]> git.saurik.com Git - redis.git/blobdiff - src/aof.c
Fixed issues with expire introduced with latest millisecond resolution feature. Many...
[redis.git] / src / aof.c
index 5417e76b14339eeb4aa84d73deb32801db4b9e38..4a463bde0664b34e39ceb5a9da9fe96bc485bdc5 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>
@@ -51,7 +52,7 @@ int startAppendOnly(void) {
     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,"User tried turning on AOF with CONFIG SET but I can't trigger a background AOF rewrite operation. Check the above logs for more info about the error.");
         return REDIS_ERR;
     }
     return REDIS_OK;
@@ -95,12 +96,13 @@ 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 +181,38 @@ 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;
@@ -214,18 +233,22 @@ 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);
     }
 
@@ -325,6 +348,8 @@ 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 +408,30 @@ 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");
+    }
+}
+
 /* Write a sequence of commands able to fully rebuild the dataset into
  * "filename". Used both by REWRITEAOF and BGREWRITEAOF. */
 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 +441,8 @@ 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 +455,17 @@ 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);
@@ -432,10 +474,10 @@ 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";
@@ -447,13 +489,13 @@ int rewriteAppendOnlyFile(char *filename) {
                     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 (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
                         if (vstr) {
-                            if (fwriteBulkString(fp,(char*)vstr,vlen) == 0)
+                            if (rioWriteBulkString(&aof,(char*)vstr,vlen) == 0)
                                 goto werr;
                         } else {
-                            if (fwriteBulkLongLong(fp,vlong) == 0)
+                            if (rioWriteBulkLongLong(&aof,vlong) == 0)
                                 goto werr;
                         }
                         p = ziplistNext(zl,p);
@@ -467,9 +509,9 @@ int rewriteAppendOnlyFile(char *filename) {
                     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;
+                        if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,eleobj) == 0) goto werr;
                     }
                 } else {
                     redisPanic("Unknown list encoding");
@@ -482,18 +524,18 @@ int rewriteAppendOnlyFile(char *filename) {
                     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;
+                        if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkLongLong(&aof,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;
+                        robj *eleobj = dictGetKey(de);
+                        if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,eleobj) == 0) goto werr;
                     }
                     dictReleaseIterator(di);
                 } else {
@@ -520,14 +562,14 @@ int rewriteAppendOnlyFile(char *filename) {
                         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 (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkDouble(&aof,score) == 0) goto werr;
                         if (vstr != NULL) {
-                            if (fwriteBulkString(fp,(char*)vstr,vlen) == 0)
+                            if (rioWriteBulkString(&aof,(char*)vstr,vlen) == 0)
                                 goto werr;
                         } else {
-                            if (fwriteBulkLongLong(fp,vll) == 0)
+                            if (rioWriteBulkLongLong(&aof,vll) == 0)
                                 goto werr;
                         }
                         zzlNext(zl,&eptr,&sptr);
@@ -538,13 +580,13 @@ int rewriteAppendOnlyFile(char *filename) {
                     dictEntry *de;
 
                     while((de = dictNext(di)) != NULL) {
-                        robj *eleobj = dictGetEntryKey(de);
-                        double *score = dictGetEntryVal(de);
+                        robj *eleobj = dictGetKey(de);
+                        double *score = dictGetVal(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;
+                        if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkDouble(&aof,*score) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,eleobj) == 0) goto werr;
                     }
                     dictReleaseIterator(di);
                 } else {
@@ -560,11 +602,11 @@ int rewriteAppendOnlyFile(char *filename) {
                     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)
+                        if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkString(&aof,(char*)field,flen) == 0)
                             goto werr;
-                        if (fwriteBulkString(fp,(char*)val,vlen) == 0)
+                        if (rioWriteBulkString(&aof,(char*)val,vlen) == 0)
                             goto werr;
                     }
                 } else {
@@ -572,13 +614,13 @@ int rewriteAppendOnlyFile(char *filename) {
                     dictEntry *de;
 
                     while((de = dictNext(di)) != NULL) {
-                        robj *field = dictGetEntryKey(de);
-                        robj *val = dictGetEntryVal(de);
+                        robj *field = dictGetKey(de);
+                        robj *val = dictGetVal(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;
+                        if (rioWrite(&aof,cmd,sizeof(cmd)-1) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,&key) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,field) == 0) goto werr;
+                        if (rioWriteBulkObject(&aof,val) == 0) goto werr;
                     }
                     dictReleaseIterator(di);
                 }
@@ -587,12 +629,12 @@ int rewriteAppendOnlyFile(char *filename) {
             }
             /* 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 +704,7 @@ 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