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