use fastlock for readwrite lock

* save replid for all masters in rdb

* expanded rdbSaveInfo to hold multiple master structs

* parse repl-masters from rdb

* recover replid info from rdb in active replica mode, attempt partial sync

* save offset from rdb into correct variable

* don't change replid based on master in active rep

* save and load psync info from correct fields

.github/workflows/endurance.yml

@@ -2,7 +2,7 @@ name: Endurance
 
 on:
   schedule:
-    - cron: '0 * * * *'
+    - cron: '30 * * * *'
 
 jobs:
 

pkg/deb/master_changelog

@@ -1,72 +1,3 @@
-keydb (6:6.2.2-1distribution_placeholder) codename_placeholder; urgency=medium
-
-  * Acquire lock in module.cpp to fix module test break
-  * Fix usage of nanosleep() on Apple hardware
-  * Fix networking deadlock (#183)
-  * Remove redundant zfree() in keydb-benchmark
-  * Issues resolved: #378, #384
-
--- Ben Schermel <ben@eqalpha.com>  Sat, 15 Jan 2022 20:00:37 +0000
-
-keydb (6:6.2.1-1distribution_placeholder) codename_placeholder; urgency=medium
-
-  * This release of KeyDB is at parity with Redis 6.2.6. In addition, the following changes were included:
-  * Systemd support for release packages 
-  * Fixes to buffer overflows in the `BITOP SHIFT` and `CRON` commands 
-  * Server times are now computed on a seperate thread to improve performance 
-  * Now enforces syslog identity and facility as soon as possible 
-  * Removed erroneous use of LOG_... flags when using syslog 
-  * Fixed erroneous `#endif` leading to build errors on some platforms 
-  * Fixed the incorrect counting of client connections 
-  * Issues resolved: #355, #370
-
- -- Ben Schermel <ben@eqalpha.com>  Wed, 17 Nov 2021 20:00:37 +0000
-
-keydb (6:6.2.0-1distribution_placeholder) codename_placeholder; urgency=medium
-
-  * Removed unused command line options(-a, -d, -P, -r, -q, --csv, -l, -I, -e, --precision, --cluster, --enable-tracking)
-  * --ms option changed to --time, added --clients, --host, --port, and --threads options
-  * Added keydb-diagnostic-tool
-  * Improved test reliability
-  * Fixed memory leak in keydb-benchmark
-  * Fixed a number of race conditions
-  * Fixed OSX build issues
-  * Feature parity with Redis 6.2.3
-  * Fixed Issues: #323 #325 #276
-
- -- Ben Schermel <ben@eqalpha.com>  Thu, 12 Aug 2021 20:00:37 +0000
-
-keydb (6:6.0.18-1distribution_placeholder) codename_placeholder; urgency=medium
-
-  * 6.0.18 - This release contains a number of improvements to stability and performance. 
-  * Additional tests were added with improvements to existing test reliability
-  * Merged Redis 6.0.10 into KeyDB
-  * Issues resolved in this release: #214, #222, #238, #257, #273, #276, #285
-
- -- Ben Schermel <ben@eqalpha.com>  Fri, 26 Mar 2021 16:00:00 +0000
-
-keydb (6:6.0.16-1distribution_placeholder) codename_placeholder; urgency=medium
-
-  * 6.0.16 - This release contains significant performance improvements as well as fixes to the following issues: #234, #236, #233, #207, #229, #231
-  * Reduced memory consumption when Active Replication is not used
-  * Reduced CPU consumption
-  * Improved TLS performance
-  * replica-quorum config for multi-master.Intended to be used with "replica-serve-stale-data no", if at least N replicas are connected we will consider our master link up and serve data.
-  * Subkey expires performance improvement for big sets
-
- -- Ben Schermel <ben@eqalpha.com>  Mon, 28 Sep 2020 16:00:00 +0000
-
-keydb (6:6.0.13-1distribution_placeholder) codename_placeholder; urgency=medium
-
-  * 6.0.13 New features added: KEYDB.MEXISTS, REPLICAOF REMOVE, multi-master-no-forward configuration option
-  * KEYDB.MEXISTS - return exactly which keys exist in the database not just the count
-  * REPLICAOF REMOVE - Remove specific masters in a multi master setup
-  * multi-master-no-forward configuration option - Reduce network traffic in mesh topology multimaster setups
-  * Redis 6.0.5 feature parity
-  * Fixed Issues: 194 Bad directive or wrong number of arguments error on 'lazyfree-lazy-user-del no'; 209 Timeout using modules; 210 Databases are not merged on multi-master sync; 211 rpm packages now signed; 212 rpm file permissions updated
-
- -- Ben Schermel <ben@eqalpha.com>  Mon, 13 Jul 2020 8:00:00 +0000
-
 keydb (6:6.0.9-1distribution_placeholder) codename_placeholder; urgency=medium
 
   * 6.0.9 Addressing issues 187 - cpu lockup with subkey expire, 190 - missing sentinel binary in keydb-tools. 

src/aof.cpp

@@ -879,7 +879,7 @@ int loadAppendOnlyFile(char *filename) {
     } else {
         /* RDB preamble. Pass loading the RDB functions. */
         rio rdb;
-        rdbSaveInfo rsi = RDB_SAVE_INFO_INIT;
+        rdbSaveInfo rsi;
 
         serverLog(LL_NOTICE,"Reading RDB preamble from AOF file...");
         if (fseek(fp,0,SEEK_SET) == -1) goto readerr;

src/module.cpp

@@ -365,7 +365,7 @@ typedef struct RedisModuleCommandFilter {
 static list *moduleCommandFilters;
 
 /* Module GIL Variables */
-static readWriteLock s_moduleGIL;
+static readWriteLock s_moduleGIL("Module GIL");
 thread_local bool g_fModuleThread = false;
 
 typedef void (*RedisModuleForkDoneHandler) (int exitcode, int bysignal, void *user_data);

src/networking.cpp

@@ -2463,8 +2463,6 @@ void processInputBuffer(client *c, int callFlags) {
 
         if (c->flags & CLIENT_EXECUTING_COMMAND) break;
 
-        if (c->flags & CLIENT_EXECUTING_COMMAND) break;
-
         /* Don't process input from the master while there is a busy script
          * condition on the replica. We want just to accumulate the replication
          * stream (instead of replying -BUSY like we do with other clients) and

src/rdb.cpp

@@ -1205,6 +1205,40 @@ int rdbSaveInfoAuxFields(rio *rdb, int rdbflags, rdbSaveInfo *rsi) {
             == -1) return -1;
         if (rdbSaveAuxFieldStrInt(rdb,"repl-offset",g_pserver->master_repl_offset)
             == -1) return -1;
+        if (g_pserver->fActiveReplica && listLength(g_pserver->masters) > 0) {
+            sdsstring val = sdsstring(sdsempty());
+            listNode *ln;
+            listIter li;
+            redisMaster* mi;
+            listRewind(g_pserver->masters,&li);
+            while ((ln = listNext(&li)) != NULL) {
+                mi = (redisMaster*)listNodeValue(ln);
+                if (!mi->master) {
+                    // If master client is not available, use info from master struct - better than nothing
+                    serverLog(LL_NOTICE, "saving master %s", mi->master_replid);
+                    if (mi->master_replid[0] == 0) {
+                        // if replid is null, there's no reason to save it
+                        continue;
+                    }
+                    val = val.catfmt("%s:%I:%s:%i;", mi->master_replid,
+                        mi->master_initial_offset,
+                        mi->masterhost,
+                        mi->masterport);
+                }
+                else {
+                    serverLog(LL_NOTICE, "saving master %s", mi->master->replid);
+                    if (mi->master->replid[0] == 0) {
+                        // if replid is null, there's no reason to save it
+                        continue;
+                    }
+                    val = val.catfmt("%s:%I:%s:%i;", mi->master->replid,
+                        mi->master->reploff,
+                        mi->masterhost,
+                        mi->masterport);
+                }
+            }
+            if (rdbSaveAuxFieldStrStr(rdb, "repl-masters",val.get()) == -1) return -1;
+        }
     }
     if (rdbSaveAuxFieldStrInt(rdb,"aof-preamble",aof_preamble) == -1) return -1;
     return 1;
@@ -2653,6 +2687,23 @@ int rdbLoadRio(rio *rdb, int rdbflags, rdbSaveInfo *rsi) {
                     memcpy(rsi->repl_id,ptrFromObj(auxval),CONFIG_RUN_ID_SIZE+1);
                     rsi->repl_id_is_set = 1;
                 }
+            } else if (!strcasecmp(szFromObj(auxkey),"repl-masters")) {
+                if (rsi) {
+                    struct redisMaster mi;
+                    char *masters = szFromObj(auxval);
+                    char *entry = strtok(masters, ":");
+                    while (entry != NULL) {
+                        memcpy(mi.master_replid, entry, sizeof(mi.master_replid));
+                        entry = strtok(NULL, ":");
+                        mi.master_initial_offset = atoi(entry);
+                        entry = strtok(NULL, ":");
+                        mi.masterhost = entry;
+                        entry = strtok(NULL, ";");
+                        mi.masterport = atoi(entry);
+                        entry = strtok(NULL, ":");
+                        rsi->addMaster(mi);
+                    }
+                }
             } else if (!strcasecmp(szFromObj(auxkey),"repl-offset")) {
                 if (rsi) rsi->repl_offset = strtoll(szFromObj(auxval),NULL,10);
             } else if (!strcasecmp(szFromObj(auxkey),"lua")) {
@@ -2817,11 +2868,11 @@ int rdbLoadRio(rio *rdb, int rdbflags, rdbSaveInfo *rsi) {
             bool fExpiredKey = iAmMaster() && !(rdbflags&RDBFLAGS_AOF_PREAMBLE) && expiretime != -1 && expiretime < now;
         
             if (fStaleMvccKey || fExpiredKey) {
-                if (fStaleMvccKey && !fExpiredKey && rsi != nullptr && rsi->mi != nullptr && rsi->mi->staleKeyMap != nullptr && lookupKeyRead(db, &keyobj) == nullptr) {
+                if (fStaleMvccKey && !fExpiredKey && rsi != nullptr && rsi->masters != nullptr && rsi->masters[0].staleKeyMap != nullptr && lookupKeyRead(db, &keyobj) == nullptr) {
                     // We have a key that we've already deleted and is not back in our database.
                     //  We'll need to inform the sending master of the delete if it is also a replica of us
                     robj_sharedptr objKeyDup(createStringObject(key, sdslen(key)));
-                    rsi->mi->staleKeyMap->operator[](db - g_pserver->db).push_back(objKeyDup);                
+                    rsi->masters[0].staleKeyMap->operator[](db - g_pserver->db).push_back(objKeyDup);                
                 }
                 fLastKeyExpired = true;
                 sdsfree(key);
@@ -3246,8 +3297,8 @@ void bgsaveCommand(client *c) {
  * is returned, and the RDB saving will not persist any replication related
  * information. */
 rdbSaveInfo *rdbPopulateSaveInfo(rdbSaveInfo *rsi) {
-    rdbSaveInfo rsi_init = RDB_SAVE_INFO_INIT;
-    *rsi = rsi_init;
+    rdbSaveInfo rsi_init;
+    *rsi = std::move(rsi_init);
 
     /* If the instance is a master, we can populate the replication info
      * only when repl_backlog is not NULL. If the repl_backlog is NULL,

src/readwritelock.h

@@ -2,22 +2,24 @@
 #include <condition_variable>
 
 class readWriteLock {
-    std::mutex m_readLock;
-    std::recursive_mutex m_writeLock;
-    std::condition_variable m_cv;
+    fastlock m_readLock;
+    fastlock m_writeLock;
+    std::condition_variable_any m_cv;
     int m_readCount = 0;
     int m_writeCount = 0;
     bool m_writeWaiting = false;
 public:
+    readWriteLock(const char *name) : m_readLock(name), m_writeLock(name) {}
+    
     void acquireRead() {
-        std::unique_lock<std::mutex> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock);
         while (m_writeCount > 0 || m_writeWaiting)
             m_cv.wait(rm);
         m_readCount++;
     }
     
     bool tryAcquireRead() {
-        std::unique_lock<std::mutex> rm(m_readLock, std::defer_lock);
+        std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
         if (!rm.try_lock())
             return false;
         if (m_writeCount > 0 || m_writeWaiting)
@@ -27,7 +29,7 @@ public:
     }
 
     void acquireWrite(bool exclusive = true) {
-        std::unique_lock<std::mutex> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock);
         m_writeWaiting = true;
         while (m_readCount > 0)
             m_cv.wait(rm);
@@ -43,7 +45,7 @@ public:
     }
 
     void upgradeWrite(bool exclusive = true) {
-        std::unique_lock<std::mutex> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock);
         m_writeWaiting = true;
         while (m_readCount > 1)
             m_cv.wait(rm);
@@ -60,7 +62,7 @@ public:
     }
 
     bool tryAcquireWrite(bool exclusive = true) {
-        std::unique_lock<std::mutex> rm(m_readLock, std::defer_lock);
+        std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
         if (!rm.try_lock())
             return false;
         if (m_readCount > 0)
@@ -73,14 +75,14 @@ public:
     }
 
     void releaseRead() {
-        std::unique_lock<std::mutex> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock);
         serverAssert(m_readCount > 0);
         m_readCount--;
         m_cv.notify_all();
     }
 
     void releaseWrite(bool exclusive = true) {
-        std::unique_lock<std::mutex> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock);
         serverAssert(m_writeCount > 0);
         if (exclusive)
             m_writeLock.unlock();
@@ -89,7 +91,7 @@ public:
     }
 
     void downgradeWrite(bool exclusive = true) {
-        std::unique_lock<std::mutex> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock);
         serverAssert(m_writeCount > 0);
         if (exclusive)
             m_writeLock.unlock();

src/replication.cpp

@@ -1987,7 +1987,7 @@ void readSyncBulkPayload(connection *conn) {
     ssize_t nread, readlen, nwritten;
     int use_diskless_load = useDisklessLoad();
     dbBackup *diskless_load_backup = NULL;
-    rdbSaveInfo rsi = RDB_SAVE_INFO_INIT;
+    rdbSaveInfo rsi;
     int empty_db_flags = g_pserver->repl_slave_lazy_flush ? EMPTYDB_ASYNC :
                                                         EMPTYDB_NO_FLAGS;
     off_t left;
@@ -2308,7 +2308,7 @@ void readSyncBulkPayload(connection *conn) {
                 mi->staleKeyMap->clear();
             else
                 mi->staleKeyMap = new (MALLOC_LOCAL) std::map<int, std::vector<robj_sharedptr>>();
-            rsi.mi = mi;
+            rsi.addMaster(*mi);
         }
         if (rdbLoadFile(rdb_filename,&rsi,RDBFLAGS_REPLICATION) != C_OK) {
             serverLog(LL_WARNING,
@@ -2359,7 +2359,7 @@ void readSyncBulkPayload(connection *conn) {
     {
         mergeReplicationId(mi->master->replid);
     }
-    else
+    else if (!g_pserver->fActiveReplica)
     {
         /* After a full resynchroniziation we use the replication ID and
         * offset of the master. The secondary ID / offset are cleared since
@@ -2555,7 +2555,7 @@ int slaveTryPartialResynchronization(redisMaster *mi, connection *conn, int read
          * client structure representing the master into g_pserver->master. */
         mi->master_initial_offset = -1;
 
-        if (mi->cached_master && !g_pserver->fActiveReplica) {
+        if (mi->cached_master) {
             psync_replid = mi->cached_master->replid;
             snprintf(psync_offset,sizeof(psync_offset),"%lld", mi->cached_master->reploff+1);
             serverLog(LL_NOTICE,"Trying a partial resynchronization (request %s:%s).", psync_replid, psync_offset);
@@ -2654,14 +2654,15 @@ int slaveTryPartialResynchronization(redisMaster *mi, connection *conn, int read
                     sizeof(g_pserver->replid2));
                 g_pserver->second_replid_offset = g_pserver->master_repl_offset+1;
 
-                /* Update the cached master ID and our own primary ID to the
-                 * new one. */
-                memcpy(g_pserver->replid,sznew,sizeof(g_pserver->replid));
-                memcpy(mi->cached_master->replid,sznew,sizeof(g_pserver->replid));
+                if (!g_pserver->fActiveReplica) {
+                    /* Update the cached master ID and our own primary ID to the
+                     * new one. */
+                    memcpy(g_pserver->replid,sznew,sizeof(g_pserver->replid));
+                    memcpy(mi->cached_master->replid,sznew,sizeof(g_pserver->replid));
 
-                /* Disconnect all the sub-slaves: they need to be notified. */
-                if (!g_pserver->fActiveReplica)
+                    /* Disconnect all the sub-slaves: they need to be notified. */
                     disconnectSlaves();
+                }
             }
         }
 
@@ -3022,18 +3023,6 @@ retry_connect:
         disconnectSlavesExcept(mi->master_uuid); /* Force our slaves to resync with us as well. */
         freeReplicationBacklog(); /* Don't allow our chained slaves to PSYNC. */
     }
-    else
-    {
-        if (listLength(g_pserver->slaves))
-        {
-            changeReplicationId();
-            clearReplicationId2();
-        }
-        else
-        {
-            freeReplicationBacklog(); /* Don't allow our chained slaves to PSYNC. */
-        }
-    }
 
     /* Fall back to SYNC if needed. Otherwise psync_result == PSYNC_FULLRESYNC
      * and the g_pserver->master_replid and master_initial_offset are
@@ -3679,6 +3668,26 @@ void replicationCacheMasterUsingMyself(redisMaster *mi) {
     mi->master = NULL;
 }
 
+/* This function is called when reloading master info from an RDB in Active Replica mode.
+ * It creates a cached master client using the info contained in the redisMaster struct.
+ *
+ * Assumes that the passed struct contains valid master info. */
+void replicationCacheMasterUsingMaster(redisMaster *mi) {
+    if (mi->cached_master) {
+        freeClient(mi->cached_master);
+    }
+
+    replicationCreateMasterClient(mi, NULL, -1);
+    std::lock_guard<decltype(mi->master->lock)> lock(mi->master->lock);
+
+    memcpy(mi->master->replid, mi->master_replid, sizeof(mi->master_replid));
+    mi->master->reploff = mi->master_initial_offset;
+
+    unlinkClient(mi->master);
+    mi->cached_master = mi->master;
+    mi->master = NULL;
+}
+
 /* Free a cached master, called when there are no longer the conditions for
  * a partial resync on reconnection. */
 void replicationDiscardCachedMaster(redisMaster *mi) {

src/sds.h

@@ -424,6 +424,21 @@ public:
         return *this;
     }
 
+    sdsstring &operator=(sdsstring &&other)
+    {
+        sds tmp = m_str;
+        m_str = other.m_str;
+        other.m_str = tmp;
+        return *this;
+    }
+
+    template<typename... Args>
+    sdsstring catfmt(const char *fmt, Args... args)
+    {
+        m_str = sdscatfmt(m_str, fmt, args...);
+        return *this;
+    }
+
     sds release() {
         sds sdsT = m_str;
         m_str = nullptr;

src/server.cpp

@@ -92,7 +92,7 @@ double R_Zero, R_PosInf, R_NegInf, R_Nan;
 /* Global vars */
 namespace GlobalHidden {
 struct redisServer server; /* Server global state */
-readWriteLock forkLock;
+readWriteLock forkLock("Fork (global)");
 }
 redisServer *g_pserver = &GlobalHidden::server;
 readWriteLock *g_forkLock = &GlobalHidden::forkLock;
@@ -2632,7 +2632,8 @@ void beforeSleep(struct aeEventLoop *eventLoop) {
         sleeping_threads++;
         serverAssert(sleeping_threads <= cserver.cthreads);
     }
-    
+
+    g_forkLock->releaseRead();    
     /* Determine whether the modules are enabled before sleeping, and use that result
        both here, and after wakeup to avoid double acquire or release of the GIL */
     serverTL->modulesEnabledThisAeLoop = !!moduleCount();
@@ -2653,6 +2654,7 @@ void afterSleep(struct aeEventLoop *eventLoop) {
        Otherwise you may double acquire the GIL and cause deadlocks in the module */
     if (!ProcessingEventsWhileBlocked) {
         if (serverTL->modulesEnabledThisAeLoop) moduleAcquireGIL(TRUE /*fServerThread*/);
+        g_forkLock->acquireRead();
         wakeTimeThread();
     }
 }
@@ -6290,7 +6292,8 @@ int redisFork(int purpose) {
         openChildInfoPipe();
     }
     long long startWriteLock = ustime();
-    g_forkLock->acquireWrite();
+    g_forkLock->releaseRead();
+    executeWithoutGlobalLock([](){ g_forkLock->acquireWrite(); });
     latencyAddSampleIfNeeded("fork-lock",(ustime()-startWriteLock)/1000);
     if ((childpid = fork()) == 0) {
         /* Child */
@@ -6301,6 +6304,7 @@ int redisFork(int purpose) {
     } else {
         /* Parent */
         g_forkLock->releaseWrite();
+        g_forkLock->acquireRead();
         g_pserver->stat_total_forks++;
         g_pserver->stat_fork_time = ustime()-start;
         g_pserver->stat_fork_rate = (double) zmalloc_used_memory() * 1000000 / g_pserver->stat_fork_time / (1024*1024*1024); /* GB per second. */
@@ -6363,7 +6367,7 @@ void loadDataFromDisk(void) {
         if (loadAppendOnlyFile(g_pserver->aof_filename) == C_OK)
             serverLog(LL_NOTICE,"DB loaded from append only file: %.3f seconds",(float)(ustime()-start)/1000000);
     } else if (g_pserver->rdb_filename != NULL || g_pserver->rdb_s3bucketpath != NULL) {
-        rdbSaveInfo rsi = RDB_SAVE_INFO_INIT;
+        rdbSaveInfo rsi;
         errno = 0; /* Prevent a stale value from affecting error checking */
         if (rdbLoad(&rsi,RDBFLAGS_NONE) == C_OK) {
             serverLog(LL_NOTICE,"DB loaded from disk: %.3f seconds",
@@ -6391,11 +6395,23 @@ void loadDataFromDisk(void) {
                 while ((ln = listNext(&li)))
                 {
                     redisMaster *mi = (redisMaster*)listNodeValue(ln);
-                    /* If we are a replica, create a cached master from this
-                    * information, in order to allow partial resynchronizations
-                    * with masters. */
-                    replicationCacheMasterUsingMyself(mi);
-                    selectDb(mi->cached_master,rsi.repl_stream_db);
+                    if (g_pserver->fActiveReplica) {
+                        for (size_t i = 0; i < rsi.numMasters(); i++) {
+                            if (!strcmp(mi->masterhost, rsi.masters[i].masterhost) && mi->masterport == rsi.masters[i].masterport) {
+                                memcpy(mi->master_replid, rsi.masters[i].master_replid, sizeof(mi->master_replid));
+                                mi->master_initial_offset = rsi.masters[i].master_initial_offset;
+                                replicationCacheMasterUsingMaster(mi);
+                                serverLog(LL_NOTICE, "Cached master recovered from RDB for %s:%d", mi->masterhost, mi->masterport);
+                            }
+                        }
+                    }
+                    else {
+                        /* If we are a replica, create a cached master from this
+                        * information, in order to allow partial resynchronizations
+                        * with masters. */
+                        replicationCacheMasterUsingMyself(mi);
+                        selectDb(mi->cached_master,rsi.repl_stream_db);
+                    }
                 }
             }
         } else if (errno != ENOENT) {
@@ -6660,6 +6676,7 @@ void *workerThreadMain(void *parg)
     }
 
     moduleAcquireGIL(true); // Normally afterSleep acquires this, but that won't be called on the first run
+    g_forkLock->acquireRead();
     aeEventLoop *el = g_pserver->rgthreadvar[iel].el;
     try
     {
@@ -6668,6 +6685,7 @@ void *workerThreadMain(void *parg)
     catch (ShutdownException)
     {
     }
+    g_forkLock->releaseRead();
     moduleReleaseGIL(true);
     serverAssert(!GlobalLocksAcquired());
     aeDeleteEventLoop(el);
@@ -6977,7 +6995,7 @@ int main(int argc, char **argv) {
             __AFL_INIT();
 #endif
             rio rdb;
-            rdbSaveInfo rsi = RDB_SAVE_INFO_INIT;
+            rdbSaveInfo rsi;
             startLoadingFile(stdin, (char*)"stdin", 0);
             rioInitWithFile(&rdb,stdin);
             rdbLoadRio(&rdb,0,&rsi);

src/server.h

@@ -1393,6 +1393,38 @@ struct redisMemOverhead {
     } *db;
 };
 
+struct redisMaster {
+    char *masteruser;               /* AUTH with this user and masterauth with master */
+    char *masterauth;               /* AUTH with this password with master */
+    char *masterhost;               /* Hostname of master */
+    int masterport;                 /* Port of master */
+    client *cached_master;          /* Cached master to be reused for PSYNC. */
+    client *master;
+    /* The following two fields is where we store master PSYNC replid/offset
+     * while the PSYNC is in progress. At the end we'll copy the fields into
+     * the server->master client structure. */
+    char master_replid[CONFIG_RUN_ID_SIZE+1];  /* Master PSYNC runid. */
+    long long master_initial_offset;           /* Master PSYNC offset. */
+
+    bool isActive = false;
+    int repl_state;          /* Replication status if the instance is a replica */
+    off_t repl_transfer_size; /* Size of RDB to read from master during sync. */
+    off_t repl_transfer_read; /* Amount of RDB read from master during sync. */
+    off_t repl_transfer_last_fsync_off; /* Offset when we fsync-ed last time. */
+    connection *repl_transfer_s;     /* Slave -> Master SYNC socket */
+    int repl_transfer_fd;    /* Slave -> Master SYNC temp file descriptor */
+    char *repl_transfer_tmpfile; /* Slave-> master SYNC temp file name */
+    time_t repl_transfer_lastio; /* Unix time of the latest read, for timeout */
+    time_t repl_down_since; /* Unix time at which link with master went down */
+
+    unsigned char master_uuid[UUID_BINARY_LEN];  /* Used during sync with master, this is our master's UUID */
+                                                /* After we've connected with our master use the UUID in g_pserver->master */
+    uint64_t mvccLastSync;
+    /* During a handshake the server may have stale keys, we track these here to share once a reciprocal connection is made */
+    std::map<int, std::vector<robj_sharedptr>> *staleKeyMap;
+    int ielReplTransfer = -1;
+};
+
 /* This structure can be optionally passed to RDB save/load functions in
  * order to implement additional functionalities, by storing and loading
  * metadata to the RDB file.
@@ -1401,7 +1433,66 @@ struct redisMemOverhead {
  * replication in order to make sure that chained slaves (slaves of slaves)
  * select the correct DB and are able to accept the stream coming from the
  * top-level master. */
-typedef struct rdbSaveInfo {
+class rdbSaveInfo {
+public:
+    rdbSaveInfo() {
+        repl_stream_db = -1;
+        repl_id_is_set = 0;
+        memcpy(repl_id, "0000000000000000000000000000000000000000", sizeof(repl_id));
+        repl_offset = -1;
+        fForceSetKey = TRUE;
+        mvccMinThreshold = 0;
+        masters = nullptr;
+        masterCount = 0;
+    }
+    rdbSaveInfo(const rdbSaveInfo &other) {
+        repl_stream_db = other.repl_stream_db;
+        repl_id_is_set = other.repl_id_is_set;
+        memcpy(repl_id, other.repl_id, sizeof(repl_id));
+        repl_offset = other.repl_offset;
+        fForceSetKey = other.fForceSetKey;
+        mvccMinThreshold = other.mvccMinThreshold;
+        masters = (struct redisMaster*)malloc(sizeof(struct redisMaster) * other.masterCount);
+        memcpy(masters, other.masters, sizeof(struct redisMaster) * other.masterCount);
+        masterCount = other.masterCount;
+    }
+    rdbSaveInfo(rdbSaveInfo &&other) : rdbSaveInfo() {
+        swap(*this, other);
+    }
+    rdbSaveInfo &operator=(rdbSaveInfo other) {
+        swap(*this, other);
+        return *this;
+    }
+    ~rdbSaveInfo() {
+        free(masters);
+    }
+    friend void swap(rdbSaveInfo &first, rdbSaveInfo &second) {
+        std::swap(first.repl_stream_db, second.repl_stream_db);
+        std::swap(first.repl_id_is_set, second.repl_id_is_set);
+        std::swap(first.repl_id, second.repl_id);
+        std::swap(first.repl_offset, second.repl_offset);
+        std::swap(first.fForceSetKey, second.fForceSetKey);
+        std::swap(first.mvccMinThreshold, second.mvccMinThreshold);
+        std::swap(first.masters, second.masters);
+        std::swap(first.masterCount, second.masterCount);
+
+    }
+
+    void addMaster(const struct redisMaster &mi) {
+        masterCount++;
+        if (masters == nullptr) {
+            masters = (struct redisMaster*)malloc(sizeof(struct redisMaster));
+        }
+        else {
+            masters = (struct redisMaster*)realloc(masters, sizeof(struct redisMaster) * masterCount);
+        }
+        memcpy(masters + masterCount - 1, &mi, sizeof(struct redisMaster));
+    }
+
+    size_t numMasters() {
+        return masterCount;
+    }
+
     /* Used saving and loading. */
     int repl_stream_db;  /* DB to select in g_pserver->master client. */
 
@@ -1411,10 +1502,11 @@ typedef struct rdbSaveInfo {
     long long repl_offset;                  /* Replication offset. */
     int fForceSetKey;
     uint64_t mvccMinThreshold;
-    struct redisMaster *mi;
-} rdbSaveInfo;
+    struct redisMaster *masters;
 
-#define RDB_SAVE_INFO_INIT {-1,0,"0000000000000000000000000000000000000000",-1, TRUE, 0, nullptr}
+private:
+    size_t masterCount;
+};
 
 struct malloc_stats {
     size_t zmalloc_used;
@@ -1503,38 +1595,6 @@ struct redisServerThreadVars {
     std::vector<client*> vecclientsProcess;
 };
 
-struct redisMaster {
-    char *masteruser;               /* AUTH with this user and masterauth with master */
-    char *masterauth;               /* AUTH with this password with master */
-    char *masterhost;               /* Hostname of master */
-    int masterport;                 /* Port of master */
-    client *cached_master;          /* Cached master to be reused for PSYNC. */
-    client *master;
-    /* The following two fields is where we store master PSYNC replid/offset
-     * while the PSYNC is in progress. At the end we'll copy the fields into
-     * the server->master client structure. */
-    char master_replid[CONFIG_RUN_ID_SIZE+1];  /* Master PSYNC runid. */
-    long long master_initial_offset;           /* Master PSYNC offset. */
-
-    bool isActive = false;
-    int repl_state;          /* Replication status if the instance is a replica */
-    off_t repl_transfer_size; /* Size of RDB to read from master during sync. */
-    off_t repl_transfer_read; /* Amount of RDB read from master during sync. */
-    off_t repl_transfer_last_fsync_off; /* Offset when we fsync-ed last time. */
-    connection *repl_transfer_s;     /* Slave -> Master SYNC socket */
-    int repl_transfer_fd;    /* Slave -> Master SYNC temp file descriptor */
-    char *repl_transfer_tmpfile; /* Slave-> master SYNC temp file name */
-    time_t repl_transfer_lastio; /* Unix time of the latest read, for timeout */
-    time_t repl_down_since; /* Unix time at which link with master went down */
-
-    unsigned char master_uuid[UUID_BINARY_LEN];  /* Used during sync with master, this is our master's UUID */
-                                                /* After we've connected with our master use the UUID in g_pserver->master */
-    uint64_t mvccLastSync;
-    /* During a handshake the server may have stale keys, we track these here to share once a reciprocal connection is made */
-    std::map<int, std::vector<robj_sharedptr>> *staleKeyMap;
-    int ielReplTransfer = -1;
-};
-
 // Const vars are not changed after worker threads are launched
 struct redisServerConst {
     pid_t pid;                  /* Main process pid. */
@@ -2446,6 +2506,7 @@ void clearReplicationId2(void);
 void mergeReplicationId(const char *);
 void chopReplicationBacklog(void);
 void replicationCacheMasterUsingMyself(struct redisMaster *mi);
+void replicationCacheMasterUsingMaster(struct redisMaster *mi);
 void feedReplicationBacklog(const void *ptr, size_t len);
 void updateMasterAuth();
 void showLatestBacklog();

src/version.h

@@ -1,4 +1,4 @@
-#define KEYDB_REAL_VERSION "6.2.2"
-#define KEYDB_VERSION_NUM 0x00060202
+#define KEYDB_REAL_VERSION "255.255.255"
+#define KEYDB_VERSION_NUM 0x00ffffff
 extern const char *KEYDB_SET_VERSION;   // Unlike real version, this can be overriden by the config
 

tests/unit/scripting.tcl

@@ -485,21 +485,6 @@ start_server {tags {"scripting"}} {
     }
 
 
-    test {EVAL with pipelined command (No crash)} {
-        r flushall
-        r config set lua-time-limit 1
-        set rd [redis_deferring_client]
-        $rd eval {for i=1,1000000 do redis.call('set', i, 'sdfdsfd') end} 0
-        $rd set testkey foo
-        $rd get testkey
-        after 1200
-        catch {r echo "foo"} err
-        assert_match {BUSY*} $err
-        $rd read
-        $rd close
-    }
-
-
     test {EVAL timeout from AOF} {
         # generate a long running script that is propagated to the AOF as script
         # make sure that the script times out during loading