t code = pFsm->FpCommitCb(pFsm, &rpcMsg, &cbMeta); ASSERT(rpcMsg.pCont == NULL); return code; } int32_t syncLogBufferValidate(SSyncLogBuffer* pBuf) { ASSERT(pBuf->startIndex <= pBuf->matchIndex); ASSERT(pBuf->commitIndex <= pBuf->matchIndex); ASSERT(pBuf->matchIndex < pBuf->endIndex); ASSERT(pBuf->endIndex - pBuf->startIndex <= pBuf->size); ASSERT(pBuf->entries[(pBuf->matchIndex + pBuf->size) % pBuf->size].pItem); return 0; } int32_t syncLogBufferCommit(SSyncLogBuffer* pBuf, SSyncNode* pNode, int64_t commitIndex) { taosThreadMutexLock(&pBuf->mutex); syncLogBufferValidate(pBuf); SSyncLogStore* pLogStore = pNode->pLogStore; SSyncFSM* pFsm = pNode->pFsm; ESyncState role = pNode->state; SyncTerm term = pNode->pRaftStore->currentTerm; SyncGroupId vgId = pNode->vgId; int32_t ret = -1; int64_t upperIndex = TMIN(commitIndex, pBuf->matchIndex); SSyncRaftEntry* pEntry = NULL; bool inBuf = false; if (commitIndex <= pBuf->commitIndex) { sDebug("vgId:%d, stale commit index. current:%" PRId64 ", notified:%" PRId64 "", vgId, pBuf->commitIndex, commitIndex); ret = 0; goto _out; } sTrace("vgId:%d, commit. log buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 "), role: %d, term: %" PRId64, pNode->vgId, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex, role, term); // execute in fsm for (int64_t index = pBuf->commitIndex + 1; index <= upperIndex; index++) { // get a log entry pEntry = syncLogBufferGetOneEntry(pBuf, pNode, index, &inBuf); if (pEntry == NULL) { goto _out; } // execute it if (!syncUtilUserCommit(pEntry->originalRpcType)) { sInfo("vgId:%d, commit sync barrier. index: %" PRId64 ", term:%" PRId64 ", type: %s", vgId, pEntry->index, pEntry->term, TMSG_INFO(pEntry->originalRpcType)); } if (syncLogFsmExecute(pNode, pFsm, role, term, pEntry) != 0) { sError("vgId:%d, failed to execute sync log entry. index:%" PRId64 ", term:%" PRId64 ", role: %d, current term: %" PRId64, vgId, pEntry->index, pEntry->term, role, term); goto _out; } pBuf->commitIndex = index; sTrace("vgId:%d, committed index: %" PRId64 ", term: %" PRId64 ", role: %d, current term: %" PRId64 "", pNode->vgId, pEntry->index, pEntry->term, role, term); if (!inBuf) { syncEntryDestroy(pEntry); pEntry = NULL; } } // recycle SyncIndex until = pBuf->commitIndex - (pBuf->size >> 4); for (SyncIndex index = pBuf->startIndex; index < until; index++) { SSyncRaftEntry* pEntry = pBuf->entries[(index + pBuf->size) % pBuf->size].pItem; ASSERT(pEntry != NULL); syncEntryDestroy(pEntry); memset(&pBuf->entries[(index + pBuf->size) % pBuf->size], 0, sizeof(pBuf->entries[0])); pBuf->startIndex = index + 1; } ret = 0; _out: // mark as restored if needed if (!pNode->restoreFinish && pBuf->commitIndex >= pNode->commitIndex) { pNode->pFsm->FpRestoreFinishCb(pNode->pFsm); pNode->restoreFinish = true; sInfo("vgId:%d, restore finished. log buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); } if (!inBuf) { syncEntryDestroy(pEntry); pEntry = NULL; } syncLogBufferValidate(pBuf); taosThreadMutexUnlock(&pBuf->mutex); return ret; } int32_t syncLogReplMgrReset(SSyncLogReplMgr* pMgr) { ASSERT(pMgr->startIndex >= 0); for (SyncIndex index = pMgr->startIndex; index < pMgr->endIndex; index++) { memset(&pMgr->states[index % pMgr->size], 0, sizeof(pMgr->states[0])); } pMgr->startIndex = 0; pMgr->matchIndex = 0; pMgr->endIndex = 0; pMgr->restored = false; pMgr->retryBackoff = 0; return 0; } int32_t syncLogReplMgrRetryOnNeed(SSyncLogReplMgr* pMgr, SSyncNode* pNode) { if (pMgr->endIndex <= pMgr->startIndex) { return 0; } SRaftId* pDestId = &pNode->replicasId[pMgr->peerId]; if (pMgr->retryBackoff == SYNC_MAX_RETRY_BACKOFF) { syncLogReplMgrReset(pMgr); sWarn("vgId:%d, reset sync log repl mgr since retry backoff exceeding limit. peer: %" PRIx64, pNode->vgId, pDestId->addr); return -1; } int32_t ret = -1; bool retried = false; int64_t retryWaitMs = syncLogGetRetryBackoffTimeMs(pMgr); int64_t nowMs = taosGetMonoTimestampMs(); int count = 0; int64_t firstIndex = -1; SyncTerm term = -1; for (SyncIndex index = pMgr->startIndex; index < pMgr->endIndex; index++) { int64_t pos = index % pMgr->size; ASSERT(!pMgr->states[pos].barrier || (index == pMgr->startIndex || index + 1 == pMgr->endIndex)); if (nowMs < pMgr->states[pos].timeMs + retryWaitMs) { break; } if (pMgr->states[pos].acked) { continue; } bool barrier = false; if (syncLogBufferReplicateOneTo(pMgr, pNode, index, &term, pDestId, &barrier) < 0) { sError("vgId:%d, failed to replicate sync log entry since %s. index: %" PRId64 ", dest: %" PRIx64 "", pNode->vgId, terrstr(), index, pDestId->addr); goto _out; } ASSERT(barrier == pMgr->states[pos].barrier); pMgr->states[pos].timeMs = nowMs; pMgr->states[pos].term = term; pMgr->states[pos].acked = false; retried = true; if (firstIndex == -1) firstIndex = index; count++; } ret = 0; _out: if (retried) { pMgr->retryBackoff = syncLogGetNextRetryBackoff(pMgr); sInfo("vgId:%d, resent %d sync log entries. dest: %" PRIx64 ", indexes: %" PRId64 " ..., terms: ... %" PRId64 ", retryWaitMs: %" PRId64 ", repl mgr: [%" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, count, pDestId->addr, firstIndex, term, retryWaitMs, pMgr->startIndex, pMgr->matchIndex, pMgr->endIndex); } return ret; } int32_t syncLogReplMgrProcessReplyInRecoveryMode(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncAppendEntriesReply* pMsg) { SSyncLogBuffer* pBuf = pNode->pLogBuf; SRaftId destId = pMsg->srcId; ASSERT(pMgr->restored == false); char host[64]; uint16_t port; syncUtilU642Addr(destId.addr, host, sizeof(host), &port); if (pMgr->endIndex == 0) { ASSERT(pMgr->startIndex == 0); ASSERT(pMgr->matchIndex == 0); if (pMsg->matchIndex < 0) { pMgr->restored = true; sInfo("vgId:%d, sync log repl mgr restored. peer: %s:%d (%" PRIx64 "), mgr: rs(%d) [%" PRId64 " %" PRId64 ", %" PRId64 "), buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, host, port, destId.addr, pMgr->restored, pMgr->startIndex, pMgr->matchIndex, pMgr->endIndex, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); return 0; } } else { if (pMsg->lastSendIndex < pMgr->startIndex || pMsg->lastSendIndex >= pMgr->endIndex) { syncLogReplMgrRetryOnNeed(pMgr, pNode); return 0; } pMgr->states[pMsg->lastSendIndex % pMgr->size].acked = true; if (pMsg->success && pMsg->matchIndex == pMsg->lastSendIndex) { pMgr->matchIndex = pMsg->matchIndex; pMgr->restored = true; sInfo("vgId:%d, sync log repl mgr restored. peer: %s:%d (%" PRIx64 "), mgr: rs(%d) [%" PRId64 " %" PRId64 ", %" PRId64 "), buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, host, port, destId.addr, pMgr->restored, pMgr->startIndex, pMgr->matchIndex, pMgr->endIndex, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); return 0; } if (pMsg->success == false && pMsg->matchIndex >= pMsg->lastSendIndex) { sWarn("vgId:%d, failed to rollback match index. peer: %s:%d, match index: %" PRId64 ", last sent: %" PRId64, pNode->vgId, host, port, pMsg->matchIndex, pMsg->lastSendIndex); if (syncNodeStartSnapshot(pNode, &destId) < 0) { sError("vgId:%d, failed to start snapshot for peer %s:%d", pNode->vgId, host, port); return -1; } sInfo("vgId:%d, snapshot replication to peer %s:%d", pNode->vgId, host, port); return 0; } } // check last match term SyncTerm term = -1; SyncIndex firstVer = pNode->pLogStore->syncLogBeginIndex(pNode->pLogStore); SyncIndex index = TMIN(pMsg->matchIndex, pNode->pLogBuf->matchIndex); if (pMsg->matchIndex < pNode->pLogBuf->matchIndex) { term = syncLogReplMgrGetPrevLogTerm(pMgr, pNode, index + 1); if (term < 0 || (term != pMsg->lastMatchTerm && (index + 1 == firstVer || index == firstVer))) { ASSERT(term >= 0 || terrno == TSDB_CODE_WAL_LOG_NOT_EXIST); if (syncNodeStartSnapshot(pNode, &destId) < 0) { sError("vgId:%d, failed to start snapshot for peer %s:%d", pNode->vgId, host, port); return -1; } sInfo("vgId:%d, snapshot replication to peer %s:%d", pNode->vgId, host, port); return 0; } ASSERT(index + 1 >= firstVer); if (term == pMsg->lastMatchTerm) { index = index + 1; ASSERT(index <= pNode->pLogBuf->matchIndex); } else { ASSERT(index > firstVer); } } // attempt to replicate the raft log at index (void)syncLogReplMgrReset(pMgr); return syncLogReplMgrReplicateProbeOnce(pMgr, pNode, index); } int32_t syncLogReplMgrProcessHeartbeatReply(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncHeartbeatReply* pMsg) { SSyncLogBuffer* pBuf = pNode->pLogBuf; taosThreadMutexLock(&pBuf->mutex); if (pMsg->startTime != 0 && pMsg->startTime != pMgr->peerStartTime) { sInfo("vgId:%d, reset sync log repl mgr in heartbeat. peer: %" PRIx64 ", start time:%" PRId64 ", old:%" PRId64 "", pNode->vgId, pMsg->srcId.addr, pMsg->startTime, pMgr->peerStartTime); syncLogReplMgrReset(pMgr); pMgr->peerStartTime = pMsg->startTime; } taosThreadMutexUnlock(&pBuf->mutex); return 0; } int32_t syncLogReplMgrProcessReply(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncAppendEntriesReply* pMsg) { SSyncLogBuffer* pBuf = pNode->pLogBuf; taosThreadMutexLock(&pBuf->mutex); if (pMsg->startTime != pMgr->peerStartTime) { sInfo("vgId:%d, reset sync log repl mgr in appendlog reply. peer: %" PRIx64 ", start time:%" PRId64 ", old:%" PRId64, pNode->vgId, pMsg->srcId.addr, pMsg->startTime, pMgr->peerStartTime); syncLogReplMgrReset(pMgr); pMgr->peerStartTime = pMsg->startTime; } if (pMgr->restored) { (void)syncLogReplMgrProcessReplyInNormalMode(pMgr, pNode, pMsg); } else { (void)syncLogReplMgrProcessReplyInRecoveryMode(pMgr, pNode, pMsg); } taosThreadMutexUnlock(&pBuf->mutex); return 0; } int32_t syncLogReplMgrReplicateOnce(SSyncLogReplMgr* pMgr, SSyncNode* pNode) { if (pMgr->restored) { (void)syncLogReplMgrReplicateAttemptedOnce(pMgr, pNode); } else { (void)syncLogReplMgrReplicateProbeOnce(pMgr, pNode, pNode->pLogBuf->matchIndex); } return 0; } int32_t syncLogReplMgrReplicateProbeOnce(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncIndex index) { ASSERT(!pMgr->restored); ASSERT(pMgr->startIndex >= 0); int64_t retryMaxWaitMs = SYNC_LOG_REPL_RETRY_WAIT_MS * (1 << SYNC_MAX_RETRY_BACKOFF); int64_t nowMs = taosGetMonoTimestampMs(); if (pMgr->endIndex > pMgr->startIndex && nowMs < pMgr->states[pMgr->startIndex % pMgr->size].timeMs + retryMaxWaitMs) { return 0; } (void)syncLogReplMgrReset(pMgr); SRaftId* pDestId = &pNode->replicasId[pMgr->peerId]; bool barrier = false; SyncTerm term = -1; if (syncLogBufferReplicateOneTo(pMgr, pNode, index, &term, pDestId, &barrier) < 0) { sError("vgId:%d, failed to replicate log entry since %s. index: %" PRId64 ", dest: 0x%016" PRIx64 "", pNode->vgId, terrstr(), index, pDestId->addr); return -1; } ASSERT(index >= 0); pMgr->states[index % pMgr->size].barrier = barrier; pMgr->states[index % pMgr->size].timeMs = nowMs; pMgr->states[index % pMgr->size].term = term; pMgr->states[index % pMgr->size].acked = false; pMgr->startIndex = index; pMgr->endIndex = index + 1; SSyncLogBuffer* pBuf = pNode->pLogBuf; sTrace("vgId:%d, attempted to probe the %d'th peer with msg of index:%" PRId64 " term: %" PRId64 ". pMgr(rs:%d): [%" PRId64 " %" PRId64 ", %" PRId64 "), pBuf: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pMgr->peerId, index, term, pMgr->restored, pMgr->startIndex, pMgr->matchIndex, pMgr->endIndex, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); return 0; } int32_t syncLogReplMgrReplicateAttemptedOnce(SSyncLogReplMgr* pMgr, SSyncNode* pNode) { ASSERT(pMgr->restored); SRaftId* pDestId = &pNode->replicasId[pMgr->peerId]; int32_t batchSize = TMAX(1, pMgr->size / 20); int32_t count = 0; int64_t nowMs = taosGetMonoTimestampMs(); int64_t limit = pMgr->size >> 1; for (SyncIndex index = pMgr->endIndex; index <= pNode->pLogBuf->matchIndex; index++) { if (batchSize < count++ || limit <= index - pMgr->startIndex) { break; } if (pMgr->startIndex + 1 < index && pMgr->states[(index - 1) % pMgr->size].barrier) { break; } int64_t pos = index % pMgr->size; SRaftId* pDestId = &pNode->replicasId[pMgr->peerId]; bool barrier = false; SyncTerm term = -1; if (syncLogBufferReplicateOneTo(pMgr, pNode, index, &term, pDestId, &barrier) < 0) { sError("vgId:%d, failed to replicate log entry since %s. index: %" PRId64 ", dest: 0x%016" PRIx64 "", pNode->vgId, terrstr(), index, pDestId->addr); return -1; } pMgr->states[pos].barrier = barrier; pMgr->states[pos].timeMs = nowMs; pMgr->states[pos].term = term; pMgr->states[pos].acked = false; pMgr->endIndex = index + 1; if (barrier) { sInfo("vgId:%d, replicated sync barrier to dest: %" PRIx64 ". index: %" PRId64 ", term: %" PRId64 ", repl mgr: rs(%d) [%" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pDestId->addr, index, term, pMgr->restored, pMgr->startIndex, pMgr->matchIndex, pMgr->endIndex); break; } } syncLogReplMgrRetryOnNeed(pMgr, pNode); SSyncLogBuffer* pBuf = pNode->pLogBuf; sTrace("vgId:%d, attempted to replicate %d msgs to the %d'th peer. pMgr(rs:%d): [%" PRId64 " %" PRId64 ", %" PRId64 "), pBuf: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, count, pMgr->peerId, pMgr->restored, pMgr->startIndex, pMgr->matchIndex, pMgr->endIndex, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); return 0; } int32_t syncLogReplMgrProcessReplyInNormalMode(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncAppendEntriesReply* pMsg) { ASSERT(pMgr->restored == true); if (pMgr->startIndex <= pMsg->lastSendIndex && pMsg->lastSendIndex < pMgr->endIndex) { if (pMgr->startIndex < pMgr->matchIndex && pMgr->retryBackoff > 0) { int64_t firstSentMs = pMgr->states[pMgr->startIndex % pMgr->size].timeMs; int64_t lastSentMs = pMgr->states[(pMgr->endIndex - 1) % pMgr->size].timeMs; int64_t timeDiffMs = lastSentMs - firstSentMs; if (timeDiffMs > 0 && timeDiffMs < (SYNC_LOG_REPL_RETRY_WAIT_MS << (pMgr->retryBackoff - 1))) { pMgr->retryBackoff -= 1; } } pMgr->states[pMsg->lastSendIndex % pMgr->size].acked = true; pMgr->matchIndex = TMAX(pMgr->matchIndex, pMsg->matchIndex); for (SyncIndex index = pMgr->startIndex; index < pMgr->matchIndex; index++) { memset(&pMgr->states[index % pMgr->size], 0, sizeof(pMgr->states[0])); } pMgr->startIndex = pMgr->matchIndex; } return syncLogReplMgrReplicateAttemptedOnce(pMgr, pNode); } SSyncLogReplMgr* syncLogReplMgrCreate() { SSyncLogReplMgr* pMgr = taosMemoryCalloc(1, sizeof(SSyncLogReplMgr)); if (pMgr == NULL) { terrno = TSDB_CODE_OUT_OF_MEMORY; return NULL; } pMgr->size = sizeof(pMgr->states) / sizeof(pMgr->states[0]); ASSERT(pMgr->size == TSDB_SYNC_LOG_BUFFER_SIZE); return pMgr; _err: taosMemoryFree(pMgr); return NULL; } void syncLogReplMgrDestroy(SSyncLogReplMgr* pMgr) { if (pMgr == NULL) { return; } (void)taosMemoryFree(pMgr); return; } int32_t syncNodeLogReplMgrInit(SSyncNode* pNode) { for (int i = 0; i < TSDB_MAX_REPLICA; i++) { ASSERT(pNode->logReplMgrs[i] == NULL); pNode->logReplMgrs[i] = syncLogReplMgrCreate(); pNode->logReplMgrs[i]->peerId = i; ASSERTS(pNode->logReplMgrs[i] != NULL, "Out of memory."); } return 0; } void syncNodeLogReplMgrDestroy(SSyncNode* pNode) { for (int i = 0; i < TSDB_MAX_REPLICA; i++) { syncLogReplMgrDestroy(pNode->logReplMgrs[i]); pNode->logReplMgrs[i] = NULL; } } SSyncLogBuffer* syncLogBufferCreate() { SSyncLogBuffer* pBuf = taosMemoryCalloc(1, sizeof(SSyncLogBuffer)); if (pBuf == NULL) { terrno = TSDB_CODE_OUT_OF_MEMORY; return NULL; } pBuf->size = sizeof(pBuf->entries) / sizeof(pBuf->entries[0]); ASSERT(pBuf->size == TSDB_SYNC_LOG_BUFFER_SIZE); if (taosThreadMutexAttrInit(&pBuf->attr) < 0) { sError("failed to init log buffer mutexattr due to %s", strerror(errno)); terrno = TAOS_SYSTEM_ERROR(errno); goto _err; } if (taosThreadMutexAttrSetType(&pBuf->attr, PTHREAD_MUTEX_RECURSIVE) < 0) { sError("failed to set log buffer mutexattr type due to %s", strerror(errno)); terrno = TAOS_SYSTEM_ERROR(errno); goto _err; } if (taosThreadMutexInit(&pBuf->mutex, &pBuf->attr) < 0) { sError("failed to init log buffer mutex due to %s", strerror(errno)); terrno = TAOS_SYSTEM_ERROR(errno); goto _err; } return pBuf; _err: taosMemoryFree(pBuf); return NULL; } void syncLogBufferClear(SSyncLogBuffer* pBuf) { taosThreadMutexLock(&pBuf->mutex); for (SyncIndex index = pBuf->startIndex; index < pBuf->endIndex; index++) { SSyncRaftEntry* pEntry = pBuf->entries[(index + pBuf->size) % pBuf->size].pItem; if (pEntry == NULL) continue; syncEntryDestroy(pEntry); pEntry = NULL; memset(&pBuf->entries[(index + pBuf->size) % pBuf->size], 0, sizeof(pBuf->entries[0])); } pBuf->startIndex = pBuf->commitIndex = pBuf->matchIndex = pBuf->endIndex = 0; taosThreadMutexUnlock(&pBuf->mutex); } void syncLogBufferDestroy(SSyncLogBuffer* pBuf) { if (pBuf == NULL) { return; } syncLogBufferClear(pBuf); (void)taosThreadMutexDestroy(&pBuf->mutex); (void)taosThreadMutexAttrDestroy(&pBuf->attr); (void)taosMemoryFree(pBuf); return; } int32_t syncLogBufferRollback(SSyncLogBuffer* pBuf, SSyncNode* pNode, SyncIndex toIndex) { ASSERT(pBuf->commitIndex < toIndex && toIndex <= pBuf->endIndex); sInfo("vgId:%d, rollback sync log buffer. toindex: %" PRId64 ", buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, toIndex, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); // trunc buffer SyncIndex index = pBuf->endIndex - 1; while (index >= toIndex) { SSyncRaftEntry* pEntry = pBuf->entries[index % pBuf->size].pItem; if (pEntry != NULL) { (void)syncEntryDestroy(pEntry); pEntry = NULL; memset(&pBuf->entries[index % pBuf->size], 0, sizeof(pBuf->entries[0])); } index--; } pBuf->endIndex = toIndex; pBuf->matchIndex = TMIN(pBuf->matchIndex, index); ASSERT(index + 1 == toIndex); // trunc wal SyncIndex lastVer = pNode->pLogStore->syncLogLastIndex(pNode->pLogStore); if (lastVer >= toIndex && pNode->pLogStore->syncLogTruncate(pNode->pLogStore, toIndex) < 0) { sError("vgId:%d, failed to truncate log store since %s. from index:%" PRId64 "", pNode->vgId, terrstr(), toIndex); return -1; } lastVer = pNode->pLogStore->syncLogLastIndex(pNode->pLogStore); ASSERT(toIndex == lastVer + 1); // refill buffer on need if (toIndex <= pBuf->startIndex) { int32_t ret = syncLogBufferInitWithoutLock(pBuf, pNode); if (ret < 0) { sError("vgId:%d, failed to refill sync log buffer since %s", pNode->vgId, terrstr()); return -1; } } ASSERT(pBuf->endIndex == toIndex); syncLogBufferValidate(pBuf); return 0; } int32_t syncLogBufferReset(SSyncLogBuffer* pBuf, SSyncNode* pNode) { taosThreadMutexLock(&pBuf->mutex); SyncIndex lastVer = pNode->pLogStore->syncLogLastIndex(pNode->pLogStore); ASSERT(lastVer == pBuf->matchIndex); SyncIndex index = pBuf->endIndex - 1; (void)syncLogBufferRollback(pBuf, pNode, pBuf->matchIndex + 1); sInfo("vgId:%d, reset sync log buffer. buffer: [%" PRId64 " %" PRId64 " %" PRId64 ", %" PRId64 ")", pNode->vgId, pBuf->startIndex, pBuf->commitIndex, pBuf->matchIndex, pBuf->endIndex); pBuf->endIndex = pBuf->matchIndex + 1; // reset repl mgr for (int i = 0; i < pNode->replicaNum; i++) { SSyncLogReplMgr* pMgr = pNode->logReplMgrs[i]; syncLogReplMgrReset(pMgr); } taosThreadMutexUnlock(&pBuf->mutex); return 0; } SSyncRaftEntry* syncLogBufferGetOneEntry(SSyncLogBuffer* pBuf, SSyncNode* pNode, SyncIndex index, bool* pInBuf) { SSyncRaftEntry* pEntry = NULL; if (index >= pBuf->endIndex) { return NULL; } if (index > pBuf->startIndex) { // startIndex might be dummy *pInBuf = true; pEntry = pBuf->entries[index % pBuf->size].pItem; } else { *pInBuf = false; if (pNode->pLogStore->syncLogGetEntry(pNode->pLogStore, index, &pEntry) < 0) { sError("vgId:%d, failed to get log entry since %s. index:%" PRId64 "", pNode->vgId, terrstr(), index); } } return pEntry; } int32_t syncLogBufferReplicateOneTo(SSyncLogReplMgr* pMgr, SSyncNode* pNode, SyncIndex index, SyncTerm* pTerm, SRaftId* pDestId, bool* pBarrier) { SSyncRaftEntry* pEntry = NULL; SRpcMsg msgOut = {0}; bool inBuf = false; int32_t ret = -1; SyncTerm prevLogTerm = -1; SSyncLogBuffer* pBuf = pNode->pLogBuf; pEntry = syncLogBufferGetOneEntry(pBuf, pNode, index, &inBuf); if (pEntry == NULL) { sError("vgId:%d, failed to get raft entry for index: %" PRId64 "", pNode->vgId, index); if (terrno == TSDB_CODE_WAL_LOG_NOT_EXIST) { SSyncLogReplMgr* pMgr = syncNodeGetLogReplMgr(pNode, pDestId); if (pMgr) { sInfo("vgId:%d, reset sync log repl mgr of peer: %" PRIx64 " since %s. index: %" PRId64, pNode->vgId, pDestId->addr, terrstr(), index); (void)syncLogReplMgrReset(pMgr); } } goto _err; } *pBarrier = syncLogIsReplicationBarrier(pEntry); prevLogTerm = syncLogReplMgrGetPrevLogTerm(pMgr, pNode, index); if (prevLogTerm < 0) { sError("vgId:%d, failed to get prev log term since %s. index: %" PRId64 "", pNode->vgId, terrstr(), index); goto _err; } if (pTerm) *pTerm = pEntry->term; int32_t code = syncBuildAppendEntriesFromRaftLog(pNode, pEntry, prevLogTerm, &msgOut); if (code < 0) { sError("vgId:%d, failed to get append entries for index:%" PRId64 "", pNode->vgId, index); goto _err; } (void)syncNodeSendAppendEntries(pNode, pDestId, &msgOut); ret = 0; sTrace("vgId:%d, replicate one msg index: %" PRId64 " term: %" PRId64 " prevterm: %" PRId64 " to dest: 0x%016" PRIx64, pNode->vgId, pEntry->index, pEntry->term, prevLogTerm, pDestId->addr); if (!inBuf) { syncEntryDestroy(pEntry); pEntry = NULL; } return 0; _err: rpcFreeCont(msgOut.pCont); msgOut.pCont = NULL; if (!inBuf) { syncEntryDestroy(pEntry); pEntry = NULL; } return -1; }