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 amount 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_DEFAULT_REQUIRE_FULL_COVERAGE 1 19 #define REDIS_CLUSTER_FAIL_REPORT_VALIDITY_MULT 2 /* Fail report validity. */ 20 #define REDIS_CLUSTER_FAIL_UNDO_TIME_MULT 2 /* Undo fail if master is back. */ 21 #define REDIS_CLUSTER_FAIL_UNDO_TIME_ADD 10 /* Some additional 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 promoted 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 /* Reasons why a slave is not able to failover. */ 66 #define REDIS_CLUSTER_CANT_FAILOVER_NONE 0 67 #define REDIS_CLUSTER_CANT_FAILOVER_DATA_AGE 1 68 #define REDIS_CLUSTER_CANT_FAILOVER_WAITING_DELAY 2 69 #define REDIS_CLUSTER_CANT_FAILOVER_EXPIRED 3 70 #define REDIS_CLUSTER_CANT_FAILOVER_WAITING_VOTES 4 71 #define REDIS_CLUSTER_CANT_FAILOVER_RELOG_PERIOD (60*5) /* seconds. */ 72 73 /* This structure represent elements of node->fail_reports. */ 74 typedef struct clusterNodeFailReport { 75 struct clusterNode *node; /* Node reporting the failure condition. */ 76 mstime_t time; /* Time of the last report from this node. */ 77 } clusterNodeFailReport; 78 79 typedef struct clusterNode { 80 mstime_t ctime; /* Node object creation time. */ 81 char name[REDIS_CLUSTER_NAMELEN]; /* Node name, hex string, sha1-size */ 82 int flags; /* REDIS_NODE_... */ 83 uint64_t configEpoch; /* Last configEpoch observed for this node */ 84 unsigned char slots[REDIS_CLUSTER_SLOTS/8]; /* slots handled by this node */ 85 int numslots; /* Number of slots handled by this node */ 86 int numslaves; /* Number of slave nodes, if this is a master */ 87 struct clusterNode **slaves; /* pointers to slave nodes */ 88 struct clusterNode *slaveof; /* pointer to the master node */ 89 mstime_t ping_sent; /* Unix time we sent latest ping */ 90 mstime_t pong_received; /* Unix time we received the pong */ 91 mstime_t fail_time; /* Unix time when FAIL flag was set */ 92 mstime_t voted_time; /* Last time we voted for a slave of this master */ 93 mstime_t repl_offset_time; /* Unix time we received offset for this node */ 94 long long repl_offset; /* Last known repl offset for this node. */ 95 char ip[REDIS_IP_STR_LEN]; /* Latest known IP address of this node */ 96 int port; /* Latest known port of this node */ 97 clusterLink *link; /* TCP/IP link with this node */ 98 list *fail_reports; /* List of nodes signaling this as failing */ 99 } clusterNode; 100 101 typedef struct clusterState { 102 clusterNode *myself; /* This node */ 103 uint64_t currentEpoch; 104 int state; /* REDIS_CLUSTER_OK, REDIS_CLUSTER_FAIL, ... */ 105 int size; /* Num of master nodes with at least one slot */ 106 dict *nodes; /* Hash table of name -> clusterNode structures */ 107 dict *nodes_black_list; /* Nodes we don't re-add for a few seconds. */ 108 clusterNode *migrating_slots_to[REDIS_CLUSTER_SLOTS]; 109 clusterNode *importing_slots_from[REDIS_CLUSTER_SLOTS]; 110 clusterNode *slots[REDIS_CLUSTER_SLOTS]; 111 zskiplist *slots_to_keys; 112 /* The following fields are used to take the slave state on elections. */ 113 mstime_t failover_auth_time; /* Time of previous or next election. */ 114 int failover_auth_count; /* Number of votes received so far. */ 115 int failover_auth_sent; /* True if we already asked for votes. */ 116 int failover_auth_rank; /* This slave rank for current auth request. */ 117 uint64_t failover_auth_epoch; /* Epoch of the current election. */ 118 int cant_failover_reason; /* Why a slave is currently not able to 119 failover. See the CANT_FAILOVER_* macros. */ 120 /* Manual failover state in common. */ 121 mstime_t mf_end; /* Manual failover time limit (ms unixtime). 122 It is zero if there is no MF in progress. */ 123 /* Manual failover state of master. */ 124 clusterNode *mf_slave; /* Slave performing the manual failover. */ 125 /* Manual failover state of slave. */ 126 long long mf_master_offset; /* Master offset the slave needs to start MF 127 or zero if stil not received. */ 128 int mf_can_start; /* If non-zero signal that the manual failover 129 can start requesting masters vote. */ 130 /* The followign fields are used by masters to take state on elections. */ 131 uint64_t lastVoteEpoch; /* Epoch of the last vote granted. */ 132 int todo_before_sleep; /* Things to do in clusterBeforeSleep(). */ 133 long long stats_bus_messages_sent; /* Num of msg sent via cluster bus. */ 134 long long stats_bus_messages_received; /* Num of msg rcvd via cluster bus.*/ 135 } clusterState; 136 137 /* clusterState todo_before_sleep flags. */ 138 #define CLUSTER_TODO_HANDLE_FAILOVER (1<<0) 139 #define CLUSTER_TODO_UPDATE_STATE (1<<1) 140 #define CLUSTER_TODO_SAVE_CONFIG (1<<2) 141 #define CLUSTER_TODO_FSYNC_CONFIG (1<<3) 142 143 /* Redis cluster messages header */ 144 145 /* Note that the PING, PONG and MEET messages are actually the same exact 146 * kind of packet. PONG is the reply to ping, in the exact format as a PING, 147 * while MEET is a special PING that forces the receiver to add the sender 148 * as a node (if it is not already in the list). */ 149 #define CLUSTERMSG_TYPE_PING 0 /* Ping */ 150 #define CLUSTERMSG_TYPE_PONG 1 /* Pong (reply to Ping) */ 151 #define CLUSTERMSG_TYPE_MEET 2 /* Meet "let's join" message */ 152 #define CLUSTERMSG_TYPE_FAIL 3 /* Mark node xxx as failing */ 153 #define CLUSTERMSG_TYPE_PUBLISH 4 /* Pub/Sub Publish propagation */ 154 #define CLUSTERMSG_TYPE_FAILOVER_AUTH_REQUEST 5 /* May I failover? */ 155 #define CLUSTERMSG_TYPE_FAILOVER_AUTH_ACK 6 /* Yes, you have my vote */ 156 #define CLUSTERMSG_TYPE_UPDATE 7 /* Another node slots configuration */ 157 #define CLUSTERMSG_TYPE_MFSTART 8 /* Pause clients for manual failover */ 158 159 /* Initially we don't know our "name", but we'll find it once we connect 160 * to the first node, using the getsockname() function. Then we'll use this 161 * address for all the next messages. */ 162 typedef struct { 163 char nodename[REDIS_CLUSTER_NAMELEN]; 164 uint32_t ping_sent; 165 uint32_t pong_received; 166 char ip[REDIS_IP_STR_LEN]; /* IP address last time it was seen */ 167 uint16_t port; /* port last time it was seen */ 168 uint16_t flags; 169 uint32_t notused; /* for 64 bit alignment */ 170 } clusterMsgDataGossip; 171 172 typedef struct { 173 char nodename[REDIS_CLUSTER_NAMELEN]; 174 } clusterMsgDataFail; 175 176 typedef struct { 177 uint32_t channel_len; 178 uint32_t message_len; 179 /* We can't reclare bulk_data as bulk_data[] since this structure is 180 * nested. The 8 bytes are removed from the count during the message 181 * length computation. */ 182 unsigned char bulk_data[8]; 183 } clusterMsgDataPublish; 184 185 typedef struct { 186 uint64_t configEpoch; /* Config epoch of the specified instance. */ 187 char nodename[REDIS_CLUSTER_NAMELEN]; /* Name of the slots owner. */ 188 unsigned char slots[REDIS_CLUSTER_SLOTS/8]; /* Slots bitmap. */ 189 } clusterMsgDataUpdate; 190 191 union clusterMsgData { 192 /* PING, MEET and PONG */ 193 struct { 194 /* Array of N clusterMsgDataGossip structures */ 195 clusterMsgDataGossip gossip[1]; 196 } ping; 197 198 /* FAIL */ 199 struct { 200 clusterMsgDataFail about; 201 } fail; 202 203 /* PUBLISH */ 204 struct { 205 clusterMsgDataPublish msg; 206 } publish; 207 208 /* UPDATE */ 209 struct { 210 clusterMsgDataUpdate nodecfg; 211 } update; 212 }; 213 214 215 typedef struct { 216 char sig[4]; /* Siganture "RCmb" (Redis Cluster message bus). */ 217 uint32_t totlen; /* Total length of this message */ 218 uint16_t ver; /* Protocol version, currently set to 0. */ 219 uint16_t notused0; /* 2 bytes not used. */ 220 uint16_t type; /* Message type */ 221 uint16_t count; /* Only used for some kind of messages. */ 222 uint64_t currentEpoch; /* The epoch accordingly to the sending node. */ 223 uint64_t configEpoch; /* The config epoch if it's a master, or the last 224 epoch advertised by its master if it is a 225 slave. */ 226 uint64_t offset; /* Master replication offset if node is a master or 227 processed replication offset if node is a slave. */ 228 char sender[REDIS_CLUSTER_NAMELEN]; /* Name of the sender node */ 229 unsigned char myslots[REDIS_CLUSTER_SLOTS/8]; 230 char slaveof[REDIS_CLUSTER_NAMELEN]; 231 char notused1[32]; /* 32 bytes reserved for future usage. */ 232 uint16_t port; /* Sender TCP base port */ 233 uint16_t flags; /* Sender node flags */ 234 unsigned char state; /* Cluster state from the POV of the sender */ 235 unsigned char mflags[3]; /* Message flags: CLUSTERMSG_FLAG[012]_... */ 236 union clusterMsgData data; 237 } clusterMsg; 238 239 #define CLUSTERMSG_MIN_LEN (sizeof(clusterMsg)-sizeof(union clusterMsgData)) 240 241 /* Message flags better specify the packet content or are used to 242 * provide some information about the node state. */ 243 #define CLUSTERMSG_FLAG0_PAUSED (1<<0) /* Master paused for manual failover. */ 244 #define CLUSTERMSG_FLAG0_FORCEACK (1<<1) /* Give ACK to AUTH_REQUEST even if 245 master is up. */ 246 247 /* ---------------------- API exported outside cluster.c -------------------- */ 248 clusterNode *getNodeByQuery(redisClient *c, struct redisCommand *cmd, robj **argv, int argc, int *hashslot, int *ask); 249 250 #endif /* __REDIS_CLUSTER_H */ 251