src/cluster.c

   1 #include "redis.h"
   2
   3 #include <arpa/inet.h>
   4 #include <fcntl.h>
   5 #include <unistd.h>
   6
   7 void clusterAcceptHandler(aeEventLoop *el, int fd, void *privdata, int mask);
   8 void clusterReadHandler(aeEventLoop *el, int fd, void *privdata, int mask);
   9 void clusterSendPing(clusterLink *link, int type);
  10 void clusterSendFail(char *nodename);
  11 void clusterUpdateState(void);
  12 int clusterNodeGetSlotBit(clusterNode *n, int slot);
  13 sds clusterGenNodesDescription(void);
  14 clusterNode *clusterLookupNode(char *name);
  15 int clusterNodeAddSlave(clusterNode *master, clusterNode *slave);
  16 int clusterAddSlot(clusterNode *n, int slot);
  17
  18 /* -----------------------------------------------------------------------------
  19  * Initialization
  20  * -------------------------------------------------------------------------- */
  21
  22 void clusterGetRandomName(char *p) {
  23     FILE *fp = fopen("/dev/urandom","r");
  24     char *charset = "0123456789abcdef";
  25     int j;
  26
  27     if (!fp) {
  28         redisLog(REDIS_WARNING,
  29             "Unrecovarable error: can't open /dev/urandom:%s" ,strerror(errno));
  30         exit(1);
  31     }
  32     fread(p,REDIS_CLUSTER_NAMELEN,1,fp);
  33     for (j = 0; j < REDIS_CLUSTER_NAMELEN; j++)
  34         p[j] = charset[p[j] & 0x0F];
  35     fclose(fp);
  36 }
  37
  38 int clusterLoadConfig(char *filename) {
  39     FILE *fp = fopen(filename,"r");
  40     char *line;
  41     int maxline, j;
  42
  43     if (fp == NULL) return REDIS_ERR;
  44
  45     /* Parse the file. Note that single liens of the cluster config file can
  46      * be really long as they include all the hash slots of the node.
  47      * This means in the worst possible case REDIS_CLUSTER_SLOTS/2 integers.
  48      * To simplify we allocate 1024+REDIS_CLUSTER_SLOTS*16 bytes per line. */
  49     maxline = 1024+REDIS_CLUSTER_SLOTS*16;
  50     line = zmalloc(maxline);
  51     while(fgets(line,maxline,fp) != NULL) {
  52         int argc;
  53         sds *argv = sdssplitargs(line,&argc);
  54         clusterNode *n, *master;
  55         char *p, *s;
  56
  57         /* Create this node if it does not exist */
  58         n = clusterLookupNode(argv[0]);
  59         if (!n) {
  60             n = createClusterNode(argv[0],0);
  61             clusterAddNode(n);
  62         }
  63         /* Address and port */
  64         if ((p = strchr(argv[1],':')) == NULL) goto fmterr;
  65         *p = '\0';
  66         memcpy(n->ip,argv[1],strlen(argv[1])+1);
  67         n->port = atoi(p+1);
  68
  69         /* Parse flags */
  70         p = s = argv[2];
  71         while(p) {
  72             p = strchr(s,',');
  73             if (p) *p = '\0';
  74             if (!strcasecmp(s,"myself")) {
  75                 redisAssert(server.cluster.myself == NULL);
  76                 server.cluster.myself = n;
  77                 n->flags |= REDIS_NODE_MYSELF;
  78             } else if (!strcasecmp(s,"master")) {
  79                 n->flags |= REDIS_NODE_MASTER;
  80             } else if (!strcasecmp(s,"slave")) {
  81                 n->flags |= REDIS_NODE_SLAVE;
  82             } else if (!strcasecmp(s,"fail?")) {
  83                 n->flags |= REDIS_NODE_PFAIL;
  84             } else if (!strcasecmp(s,"fail")) {
  85                 n->flags |= REDIS_NODE_FAIL;
  86             } else if (!strcasecmp(s,"handshake")) {
  87                 n->flags |= REDIS_NODE_HANDSHAKE;
  88             } else if (!strcasecmp(s,"noaddr")) {
  89                 n->flags |= REDIS_NODE_NOADDR;
  90             } else if (!strcasecmp(s,"noflags")) {
  91                 /* nothing to do */
  92             } else {
  93                 redisPanic("Unknown flag in redis cluster config file");
  94             }
  95             if (p) s = p+1;
  96         }
  97
  98         /* Get master if any. Set the master and populate master's
  99          * slave list. */
 100         if (argv[3][0] != '-') {
 101             master = clusterLookupNode(argv[3]);
 102             if (!master) {
 103                 master = createClusterNode(argv[3],0);
 104                 clusterAddNode(master);
 105             }
 106             n->slaveof = master;
 107             clusterNodeAddSlave(master,n);
 108         }
 109
 110         /* Set ping sent / pong received timestamps */
 111         if (atoi(argv[4])) n->ping_sent = time(NULL);
 112         if (atoi(argv[5])) n->pong_received = time(NULL);
 113
 114         /* Populate hash slots served by this instance. */
 115         for (j = 7; j < argc; j++) {
 116             int start, stop;
 117
 118             if ((p = strchr(argv[j],'-')) != NULL) {
 119                 *p = '\0';
 120                 start = atoi(argv[j]);
 121                 stop = atoi(p+1);
 122             } else {
 123                 start = stop = atoi(argv[j]);
 124             }
 125             while(start <= stop) clusterAddSlot(n, start++);
 126         }
 127
 128         sdssplitargs_free(argv,argc);
 129     }
 130     zfree(line);
 131     fclose(fp);
 132
 133     /* Config sanity check */
 134     redisAssert(server.cluster.myself != NULL);
 135     redisLog(REDIS_NOTICE,"Node configuration loaded, I'm %.40s",
 136         server.cluster.myself->name);
 137     clusterUpdateState();
 138     return REDIS_OK;
 139
 140 fmterr:
 141     redisLog(REDIS_WARNING,"Unrecovarable error: corrupted cluster config file.");
 142     fclose(fp);
 143     exit(1);
 144 }
 145
 146 /* Cluster node configuration is exactly the same as CLUSTER NODES output.
 147  *
 148  * This function writes the node config and returns 0, on error -1
 149  * is returned. */
 150 int clusterSaveConfig(void) {
 151     sds ci = clusterGenNodesDescription();
 152     int fd;
 153
 154     if ((fd = open(server.cluster.configfile,O_WRONLY|O_CREAT|O_TRUNC,0644))
 155         == -1) goto err;
 156     if (write(fd,ci,sdslen(ci)) != (ssize_t)sdslen(ci)) goto err;
 157     close(fd);
 158     sdsfree(ci);
 159     return 0;
 160
 161 err:
 162     sdsfree(ci);
 163     return -1;
 164 }
 165
 166 void clusterSaveConfigOrDie(void) {
 167     if (clusterSaveConfig() == -1) {
 168         redisLog(REDIS_WARNING,"Fatal: can't update cluster config file.");
 169         exit(1);
 170     }
 171 }
 172
 173 void clusterInit(void) {
 174     int saveconf = 0;
 175
 176     server.cluster.myself = NULL;
 177     server.cluster.state = REDIS_CLUSTER_FAIL;
 178     server.cluster.nodes = dictCreate(&clusterNodesDictType,NULL);
 179     server.cluster.node_timeout = 15;
 180     memset(server.cluster.migrating_slots_to,0,
 181         sizeof(server.cluster.migrating_slots_to));
 182     memset(server.cluster.importing_slots_from,0,
 183         sizeof(server.cluster.importing_slots_from));
 184     memset(server.cluster.slots,0,
 185         sizeof(server.cluster.slots));
 186     if (clusterLoadConfig(server.cluster.configfile) == REDIS_ERR) {
 187         /* No configuration found. We will just use the random name provided
 188          * by the createClusterNode() function. */
 189         server.cluster.myself = createClusterNode(NULL,REDIS_NODE_MYSELF);
 190         redisLog(REDIS_NOTICE,"No cluster configuration found, I'm %.40s",
 191             server.cluster.myself->name);
 192         clusterAddNode(server.cluster.myself);
 193         saveconf = 1;
 194     }
 195     if (saveconf) clusterSaveConfigOrDie();
 196     /* We need a listening TCP port for our cluster messaging needs */
 197     server.cfd = anetTcpServer(server.neterr,
 198             server.port+REDIS_CLUSTER_PORT_INCR, server.bindaddr);
 199     if (server.cfd == -1) {
 200         redisLog(REDIS_WARNING, "Opening cluster TCP port: %s", server.neterr);
 201         exit(1);
 202     }
 203     if (aeCreateFileEvent(server.el, server.cfd, AE_READABLE,
 204         clusterAcceptHandler, NULL) == AE_ERR) oom("creating file event");
 205 }
 206
 207 /* -----------------------------------------------------------------------------
 208  * CLUSTER communication link
 209  * -------------------------------------------------------------------------- */
 210
 211 clusterLink *createClusterLink(clusterNode *node) {
 212     clusterLink *link = zmalloc(sizeof(*link));
 213     link->sndbuf = sdsempty();
 214     link->rcvbuf = sdsempty();
 215     link->node = node;
 216     link->fd = -1;
 217     return link;
 218 }
 219
 220 /* Free a cluster link, but does not free the associated node of course.
 221  * Just this function will make sure that the original node associated
 222  * with this link will have the 'link' field set to NULL. */
 223 void freeClusterLink(clusterLink *link) {
 224     if (link->fd != -1) {
 225         aeDeleteFileEvent(server.el, link->fd, AE_WRITABLE);
 226         aeDeleteFileEvent(server.el, link->fd, AE_READABLE);
 227     }
 228     sdsfree(link->sndbuf);
 229     sdsfree(link->rcvbuf);
 230     if (link->node)
 231         link->node->link = NULL;
 232     close(link->fd);
 233     zfree(link);
 234 }
 235
 236 void clusterAcceptHandler(aeEventLoop *el, int fd, void *privdata, int mask) {
 237     int cport, cfd;
 238     char cip[128];
 239     clusterLink *link;
 240     REDIS_NOTUSED(el);
 241     REDIS_NOTUSED(mask);
 242     REDIS_NOTUSED(privdata);
 243
 244     cfd = anetTcpAccept(server.neterr, fd, cip, &cport);
 245     if (cfd == AE_ERR) {
 246         redisLog(REDIS_VERBOSE,"Accepting cluster node: %s", server.neterr);
 247         return;
 248     }
 249     redisLog(REDIS_VERBOSE,"Accepted cluster node %s:%d", cip, cport);
 250     /* We need to create a temporary node in order to read the incoming
 251      * packet in a valid contest. This node will be released once we
 252      * read the packet and reply. */
 253     link = createClusterLink(NULL);
 254     link->fd = cfd;
 255     aeCreateFileEvent(server.el,cfd,AE_READABLE,clusterReadHandler,link);
 256 }
 257
 258 /* -----------------------------------------------------------------------------
 259  * Key space handling
 260  * -------------------------------------------------------------------------- */
 261
 262 /* We have 4096 hash slots. The hash slot of a given key is obtained
 263  * as the least significant 12 bits of the crc16 of the key. */
 264 unsigned int keyHashSlot(char *key, int keylen) {
 265     return crc16(key,keylen) & 0x0FFF;
 266 }
 267
 268 /* -----------------------------------------------------------------------------
 269  * CLUSTER node API
 270  * -------------------------------------------------------------------------- */
 271
 272 /* Create a new cluster node, with the specified flags.
 273  * If "nodename" is NULL this is considered a first handshake and a random
 274  * node name is assigned to this node (it will be fixed later when we'll
 275  * receive the first pong).
 276  *
 277  * The node is created and returned to the user, but it is not automatically
 278  * added to the nodes hash table. */
 279 clusterNode *createClusterNode(char *nodename, int flags) {
 280     clusterNode *node = zmalloc(sizeof(*node));
 281
 282     if (nodename)
 283         memcpy(node->name, nodename, REDIS_CLUSTER_NAMELEN);
 284     else
 285         clusterGetRandomName(node->name);
 286     node->flags = flags;
 287     memset(node->slots,0,sizeof(node->slots));
 288     node->numslaves = 0;
 289     node->slaves = NULL;
 290     node->slaveof = NULL;
 291     node->ping_sent = node->pong_received = 0;
 292     node->configdigest = NULL;
 293     node->configdigest_ts = 0;
 294     node->link = NULL;
 295     return node;
 296 }
 297
 298 int clusterNodeRemoveSlave(clusterNode *master, clusterNode *slave) {
 299     int j;
 300
 301     for (j = 0; j < master->numslaves; j++) {
 302         if (master->slaves[j] == slave) {
 303             memmove(master->slaves+j,master->slaves+(j+1),
 304                 (master->numslaves-1)-j);
 305             master->numslaves--;
 306             return REDIS_OK;
 307         }
 308     }
 309     return REDIS_ERR;
 310 }
 311
 312 int clusterNodeAddSlave(clusterNode *master, clusterNode *slave) {
 313     int j;
 314
 315     /* If it's already a slave, don't add it again. */
 316     for (j = 0; j < master->numslaves; j++)
 317         if (master->slaves[j] == slave) return REDIS_ERR;
 318     master->slaves = zrealloc(master->slaves,
 319         sizeof(clusterNode*)*(master->numslaves+1));
 320     master->slaves[master->numslaves] = slave;
 321     master->numslaves++;
 322     return REDIS_OK;
 323 }
 324
 325 void clusterNodeResetSlaves(clusterNode *n) {
 326     zfree(n->slaves);
 327     n->numslaves = 0;
 328 }
 329
 330 void freeClusterNode(clusterNode *n) {
 331     sds nodename;
 332
 333     nodename = sdsnewlen(n->name, REDIS_CLUSTER_NAMELEN);
 334     redisAssert(dictDelete(server.cluster.nodes,nodename) == DICT_OK);
 335     sdsfree(nodename);
 336     if (n->slaveof) clusterNodeRemoveSlave(n->slaveof, n);
 337     if (n->link) freeClusterLink(n->link);
 338     zfree(n);
 339 }
 340
 341 /* Add a node to the nodes hash table */
 342 int clusterAddNode(clusterNode *node) {
 343     int retval;
 344
 345     retval = dictAdd(server.cluster.nodes,
 346             sdsnewlen(node->name,REDIS_CLUSTER_NAMELEN), node);
 347     return (retval == DICT_OK) ? REDIS_OK : REDIS_ERR;
 348 }
 349
 350 /* Node lookup by name */
 351 clusterNode *clusterLookupNode(char *name) {
 352     sds s = sdsnewlen(name, REDIS_CLUSTER_NAMELEN);
 353     struct dictEntry *de;
 354
 355     de = dictFind(server.cluster.nodes,s);
 356     sdsfree(s);
 357     if (de == NULL) return NULL;
 358     return dictGetEntryVal(de);
 359 }
 360
 361 /* This is only used after the handshake. When we connect a given IP/PORT
 362  * as a result of CLUSTER MEET we don't have the node name yet, so we
 363  * pick a random one, and will fix it when we receive the PONG request using
 364  * this function. */
 365 void clusterRenameNode(clusterNode *node, char *newname) {
 366     int retval;
 367     sds s = sdsnewlen(node->name, REDIS_CLUSTER_NAMELEN);
 368
 369     redisLog(REDIS_DEBUG,"Renaming node %.40s into %.40s",
 370         node->name, newname);
 371     retval = dictDelete(server.cluster.nodes, s);
 372     sdsfree(s);
 373     redisAssert(retval == DICT_OK);
 374     memcpy(node->name, newname, REDIS_CLUSTER_NAMELEN);
 375     clusterAddNode(node);
 376 }
 377
 378 /* -----------------------------------------------------------------------------
 379  * CLUSTER messages exchange - PING/PONG and gossip
 380  * -------------------------------------------------------------------------- */
 381
 382 /* Process the gossip section of PING or PONG packets.
 383  * Note that this function assumes that the packet is already sanity-checked
 384  * by the caller, not in the content of the gossip section, but in the
 385  * length. */
 386 void clusterProcessGossipSection(clusterMsg *hdr, clusterLink *link) {
 387     uint16_t count = ntohs(hdr->count);
 388     clusterMsgDataGossip *g = (clusterMsgDataGossip*) hdr->data.ping.gossip;
 389     clusterNode *sender = link->node ? link->node : clusterLookupNode(hdr->sender);
 390
 391     while(count--) {
 392         sds ci = sdsempty();
 393         uint16_t flags = ntohs(g->flags);
 394         clusterNode *node;
 395
 396         if (flags == 0) ci = sdscat(ci,"noflags,");
 397         if (flags & REDIS_NODE_MYSELF) ci = sdscat(ci,"myself,");
 398         if (flags & REDIS_NODE_MASTER) ci = sdscat(ci,"master,");
 399         if (flags & REDIS_NODE_SLAVE) ci = sdscat(ci,"slave,");
 400         if (flags & REDIS_NODE_PFAIL) ci = sdscat(ci,"fail?,");
 401         if (flags & REDIS_NODE_FAIL) ci = sdscat(ci,"fail,");
 402         if (flags & REDIS_NODE_HANDSHAKE) ci = sdscat(ci,"handshake,");
 403         if (flags & REDIS_NODE_NOADDR) ci = sdscat(ci,"noaddr,");
 404         if (ci[sdslen(ci)-1] == ',') ci[sdslen(ci)-1] = ' ';
 405
 406         redisLog(REDIS_DEBUG,"GOSSIP %.40s %s:%d %s",
 407             g->nodename,
 408             g->ip,
 409             ntohs(g->port),
 410             ci);
 411         sdsfree(ci);
 412
 413         /* Update our state accordingly to the gossip sections */
 414         node = clusterLookupNode(g->nodename);
 415         if (node != NULL) {
 416             /* We already know this node. Let's start updating the last
 417              * time PONG figure if it is newer than our figure.
 418              * Note that it's not a problem if we have a PING already
 419              * in progress against this node. */
 420             if (node->pong_received < ntohl(g->pong_received)) {
 421                  redisLog(REDIS_DEBUG,"Node pong_received updated by gossip");
 422                 node->pong_received = ntohl(g->pong_received);
 423             }
 424             /* Mark this node as FAILED if we think it is possibly failing
 425              * and another node also thinks it's failing. */
 426             if (node->flags & REDIS_NODE_PFAIL &&
 427                 (flags & (REDIS_NODE_FAIL|REDIS_NODE_PFAIL)))
 428             {
 429                 redisLog(REDIS_NOTICE,"Received a PFAIL acknowledge from node %.40s, marking node %.40s as FAIL!", hdr->sender, node->name);
 430                 node->flags &= ~REDIS_NODE_PFAIL;
 431                 node->flags |= REDIS_NODE_FAIL;
 432                 /* Broadcast the failing node name to everybody */
 433                 clusterSendFail(node->name);
 434                 clusterUpdateState();
 435                 clusterSaveConfigOrDie();
 436             }
 437         } else {
 438             /* If it's not in NOADDR state and we don't have it, we
 439              * start an handshake process against this IP/PORT pairs.
 440              *
 441              * Note that we require that the sender of this gossip message
 442              * is a well known node in our cluster, otherwise we risk
 443              * joining another cluster. */
 444             if (sender && !(flags & REDIS_NODE_NOADDR)) {
 445                 clusterNode *newnode;
 446
 447                 redisLog(REDIS_DEBUG,"Adding the new node");
 448                 newnode = createClusterNode(NULL,REDIS_NODE_HANDSHAKE);
 449                 memcpy(newnode->ip,g->ip,sizeof(g->ip));
 450                 newnode->port = ntohs(g->port);
 451                 clusterAddNode(newnode);
 452             }
 453         }
 454
 455         /* Next node */
 456         g++;
 457     }
 458 }
 459
 460 /* IP -> string conversion. 'buf' is supposed to at least be 16 bytes. */
 461 void nodeIp2String(char *buf, clusterLink *link) {
 462     struct sockaddr_in sa;
 463     socklen_t salen = sizeof(sa);
 464
 465     if (getpeername(link->fd, (struct sockaddr*) &sa, &salen) == -1)
 466         redisPanic("getpeername() failed.");
 467     strncpy(buf,inet_ntoa(sa.sin_addr),sizeof(link->node->ip));
 468 }
 469
 470
 471 /* Update the node address to the IP address that can be extracted
 472  * from link->fd, and at the specified port. */
 473 void nodeUpdateAddress(clusterNode *node, clusterLink *link, int port) {
 474 }
 475
 476 /* When this function is called, there is a packet to process starting
 477  * at node->rcvbuf. Releasing the buffer is up to the caller, so this
 478  * function should just handle the higher level stuff of processing the
 479  * packet, modifying the cluster state if needed.
 480  *
 481  * The function returns 1 if the link is still valid after the packet
 482  * was processed, otherwise 0 if the link was freed since the packet
 483  * processing lead to some inconsistency error (for instance a PONG
 484  * received from the wrong sender ID). */
 485 int clusterProcessPacket(clusterLink *link) {
 486     clusterMsg *hdr = (clusterMsg*) link->rcvbuf;
 487     uint32_t totlen = ntohl(hdr->totlen);
 488     uint16_t type = ntohs(hdr->type);
 489     clusterNode *sender;
 490
 491     redisLog(REDIS_DEBUG,"--- packet to process %lu bytes (%lu) ---",
 492         (unsigned long) totlen, sdslen(link->rcvbuf));
 493     if (totlen < 8) return 1;
 494     if (totlen > sdslen(link->rcvbuf)) return 1;
 495     if (type == CLUSTERMSG_TYPE_PING || type == CLUSTERMSG_TYPE_PONG ||
 496         type == CLUSTERMSG_TYPE_MEET)
 497     {
 498         uint16_t count = ntohs(hdr->count);
 499         uint32_t explen; /* expected length of this packet */
 500
 501         explen = sizeof(clusterMsg)-sizeof(union clusterMsgData);
 502         explen += (sizeof(clusterMsgDataGossip)*count);
 503         if (totlen != explen) return 1;
 504     }
 505     if (type == CLUSTERMSG_TYPE_FAIL) {
 506         uint32_t explen = sizeof(clusterMsg)-sizeof(union clusterMsgData);
 507
 508         explen += sizeof(clusterMsgDataFail);
 509         if (totlen != explen) return 1;
 510     }
 511
 512     sender = clusterLookupNode(hdr->sender);
 513     if (type == CLUSTERMSG_TYPE_PING || type == CLUSTERMSG_TYPE_MEET) {
 514         int update_config = 0;
 515         redisLog(REDIS_DEBUG,"Ping packet received: %p", link->node);
 516
 517         /* Add this node if it is new for us and the msg type is MEET.
 518          * In this stage we don't try to add the node with the right
 519          * flags, slaveof pointer, and so forth, as this details will be
 520          * resolved when we'll receive PONGs from the server. */
 521         if (!sender && type == CLUSTERMSG_TYPE_MEET) {
 522             clusterNode *node;
 523
 524             node = createClusterNode(NULL,REDIS_NODE_HANDSHAKE);
 525             nodeIp2String(node->ip,link);
 526             node->port = ntohs(hdr->port);
 527             clusterAddNode(node);
 528             update_config = 1;
 529         }
 530
 531         /* Get info from the gossip section */
 532         clusterProcessGossipSection(hdr,link);
 533
 534         /* Anyway reply with a PONG */
 535         clusterSendPing(link,CLUSTERMSG_TYPE_PONG);
 536
 537         /* Update config if needed */
 538         if (update_config) clusterSaveConfigOrDie();
 539     } else if (type == CLUSTERMSG_TYPE_PONG) {
 540         int update_state = 0;
 541         int update_config = 0;
 542
 543         redisLog(REDIS_DEBUG,"Pong packet received: %p", link->node);
 544         if (link->node) {
 545             if (link->node->flags & REDIS_NODE_HANDSHAKE) {
 546                 /* If we already have this node, try to change the
 547                  * IP/port of the node with the new one. */
 548                 if (sender) {
 549                     redisLog(REDIS_WARNING,
 550                         "Handshake error: we already know node %.40s, updating the address if needed.", sender->name);
 551                     nodeUpdateAddress(sender,link,ntohs(hdr->port));
 552                     freeClusterNode(link->node); /* will free the link too */
 553                     return 0;
 554                 }
 555
 556                 /* First thing to do is replacing the random name with the
 557                  * right node name if this was an handshake stage. */
 558                 clusterRenameNode(link->node, hdr->sender);
 559                 redisLog(REDIS_DEBUG,"Handshake with node %.40s completed.",
 560                     link->node->name);
 561                 link->node->flags &= ~REDIS_NODE_HANDSHAKE;
 562                 update_config = 1;
 563             } else if (memcmp(link->node->name,hdr->sender,
 564                         REDIS_CLUSTER_NAMELEN) != 0)
 565             {
 566                 /* If the reply has a non matching node ID we
 567                  * disconnect this node and set it as not having an associated
 568                  * address. */
 569                 redisLog(REDIS_DEBUG,"PONG contains mismatching sender ID");
 570                 link->node->flags |= REDIS_NODE_NOADDR;
 571                 freeClusterLink(link);
 572                 update_config = 1;
 573                 /* FIXME: remove this node if we already have it.
 574                  *
 575                  * If we already have it but the IP is different, use
 576                  * the new one if the old node is in FAIL, PFAIL, or NOADDR
 577                  * status... */
 578                 return 0;
 579             }
 580         }
 581         /* Update our info about the node */
 582         link->node->pong_received = time(NULL);
 583
 584         /* Update master/slave info */
 585         if (sender) {
 586             if (!memcmp(hdr->slaveof,REDIS_NODE_NULL_NAME,
 587                 sizeof(hdr->slaveof)))
 588             {
 589                 sender->flags &= ~REDIS_NODE_SLAVE;
 590                 sender->flags |= REDIS_NODE_MASTER;
 591                 sender->slaveof = NULL;
 592             } else {
 593                 clusterNode *master = clusterLookupNode(hdr->slaveof);
 594
 595                 sender->flags &= ~REDIS_NODE_MASTER;
 596                 sender->flags |= REDIS_NODE_SLAVE;
 597                 if (sender->numslaves) clusterNodeResetSlaves(sender);
 598                 if (master) clusterNodeAddSlave(master,sender);
 599             }
 600         }
 601
 602         /* Update our info about served slots if this new node is serving
 603          * slots that are not served from our point of view. */
 604         if (sender && sender->flags & REDIS_NODE_MASTER) {
 605             int newslots, j;
 606
 607             newslots =
 608                 memcmp(sender->slots,hdr->myslots,sizeof(hdr->myslots)) != 0;
 609             memcpy(sender->slots,hdr->myslots,sizeof(hdr->myslots));
 610             if (newslots) {
 611                 for (j = 0; j < REDIS_CLUSTER_SLOTS; j++) {
 612                     if (clusterNodeGetSlotBit(sender,j)) {
 613                         if (server.cluster.slots[j] == sender) continue;
 614                         if (server.cluster.slots[j] == NULL ||
 615                             server.cluster.slots[j]->flags & REDIS_NODE_FAIL)
 616                         {
 617                             server.cluster.slots[j] = sender;
 618                             update_state = update_config = 1;
 619                         }
 620                     }
 621                 }
 622             }
 623         }
 624
 625         /* Get info from the gossip section */
 626         clusterProcessGossipSection(hdr,link);
 627
 628         /* Update the cluster state if needed */
 629         if (update_state) clusterUpdateState();
 630         if (update_config) clusterSaveConfigOrDie();
 631     } else if (type == CLUSTERMSG_TYPE_FAIL && sender) {
 632         clusterNode *failing;
 633
 634         failing = clusterLookupNode(hdr->data.fail.about.nodename);
 635         if (failing && !(failing->flags & (REDIS_NODE_FAIL|REDIS_NODE_MYSELF)))
 636         {
 637             redisLog(REDIS_NOTICE,
 638                 "FAIL message received from %.40s about %.40s",
 639                 hdr->sender, hdr->data.fail.about.nodename);
 640             failing->flags |= REDIS_NODE_FAIL;
 641             failing->flags &= ~REDIS_NODE_PFAIL;
 642             clusterUpdateState();
 643             clusterSaveConfigOrDie();
 644         }
 645     } else {
 646         redisLog(REDIS_NOTICE,"Received unknown packet type: %d", type);
 647     }
 648     return 1;
 649 }
 650
 651 /* This function is called when we detect the link with this node is lost.
 652    We set the node as no longer connected. The Cluster Cron will detect
 653    this connection and will try to get it connected again.
 654
 655    Instead if the node is a temporary node used to accept a query, we
 656    completely free the node on error. */
 657 void handleLinkIOError(clusterLink *link) {
 658     freeClusterLink(link);
 659 }
 660
 661 /* Send data. This is handled using a trivial send buffer that gets
 662  * consumed by write(). We don't try to optimize this for speed too much
 663  * as this is a very low traffic channel. */
 664 void clusterWriteHandler(aeEventLoop *el, int fd, void *privdata, int mask) {
 665     clusterLink *link = (clusterLink*) privdata;
 666     ssize_t nwritten;
 667     REDIS_NOTUSED(el);
 668     REDIS_NOTUSED(mask);
 669
 670     nwritten = write(fd, link->sndbuf, sdslen(link->sndbuf));
 671     if (nwritten <= 0) {
 672         redisLog(REDIS_NOTICE,"I/O error writing to node link: %s",
 673             strerror(errno));
 674         handleLinkIOError(link);
 675         return;
 676     }
 677     link->sndbuf = sdsrange(link->sndbuf,nwritten,-1);
 678     if (sdslen(link->sndbuf) == 0)
 679         aeDeleteFileEvent(server.el, link->fd, AE_WRITABLE);
 680 }
 681
 682 /* Read data. Try to read the first field of the header first to check the
 683  * full length of the packet. When a whole packet is in memory this function
 684  * will call the function to process the packet. And so forth. */
 685 void clusterReadHandler(aeEventLoop *el, int fd, void *privdata, int mask) {
 686     char buf[1024];
 687     ssize_t nread;
 688     clusterMsg *hdr;
 689     clusterLink *link = (clusterLink*) privdata;
 690     int readlen;
 691     REDIS_NOTUSED(el);
 692     REDIS_NOTUSED(mask);
 693
 694 again:
 695     if (sdslen(link->rcvbuf) >= 4) {
 696         hdr = (clusterMsg*) link->rcvbuf;
 697         readlen = ntohl(hdr->totlen) - sdslen(link->rcvbuf);
 698     } else {
 699         readlen = 4 - sdslen(link->rcvbuf);
 700     }
 701
 702     nread = read(fd,buf,readlen);
 703     if (nread == -1 && errno == EAGAIN) return; /* Just no data */
 704
 705     if (nread <= 0) {
 706         /* I/O error... */
 707         redisLog(REDIS_NOTICE,"I/O error reading from node link: %s",
 708             (nread == 0) ? "connection closed" : strerror(errno));
 709         handleLinkIOError(link);
 710         return;
 711     } else {
 712         /* Read data and recast the pointer to the new buffer. */
 713         link->rcvbuf = sdscatlen(link->rcvbuf,buf,nread);
 714         hdr = (clusterMsg*) link->rcvbuf;
 715     }
 716
 717     /* Total length obtained? read the payload now instead of burning
 718      * cycles waiting for a new event to fire. */
 719     if (sdslen(link->rcvbuf) == 4) goto again;
 720
 721     /* Whole packet in memory? We can process it. */
 722     if (sdslen(link->rcvbuf) == ntohl(hdr->totlen)) {
 723         if (clusterProcessPacket(link)) {
 724             sdsfree(link->rcvbuf);
 725             link->rcvbuf = sdsempty();
 726         }
 727     }
 728 }
 729
 730 /* Put stuff into the send buffer. */
 731 void clusterSendMessage(clusterLink *link, unsigned char *msg, size_t msglen) {
 732     if (sdslen(link->sndbuf) == 0 && msglen != 0)
 733         aeCreateFileEvent(server.el,link->fd,AE_WRITABLE,
 734                     clusterWriteHandler,link);
 735
 736     link->sndbuf = sdscatlen(link->sndbuf, msg, msglen);
 737 }
 738
 739 /* Build the message header */
 740 void clusterBuildMessageHdr(clusterMsg *hdr, int type) {
 741     int totlen;
 742
 743     memset(hdr,0,sizeof(*hdr));
 744     hdr->type = htons(type);
 745     memcpy(hdr->sender,server.cluster.myself->name,REDIS_CLUSTER_NAMELEN);
 746     memcpy(hdr->myslots,server.cluster.myself->slots,
 747         sizeof(hdr->myslots));
 748     memset(hdr->slaveof,0,REDIS_CLUSTER_NAMELEN);
 749     if (server.cluster.myself->slaveof != NULL) {
 750         memcpy(hdr->slaveof,server.cluster.myself->slaveof->name,
 751                                     REDIS_CLUSTER_NAMELEN);
 752     }
 753     hdr->port = htons(server.port);
 754     hdr->state = server.cluster.state;
 755     memset(hdr->configdigest,0,32); /* FIXME: set config digest */
 756
 757     if (type == CLUSTERMSG_TYPE_FAIL) {
 758         totlen = sizeof(clusterMsg)-sizeof(union clusterMsgData);
 759         totlen += sizeof(clusterMsgDataFail);
 760     }
 761     hdr->totlen = htonl(totlen);
 762     /* For PING, PONG, and MEET, fixing the totlen field is up to the caller */
 763 }
 764
 765 /* Send a PING or PONG packet to the specified node, making sure to add enough
 766  * gossip informations. */
 767 void clusterSendPing(clusterLink *link, int type) {
 768     unsigned char buf[1024];
 769     clusterMsg *hdr = (clusterMsg*) buf;
 770     int gossipcount = 0, totlen;
 771     /* freshnodes is the number of nodes we can still use to populate the
 772      * gossip section of the ping packet. Basically we start with the nodes
 773      * we have in memory minus two (ourself and the node we are sending the
 774      * message to). Every time we add a node we decrement the counter, so when
 775      * it will drop to <= zero we know there is no more gossip info we can
 776      * send. */
 777     int freshnodes = dictSize(server.cluster.nodes)-2;
 778
 779     if (link->node && type == CLUSTERMSG_TYPE_PING)
 780         link->node->ping_sent = time(NULL);
 781     clusterBuildMessageHdr(hdr,type);
 782
 783     /* Populate the gossip fields */
 784     while(freshnodes > 0 && gossipcount < 3) {
 785         struct dictEntry *de = dictGetRandomKey(server.cluster.nodes);
 786         clusterNode *this = dictGetEntryVal(de);
 787         clusterMsgDataGossip *gossip;
 788         int j;
 789
 790         /* Not interesting to gossip about ourself.
 791          * Nor to send gossip info about HANDSHAKE state nodes (zero info). */
 792         if (this == server.cluster.myself ||
 793             this->flags & REDIS_NODE_HANDSHAKE) {
 794                 freshnodes--; /* otherwise we may loop forever. */
 795                 continue;
 796         }
 797
 798         /* Check if we already added this node */
 799         for (j = 0; j < gossipcount; j++) {
 800             if (memcmp(hdr->data.ping.gossip[j].nodename,this->name,
 801                     REDIS_CLUSTER_NAMELEN) == 0) break;
 802         }
 803         if (j != gossipcount) continue;
 804
 805         /* Add it */
 806         freshnodes--;
 807         gossip = &(hdr->data.ping.gossip[gossipcount]);
 808         memcpy(gossip->nodename,this->name,REDIS_CLUSTER_NAMELEN);
 809         gossip->ping_sent = htonl(this->ping_sent);
 810         gossip->pong_received = htonl(this->pong_received);
 811         memcpy(gossip->ip,this->ip,sizeof(this->ip));
 812         gossip->port = htons(this->port);
 813         gossip->flags = htons(this->flags);
 814         gossipcount++;
 815     }
 816     totlen = sizeof(clusterMsg)-sizeof(union clusterMsgData);
 817     totlen += (sizeof(clusterMsgDataGossip)*gossipcount);
 818     hdr->count = htons(gossipcount);
 819     hdr->totlen = htonl(totlen);
 820     clusterSendMessage(link,buf,totlen);
 821 }
 822
 823 /* Send a message to all the nodes with a reliable link */
 824 void clusterBroadcastMessage(void *buf, size_t len) {
 825     dictIterator *di;
 826     dictEntry *de;
 827
 828     di = dictGetIterator(server.cluster.nodes);
 829     while((de = dictNext(di)) != NULL) {
 830         clusterNode *node = dictGetEntryVal(de);
 831
 832         if (!node->link) continue;
 833         if (node->flags & (REDIS_NODE_MYSELF|REDIS_NODE_NOADDR)) continue;
 834         clusterSendMessage(node->link,buf,len);
 835     }
 836     dictReleaseIterator(di);
 837 }
 838
 839 /* Send a FAIL message to all the nodes we are able to contact.
 840  * The FAIL message is sent when we detect that a node is failing
 841  * (REDIS_NODE_PFAIL) and we also receive a gossip confirmation of this:
 842  * we switch the node state to REDIS_NODE_FAIL and ask all the other
 843  * nodes to do the same ASAP. */
 844 void clusterSendFail(char *nodename) {
 845     unsigned char buf[1024];
 846     clusterMsg *hdr = (clusterMsg*) buf;
 847
 848     clusterBuildMessageHdr(hdr,CLUSTERMSG_TYPE_FAIL);
 849     memcpy(hdr->data.fail.about.nodename,nodename,REDIS_CLUSTER_NAMELEN);
 850     clusterBroadcastMessage(buf,ntohl(hdr->totlen));
 851 }
 852
 853 /* -----------------------------------------------------------------------------
 854  * CLUSTER cron job
 855  * -------------------------------------------------------------------------- */
 856
 857 /* This is executed 1 time every second */
 858 void clusterCron(void) {
 859     dictIterator *di;
 860     dictEntry *de;
 861     int j;
 862     time_t min_ping_sent = 0;
 863     clusterNode *min_ping_node = NULL;
 864
 865     /* Check if we have disconnected nodes and reestablish the connection. */
 866     di = dictGetIterator(server.cluster.nodes);
 867     while((de = dictNext(di)) != NULL) {
 868         clusterNode *node = dictGetEntryVal(de);
 869
 870         if (node->flags & (REDIS_NODE_MYSELF|REDIS_NODE_NOADDR)) continue;
 871         if (node->link == NULL) {
 872             int fd;
 873             clusterLink *link;
 874
 875             fd = anetTcpNonBlockConnect(server.neterr, node->ip,
 876                 node->port+REDIS_CLUSTER_PORT_INCR);
 877             if (fd == -1) continue;
 878             link = createClusterLink(node);
 879             link->fd = fd;
 880             node->link = link;
 881             aeCreateFileEvent(server.el,link->fd,AE_READABLE,clusterReadHandler,link);
 882             /* If the node is flagged as MEET, we send a MEET message instead
 883              * of a PING one, to force the receiver to add us in its node
 884              * table. */
 885             clusterSendPing(link, node->flags & REDIS_NODE_MEET ?
 886                     CLUSTERMSG_TYPE_MEET : CLUSTERMSG_TYPE_PING);
 887             /* We can clear the flag after the first packet is sent.
 888              * If we'll never receive a PONG, we'll never send new packets
 889              * to this node. Instead after the PONG is received and we
 890              * are no longer in meet/handshake status, we want to send
 891              * normal PING packets. */
 892             node->flags &= ~REDIS_NODE_MEET;
 893
 894             redisLog(REDIS_NOTICE,"Connecting with Node %.40s at %s:%d", node->name, node->ip, node->port+REDIS_CLUSTER_PORT_INCR);
 895         }
 896     }
 897     dictReleaseIterator(di);
 898
 899     /* Ping some random node. Check a few random nodes and ping the one with
 900      * the oldest ping_sent time */
 901     for (j = 0; j < 5; j++) {
 902         de = dictGetRandomKey(server.cluster.nodes);
 903         clusterNode *this = dictGetEntryVal(de);
 904
 905         if (this->link == NULL) continue;
 906         if (this->flags & (REDIS_NODE_MYSELF|REDIS_NODE_HANDSHAKE)) continue;
 907         if (min_ping_node == NULL || min_ping_sent > this->ping_sent) {
 908             min_ping_node = this;
 909             min_ping_sent = this->ping_sent;
 910         }
 911     }
 912     if (min_ping_node) {
 913         redisLog(REDIS_DEBUG,"Pinging node %40s", min_ping_node->name);
 914         clusterSendPing(min_ping_node->link, CLUSTERMSG_TYPE_PING);
 915     }
 916
 917     /* Iterate nodes to check if we need to flag something as failing */
 918     di = dictGetIterator(server.cluster.nodes);
 919     while((de = dictNext(di)) != NULL) {
 920         clusterNode *node = dictGetEntryVal(de);
 921         int delay;
 922
 923         if (node->flags &
 924             (REDIS_NODE_MYSELF|REDIS_NODE_NOADDR|REDIS_NODE_HANDSHAKE))
 925                 continue;
 926         /* Check only if we already sent a ping and did not received
 927          * a reply yet. */
 928         if (node->ping_sent == 0 ||
 929             node->ping_sent <= node->pong_received) continue;
 930
 931         delay = time(NULL) - node->pong_received;
 932         if (delay < server.cluster.node_timeout) {
 933             /* The PFAIL condition can be reversed without external
 934              * help if it is not transitive (that is, if it does not
 935              * turn into a FAIL state).
 936              *
 937              * The FAIL condition is also reversible if there are no slaves
 938              * for this host, so no slave election should be in progress.
 939              *
 940              * TODO: consider all the implications of resurrecting a
 941              * FAIL node. */
 942             if (node->flags & REDIS_NODE_PFAIL) {
 943                 node->flags &= ~REDIS_NODE_PFAIL;
 944             } else if (node->flags & REDIS_NODE_FAIL && !node->numslaves) {
 945                 node->flags &= ~REDIS_NODE_FAIL;
 946                 clusterUpdateState();
 947             }
 948         } else {
 949             /* Timeout reached. Set the noad se possibly failing if it is
 950              * not already in this state. */
 951             if (!(node->flags & (REDIS_NODE_PFAIL|REDIS_NODE_FAIL))) {
 952                 redisLog(REDIS_DEBUG,"*** NODE %.40s possibly failing",
 953                     node->name);
 954                 node->flags |= REDIS_NODE_PFAIL;
 955             }
 956         }
 957     }
 958     dictReleaseIterator(di);
 959 }
 960
 961 /* -----------------------------------------------------------------------------
 962  * Slots management
 963  * -------------------------------------------------------------------------- */
 964
 965 /* Set the slot bit and return the old value. */
 966 int clusterNodeSetSlotBit(clusterNode *n, int slot) {
 967     off_t byte = slot/8;
 968     int bit = slot&7;
 969     int old = (n->slots[byte] & (1<<bit)) != 0;
 970     n->slots[byte] |= 1<<bit;
 971     return old;
 972 }
 973
 974 /* Clear the slot bit and return the old value. */
 975 int clusterNodeClearSlotBit(clusterNode *n, int slot) {
 976     off_t byte = slot/8;
 977     int bit = slot&7;
 978     int old = (n->slots[byte] & (1<<bit)) != 0;
 979     n->slots[byte] &= ~(1<<bit);
 980     return old;
 981 }
 982
 983 /* Return the slot bit from the cluster node structure. */
 984 int clusterNodeGetSlotBit(clusterNode *n, int slot) {
 985     off_t byte = slot/8;
 986     int bit = slot&7;
 987     return (n->slots[byte] & (1<<bit)) != 0;
 988 }
 989
 990 /* Add the specified slot to the list of slots that node 'n' will
 991  * serve. Return REDIS_OK if the operation ended with success.
 992  * If the slot is already assigned to another instance this is considered
 993  * an error and REDIS_ERR is returned. */
 994 int clusterAddSlot(clusterNode *n, int slot) {
 995     redisAssert(clusterNodeSetSlotBit(n,slot) == 0);
 996     server.cluster.slots[slot] = n;
 997     printf("SLOT %d added to %.40s\n", slot, n->name);
 998     return REDIS_OK;
 999 }
1000
1001 /* -----------------------------------------------------------------------------
1002  * Cluster state evaluation function
1003  * -------------------------------------------------------------------------- */
1004 void clusterUpdateState(void) {
1005     int ok = 1;
1006     int j;
1007
1008     for (j = 0; j < REDIS_CLUSTER_SLOTS; j++) {
1009         if (server.cluster.slots[j] == NULL ||
1010             server.cluster.slots[j]->flags & (REDIS_NODE_FAIL))
1011         {
1012             ok = 0;
1013             break;
1014         }
1015     }
1016     if (ok) {
1017         if (server.cluster.state == REDIS_CLUSTER_NEEDHELP) {
1018             server.cluster.state = REDIS_CLUSTER_NEEDHELP;
1019         } else {
1020             server.cluster.state = REDIS_CLUSTER_OK;
1021         }
1022     } else {
1023         server.cluster.state = REDIS_CLUSTER_FAIL;
1024     }
1025 }
1026
1027 /* -----------------------------------------------------------------------------
1028  * CLUSTER command
1029  * -------------------------------------------------------------------------- */
1030
1031 sds clusterGenNodesDescription(void) {
1032     sds ci = sdsempty();
1033     dictIterator *di;
1034     dictEntry *de;
1035     int j, start;
1036
1037     di = dictGetIterator(server.cluster.nodes);
1038     while((de = dictNext(di)) != NULL) {
1039         clusterNode *node = dictGetEntryVal(de);
1040
1041         /* Node coordinates */
1042         ci = sdscatprintf(ci,"%.40s %s:%d ",
1043             node->name,
1044             node->ip,
1045             node->port);
1046
1047         /* Flags */
1048         if (node->flags == 0) ci = sdscat(ci,"noflags,");
1049         if (node->flags & REDIS_NODE_MYSELF) ci = sdscat(ci,"myself,");
1050         if (node->flags & REDIS_NODE_MASTER) ci = sdscat(ci,"master,");
1051         if (node->flags & REDIS_NODE_SLAVE) ci = sdscat(ci,"slave,");
1052         if (node->flags & REDIS_NODE_PFAIL) ci = sdscat(ci,"fail?,");
1053         if (node->flags & REDIS_NODE_FAIL) ci = sdscat(ci,"fail,");
1054         if (node->flags & REDIS_NODE_HANDSHAKE) ci =sdscat(ci,"handshake,");
1055         if (node->flags & REDIS_NODE_NOADDR) ci = sdscat(ci,"noaddr,");
1056         if (ci[sdslen(ci)-1] == ',') ci[sdslen(ci)-1] = ' ';
1057
1058         /* Slave of... or just "-" */
1059         if (node->slaveof)
1060             ci = sdscatprintf(ci,"%.40s ",node->slaveof->name);
1061         else
1062             ci = sdscatprintf(ci,"- ");
1063
1064         /* Latency from the POV of this node, link status */
1065         ci = sdscatprintf(ci,"%ld %ld %s",
1066             (long) node->ping_sent,
1067             (long) node->pong_received,
1068             node->link ? "connected" : "disconnected");
1069
1070         /* Slots served by this instance */
1071         start = -1;
1072         for (j = 0; j < REDIS_CLUSTER_SLOTS; j++) {
1073             int bit;
1074
1075             if ((bit = clusterNodeGetSlotBit(node,j)) != 0) {
1076                 if (start == -1) start = j;
1077             }
1078             if (start != -1 && (!bit || j == REDIS_CLUSTER_SLOTS-1)) {
1079                 if (j == REDIS_CLUSTER_SLOTS-1) j++;
1080
1081                 if (start == j-1) {
1082                     ci = sdscatprintf(ci," %d",start);
1083                 } else {
1084                     ci = sdscatprintf(ci," %d-%d",start,j-1);
1085                 }
1086                 start = -1;
1087             }
1088         }
1089         ci = sdscatlen(ci,"\n",1);
1090     }
1091     dictReleaseIterator(di);
1092     return ci;
1093 }
1094
1095 void clusterCommand(redisClient *c) {
1096     if (server.cluster_enabled == 0) {
1097         addReplyError(c,"This instance has cluster support disabled");
1098         return;
1099     }
1100
1101     if (!strcasecmp(c->argv[1]->ptr,"meet") && c->argc == 4) {
1102         clusterNode *n;
1103         struct sockaddr_in sa;
1104         long port;
1105
1106         /* Perform sanity checks on IP/port */
1107         if (inet_aton(c->argv[2]->ptr,&sa.sin_addr) == 0) {
1108             addReplyError(c,"Invalid IP address in MEET");
1109             return;
1110         }
1111         if (getLongFromObjectOrReply(c, c->argv[3], &port, NULL) != REDIS_OK ||
1112                     port < 0 || port > (65535-REDIS_CLUSTER_PORT_INCR))
1113         {
1114             addReplyError(c,"Invalid TCP port specified");
1115             return;
1116         }
1117
1118         /* Finally add the node to the cluster with a random name, this
1119          * will get fixed in the first handshake (ping/pong). */
1120         n = createClusterNode(NULL,REDIS_NODE_HANDSHAKE|REDIS_NODE_MEET);
1121         strncpy(n->ip,inet_ntoa(sa.sin_addr),sizeof(n->ip));
1122         n->port = port;
1123         clusterAddNode(n);
1124         addReply(c,shared.ok);
1125     } else if (!strcasecmp(c->argv[1]->ptr,"nodes") && c->argc == 2) {
1126         robj *o;
1127         sds ci = clusterGenNodesDescription();
1128
1129         o = createObject(REDIS_STRING,ci);
1130         addReplyBulk(c,o);
1131         decrRefCount(o);
1132     } else if (!strcasecmp(c->argv[1]->ptr,"addslots") && c->argc >= 3) {
1133         int j;
1134         long long slot;
1135         unsigned char *slots = zmalloc(REDIS_CLUSTER_SLOTS);
1136
1137         memset(slots,0,REDIS_CLUSTER_SLOTS);
1138         /* Check that all the arguments are parsable and that all the
1139          * slots are not already busy. */
1140         for (j = 2; j < c->argc; j++) {
1141             if (getLongLongFromObject(c->argv[j],&slot) != REDIS_OK ||
1142                 slot < 0 || slot > REDIS_CLUSTER_SLOTS)
1143             {
1144                 addReplyError(c,"Invalid or out of range slot index");
1145                 zfree(slots);
1146                 return;
1147             }
1148             if (server.cluster.slots[slot]) {
1149                 addReplyErrorFormat(c,"Slot %lld is already busy", slot);
1150                 zfree(slots);
1151                 return;
1152             }
1153             if (slots[slot]++ == 1) {
1154                 addReplyErrorFormat(c,"Slot %d specified multiple times",
1155                     (int)slot);
1156                 zfree(slots);
1157                 return;
1158             }
1159         }
1160         for (j = 0; j < REDIS_CLUSTER_SLOTS; j++) {
1161             if (slots[j]) {
1162                 int retval = clusterAddSlot(server.cluster.myself,j);
1163
1164                 redisAssert(retval == REDIS_OK);
1165             }
1166         }
1167         zfree(slots);
1168         clusterUpdateState();
1169         clusterSaveConfigOrDie();
1170         addReply(c,shared.ok);
1171     } else if (!strcasecmp(c->argv[1]->ptr,"info") && c->argc == 2) {
1172         char *statestr[] = {"ok","fail","needhelp"};
1173         int slots_assigned = 0, slots_ok = 0, slots_pfail = 0, slots_fail = 0;
1174         int j;
1175
1176         for (j = 0; j < REDIS_CLUSTER_SLOTS; j++) {
1177             clusterNode *n = server.cluster.slots[j];
1178
1179             if (n == NULL) continue;
1180             slots_assigned++;
1181             if (n->flags & REDIS_NODE_FAIL) {
1182                 slots_fail++;
1183             } else if (n->flags & REDIS_NODE_PFAIL) {
1184                 slots_pfail++;
1185             } else {
1186                 slots_ok++;
1187             }
1188         }
1189
1190         sds info = sdscatprintf(sdsempty(),
1191             "cluster_state:%s\r\n"
1192             "cluster_slots_assigned:%d\r\n"
1193             "cluster_slots_ok:%d\r\n"
1194             "cluster_slots_pfail:%d\r\n"
1195             "cluster_slots_fail:%d\r\n"
1196             , statestr[server.cluster.state],
1197             slots_assigned,
1198             slots_ok,
1199             slots_pfail,
1200             slots_fail
1201         );
1202         addReplySds(c,sdscatprintf(sdsempty(),"$%lu\r\n",
1203             (unsigned long)sdslen(info)));
1204         addReplySds(c,info);
1205         addReply(c,shared.crlf);
1206     } else {
1207         addReplyError(c,"Wrong CLUSTER subcommand or number of arguments");
1208     }
1209 }
1210
1211 /* -----------------------------------------------------------------------------
1212  * RESTORE and MIGRATE commands
1213  * -------------------------------------------------------------------------- */
1214
1215 /* RESTORE key ttl serialized-value */
1216 void restoreCommand(redisClient *c) {
1217     FILE *fp;
1218     char buf[64];
1219     robj *o;
1220     unsigned char *data;
1221     long ttl;
1222
1223     /* Make sure this key does not already exist here... */
1224     if (dbExists(c->db,c->argv[1])) {
1225         addReplyError(c,"Target key name is busy.");
1226         return;
1227     }
1228
1229     /* Check if the TTL value makes sense */
1230     if (getLongFromObjectOrReply(c,c->argv[2],&ttl,NULL) != REDIS_OK) {
1231         return;
1232     } else if (ttl < 0) {
1233         addReplyError(c,"Invalid TTL value, must be >= 0");
1234         return;
1235     }
1236
1237     /* rdbLoadObject() only works against file descriptors so we need to
1238      * dump the serialized object into a file and reload. */
1239     snprintf(buf,sizeof(buf),"redis-restore-%d.tmp",getpid());
1240     fp = fopen(buf,"w+");
1241     if (!fp) {
1242         redisLog(REDIS_WARNING,"Can't open tmp file for RESTORE: %s",
1243             strerror(errno));
1244         addReplyErrorFormat(c,"RESTORE failed, tmp file creation error: %s",
1245             strerror(errno));
1246         return;
1247     }
1248     unlink(buf);
1249
1250     /* Write the actual data and rewind the file */
1251     data = (unsigned char*) c->argv[3]->ptr;
1252     if (fwrite(data+1,sdslen((sds)data)-1,1,fp) != 1) {
1253         redisLog(REDIS_WARNING,"Can't write against tmp file for RESTORE: %s",
1254             strerror(errno));
1255         addReplyError(c,"RESTORE failed, tmp file I/O error.");
1256         fclose(fp);
1257         return;
1258     }
1259     rewind(fp);
1260
1261     /* Finally create the object from the serialized dump and
1262      * store it at the specified key. */
1263     if ((data[0] > 4 && data[0] < 9) ||
1264          data[0] > 11 ||
1265         (o = rdbLoadObject(data[0],fp)) == NULL)
1266     {
1267         addReplyError(c,"Bad data format.");
1268         fclose(fp);
1269         return;
1270     }
1271     fclose(fp);
1272
1273     /* Create the key and set the TTL if any */
1274     dbAdd(c->db,c->argv[1],o);
1275     if (ttl) setExpire(c->db,c->argv[1],time(NULL)+ttl);
1276     addReply(c,shared.ok);
1277 }
1278
1279 /* MIGRATE host port key dbid timeout */
1280 void migrateCommand(redisClient *c) {
1281     int fd;
1282     long timeout;
1283     long dbid;
1284     char buf[64];
1285     FILE *fp;
1286     time_t ttl;
1287     robj *o;
1288     unsigned char type;
1289     off_t payload_len;
1290
1291     /* Sanity check */
1292     if (getLongFromObjectOrReply(c,c->argv[5],&timeout,NULL) != REDIS_OK)
1293         return;
1294     if (getLongFromObjectOrReply(c,c->argv[4],&dbid,NULL) != REDIS_OK)
1295         return;
1296     if (timeout <= 0) timeout = 1;
1297
1298     /* Check if the key is here. If not we reply with success as there is
1299      * nothing to migrate (for instance the key expired in the meantime), but
1300      * we include such information in the reply string. */
1301     if ((o = lookupKeyRead(c->db,c->argv[3])) == NULL) {
1302         addReplySds(c,sdsnew("+NOKEY"));
1303         return;
1304     }
1305
1306     /* Connect */
1307     fd = anetTcpNonBlockConnect(server.neterr,c->argv[1]->ptr,
1308                 atoi(c->argv[2]->ptr));
1309     if (fd == -1) {
1310         addReplyErrorFormat(c,"Can't connect to target node: %s",
1311             server.neterr);
1312         return;
1313     }
1314     if ((aeWait(fd,AE_WRITABLE,timeout*1000) & AE_WRITABLE) == 0) {
1315         addReplyError(c,"Timeout connecting to the client");
1316         return;
1317     }
1318
1319     /* Create temp file */
1320     snprintf(buf,sizeof(buf),"redis-migrate-%d.tmp",getpid());
1321     fp = fopen(buf,"w+");
1322     if (!fp) {
1323         redisLog(REDIS_WARNING,"Can't open tmp file for MIGRATE: %s",
1324             strerror(errno));
1325         addReplyErrorFormat(c,"MIGRATE failed, tmp file creation error: %s.",
1326             strerror(errno));
1327         return;
1328     }
1329     unlink(buf);
1330
1331     /* Build the SELECT + RESTORE query writing it in our temp file. */
1332     if (fwriteBulkCount(fp,'*',2) == 0) goto file_wr_err;
1333     if (fwriteBulkString(fp,"SELECT",6) == 0) goto file_wr_err;
1334     if (fwriteBulkLongLong(fp,dbid) == 0) goto file_wr_err;
1335
1336     ttl = getExpire(c->db,c->argv[3]);
1337     type = o->type;
1338     if (fwriteBulkCount(fp,'*',4) == 0) goto file_wr_err;
1339     if (fwriteBulkString(fp,"RESTORE",7) == 0) goto file_wr_err;
1340     if (fwriteBulkObject(fp,c->argv[3]) == 0) goto file_wr_err;
1341     if (fwriteBulkLongLong(fp, (ttl == -1) ? 0 : ttl) == 0) goto file_wr_err;
1342
1343     /* Finally the last argument that is the serailized object payload
1344      * in the form: <type><rdb-serailized-object>. */
1345     payload_len = rdbSavedObjectLen(o);
1346     if (fwriteBulkCount(fp,'$',payload_len+1) == 0) goto file_wr_err;
1347     if (fwrite(&type,1,1,fp) == 0) goto file_wr_err;
1348     if (rdbSaveObject(fp,o) == -1) goto file_wr_err;
1349     if (fwrite("\r\n",2,1,fp) == 0) goto file_wr_err;
1350
1351     /* Tranfer the query to the other node */
1352     rewind(fp);
1353     {
1354         char buf[4096];
1355         size_t nread;
1356
1357         while ((nread = fread(buf,1,sizeof(buf),fp)) != 0) {
1358             int nwritten;
1359
1360             nwritten = syncWrite(fd,buf,nread,timeout);
1361             if (nwritten != (signed)nread) goto socket_wr_err;
1362         }
1363         if (ferror(fp)) goto file_rd_err;
1364     }
1365
1366     /* Read back the reply */
1367     {
1368         char buf1[1024];
1369         char buf2[1024];
1370
1371         /* Read the two replies */
1372         if (syncReadLine(fd, buf1, sizeof(buf1), timeout) <= 0)
1373             goto socket_rd_err;
1374         if (syncReadLine(fd, buf2, sizeof(buf2), timeout) <= 0)
1375                 goto socket_rd_err;
1376         if (buf1[0] == '-' || buf2[0] == '-') {
1377             addReplyErrorFormat(c,"Target instance replied with error: %s",
1378                 (buf1[0] == '-') ? buf1+1 : buf2+1);
1379         } else {
1380             dbDelete(c->db,c->argv[3]);
1381             addReply(c,shared.ok);
1382         }
1383     }
1384     fclose(fp);
1385     close(fd);
1386     return;
1387
1388 file_wr_err:
1389     redisLog(REDIS_WARNING,"Can't write on tmp file for MIGRATE: %s",
1390         strerror(errno));
1391     addReplyErrorFormat(c,"MIGRATE failed, tmp file write error: %s.",
1392         strerror(errno));
1393     fclose(fp);
1394     close(fd);
1395     return;
1396
1397 file_rd_err:
1398     redisLog(REDIS_WARNING,"Can't read from tmp file for MIGRATE: %s",
1399         strerror(errno));
1400     addReplyErrorFormat(c,"MIGRATE failed, tmp file read error: %s.",
1401         strerror(errno));
1402     fclose(fp);
1403     close(fd);
1404     return;
1405
1406 socket_wr_err:
1407     redisLog(REDIS_NOTICE,"Can't write to target node for MIGRATE: %s",
1408         strerror(errno));
1409     addReplyErrorFormat(c,"MIGRATE failed, writing to target node: %s.",
1410         strerror(errno));
1411     fclose(fp);
1412     close(fd);
1413     return;
1414
1415 socket_rd_err:
1416     redisLog(REDIS_NOTICE,"Can't read from target node for MIGRATE: %s",
1417         strerror(errno));
1418     addReplyErrorFormat(c,"MIGRATE failed, reading from target node: %s.",
1419         strerror(errno));
1420     fclose(fp);
1421     close(fd);
1422     return;
1423 }
1424
1425 /* DUMP keyname
1426  * DUMP is actually not used by Redis Cluster but it is the obvious
1427  * complement of RESTORE and can be useful for different applications. */
1428 void dumpCommand(redisClient *c) {
1429     char buf[64];
1430     FILE *fp;
1431     robj *o, *dumpobj;
1432     sds dump = NULL;
1433     off_t payload_len;
1434     unsigned int type;
1435
1436     /* Check if the key is here. */
1437     if ((o = lookupKeyRead(c->db,c->argv[1])) == NULL) {
1438         addReply(c,shared.nullbulk);
1439         return;
1440     }
1441
1442     /* Create temp file */
1443     snprintf(buf,sizeof(buf),"redis-dump-%d.tmp",getpid());
1444     fp = fopen(buf,"w+");
1445     if (!fp) {
1446         redisLog(REDIS_WARNING,"Can't open tmp file for MIGRATE: %s",
1447             strerror(errno));
1448         addReplyErrorFormat(c,"DUMP failed, tmp file creation error: %s.",
1449             strerror(errno));
1450         return;
1451     }
1452     unlink(buf);
1453
1454     /* Dump the serailized object and read it back in memory.
1455      * We prefix it with a one byte containing the type ID.
1456      * This is the serialization format understood by RESTORE. */
1457     if (rdbSaveObject(fp,o) == -1) goto file_wr_err;
1458     payload_len = ftello(fp);
1459     if (fseeko(fp,0,SEEK_SET) == -1) goto file_rd_err;
1460     dump = sdsnewlen(NULL,payload_len+1);
1461     if (payload_len && fread(dump+1,payload_len,1,fp) != 1) goto file_rd_err;
1462     fclose(fp);
1463     type = o->type;
1464     if (type == REDIS_LIST && o->encoding == REDIS_ENCODING_ZIPLIST)
1465         type = REDIS_LIST_ZIPLIST;
1466     else if (type == REDIS_HASH && o->encoding == REDIS_ENCODING_ZIPMAP)
1467         type = REDIS_HASH_ZIPMAP;
1468     else if (type == REDIS_SET && o->encoding == REDIS_ENCODING_INTSET)
1469         type = REDIS_SET_INTSET;
1470     else
1471         type = o->type;
1472     dump[0] = type;
1473
1474     /* Transfer to the client */
1475     dumpobj = createObject(REDIS_STRING,dump);
1476     addReplyBulk(c,dumpobj);
1477     decrRefCount(dumpobj);
1478     return;
1479
1480 file_wr_err:
1481     redisLog(REDIS_WARNING,"Can't write on tmp file for DUMP: %s",
1482         strerror(errno));
1483     addReplyErrorFormat(c,"DUMP failed, tmp file write error: %s.",
1484         strerror(errno));
1485     sdsfree(dump);
1486     fclose(fp);
1487     return;
1488
1489 file_rd_err:
1490     redisLog(REDIS_WARNING,"Can't read from tmp file for DUMP: %s",
1491         strerror(errno));
1492     addReplyErrorFormat(c,"DUMP failed, tmp file read error: %s.",
1493         strerror(errno));
1494     sdsfree(dump);
1495     fclose(fp);
1496     return;
1497 }
1498
1499 /* -----------------------------------------------------------------------------
1500  * Cluster functions related to serving / redirecting clients
1501  * -------------------------------------------------------------------------- */
1502
1503 /* Return the pointer to the cluster node that is able to serve the query
1504  * as all the keys belong to hash slots for which the node is in charge.
1505  *
1506  * If keys in query spawn multiple nodes NULL is returned. */
1507 clusterNode *getNodeByQuery(redisClient *c, struct redisCommand *cmd, robj **argv, int argc, int *hashslot) {
1508     clusterNode *n = NULL;
1509     multiState *ms, _ms;
1510     multiCmd mc;
1511     int i;
1512
1513     /* We handle all the cases as if they were EXEC commands, so we have
1514      * a common code path for everything */
1515     if (cmd->proc == execCommand) {
1516         /* If REDIS_MULTI flag is not set EXEC is just going to return an
1517          * error. */
1518         if (!(c->flags & REDIS_MULTI)) return server.cluster.myself;
1519         ms = &c->mstate;
1520     } else {
1521         /* Create a fake Multi State structure, with just one command */
1522         ms = &_ms;
1523         _ms.commands = &mc;
1524         _ms.count = 1;
1525         mc.argv = argv;
1526         mc.argc = argc;
1527         mc.cmd = cmd;
1528     }
1529
1530     for (i = 0; i < ms->count; i++) {
1531         struct redisCommand *mcmd;
1532         robj **margv;
1533         int margc, *keyindex, numkeys, j;
1534
1535         mcmd = ms->commands[i].cmd;
1536         margc = ms->commands[i].argc;
1537         margv = ms->commands[i].argv;
1538
1539         keyindex = getKeysFromCommand(mcmd,margv,margc,&numkeys,
1540                                       REDIS_GETKEYS_PRELOAD);
1541         for (j = 0; j < numkeys; j++) {
1542             int slot = keyHashSlot((char*)margv[keyindex[j]]->ptr,
1543                 sdslen(margv[keyindex[j]]->ptr));
1544             struct clusterNode *slotnode;
1545
1546             slotnode = server.cluster.slots[slot];
1547             if (hashslot) *hashslot = slot;
1548             /* Node not assigned? (Should never happen actually
1549              * if we reached this function).
1550              * Different node than the previous one?
1551              * Return NULL, the cluster can't serve multi-node requests */
1552             if (slotnode == NULL || (n && slotnode != n)) {
1553                 getKeysFreeResult(keyindex);
1554                 return NULL;
1555             } else {
1556                 n = slotnode;
1557             }
1558         }
1559         getKeysFreeResult(keyindex);
1560     }
1561     return (n == NULL) ? server.cluster.myself : n;
1562 }