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