1 #ifndef __CLUSTER_H 2 #define __CLUSTER_H 3 4 /*----------------------------------------------------------------------------- 5 * Redis cluster data structures, defines, exported API. 6 *----------------------------------------------------------------------------*/ 7 8 #define CLUSTER_SLOTS 16384 9 #define CLUSTER_OK 0 /* Everything looks ok */ 10 #define CLUSTER_FAIL 1 /* The cluster can't work */ 11 #define CLUSTER_NAMELEN 40 /* sha1 hex length */ 12 #define CLUSTER_PORT_INCR 10000 /* Cluster port = baseport + PORT_INCR */ 13 14 /* The following defines are amount of time, sometimes expressed as 15 * multiplicators of the node timeout value (when ending with MULT). */ 16 #define CLUSTER_DEFAULT_NODE_TIMEOUT 15000 17 #define CLUSTER_DEFAULT_SLAVE_VALIDITY 10 /* Slave max data age factor. */ 18 #define CLUSTER_DEFAULT_REQUIRE_FULL_COVERAGE 1 19 #define CLUSTER_FAIL_REPORT_VALIDITY_MULT 2 /* Fail report validity. */ 20 #define CLUSTER_FAIL_UNDO_TIME_MULT 2 /* Undo fail if master is back. */ 21 #define CLUSTER_FAIL_UNDO_TIME_ADD 10 /* Some additional time. */ 22 #define CLUSTER_FAILOVER_DELAY 5 /* Seconds */ 23 #define CLUSTER_DEFAULT_MIGRATION_BARRIER 1 24 #define CLUSTER_MF_TIMEOUT 5000 /* Milliseconds to do a manual failover. */ 25 #define CLUSTER_MF_PAUSE_MULT 2 /* Master pause manual failover mult. */ 26 27 /* Redirection errors returned by getNodeByQuery(). */ 28 #define CLUSTER_REDIR_NONE 0 /* Node can serve the request. */ 29 #define CLUSTER_REDIR_CROSS_SLOT 1 /* -CROSSSLOT request. */ 30 #define CLUSTER_REDIR_UNSTABLE 2 /* -TRYAGAIN redirection required */ 31 #define CLUSTER_REDIR_ASK 3 /* -ASK redirection required. */ 32 #define CLUSTER_REDIR_MOVED 4 /* -MOVED redirection required. */ 33 #define CLUSTER_REDIR_DOWN_STATE 5 /* -CLUSTERDOWN, global state. */ 34 #define CLUSTER_REDIR_DOWN_UNBOUND 6 /* -CLUSTERDOWN, unbound slot. */ 35 36 struct clusterNode; 37 38 /* clusterLink encapsulates everything needed to talk with a remote node. */ 39 typedef struct clusterLink { 40 mstime_t ctime; /* Link creation time */ 41 int fd; /* TCP socket file descriptor */ 42 sds sndbuf; /* Packet send buffer */ 43 sds rcvbuf; /* Packet reception buffer */ 44 struct clusterNode *node; /* Node related to this link if any, or NULL */ 45 } clusterLink; 46 47 /* Cluster node flags and macros. */ 48 #define CLUSTER_NODE_MASTER 1 /* The node is a master */ 49 #define CLUSTER_NODE_SLAVE 2 /* The node is a slave */ 50 #define CLUSTER_NODE_PFAIL 4 /* Failure? Need acknowledge */ 51 #define CLUSTER_NODE_FAIL 8 /* The node is believed to be malfunctioning */ 52 #define CLUSTER_NODE_MYSELF 16 /* This node is myself */ 53 #define CLUSTER_NODE_HANDSHAKE 32 /* We have still to exchange the first ping */ 54 #define CLUSTER_NODE_NOADDR 64 /* We don't know the address of this node */ 55 #define CLUSTER_NODE_MEET 128 /* Send a MEET message to this node */ 56 #define CLUSTER_NODE_MIGRATE_TO 256 /* Master elegible for replica migration. */ 57 #define CLUSTER_NODE_NULL_NAME "\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000\000" 58 59 #define nodeIsMaster(n) ((n)->flags & CLUSTER_NODE_MASTER) 60 #define nodeIsSlave(n) ((n)->flags & CLUSTER_NODE_SLAVE) 61 #define nodeInHandshake(n) ((n)->flags & CLUSTER_NODE_HANDSHAKE) 62 #define nodeHasAddr(n) (!((n)->flags & CLUSTER_NODE_NOADDR)) 63 #define nodeWithoutAddr(n) ((n)->flags & CLUSTER_NODE_NOADDR) 64 #define nodeTimedOut(n) ((n)->flags & CLUSTER_NODE_PFAIL) 65 #define nodeFailed(n) ((n)->flags & CLUSTER_NODE_FAIL) 66 67 /* Reasons why a slave is not able to failover. */ 68 #define CLUSTER_CANT_FAILOVER_NONE 0 69 #define CLUSTER_CANT_FAILOVER_DATA_AGE 1 70 #define CLUSTER_CANT_FAILOVER_WAITING_DELAY 2 71 #define CLUSTER_CANT_FAILOVER_EXPIRED 3 72 #define CLUSTER_CANT_FAILOVER_WAITING_VOTES 4 73 #define CLUSTER_CANT_FAILOVER_RELOG_PERIOD (60*5) /* seconds. */ 74 75 /* This structure represent elements of node->fail_reports. */ 76 typedef struct clusterNodeFailReport { 77 struct clusterNode *node; /* Node reporting the failure condition. */ 78 mstime_t time; /* Time of the last report from this node. */ 79 } clusterNodeFailReport; 80 81 typedef struct clusterNode { 82 mstime_t ctime; /* Node object creation time. */ 83 char name[CLUSTER_NAMELEN]; /* Node name, hex string, sha1-size */ 84 int flags; /* CLUSTER_NODE_... */ 85 uint64_t configEpoch; /* Last configEpoch observed for this node */ 86 unsigned char slots[CLUSTER_SLOTS/8]; /* slots handled by this node */ 87 int numslots; /* Number of slots handled by this node */ 88 int numslaves; /* Number of slave nodes, if this is a master */ 89 struct clusterNode **slaves; /* pointers to slave nodes */ 90 struct clusterNode *slaveof; /* pointer to the master node */ 91 mstime_t ping_sent; /* Unix time we sent latest ping */ 92 mstime_t pong_received; /* Unix time we received the pong */ 93 mstime_t fail_time; /* Unix time when FAIL flag was set */ 94 mstime_t voted_time; /* Last time we voted for a slave of this master */ 95 mstime_t repl_offset_time; /* Unix time we received offset for this node */ 96 long long repl_offset; /* Last known repl offset for this node. */ 97 char ip[NET_IP_STR_LEN]; /* Latest known IP address of this node */ 98 int port; /* Latest known port of this node */ 99 clusterLink *link; /* TCP/IP link with this node */ 100 list *fail_reports; /* List of nodes signaling this as failing */ 101 } clusterNode; 102 103 typedef struct clusterState { 104 clusterNode *myself; /* This node */ 105 uint64_t currentEpoch; 106 int state; /* CLUSTER_OK, CLUSTER_FAIL, ... */ 107 int size; /* Num of master nodes with at least one slot */ 108 dict *nodes; /* Hash table of name -> clusterNode structures */ 109 dict *nodes_black_list; /* Nodes we don't re-add for a few seconds. */ 110 clusterNode *migrating_slots_to[CLUSTER_SLOTS]; 111 clusterNode *importing_slots_from[CLUSTER_SLOTS]; 112 clusterNode *slots[CLUSTER_SLOTS]; 113 zskiplist *slots_to_keys; 114 /* The following fields are used to take the slave state on elections. */ 115 mstime_t failover_auth_time; /* Time of previous or next election. */ 116 int failover_auth_count; /* Number of votes received so far. */ 117 int failover_auth_sent; /* True if we already asked for votes. */ 118 int failover_auth_rank; /* This slave rank for current auth request. */ 119 uint64_t failover_auth_epoch; /* Epoch of the current election. */ 120 int cant_failover_reason; /* Why a slave is currently not able to 121 failover. See the CANT_FAILOVER_* macros. */ 122 /* Manual failover state in common. */ 123 mstime_t mf_end; /* Manual failover time limit (ms unixtime). 124 It is zero if there is no MF in progress. */ 125 /* Manual failover state of master. */ 126 clusterNode *mf_slave; /* Slave performing the manual failover. */ 127 /* Manual failover state of slave. */ 128 long long mf_master_offset; /* Master offset the slave needs to start MF 129 or zero if stil not received. */ 130 int mf_can_start; /* If non-zero signal that the manual failover 131 can start requesting masters vote. */ 132 /* The followign fields are used by masters to take state on elections. */ 133 uint64_t lastVoteEpoch; /* Epoch of the last vote granted. */ 134 int todo_before_sleep; /* Things to do in clusterBeforeSleep(). */ 135 long long stats_bus_messages_sent; /* Num of msg sent via cluster bus. */ 136 long long stats_bus_messages_received; /* Num of msg rcvd via cluster bus.*/ 137 } clusterState; 138 139 /* clusterState todo_before_sleep flags. */ 140 #define CLUSTER_TODO_HANDLE_FAILOVER (1<<0) 141 #define CLUSTER_TODO_UPDATE_STATE (1<<1) 142 #define CLUSTER_TODO_SAVE_CONFIG (1<<2) 143 #define CLUSTER_TODO_FSYNC_CONFIG (1<<3) 144 145 /* Redis cluster messages header */ 146 147 /* Note that the PING, PONG and MEET messages are actually the same exact 148 * kind of packet. PONG is the reply to ping, in the exact format as a PING, 149 * while MEET is a special PING that forces the receiver to add the sender 150 * as a node (if it is not already in the list). */ 151 #define CLUSTERMSG_TYPE_PING 0 /* Ping */ 152 #define CLUSTERMSG_TYPE_PONG 1 /* Pong (reply to Ping) */ 153 #define CLUSTERMSG_TYPE_MEET 2 /* Meet "let's join" message */ 154 #define CLUSTERMSG_TYPE_FAIL 3 /* Mark node xxx as failing */ 155 #define CLUSTERMSG_TYPE_PUBLISH 4 /* Pub/Sub Publish propagation */ 156 #define CLUSTERMSG_TYPE_FAILOVER_AUTH_REQUEST 5 /* May I failover? */ 157 #define CLUSTERMSG_TYPE_FAILOVER_AUTH_ACK 6 /* Yes, you have my vote */ 158 #define CLUSTERMSG_TYPE_UPDATE 7 /* Another node slots configuration */ 159 #define CLUSTERMSG_TYPE_MFSTART 8 /* Pause clients for manual failover */ 160 161 /* Initially we don't know our "name", but we'll find it once we connect 162 * to the first node, using the getsockname() function. Then we'll use this 163 * address for all the next messages. */ 164 typedef struct { 165 char nodename[CLUSTER_NAMELEN]; 166 uint32_t ping_sent; 167 uint32_t pong_received; 168 char ip[NET_IP_STR_LEN]; /* IP address last time it was seen */ 169 uint16_t port; /* port last time it was seen */ 170 uint16_t flags; /* node->flags copy */ 171 uint16_t notused1; /* Some room for future improvements. */ 172 uint32_t notused2; 173 } clusterMsgDataGossip; 174 175 typedef struct { 176 char nodename[CLUSTER_NAMELEN]; 177 } clusterMsgDataFail; 178 179 typedef struct { 180 uint32_t channel_len; 181 uint32_t message_len; 182 /* We can't reclare bulk_data as bulk_data[] since this structure is 183 * nested. The 8 bytes are removed from the count during the message 184 * length computation. */ 185 unsigned char bulk_data[8]; 186 } clusterMsgDataPublish; 187 188 typedef struct { 189 uint64_t configEpoch; /* Config epoch of the specified instance. */ 190 char nodename[CLUSTER_NAMELEN]; /* Name of the slots owner. */ 191 unsigned char slots[CLUSTER_SLOTS/8]; /* Slots bitmap. */ 192 } clusterMsgDataUpdate; 193 194 union clusterMsgData { 195 /* PING, MEET and PONG */ 196 struct { 197 /* Array of N clusterMsgDataGossip structures */ 198 clusterMsgDataGossip gossip[1]; 199 } ping; 200 201 /* FAIL */ 202 struct { 203 clusterMsgDataFail about; 204 } fail; 205 206 /* PUBLISH */ 207 struct { 208 clusterMsgDataPublish msg; 209 } publish; 210 211 /* UPDATE */ 212 struct { 213 clusterMsgDataUpdate nodecfg; 214 } update; 215 }; 216 217 #define CLUSTER_PROTO_VER 0 /* Cluster bus protocol version. */ 218 219 typedef struct { 220 char sig[4]; /* Siganture "RCmb" (Redis Cluster message bus). */ 221 uint32_t totlen; /* Total length of this message */ 222 uint16_t ver; /* Protocol version, currently set to 0. */ 223 uint16_t notused0; /* 2 bytes not used. */ 224 uint16_t type; /* Message type */ 225 uint16_t count; /* Only used for some kind of messages. */ 226 uint64_t currentEpoch; /* The epoch accordingly to the sending node. */ 227 uint64_t configEpoch; /* The config epoch if it's a master, or the last 228 epoch advertised by its master if it is a 229 slave. */ 230 uint64_t offset; /* Master replication offset if node is a master or 231 processed replication offset if node is a slave. */ 232 char sender[CLUSTER_NAMELEN]; /* Name of the sender node */ 233 unsigned char myslots[CLUSTER_SLOTS/8]; 234 char slaveof[CLUSTER_NAMELEN]; 235 char notused1[32]; /* 32 bytes reserved for future usage. */ 236 uint16_t port; /* Sender TCP base port */ 237 uint16_t flags; /* Sender node flags */ 238 unsigned char state; /* Cluster state from the POV of the sender */ 239 unsigned char mflags[3]; /* Message flags: CLUSTERMSG_FLAG[012]_... */ 240 union clusterMsgData data; 241 } clusterMsg; 242 243 #define CLUSTERMSG_MIN_LEN (sizeof(clusterMsg)-sizeof(union clusterMsgData)) 244 245 /* Message flags better specify the packet content or are used to 246 * provide some information about the node state. */ 247 #define CLUSTERMSG_FLAG0_PAUSED (1<<0) /* Master paused for manual failover. */ 248 #define CLUSTERMSG_FLAG0_FORCEACK (1<<1) /* Give ACK to AUTH_REQUEST even if 249 master is up. */ 250 251 /* ---------------------- API exported outside cluster.c -------------------- */ 252 clusterNode *getNodeByQuery(client *c, struct redisCommand *cmd, robj **argv, int argc, int *hashslot, int *ask); 253 int clusterRedirectBlockedClientIfNeeded(client *c); 254 void clusterRedirectClient(client *c, clusterNode *n, int hashslot, int error_code); 255 256 #endif /* __CLUSTER_H */ 257