src/replication.c

   1 #include "redis.h"
   2
   3 #include <sys/time.h>
   4 #include <unistd.h>
   5 #include <fcntl.h>
   6 #include <sys/stat.h>
   7
   8 /* ---------------------------------- MASTER -------------------------------- */
   9
  10 void replicationFeedSlaves(list *slaves, int dictid, robj **argv, int argc) {
  11     listNode *ln;
  12     listIter li;
  13     int j;
  14
  15     listRewind(slaves,&li);
  16     while((ln = listNext(&li))) {
  17         redisClient *slave = ln->value;
  18
  19         /* Don't feed slaves that are still waiting for BGSAVE to start */
  20         if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_START) continue;
  21
  22         /* Feed slaves that are waiting for the initial SYNC (so these commands
  23          * are queued in the output buffer until the intial SYNC completes),
  24          * or are already in sync with the master. */
  25         if (slave->slaveseldb != dictid) {
  26             robj *selectcmd;
  27
  28             switch(dictid) {
  29             case 0: selectcmd = shared.select0; break;
  30             case 1: selectcmd = shared.select1; break;
  31             case 2: selectcmd = shared.select2; break;
  32             case 3: selectcmd = shared.select3; break;
  33             case 4: selectcmd = shared.select4; break;
  34             case 5: selectcmd = shared.select5; break;
  35             case 6: selectcmd = shared.select6; break;
  36             case 7: selectcmd = shared.select7; break;
  37             case 8: selectcmd = shared.select8; break;
  38             case 9: selectcmd = shared.select9; break;
  39             default:
  40                 selectcmd = createObject(REDIS_STRING,
  41                     sdscatprintf(sdsempty(),"select %d\r\n",dictid));
  42                 selectcmd->refcount = 0;
  43                 break;
  44             }
  45             addReply(slave,selectcmd);
  46             slave->slaveseldb = dictid;
  47         }
  48         addReplyMultiBulkLen(slave,argc);
  49         for (j = 0; j < argc; j++) addReplyBulk(slave,argv[j]);
  50     }
  51 }
  52
  53 void replicationFeedMonitors(list *monitors, int dictid, robj **argv, int argc) {
  54     listNode *ln;
  55     listIter li;
  56     int j;
  57     sds cmdrepr = sdsnew("+");
  58     robj *cmdobj;
  59     struct timeval tv;
  60
  61     gettimeofday(&tv,NULL);
  62     cmdrepr = sdscatprintf(cmdrepr,"%ld.%06ld ",(long)tv.tv_sec,(long)tv.tv_usec);
  63     if (dictid != 0) cmdrepr = sdscatprintf(cmdrepr,"(db %d) ", dictid);
  64
  65     for (j = 0; j < argc; j++) {
  66         if (argv[j]->encoding == REDIS_ENCODING_INT) {
  67             cmdrepr = sdscatprintf(cmdrepr, "\"%ld\"", (long)argv[j]->ptr);
  68         } else {
  69             cmdrepr = sdscatrepr(cmdrepr,(char*)argv[j]->ptr,
  70                         sdslen(argv[j]->ptr));
  71         }
  72         if (j != argc-1)
  73             cmdrepr = sdscatlen(cmdrepr," ",1);
  74     }
  75     cmdrepr = sdscatlen(cmdrepr,"\r\n",2);
  76     cmdobj = createObject(REDIS_STRING,cmdrepr);
  77
  78     listRewind(monitors,&li);
  79     while((ln = listNext(&li))) {
  80         redisClient *monitor = ln->value;
  81         addReply(monitor,cmdobj);
  82     }
  83     decrRefCount(cmdobj);
  84 }
  85
  86 void syncCommand(redisClient *c) {
  87     /* ignore SYNC if aleady slave or in monitor mode */
  88     if (c->flags & REDIS_SLAVE) return;
  89
  90     /* Refuse SYNC requests if we are a slave but the link with our master
  91      * is not ok... */
  92     if (server.masterhost && server.replstate != REDIS_REPL_CONNECTED) {
  93         addReplyError(c,"Can't SYNC while not connected with my master");
  94         return;
  95     }
  96
  97     /* SYNC can't be issued when the server has pending data to send to
  98      * the client about already issued commands. We need a fresh reply
  99      * buffer registering the differences between the BGSAVE and the current
 100      * dataset, so that we can copy to other slaves if needed. */
 101     if (listLength(c->reply) != 0) {
 102         addReplyError(c,"SYNC is invalid with pending input");
 103         return;
 104     }
 105
 106     redisLog(REDIS_NOTICE,"Slave ask for synchronization");
 107     /* Here we need to check if there is a background saving operation
 108      * in progress, or if it is required to start one */
 109     if (server.bgsavechildpid != -1) {
 110         /* Ok a background save is in progress. Let's check if it is a good
 111          * one for replication, i.e. if there is another slave that is
 112          * registering differences since the server forked to save */
 113         redisClient *slave;
 114         listNode *ln;
 115         listIter li;
 116
 117         listRewind(server.slaves,&li);
 118         while((ln = listNext(&li))) {
 119             slave = ln->value;
 120             if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_END) break;
 121         }
 122         if (ln) {
 123             /* Perfect, the server is already registering differences for
 124              * another slave. Set the right state, and copy the buffer. */
 125             listRelease(c->reply);
 126             c->reply = listDup(slave->reply);
 127             c->replstate = REDIS_REPL_WAIT_BGSAVE_END;
 128             redisLog(REDIS_NOTICE,"Waiting for end of BGSAVE for SYNC");
 129         } else {
 130             /* No way, we need to wait for the next BGSAVE in order to
 131              * register differences */
 132             c->replstate = REDIS_REPL_WAIT_BGSAVE_START;
 133             redisLog(REDIS_NOTICE,"Waiting for next BGSAVE for SYNC");
 134         }
 135     } else {
 136         /* Ok we don't have a BGSAVE in progress, let's start one */
 137         redisLog(REDIS_NOTICE,"Starting BGSAVE for SYNC");
 138         if (rdbSaveBackground(server.dbfilename) != REDIS_OK) {
 139             redisLog(REDIS_NOTICE,"Replication failed, can't BGSAVE");
 140             addReplyError(c,"Unable to perform background save");
 141             return;
 142         }
 143         c->replstate = REDIS_REPL_WAIT_BGSAVE_END;
 144     }
 145     c->repldbfd = -1;
 146     c->flags |= REDIS_SLAVE;
 147     c->slaveseldb = 0;
 148     listAddNodeTail(server.slaves,c);
 149     return;
 150 }
 151
 152 void sendBulkToSlave(aeEventLoop *el, int fd, void *privdata, int mask) {
 153     redisClient *slave = privdata;
 154     REDIS_NOTUSED(el);
 155     REDIS_NOTUSED(mask);
 156     char buf[REDIS_IOBUF_LEN];
 157     ssize_t nwritten, buflen;
 158
 159     if (slave->repldboff == 0) {
 160         /* Write the bulk write count before to transfer the DB. In theory here
 161          * we don't know how much room there is in the output buffer of the
 162          * socket, but in pratice SO_SNDLOWAT (the minimum count for output
 163          * operations) will never be smaller than the few bytes we need. */
 164         sds bulkcount;
 165
 166         bulkcount = sdscatprintf(sdsempty(),"$%lld\r\n",(unsigned long long)
 167             slave->repldbsize);
 168         if (write(fd,bulkcount,sdslen(bulkcount)) != (signed)sdslen(bulkcount))
 169         {
 170             sdsfree(bulkcount);
 171             freeClient(slave);
 172             return;
 173         }
 174         sdsfree(bulkcount);
 175     }
 176     lseek(slave->repldbfd,slave->repldboff,SEEK_SET);
 177     buflen = read(slave->repldbfd,buf,REDIS_IOBUF_LEN);
 178     if (buflen <= 0) {
 179         redisLog(REDIS_WARNING,"Read error sending DB to slave: %s",
 180             (buflen == 0) ? "premature EOF" : strerror(errno));
 181         freeClient(slave);
 182         return;
 183     }
 184     if ((nwritten = write(fd,buf,buflen)) == -1) {
 185         redisLog(REDIS_VERBOSE,"Write error sending DB to slave: %s",
 186             strerror(errno));
 187         freeClient(slave);
 188         return;
 189     }
 190     slave->repldboff += nwritten;
 191     if (slave->repldboff == slave->repldbsize) {
 192         close(slave->repldbfd);
 193         slave->repldbfd = -1;
 194         aeDeleteFileEvent(server.el,slave->fd,AE_WRITABLE);
 195         slave->replstate = REDIS_REPL_ONLINE;
 196         if (aeCreateFileEvent(server.el, slave->fd, AE_WRITABLE,
 197             sendReplyToClient, slave) == AE_ERR) {
 198             freeClient(slave);
 199             return;
 200         }
 201         addReplySds(slave,sdsempty());
 202         redisLog(REDIS_NOTICE,"Synchronization with slave succeeded");
 203     }
 204 }
 205
 206 /* This function is called at the end of every backgrond saving.
 207  * The argument bgsaveerr is REDIS_OK if the background saving succeeded
 208  * otherwise REDIS_ERR is passed to the function.
 209  *
 210  * The goal of this function is to handle slaves waiting for a successful
 211  * background saving in order to perform non-blocking synchronization. */
 212 void updateSlavesWaitingBgsave(int bgsaveerr) {
 213     listNode *ln;
 214     int startbgsave = 0;
 215     listIter li;
 216
 217     listRewind(server.slaves,&li);
 218     while((ln = listNext(&li))) {
 219         redisClient *slave = ln->value;
 220
 221         if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_START) {
 222             startbgsave = 1;
 223             slave->replstate = REDIS_REPL_WAIT_BGSAVE_END;
 224         } else if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_END) {
 225             struct redis_stat buf;
 226
 227             if (bgsaveerr != REDIS_OK) {
 228                 freeClient(slave);
 229                 redisLog(REDIS_WARNING,"SYNC failed. BGSAVE child returned an error");
 230                 continue;
 231             }
 232             if ((slave->repldbfd = open(server.dbfilename,O_RDONLY)) == -1 ||
 233                 redis_fstat(slave->repldbfd,&buf) == -1) {
 234                 freeClient(slave);
 235                 redisLog(REDIS_WARNING,"SYNC failed. Can't open/stat DB after BGSAVE: %s", strerror(errno));
 236                 continue;
 237             }
 238             slave->repldboff = 0;
 239             slave->repldbsize = buf.st_size;
 240             slave->replstate = REDIS_REPL_SEND_BULK;
 241             aeDeleteFileEvent(server.el,slave->fd,AE_WRITABLE);
 242             if (aeCreateFileEvent(server.el, slave->fd, AE_WRITABLE, sendBulkToSlave, slave) == AE_ERR) {
 243                 freeClient(slave);
 244                 continue;
 245             }
 246         }
 247     }
 248     if (startbgsave) {
 249         if (rdbSaveBackground(server.dbfilename) != REDIS_OK) {
 250             listIter li;
 251
 252             listRewind(server.slaves,&li);
 253             redisLog(REDIS_WARNING,"SYNC failed. BGSAVE failed");
 254             while((ln = listNext(&li))) {
 255                 redisClient *slave = ln->value;
 256
 257                 if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_START)
 258                     freeClient(slave);
 259             }
 260         }
 261     }
 262 }
 263
 264 /* ----------------------------------- SLAVE -------------------------------- */
 265
 266 /* Abort the async download of the bulk dataset while SYNC-ing with master */
 267 void replicationAbortSyncTransfer(void) {
 268     redisAssert(server.replstate == REDIS_REPL_TRANSFER);
 269
 270     aeDeleteFileEvent(server.el,server.repl_transfer_s,AE_READABLE);
 271     close(server.repl_transfer_s);
 272     close(server.repl_transfer_fd);
 273     unlink(server.repl_transfer_tmpfile);
 274     zfree(server.repl_transfer_tmpfile);
 275     server.replstate = REDIS_REPL_CONNECT;
 276 }
 277
 278 /* Asynchronously read the SYNC payload we receive from a master */
 279 void readSyncBulkPayload(aeEventLoop *el, int fd, void *privdata, int mask) {
 280     char buf[4096];
 281     ssize_t nread, readlen;
 282     REDIS_NOTUSED(el);
 283     REDIS_NOTUSED(privdata);
 284     REDIS_NOTUSED(mask);
 285
 286     /* If repl_transfer_left == -1 we still have to read the bulk length
 287      * from the master reply. */
 288     if (server.repl_transfer_left == -1) {
 289         if (syncReadLine(fd,buf,1024,server.repl_syncio_timeout) == -1) {
 290             redisLog(REDIS_WARNING,
 291                 "I/O error reading bulk count from MASTER: %s",
 292                 strerror(errno));
 293             goto error;
 294         }
 295
 296         if (buf[0] == '-') {
 297             redisLog(REDIS_WARNING,
 298                 "MASTER aborted replication with an error: %s",
 299                 buf+1);
 300             goto error;
 301         } else if (buf[0] == '\0') {
 302             /* At this stage just a newline works as a PING in order to take
 303              * the connection live. So we refresh our last interaction
 304              * timestamp. */
 305             server.repl_transfer_lastio = time(NULL);
 306             return;
 307         } else if (buf[0] != '$') {
 308             redisLog(REDIS_WARNING,"Bad protocol from MASTER, the first byte is not '$', are you sure the host and port are right?");
 309             goto error;
 310         }
 311         server.repl_transfer_left = strtol(buf+1,NULL,10);
 312         redisLog(REDIS_NOTICE,
 313             "MASTER <-> SLAVE sync: receiving %ld bytes from master",
 314             server.repl_transfer_left);
 315         return;
 316     }
 317
 318     /* Read bulk data */
 319     readlen = (server.repl_transfer_left < (signed)sizeof(buf)) ?
 320         server.repl_transfer_left : (signed)sizeof(buf);
 321     nread = read(fd,buf,readlen);
 322     if (nread <= 0) {
 323         redisLog(REDIS_WARNING,"I/O error trying to sync with MASTER: %s",
 324             (nread == -1) ? strerror(errno) : "connection lost");
 325         replicationAbortSyncTransfer();
 326         return;
 327     }
 328     server.repl_transfer_lastio = time(NULL);
 329     if (write(server.repl_transfer_fd,buf,nread) != nread) {
 330         redisLog(REDIS_WARNING,"Write error or short write writing to the DB dump file needed for MASTER <-> SLAVE synchrnonization: %s", strerror(errno));
 331         goto error;
 332     }
 333     server.repl_transfer_left -= nread;
 334     /* Check if the transfer is now complete */
 335     if (server.repl_transfer_left == 0) {
 336         if (rename(server.repl_transfer_tmpfile,server.dbfilename) == -1) {
 337             redisLog(REDIS_WARNING,"Failed trying to rename the temp DB into dump.rdb in MASTER <-> SLAVE synchronization: %s", strerror(errno));
 338             replicationAbortSyncTransfer();
 339             return;
 340         }
 341         redisLog(REDIS_NOTICE, "MASTER <-> SLAVE sync: Loading DB in memory");
 342         emptyDb();
 343         /* Before loading the DB into memory we need to delete the readable
 344          * handler, otherwise it will get called recursively since
 345          * rdbLoad() will call the event loop to process events from time to
 346          * time for non blocking loading. */
 347         aeDeleteFileEvent(server.el,server.repl_transfer_s,AE_READABLE);
 348         if (rdbLoad(server.dbfilename) != REDIS_OK) {
 349             redisLog(REDIS_WARNING,"Failed trying to load the MASTER synchronization DB from disk");
 350             replicationAbortSyncTransfer();
 351             return;
 352         }
 353         /* Final setup of the connected slave <- master link */
 354         zfree(server.repl_transfer_tmpfile);
 355         close(server.repl_transfer_fd);
 356         server.master = createClient(server.repl_transfer_s);
 357         server.master->flags |= REDIS_MASTER;
 358         server.master->authenticated = 1;
 359         server.replstate = REDIS_REPL_CONNECTED;
 360         redisLog(REDIS_NOTICE, "MASTER <-> SLAVE sync: Finished with success");
 361         /* Rewrite the AOF file now that the dataset changed. */
 362         if (server.appendonly) rewriteAppendOnlyFileBackground();
 363     }
 364
 365     return;
 366
 367 error:
 368     replicationAbortSyncTransfer();
 369     return;
 370 }
 371
 372 void syncWithMaster(aeEventLoop *el, int fd, void *privdata, int mask) {
 373     char buf[1024], tmpfile[256];
 374     int dfd, maxtries = 5;
 375     REDIS_NOTUSED(el);
 376     REDIS_NOTUSED(privdata);
 377     REDIS_NOTUSED(mask);
 378
 379     redisLog(REDIS_NOTICE,"Non blocking connect for SYNC fired the event.");
 380     /* This event should only be triggered once since it is used to have a
 381      * non-blocking connect(2) to the master. It has been triggered when this
 382      * function is called, so we can delete it. */
 383     aeDeleteFileEvent(server.el,fd,AE_READABLE|AE_WRITABLE);
 384
 385     /* AUTH with the master if required. */
 386     if(server.masterauth) {
 387         char authcmd[1024];
 388         size_t authlen;
 389
 390         authlen = snprintf(authcmd,sizeof(authcmd),"AUTH %s\r\n",server.masterauth);
 391         if (syncWrite(fd,authcmd,authlen,server.repl_syncio_timeout) == -1) {
 392             redisLog(REDIS_WARNING,"Unable to AUTH to MASTER: %s",
 393                 strerror(errno));
 394             goto error;
 395         }
 396         /* Read the AUTH result.  */
 397         if (syncReadLine(fd,buf,1024,server.repl_syncio_timeout) == -1) {
 398             redisLog(REDIS_WARNING,"I/O error reading auth result from MASTER: %s",
 399                 strerror(errno));
 400             goto error;
 401         }
 402         if (buf[0] != '+') {
 403             redisLog(REDIS_WARNING,"Cannot AUTH to MASTER, is the masterauth password correct?");
 404             goto error;
 405         }
 406     }
 407
 408     /* Issue the SYNC command */
 409     if (syncWrite(fd,"SYNC \r\n",7,server.repl_syncio_timeout) == -1) {
 410         redisLog(REDIS_WARNING,"I/O error writing to MASTER: %s",
 411             strerror(errno));
 412         goto error;
 413     }
 414
 415     /* Prepare a suitable temp file for bulk transfer */
 416     while(maxtries--) {
 417         snprintf(tmpfile,256,
 418             "temp-%d.%ld.rdb",(int)time(NULL),(long int)getpid());
 419         dfd = open(tmpfile,O_CREAT|O_WRONLY|O_EXCL,0644);
 420         if (dfd != -1) break;
 421         sleep(1);
 422     }
 423     if (dfd == -1) {
 424         redisLog(REDIS_WARNING,"Opening the temp file needed for MASTER <-> SLAVE synchronization: %s",strerror(errno));
 425         goto error;
 426     }
 427
 428     /* Setup the non blocking download of the bulk file. */
 429     if (aeCreateFileEvent(server.el,fd, AE_READABLE,readSyncBulkPayload,NULL)
 430             == AE_ERR)
 431     {
 432         redisLog(REDIS_WARNING,"Can't create readable event for SYNC");
 433         goto error;
 434     }
 435
 436     server.replstate = REDIS_REPL_TRANSFER;
 437     server.repl_transfer_left = -1;
 438     server.repl_transfer_fd = dfd;
 439     server.repl_transfer_lastio = time(NULL);
 440     server.repl_transfer_tmpfile = zstrdup(tmpfile);
 441     return;
 442
 443 error:
 444     server.replstate = REDIS_REPL_CONNECT;
 445     close(fd);
 446     return;
 447 }
 448
 449 int connectWithMaster(void) {
 450     int fd;
 451
 452     fd = anetTcpNonBlockConnect(NULL,server.masterhost,server.masterport);
 453     if (fd == -1) {
 454         redisLog(REDIS_WARNING,"Unable to connect to MASTER: %s",
 455             strerror(errno));
 456         return REDIS_ERR;
 457     }
 458
 459     if (aeCreateFileEvent(server.el,fd,AE_READABLE|AE_WRITABLE,syncWithMaster,NULL) ==
 460             AE_ERR)
 461     {
 462         close(fd);
 463         redisLog(REDIS_WARNING,"Can't create readable event for SYNC");
 464         return REDIS_ERR;
 465     }
 466
 467     server.repl_transfer_s = fd;
 468     server.replstate = REDIS_REPL_CONNECTING;
 469     return REDIS_OK;
 470 }
 471
 472 void slaveofCommand(redisClient *c) {
 473     if (!strcasecmp(c->argv[1]->ptr,"no") &&
 474         !strcasecmp(c->argv[2]->ptr,"one")) {
 475         if (server.masterhost) {
 476             sdsfree(server.masterhost);
 477             server.masterhost = NULL;
 478             if (server.master) freeClient(server.master);
 479             if (server.replstate == REDIS_REPL_TRANSFER)
 480                 replicationAbortSyncTransfer();
 481             server.replstate = REDIS_REPL_NONE;
 482             redisLog(REDIS_NOTICE,"MASTER MODE enabled (user request)");
 483         }
 484     } else {
 485         sdsfree(server.masterhost);
 486         server.masterhost = sdsdup(c->argv[1]->ptr);
 487         server.masterport = atoi(c->argv[2]->ptr);
 488         if (server.master) freeClient(server.master);
 489         if (server.replstate == REDIS_REPL_TRANSFER)
 490             replicationAbortSyncTransfer();
 491         server.replstate = REDIS_REPL_CONNECT;
 492         redisLog(REDIS_NOTICE,"SLAVE OF %s:%d enabled (user request)",
 493             server.masterhost, server.masterport);
 494     }
 495     addReply(c,shared.ok);
 496 }
 497
 498 /* --------------------------- REPLICATION CRON  ---------------------------- */
 499
 500 #define REDIS_REPL_TIMEOUT 60
 501 #define REDIS_REPL_PING_SLAVE_PERIOD 10
 502
 503 void replicationCron(void) {
 504     /* Bulk transfer I/O timeout? */
 505     if (server.masterhost && server.replstate == REDIS_REPL_TRANSFER &&
 506         (time(NULL)-server.repl_transfer_lastio) > REDIS_REPL_TIMEOUT)
 507     {
 508         redisLog(REDIS_WARNING,"Timeout receiving bulk data from MASTER...");
 509         replicationAbortSyncTransfer();
 510     }
 511
 512     /* Timed out master when we are an already connected slave? */
 513     if (server.masterhost && server.replstate == REDIS_REPL_CONNECTED &&
 514         (time(NULL)-server.master->lastinteraction) > REDIS_REPL_TIMEOUT)
 515     {
 516         redisLog(REDIS_WARNING,"MASTER time out: no data nor PING received...");
 517         freeClient(server.master);
 518     }
 519
 520     /* Check if we should connect to a MASTER */
 521     if (server.replstate == REDIS_REPL_CONNECT) {
 522         redisLog(REDIS_NOTICE,"Connecting to MASTER...");
 523         if (connectWithMaster() == REDIS_OK) {
 524             redisLog(REDIS_NOTICE,"MASTER <-> SLAVE sync started");
 525         }
 526     }
 527
 528     /* If we have attached slaves, PING them from time to time.
 529      * So slaves can implement an explicit timeout to masters, and will
 530      * be able to detect a link disconnection even if the TCP connection
 531      * will not actually go down. */
 532     if (!(server.cronloops % (REDIS_REPL_PING_SLAVE_PERIOD*10))) {
 533         listIter li;
 534         listNode *ln;
 535
 536         listRewind(server.slaves,&li);
 537         while((ln = listNext(&li))) {
 538             redisClient *slave = ln->value;
 539
 540             /* Don't ping slaves that are in the middle of a bulk transfer
 541              * with the master for first synchronization. */
 542             if (slave->replstate == REDIS_REPL_SEND_BULK) continue;
 543             if (slave->replstate == REDIS_REPL_ONLINE) {
 544                 /* If the slave is online send a normal ping */
 545                 addReplySds(slave,sdsnew("PING\r\n"));
 546             } else {
 547                 /* Otherwise we are in the pre-synchronization stage.
 548                  * Just a newline will do the work of refreshing the
 549                  * connection last interaction time, and at the same time
 550                  * we'll be sure that being a single char there are no
 551                  * short-write problems. */
 552                 if (write(slave->fd, "\n", 1) == -1) {
 553                     /* Don't worry, it's just a ping. */
 554                 }
 555             }
 556         }
 557     }
 558 }