]> git.saurik.com Git - redis.git/blobdiff - src/aof.c
Fix: when aof_write_rewrite is true don't append on the AOF buffer but accumulate...
[redis.git] / src / aof.c
index 517a19d358588c0f04b549cf0cc38ad53d23d282..15a456f9de00269d4402682e1cfa39b112240314 100644 (file)
--- a/src/aof.c
+++ b/src/aof.c
@@ -1,5 +1,6 @@
 #include "redis.h"
 #include "bio.h"
+#include "rio.h"
 
 #include <signal.h>
 #include <fcntl.h>
@@ -25,6 +26,7 @@ void stopAppendOnly(void) {
     server.appendfd = -1;
     server.appendseldb = -1;
     server.appendonly = 0;
+    server.aof_wait_rewrite = 0;
     /* rewrite operation in progress? kill it, wait child exit */
     if (server.bgrewritechildpid != -1) {
         int statloc;
@@ -34,6 +36,7 @@ void stopAppendOnly(void) {
         /* reset the buffer accumulating changes while the child saves */
         sdsfree(server.bgrewritebuf);
         server.bgrewritebuf = sdsempty();
+        aofRemoveTempFile(server.bgrewritechildpid);
         server.bgrewritechildpid = -1;
     }
 }
@@ -45,15 +48,18 @@ int startAppendOnly(void) {
     server.lastfsync = time(NULL);
     server.appendfd = open(server.appendfilename,O_WRONLY|O_APPEND|O_CREAT,0644);
     if (server.appendfd == -1) {
-        redisLog(REDIS_WARNING,"Used tried to switch on AOF via CONFIG, but I can't open the AOF file: %s",strerror(errno));
+        redisLog(REDIS_WARNING,"Redis needs to enable the AOF but can't open the append only file: %s",strerror(errno));
         return REDIS_ERR;
     }
     if (rewriteAppendOnlyFileBackground() == REDIS_ERR) {
         server.appendonly = 0;
         close(server.appendfd);
-        redisLog(REDIS_WARNING,"Used tried to switch on AOF via CONFIG, I can't trigger a background AOF rewrite operation. Check the above logs for more info about the error.",strerror(errno));
+        redisLog(REDIS_WARNING,"Redis needs to enable the AOF but can't trigger a background AOF rewrite operation. Check the above logs for more info about the error.");
         return REDIS_ERR;
     }
+    /* We correctly switched on AOF, now wait for the rerwite to be complete
+     * in order to append data on disk. */
+    server.aof_wait_rewrite = 1;
     return REDIS_OK;
 }
 
@@ -95,12 +101,13 @@ void flushAppendOnlyFile(int force) {
                 server.aof_flush_postponed_start = server.unixtime;
                 return;
             } else if (server.unixtime - server.aof_flush_postponed_start < 2) {
-                /* We were already writing for fsync to finish, but for less
+                /* We were already waiting for fsync to finish, but for less
                  * than two seconds this is still ok. Postpone again. */
                 return;
             }
             /* Otherwise fall trough, and go write since we can't wait
              * over two seconds. */
+            redisLog(REDIS_NOTICE,"Asynchronous AOF fsync is taking too long (disk is busy?). Writing the AOF buffer without waiting for fsync to complete, this may slow down Redis.");
         }
     }
     /* If you are following this code path, then we are going to write so
@@ -179,21 +186,38 @@ sds catAppendOnlyGenericCommand(sds dst, int argc, robj **argv) {
     return dst;
 }
 
-sds catAppendOnlyExpireAtCommand(sds buf, robj *key, robj *seconds) {
-    int argc = 3;
-    long when;
+/* Create the sds representation of an PEXPIREAT command, using
+ * 'seconds' as time to live and 'cmd' to understand what command
+ * we are translating into a PEXPIREAT.
+ *
+ * This command is used in order to translate EXPIRE and PEXPIRE commands
+ * into PEXPIREAT command so that we retain precision in the append only
+ * file, and the time is always absolute and not relative. */
+sds catAppendOnlyExpireAtCommand(sds buf, struct redisCommand *cmd, robj *key, robj *seconds) {
+    long long when;
     robj *argv[3];
 
     /* Make sure we can use strtol */
     seconds = getDecodedObject(seconds);
-    when = time(NULL)+strtol(seconds->ptr,NULL,10);
+    when = strtoll(seconds->ptr,NULL,10);
+    /* Convert argument into milliseconds for EXPIRE, SETEX, EXPIREAT */
+    if (cmd->proc == expireCommand || cmd->proc == setexCommand ||
+        cmd->proc == expireatCommand)
+    {
+        when *= 1000;
+    }
+    /* Convert into absolute time for EXPIRE, PEXPIRE, SETEX, PSETEX */
+    if (cmd->proc == expireCommand || cmd->proc == pexpireCommand ||
+        cmd->proc == setexCommand || cmd->proc == psetexCommand)
+    {
+        when += mstime();
+    }
     decrRefCount(seconds);
 
-    argv[0] = createStringObject("EXPIREAT",8);
+    argv[0] = createStringObject("PEXPIREAT",9);
     argv[1] = key;
-    argv[2] = createObject(REDIS_STRING,
-        sdscatprintf(sdsempty(),"%ld",when));
-    buf = catAppendOnlyGenericCommand(buf, argc, argv);
+    argv[2] = createStringObjectFromLongLong(when);
+    buf = catAppendOnlyGenericCommand(buf, 3, argv);
     decrRefCount(argv[0]);
     decrRefCount(argv[2]);
     return buf;
@@ -214,25 +238,36 @@ void feedAppendOnlyFile(struct redisCommand *cmd, int dictid, robj **argv, int a
         server.appendseldb = dictid;
     }
 
-    if (cmd->proc == expireCommand) {
-        /* Translate EXPIRE into EXPIREAT */
-        buf = catAppendOnlyExpireAtCommand(buf,argv[1],argv[2]);
-    } else if (cmd->proc == setexCommand) {
-        /* Translate SETEX to SET and EXPIREAT */
+    if (cmd->proc == expireCommand || cmd->proc == pexpireCommand ||
+        cmd->proc == expireatCommand) {
+        /* Translate EXPIRE/PEXPIRE/EXPIREAT into PEXPIREAT */
+        buf = catAppendOnlyExpireAtCommand(buf,cmd,argv[1],argv[2]);
+    } else if (cmd->proc == setexCommand || cmd->proc == psetexCommand) {
+        /* Translate SETEX/PSETEX to SET and PEXPIREAT */
         tmpargv[0] = createStringObject("SET",3);
         tmpargv[1] = argv[1];
         tmpargv[2] = argv[3];
         buf = catAppendOnlyGenericCommand(buf,3,tmpargv);
         decrRefCount(tmpargv[0]);
-        buf = catAppendOnlyExpireAtCommand(buf,argv[1],argv[2]);
+        buf = catAppendOnlyExpireAtCommand(buf,cmd,argv[1],argv[2]);
     } else {
+        /* All the other commands don't need translation or need the
+         * same translation already operated in the command vector
+         * for the replication itself. */
         buf = catAppendOnlyGenericCommand(buf,argc,argv);
     }
 
     /* Append to the AOF buffer. This will be flushed on disk just before
      * of re-entering the event loop, so before the client will get a
-     * positive reply about the operation performed. */
-    server.aofbuf = sdscatlen(server.aofbuf,buf,sdslen(buf));
+     * positive reply about the operation performed.
+     *
+     * Note, we don't add stuff in the AOF buffer if aof_wait_rewrite is
+     * non zero, as this means we are starting with a new AOF and the
+     * current one is meaningless (this happens for instance after
+     * a slave resyncs with its master). */
+    if (!server.aof_wait_rewrite) {
+        server.aofbuf = sdscatlen(server.aofbuf,buf,sdslen(buf));
+    }
 
     /* If a background append only file rewriting is in progress we want to
      * accumulate the differences between the child DB and the current one
@@ -325,6 +360,8 @@ int loadAppendOnlyFile(char *filename) {
         }
         if (buf[0] != '*') goto fmterr;
         argc = atoi(buf+1);
+        if (argc < 1) goto fmterr;
+
         argv = zmalloc(sizeof(robj*)*argc);
         for (j = 0; j < argc; j++) {
             if (fgets(buf,sizeof(buf),fp) == NULL) goto readerr;
@@ -383,15 +420,258 @@ fmterr:
     exit(1);
 }
 
+/* Delegate writing an object to writing a bulk string or bulk long long.
+ * This is not placed in rio.c since that adds the redis.h dependency. */
+int rioWriteBulkObject(rio *r, robj *obj) {
+    /* Avoid using getDecodedObject to help copy-on-write (we are often
+     * in a child process when this function is called). */
+    if (obj->encoding == REDIS_ENCODING_INT) {
+        return rioWriteBulkLongLong(r,(long)obj->ptr);
+    } else if (obj->encoding == REDIS_ENCODING_RAW) {
+        return rioWriteBulkString(r,obj->ptr,sdslen(obj->ptr));
+    } else {
+        redisPanic("Unknown string encoding");
+    }
+}
+
+/* Emit the commands needed to rebuild a list object.
+ * The function returns 0 on error, 1 on success. */
+int rewriteListObject(rio *r, robj *key, robj *o) {
+    long long count = 0, items = listTypeLength(o);
+
+    if (o->encoding == REDIS_ENCODING_ZIPLIST) {
+        unsigned char *zl = o->ptr;
+        unsigned char *p = ziplistIndex(zl,0);
+        unsigned char *vstr;
+        unsigned int vlen;
+        long long vlong;
+
+        while(ziplistGet(p,&vstr,&vlen,&vlong)) {
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items) == 0) return 0;
+                if (rioWriteBulkString(r,"RPUSH",5) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (vstr) {
+                if (rioWriteBulkString(r,(char*)vstr,vlen) == 0) return 0;
+            } else {
+                if (rioWriteBulkLongLong(r,vlong) == 0) return 0;
+            }
+            p = ziplistNext(zl,p);
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+    } else if (o->encoding == REDIS_ENCODING_LINKEDLIST) {
+        list *list = o->ptr;
+        listNode *ln;
+        listIter li;
+
+        listRewind(list,&li);
+        while((ln = listNext(&li))) {
+            robj *eleobj = listNodeValue(ln);
+
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items) == 0) return 0;
+                if (rioWriteBulkString(r,"RPUSH",5) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkObject(r,eleobj) == 0) return 0;
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+    } else {
+        redisPanic("Unknown list encoding");
+    }
+    return 1;
+}
+
+/* Emit the commands needed to rebuild a set object.
+ * The function returns 0 on error, 1 on success. */
+int rewriteSetObject(rio *r, robj *key, robj *o) {
+    long long count = 0, items = setTypeSize(o);
+
+    if (o->encoding == REDIS_ENCODING_INTSET) {
+        int ii = 0;
+        int64_t llval;
+
+        while(intsetGet(o->ptr,ii++,&llval)) {
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items) == 0) return 0;
+                if (rioWriteBulkString(r,"SADD",4) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkLongLong(r,llval) == 0) return 0;
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+    } else if (o->encoding == REDIS_ENCODING_HT) {
+        dictIterator *di = dictGetIterator(o->ptr);
+        dictEntry *de;
+
+        while((de = dictNext(di)) != NULL) {
+            robj *eleobj = dictGetKey(de);
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items) == 0) return 0;
+                if (rioWriteBulkString(r,"SADD",4) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkObject(r,eleobj) == 0) return 0;
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+        dictReleaseIterator(di);
+    } else {
+        redisPanic("Unknown set encoding");
+    }
+    return 1;
+}
+
+/* Emit the commands needed to rebuild a sorted set object.
+ * The function returns 0 on error, 1 on success. */
+int rewriteSortedSetObject(rio *r, robj *key, robj *o) {
+    long long count = 0, items = zsetLength(o);
+
+    if (o->encoding == REDIS_ENCODING_ZIPLIST) {
+        unsigned char *zl = o->ptr;
+        unsigned char *eptr, *sptr;
+        unsigned char *vstr;
+        unsigned int vlen;
+        long long vll;
+        double score;
+
+        eptr = ziplistIndex(zl,0);
+        redisAssert(eptr != NULL);
+        sptr = ziplistNext(zl,eptr);
+        redisAssert(sptr != NULL);
+
+        while (eptr != NULL) {
+            redisAssert(ziplistGet(eptr,&vstr,&vlen,&vll));
+            score = zzlGetScore(sptr);
+
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items*2) == 0) return 0;
+                if (rioWriteBulkString(r,"ZADD",4) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkDouble(r,score) == 0) return 0;
+            if (vstr != NULL) {
+                if (rioWriteBulkString(r,(char*)vstr,vlen) == 0) return 0;
+            } else {
+                if (rioWriteBulkLongLong(r,vll) == 0) return 0;
+            }
+            zzlNext(zl,&eptr,&sptr);
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+    } else if (o->encoding == REDIS_ENCODING_SKIPLIST) {
+        zset *zs = o->ptr;
+        dictIterator *di = dictGetIterator(zs->dict);
+        dictEntry *de;
+
+        while((de = dictNext(di)) != NULL) {
+            robj *eleobj = dictGetKey(de);
+            double *score = dictGetVal(de);
+
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items*2) == 0) return 0;
+                if (rioWriteBulkString(r,"ZADD",4) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkDouble(r,*score) == 0) return 0;
+            if (rioWriteBulkObject(r,eleobj) == 0) return 0;
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+        dictReleaseIterator(di);
+    } else {
+        redisPanic("Unknown sorted zset encoding");
+    }
+    return 1;
+}
+
+/* Emit the commands needed to rebuild a hash object.
+ * The function returns 0 on error, 1 on success. */
+int rewriteHashObject(rio *r, robj *key, robj *o) {
+    long long count = 0, items = hashTypeLength(o);
+
+    if (o->encoding == REDIS_ENCODING_ZIPMAP) {
+        unsigned char *p = zipmapRewind(o->ptr);
+        unsigned char *field, *val;
+        unsigned int flen, vlen;
+
+        while((p = zipmapNext(p,&field,&flen,&val,&vlen)) != NULL) {
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items*2) == 0) return 0;
+                if (rioWriteBulkString(r,"HMSET",5) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkString(r,(char*)field,flen) == 0) return 0;
+            if (rioWriteBulkString(r,(char*)val,vlen) == 0) return 0;
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+    } else {
+        dictIterator *di = dictGetIterator(o->ptr);
+        dictEntry *de;
+
+        while((de = dictNext(di)) != NULL) {
+            robj *field = dictGetKey(de);
+            robj *val = dictGetVal(de);
+
+            if (count == 0) {
+                int cmd_items = (items > REDIS_AOFREWRITE_ITEMS_PER_CMD) ?
+                    REDIS_AOFREWRITE_ITEMS_PER_CMD : items;
+
+                if (rioWriteBulkCount(r,'*',2+cmd_items*2) == 0) return 0;
+                if (rioWriteBulkString(r,"HMSET",5) == 0) return 0;
+                if (rioWriteBulkObject(r,key) == 0) return 0;
+            }
+            if (rioWriteBulkObject(r,field) == 0) return 0;
+            if (rioWriteBulkObject(r,val) == 0) return 0;
+            if (++count == REDIS_AOFREWRITE_ITEMS_PER_CMD) count = 0;
+            items--;
+        }
+        dictReleaseIterator(di);
+    }
+    return 1;
+}
+
 /* Write a sequence of commands able to fully rebuild the dataset into
- * "filename". Used both by REWRITEAOF and BGREWRITEAOF. */
+ * "filename". Used both by REWRITEAOF and BGREWRITEAOF.
+ *
+ * In order to minimize the number of commands needed in the rewritten
+ * log Redis uses variadic commands when possible, such as RPUSH, SADD
+ * and ZADD. However at max REDIS_AOFREWRITE_ITEMS_PER_CMD items per time
+ * are inserted using a single command. */
 int rewriteAppendOnlyFile(char *filename) {
     dictIterator *di = NULL;
     dictEntry *de;
+    rio aof;
     FILE *fp;
     char tmpfile[256];
     int j;
-    time_t now = time(NULL);
+    long long now = mstime();
 
     /* Note that we have to use a different temp name here compared to the
      * one used by rewriteAppendOnlyFileBackground() function. */
@@ -401,6 +681,8 @@ int rewriteAppendOnlyFile(char *filename) {
         redisLog(REDIS_WARNING, "Failed rewriting the append only file: %s", strerror(errno));
         return REDIS_ERR;
     }
+
+    rioInitWithFile(&aof,fp);
     for (j = 0; j < server.dbnum; j++) {
         char selectcmd[] = "*2\r\n$6\r\nSELECT\r\n";
         redisDb *db = server.db+j;
@@ -413,17 +695,17 @@ int rewriteAppendOnlyFile(char *filename) {
         }
 
         /* SELECT the new DB */
-        if (fwrite(selectcmd,sizeof(selectcmd)-1,1,fp) == 0) goto werr;
-        if (fwriteBulkLongLong(fp,j) == 0) goto werr;
+        if (rioWrite(&aof,selectcmd,sizeof(selectcmd)-1) == 0) goto werr;
+        if (rioWriteBulkLongLong(&aof,j) == 0) goto werr;
 
         /* Iterate this DB writing every entry */
         while((de = dictNext(di)) != NULL) {
             sds keystr;
             robj key, *o;
-            time_t expiretime;
+            long long expiretime;
 
-            keystr = dictGetEntryKey(de);
-            o = dictGetEntryVal(de);
+            keystr = dictGetKey(de);
+            o = dictGetVal(de);
             initStaticStringObject(key,keystr);
 
             expiretime = getExpire(db,&key);
@@ -432,167 +714,29 @@ int rewriteAppendOnlyFile(char *filename) {
             if (o->type == REDIS_STRING) {
                 /* Emit a SET command */
                 char cmd[]="*3\r\n$3\r\nSET\r\n";
-                if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
+                if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
                 /* Key and value */
-                if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                if (fwriteBulkObject(fp,o) == 0) goto werr;
+                if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                if (rioWriteBulkObject(&aof,o) == 0) goto werr;
             } else if (o->type == REDIS_LIST) {
-                /* Emit the RPUSHes needed to rebuild the list */
-                char cmd[]="*3\r\n$5\r\nRPUSH\r\n";
-                if (o->encoding == REDIS_ENCODING_ZIPLIST) {
-                    unsigned char *zl = o->ptr;
-                    unsigned char *p = ziplistIndex(zl,0);
-                    unsigned char *vstr;
-                    unsigned int vlen;
-                    long long vlong;
-
-                    while(ziplistGet(p,&vstr,&vlen,&vlong)) {
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (vstr) {
-                            if (fwriteBulkString(fp,(char*)vstr,vlen) == 0)
-                                goto werr;
-                        } else {
-                            if (fwriteBulkLongLong(fp,vlong) == 0)
-                                goto werr;
-                        }
-                        p = ziplistNext(zl,p);
-                    }
-                } else if (o->encoding == REDIS_ENCODING_LINKEDLIST) {
-                    list *list = o->ptr;
-                    listNode *ln;
-                    listIter li;
-
-                    listRewind(list,&li);
-                    while((ln = listNext(&li))) {
-                        robj *eleobj = listNodeValue(ln);
-
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkObject(fp,eleobj) == 0) goto werr;
-                    }
-                } else {
-                    redisPanic("Unknown list encoding");
-                }
+                if (rewriteListObject(&aof,&key,o) == 0) goto werr;
             } else if (o->type == REDIS_SET) {
-                char cmd[]="*3\r\n$4\r\nSADD\r\n";
-
-                /* Emit the SADDs needed to rebuild the set */
-                if (o->encoding == REDIS_ENCODING_INTSET) {
-                    int ii = 0;
-                    int64_t llval;
-                    while(intsetGet(o->ptr,ii++,&llval)) {
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkLongLong(fp,llval) == 0) goto werr;
-                    }
-                } else if (o->encoding == REDIS_ENCODING_HT) {
-                    dictIterator *di = dictGetIterator(o->ptr);
-                    dictEntry *de;
-                    while((de = dictNext(di)) != NULL) {
-                        robj *eleobj = dictGetEntryKey(de);
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkObject(fp,eleobj) == 0) goto werr;
-                    }
-                    dictReleaseIterator(di);
-                } else {
-                    redisPanic("Unknown set encoding");
-                }
+                if (rewriteSetObject(&aof,&key,o) == 0) goto werr;
             } else if (o->type == REDIS_ZSET) {
-                /* Emit the ZADDs needed to rebuild the sorted set */
-                char cmd[]="*4\r\n$4\r\nZADD\r\n";
-
-                if (o->encoding == REDIS_ENCODING_ZIPLIST) {
-                    unsigned char *zl = o->ptr;
-                    unsigned char *eptr, *sptr;
-                    unsigned char *vstr;
-                    unsigned int vlen;
-                    long long vll;
-                    double score;
-
-                    eptr = ziplistIndex(zl,0);
-                    redisAssert(eptr != NULL);
-                    sptr = ziplistNext(zl,eptr);
-                    redisAssert(sptr != NULL);
-
-                    while (eptr != NULL) {
-                        redisAssert(ziplistGet(eptr,&vstr,&vlen,&vll));
-                        score = zzlGetScore(sptr);
-
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkDouble(fp,score) == 0) goto werr;
-                        if (vstr != NULL) {
-                            if (fwriteBulkString(fp,(char*)vstr,vlen) == 0)
-                                goto werr;
-                        } else {
-                            if (fwriteBulkLongLong(fp,vll) == 0)
-                                goto werr;
-                        }
-                        zzlNext(zl,&eptr,&sptr);
-                    }
-                } else if (o->encoding == REDIS_ENCODING_SKIPLIST) {
-                    zset *zs = o->ptr;
-                    dictIterator *di = dictGetIterator(zs->dict);
-                    dictEntry *de;
-
-                    while((de = dictNext(di)) != NULL) {
-                        robj *eleobj = dictGetEntryKey(de);
-                        double *score = dictGetEntryVal(de);
-
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkDouble(fp,*score) == 0) goto werr;
-                        if (fwriteBulkObject(fp,eleobj) == 0) goto werr;
-                    }
-                    dictReleaseIterator(di);
-                } else {
-                    redisPanic("Unknown sorted set encoding");
-                }
+                if (rewriteSortedSetObject(&aof,&key,o) == 0) goto werr;
             } else if (o->type == REDIS_HASH) {
-                char cmd[]="*4\r\n$4\r\nHSET\r\n";
-
-                /* Emit the HSETs needed to rebuild the hash */
-                if (o->encoding == REDIS_ENCODING_ZIPMAP) {
-                    unsigned char *p = zipmapRewind(o->ptr);
-                    unsigned char *field, *val;
-                    unsigned int flen, vlen;
-
-                    while((p = zipmapNext(p,&field,&flen,&val,&vlen)) != NULL) {
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkString(fp,(char*)field,flen) == 0)
-                            goto werr;
-                        if (fwriteBulkString(fp,(char*)val,vlen) == 0)
-                            goto werr;
-                    }
-                } else {
-                    dictIterator *di = dictGetIterator(o->ptr);
-                    dictEntry *de;
-
-                    while((de = dictNext(di)) != NULL) {
-                        robj *field = dictGetEntryKey(de);
-                        robj *val = dictGetEntryVal(de);
-
-                        if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                        if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                        if (fwriteBulkObject(fp,field) == 0) goto werr;
-                        if (fwriteBulkObject(fp,val) == 0) goto werr;
-                    }
-                    dictReleaseIterator(di);
-                }
+                if (rewriteHashObject(&aof,&key,o) == 0) goto werr;
             } else {
                 redisPanic("Unknown object type");
             }
             /* Save the expire time */
             if (expiretime != -1) {
-                char cmd[]="*3\r\n$8\r\nEXPIREAT\r\n";
+                char cmd[]="*3\r\n$9\r\nPEXPIREAT\r\n";
                 /* If this key is already expired skip it */
                 if (expiretime < now) continue;
-                if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
-                if (fwriteBulkObject(fp,&key) == 0) goto werr;
-                if (fwriteBulkLongLong(fp,expiretime) == 0) goto werr;
+                if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                if (rioWriteBulkLongLong(&aof,expiretime) == 0) goto werr;
             }
         }
         dictReleaseIterator(di);
@@ -662,6 +806,7 @@ int rewriteAppendOnlyFileBackground(void) {
         }
         redisLog(REDIS_NOTICE,
             "Background append only file rewriting started by pid %d",childpid);
+        server.aofrewrite_scheduled = 0;
         server.bgrewritechildpid = childpid;
         updateDictResizePolicy();
         /* We set appendseldb to -1 in order to force the next call to the
@@ -812,9 +957,15 @@ void backgroundRewriteDoneHandler(int exitcode, int bysignal) {
             server.appendseldb = -1; /* Make sure SELECT is re-issued */
             aofUpdateCurrentSize();
             server.auto_aofrewrite_base_size = server.appendonly_current_size;
+
+            /* Clear regular AOF buffer since its contents was just written to
+             * the new AOF from the background rewrite buffer. */
+            sdsfree(server.aofbuf);
+            server.aofbuf = sdsempty();
         }
 
         redisLog(REDIS_NOTICE, "Background AOF rewrite successful");
+        server.aof_wait_rewrite = 0;
 
         /* Asynchronously close the overwritten AOF. */
         if (oldfd != -1) bioCreateBackgroundJob(REDIS_BIO_CLOSE_FILE,(void*)(long)oldfd,NULL,NULL);
@@ -834,4 +985,10 @@ cleanup:
     server.bgrewritebuf = sdsempty();
     aofRemoveTempFile(server.bgrewritechildpid);
     server.bgrewritechildpid = -1;
+    /* If we were waiting for an AOF rewrite before to start appending
+     * to the AOF again (this happens both when the user switches on
+     * AOF with CONFIG SET, and after a slave with AOF enabled syncs with
+     * the master), but the rewrite failed (otherwise aof_wait_rewrite
+     * would be zero), we need to schedule a new one. */
+    if (server.aof_wait_rewrite) server.aofrewrite_scheduled = 1;
 }