CLUSTER GETKEYSINSLOT implemented

[redis.git] / src / db.c

diff --git a/src/db.c b/src/db.c
index 7ae23fa0059ae51ac841bc491ec2f75f57aeea18..670e2bce4e850496a783dc8992d1dc1414bb1caf 100644 (file)
--- a/src/db.c
+++ b/src/db.c
@@ -2,10 +2,42 @@
 
 #include <signal.h>
 
 
 #include <signal.h>
 

+void SlotToKeyAdd(robj *key);
+void SlotToKeyDel(robj *key);
+

 /*-----------------------------------------------------------------------------
  * C-level DB API
  *----------------------------------------------------------------------------*/
 
 /*-----------------------------------------------------------------------------
  * C-level DB API
  *----------------------------------------------------------------------------*/
 

+/* Important notes on lookup and disk store.
+ *
+ * When disk store is enabled on lookup we can have different cases.
+ *
+ * a) The key is in memory:
+ *    - If the key is not in IO_SAVEINPROG state we can access it.
+ *      As if it's just IO_SAVE this means we have the key in the IO queue
+ *      but can't be accessed by the IO thread (it requires to be
+ *      translated into an IO Job by the cache cron function.)
+ *    - If the key is in IO_SAVEINPROG we can't touch the key and have
+ *      to blocking wait completion of operations.
+ * b) The key is not in memory:
+ *    - If it's marked as non existing on disk as well (negative cache)
+ *      we don't need to perform the disk access.
+ *    - if the key MAY EXIST, but is not in memory, and it is marked as IO_SAVE
+ *      then the key can only be a deleted one. As IO_SAVE keys are never
+ *      evicted (dirty state), so the only possibility is that key was deleted.
+ *    - if the key MAY EXIST we need to blocking load it.
+ *      We check that the key is not in IO_SAVEINPROG state before accessing
+ *      the disk object. If it is in this state, we wait.
+ */
+
+void lookupWaitBusyKey(redisDb *db, robj *key) {
+    /* FIXME: wait just for this key, not everything */
+    waitEmptyIOJobsQueue();
+    processAllPendingIOJobs();
+    redisAssert((cacheScheduleIOGetFlags(db,key) & REDIS_IO_SAVEINPROG) == 0);
+}
+

 robj *lookupKey(redisDb *db, robj *key) {
     dictEntry *de = dictFind(db->dict,key->ptr);
     if (de) {
 robj *lookupKey(redisDb *db, robj *key) {
     dictEntry *de = dictFind(db->dict,key->ptr);
     if (de) {


@@ -17,12 +49,12 @@ robj *lookupKey(redisDb *db, robj *key) {
         if (server.bgsavechildpid == -1 && server.bgrewritechildpid == -1)
             val->lru = server.lruclock;
 
         if (server.bgsavechildpid == -1 && server.bgrewritechildpid == -1)
             val->lru = server.lruclock;
 

-        if (server.ds_enabled && val->storage == REDIS_DS_SAVING) {
-            /* FIXME: change this code to just wait for our object to
-             * get out of the IO Job. */
-            waitEmptyIOJobsQueue();
-            processAllPendingIOJobs();
-            redisAssert(val->storage != REDIS_DS_SAVING);
+        if (server.ds_enabled &&
+            cacheScheduleIOGetFlags(db,key) & REDIS_IO_SAVEINPROG)
+        {
+            /* Need to wait for the key to get unbusy */
+            redisLog(REDIS_DEBUG,"Lookup found a key in SAVEINPROG state. Waiting. (Key was in the cache)");
+            lookupWaitBusyKey(db,key);

         }
         server.stat_keyspace_hits++;
         return val;
         }
         server.stat_keyspace_hits++;
         return val;


@@ -32,12 +64,26 @@ robj *lookupKey(redisDb *db, robj *key) {
 
         /* Key not found in the in memory hash table, but if disk store is
          * enabled we may have this key on disk. If so load it in memory
 
         /* Key not found in the in memory hash table, but if disk store is
          * enabled we may have this key on disk. If so load it in memory

-         * in a blocking way.
-         * 
-         * FIXME: race condition here. If there was an already scheduled
-         * async loading of this key, what may happen is that the old
-         * key is loaded in memory if this gets deleted in the meantime. */
+         * in a blocking way. */

         if (server.ds_enabled && cacheKeyMayExist(db,key)) {
         if (server.ds_enabled && cacheKeyMayExist(db,key)) {

+            long flags = cacheScheduleIOGetFlags(db,key);
+
+            /* They key is not in cache, but it has a SAVE op in queue?
+             * The only possibility is that the key was deleted, since
+             * dirty keys are not evicted. */
+            if (flags & REDIS_IO_SAVE) {
+                server.stat_keyspace_misses++;
+                return NULL;
+            }
+
+            /* At this point we need to blocking load the key in memory.
+             * The first thing we do is waiting here if the key is busy. */
+            if (flags & REDIS_IO_SAVEINPROG) {
+                redisLog(REDIS_DEBUG,"Lookup found a key in SAVEINPROG state. Waiting (while force loading).");
+                lookupWaitBusyKey(db,key);
+            }
+
+            redisLog(REDIS_DEBUG,"Force loading key %s via lookup", key->ptr);

             val = dsGet(db,key,&expire);
             if (val) {
                 int retval = dbAdd(db,key,val);
             val = dsGet(db,key,&expire);
             if (val) {
                 int retval = dbAdd(db,key,val);


@@ -45,6 +91,8 @@ robj *lookupKey(redisDb *db, robj *key) {
                 if (expire != -1) setExpire(db,key,expire);
                 server.stat_keyspace_hits++;
                 return val;
                 if (expire != -1) setExpire(db,key,expire);
                 server.stat_keyspace_hits++;
                 return val;

+            } else {
+                cacheSetKeyDoesNotExist(db,key);

             }
         }
         server.stat_keyspace_misses++;
             }
         }
         server.stat_keyspace_misses++;


@@ -85,9 +133,8 @@ int dbAdd(redisDb *db, robj *key, robj *val) {
     } else {
         sds copy = sdsdup(key->ptr);
         dictAdd(db->dict, copy, val);
     } else {
         sds copy = sdsdup(key->ptr);
         dictAdd(db->dict, copy, val);

-        if (server.ds_enabled) {
-            /* FIXME: remove entry from negative cache */
-        }
+        if (server.ds_enabled) cacheSetKeyMayExist(db,key);
+        if (server.cluster_enabled) SlotToKeyAdd(key);

         return REDIS_OK;
     }
 }
         return REDIS_OK;
     }
 }


@@ -97,14 +144,20 @@ int dbAdd(redisDb *db, robj *key, robj *val) {
  *
  * On update (key already existed) 0 is returned. Otherwise 1. */
 int dbReplace(redisDb *db, robj *key, robj *val) {
  *
  * On update (key already existed) 0 is returned. Otherwise 1. */
 int dbReplace(redisDb *db, robj *key, robj *val) {

-    if (dictFind(db->dict,key->ptr) == NULL) {
+    robj *oldval;
+    int retval;
+
+    if ((oldval = dictFetchValue(db->dict,key->ptr)) == NULL) {

         sds copy = sdsdup(key->ptr);
         dictAdd(db->dict, copy, val);
         sds copy = sdsdup(key->ptr);
         dictAdd(db->dict, copy, val);

-        return 1;
+        if (server.cluster_enabled) SlotToKeyAdd(key);
+        retval = 1;

     } else {
         dictReplace(db->dict, key->ptr, val);
     } else {
         dictReplace(db->dict, key->ptr, val);

-        return 0;
+        retval = 0;

     }
     }

+    if (server.ds_enabled) cacheSetKeyMayExist(db,key);
+    return retval;

 }
 
 int dbExists(redisDb *db, robj *key) {
 }
 
 int dbExists(redisDb *db, robj *key) {


@@ -139,22 +192,27 @@ robj *dbRandomKey(redisDb *db) {
 
 /* Delete a key, value, and associated expiration entry if any, from the DB */
 int dbDelete(redisDb *db, robj *key) {
 
 /* Delete a key, value, and associated expiration entry if any, from the DB */
 int dbDelete(redisDb *db, robj *key) {

-    /* If VM is enabled make sure to awake waiting clients for this key:
-     * deleting the key will kill the I/O thread bringing the key from swap
-     * to memory, so the client will never be notified and unblocked if we
-     * don't do it now. */
-    if (server.ds_enabled) handleClientsBlockedOnSwappedKey(db,key);
-
-    /* FIXME: we need to delete the IO Job loading the key, or simply we can
-     * wait for it to finish. */
+    /* If diskstore is enabled make sure to awake waiting clients for this key
+     * as it is not really useful to wait for a key already deleted to be
+     * loaded from disk. */
+    if (server.ds_enabled) {
+        handleClientsBlockedOnSwappedKey(db,key);
+        cacheSetKeyDoesNotExist(db,key);
+    }

 
     /* Deleting an entry from the expires dict will not free the sds of
      * the key, because it is shared with the main dictionary. */
     if (dictSize(db->expires) > 0) dictDelete(db->expires,key->ptr);
 
     /* Deleting an entry from the expires dict will not free the sds of
      * the key, because it is shared with the main dictionary. */
     if (dictSize(db->expires) > 0) dictDelete(db->expires,key->ptr);

-    return dictDelete(db->dict,key->ptr) == DICT_OK;
+    if (dictDelete(db->dict,key->ptr) == DICT_OK) {
+        if (server.cluster_enabled) SlotToKeyDel(key);
+        return 1;
+    } else {
+        return 0;
+    }

 }
 
 }
 

-/* Empty the whole database */
+/* Empty the whole database.
+ * If diskstore is enabled this function will just flush the in-memory cache. */

 long long emptyDb() {
     int j;
     long long removed = 0;
 long long emptyDb() {
     int j;
     long long removed = 0;


@@ -163,6 +221,7 @@ long long emptyDb() {
         removed += dictSize(server.db[j].dict);
         dictEmpty(server.db[j].dict);
         dictEmpty(server.db[j].expires);
         removed += dictSize(server.db[j].dict);
         dictEmpty(server.db[j].dict);
         dictEmpty(server.db[j].expires);

+        if (server.ds_enabled) dictEmpty(server.db[j].io_negcache);

     }
     return removed;
 }
     }
     return removed;
 }


@@ -186,13 +245,11 @@ int selectDb(redisClient *c, int id) {
 void signalModifiedKey(redisDb *db, robj *key) {
     touchWatchedKey(db,key);
     if (server.ds_enabled)
 void signalModifiedKey(redisDb *db, robj *key) {
     touchWatchedKey(db,key);
     if (server.ds_enabled)

-        cacheScheduleForFlush(db,key);
+        cacheScheduleIO(db,key,REDIS_IO_SAVE);

 }
 
 void signalFlushedDb(int dbid) {
     touchWatchedKeysOnFlush(dbid);
 }
 
 void signalFlushedDb(int dbid) {
     touchWatchedKeysOnFlush(dbid);

-    if (server.ds_enabled)
-        dsFlushDb(dbid);

 }
 
 /*-----------------------------------------------------------------------------
 }
 
 /*-----------------------------------------------------------------------------


@@ -204,6 +261,7 @@ void flushdbCommand(redisClient *c) {
     signalFlushedDb(c->db->id);
     dictEmpty(c->db->dict);
     dictEmpty(c->db->expires);
     signalFlushedDb(c->db->id);
     dictEmpty(c->db->dict);
     dictEmpty(c->db->expires);

+    if (server.ds_enabled) dsFlushDb(c->db->id);

     addReply(c,shared.ok);
 }
 
     addReply(c,shared.ok);
 }
 


@@ -215,7 +273,10 @@ void flushallCommand(redisClient *c) {
         kill(server.bgsavechildpid,SIGKILL);
         rdbRemoveTempFile(server.bgsavechildpid);
     }
         kill(server.bgsavechildpid,SIGKILL);
         rdbRemoveTempFile(server.bgsavechildpid);
     }

-    rdbSave(server.dbfilename);
+    if (server.ds_enabled)
+        dsFlushDb(-1);
+    else
+        rdbSave(server.dbfilename);

     server.dirty++;
 }
 
     server.dirty++;
 }
 


@@ -236,7 +297,7 @@ void delCommand(redisClient *c) {
             if (cacheKeyMayExist(c->db,c->argv[j]) &&
                 dsExists(c->db,c->argv[j]))
             {
             if (cacheKeyMayExist(c->db,c->argv[j]) &&
                 dsExists(c->db,c->argv[j]))
             {

-                cacheScheduleForFlush(c->db,c->argv[j]);
+                cacheScheduleIO(c->db,c->argv[j],REDIS_IO_SAVE);

                 deleted = 1;
             }
         }
                 deleted = 1;
             }
         }


@@ -256,6 +317,10 @@ void existsCommand(redisClient *c) {
 void selectCommand(redisClient *c) {
     int id = atoi(c->argv[1]->ptr);
 
 void selectCommand(redisClient *c) {
     int id = atoi(c->argv[1]->ptr);
 

+    if (server.cluster_enabled) {
+        addReplyError(c,"SELECT is not allowed in cluster mode");
+        return;
+    }

     if (selectDb(c,id) == REDIS_ERR) {
         addReplyError(c,"invalid DB index");
     } else {
     if (selectDb(c,id) == REDIS_ERR) {
         addReplyError(c,"invalid DB index");
     } else {


@@ -330,30 +395,6 @@ void typeCommand(redisClient *c) {
     addReplyStatus(c,type);
 }
 
     addReplyStatus(c,type);
 }
 

-void saveCommand(redisClient *c) {
-    if (server.bgsavechildpid != -1) {
-        addReplyError(c,"Background save already in progress");
-        return;
-    }
-    if (rdbSave(server.dbfilename) == REDIS_OK) {
-        addReply(c,shared.ok);
-    } else {
-        addReply(c,shared.err);
-    }
-}
-
-void bgsaveCommand(redisClient *c) {
-    if (server.bgsavechildpid != -1) {
-        addReplyError(c,"Background save already in progress");
-        return;
-    }
-    if (rdbSaveBackground(server.dbfilename) == REDIS_OK) {
-        addReplyStatus(c,"Background saving started");
-    } else {
-        addReply(c,shared.err);
-    }
-}
-

 void shutdownCommand(redisClient *c) {
     if (prepareForShutdown() == REDIS_OK)
         exit(0);
 void shutdownCommand(redisClient *c) {
     if (prepareForShutdown() == REDIS_OK)
         exit(0);


@@ -401,6 +442,11 @@ void moveCommand(redisClient *c) {
     redisDb *src, *dst;
     int srcid;
 
     redisDb *src, *dst;
     int srcid;
 

+    if (server.cluster_enabled) {
+        addReplyError(c,"MOVE is not allowed in cluster mode");
+        return;
+    }
+

     /* Obtain source and target DB pointers */
     src = c->db;
     srcid = c->db->id;
     /* Obtain source and target DB pointers */
     src = c->db;
     srcid = c->db->id;


@@ -500,6 +546,8 @@ void propagateExpire(redisDb *db, robj *key) {
 int expireIfNeeded(redisDb *db, robj *key) {
     time_t when = getExpire(db,key);
 
 int expireIfNeeded(redisDb *db, robj *key) {
     time_t when = getExpire(db,key);
 

+    if (when < 0) return 0; /* No expire for this key */
+

     /* If we are running in the context of a slave, return ASAP:
      * the slave key expiration is controlled by the master that will
      * send us synthesized DEL operations for expired keys.
     /* If we are running in the context of a slave, return ASAP:
      * the slave key expiration is controlled by the master that will
      * send us synthesized DEL operations for expired keys.


@@ -511,8 +559,6 @@ int expireIfNeeded(redisDb *db, robj *key) {
         return time(NULL) > when;
     }
 
         return time(NULL) > when;
     }
 

-    if (when < 0) return 0;
-

     /* Return when this key has not expired */
     if (time(NULL) <= when) return 0;
 
     /* Return when this key has not expired */
     if (time(NULL) <= when) return 0;
 


@@ -565,6 +611,7 @@ void expireatCommand(redisClient *c) {
 void ttlCommand(redisClient *c) {
     time_t expire, ttl = -1;
 
 void ttlCommand(redisClient *c) {
     time_t expire, ttl = -1;
 

+    if (server.ds_enabled) lookupKeyRead(c->db,c->argv[1]);

     expire = getExpire(c->db,c->argv[1]);
     if (expire != -1) {
         ttl = (expire-time(NULL));
     expire = getExpire(c->db,c->argv[1]);
     if (expire != -1) {
         ttl = (expire-time(NULL));


@@ -588,3 +635,108 @@ void persistCommand(redisClient *c) {
         }
     }
 }
         }
     }
 }

+
+/* -----------------------------------------------------------------------------
+ * API to get key arguments from commands
+ * ---------------------------------------------------------------------------*/
+
+int *getKeysUsingCommandTable(struct redisCommand *cmd,robj **argv, int argc, int *numkeys) {
+    int j, i = 0, last, *keys;
+    REDIS_NOTUSED(argv);
+
+    if (cmd->firstkey == 0) {
+        *numkeys = 0;
+        return NULL;
+    }
+    last = cmd->lastkey;
+    if (last < 0) last = argc+last;
+    keys = zmalloc(sizeof(int)*((last - cmd->firstkey)+1));
+    for (j = cmd->firstkey; j <= last; j += cmd->keystep) {
+        redisAssert(j < argc);
+        keys[i++] = j;
+    }
+    *numkeys = i;
+    return keys;
+}
+
+int *getKeysFromCommand(struct redisCommand *cmd,robj **argv, int argc, int *numkeys, int flags) {
+    if (cmd->getkeys_proc) {
+        return cmd->getkeys_proc(cmd,argv,argc,numkeys,flags);
+    } else {
+        return getKeysUsingCommandTable(cmd,argv,argc,numkeys);
+    }
+}
+
+void getKeysFreeResult(int *result) {
+    zfree(result);
+}
+
+int *noPreloadGetKeys(struct redisCommand *cmd,robj **argv, int argc, int *numkeys, int flags) {
+    if (flags & REDIS_GETKEYS_PRELOAD) {
+        *numkeys = 0;
+        return NULL;
+    } else {
+        return getKeysUsingCommandTable(cmd,argv,argc,numkeys);
+    }
+}
+
+int *renameGetKeys(struct redisCommand *cmd,robj **argv, int argc, int *numkeys, int flags) {
+    if (flags & REDIS_GETKEYS_PRELOAD) {
+        int *keys = zmalloc(sizeof(int));
+        *numkeys = 1;
+        keys[0] = 1;
+        return keys;
+    } else {
+        return getKeysUsingCommandTable(cmd,argv,argc,numkeys);
+    }
+}
+
+int *zunionInterGetKeys(struct redisCommand *cmd,robj **argv, int argc, int *numkeys, int flags) {
+    int i, num, *keys;
+    REDIS_NOTUSED(cmd);
+    REDIS_NOTUSED(flags);
+
+    num = atoi(argv[2]->ptr);
+    /* Sanity check. Don't return any key if the command is going to
+     * reply with syntax error. */
+    if (num > (argc-3)) {
+        *numkeys = 0;
+        return NULL;
+    }
+    keys = zmalloc(sizeof(int)*num);
+    for (i = 0; i < num; i++) keys[i] = 3+i;
+    *numkeys = num;
+    return keys;
+}
+
+/* Slot to Key API. This is used by Redis Cluster in order to obtain in
+ * a fast way a key that belongs to a specified hash slot. This is useful
+ * while rehashing the cluster. */
+void SlotToKeyAdd(robj *key) {
+    unsigned int hashslot = keyHashSlot(key->ptr,sdslen(key->ptr));
+
+    zslInsert(server.cluster.slots_to_keys,hashslot,key);
+    incrRefCount(key);
+}
+
+void SlotToKeyDel(robj *key) {
+    unsigned int hashslot = keyHashSlot(key->ptr,sdslen(key->ptr));
+
+    zslDelete(server.cluster.slots_to_keys,hashslot,key);
+}
+
+unsigned int GetKeysInSlot(unsigned int hashslot, robj **keys, unsigned int count) {
+    zskiplistNode *n;
+    zrangespec range;
+    int j = 0;
+
+    range.min = range.max = hashslot;
+    range.minex = range.maxex = 0;
+    
+    n = zslFirstInRange(server.cluster.slots_to_keys, range);
+    while(n && n->score == hashslot && count--) {
+        keys[j++] = n->obj;
+        n = n->level[0].forward;
+    }
+    return j;
+}