src/aof.c

   1 #include "redis.h"
   2
   3 #include <signal.h>
   4 #include <fcntl.h>
   5 #include <sys/stat.h>
   6 #include <sys/types.h>
   7 #include <sys/time.h>
   8 #include <sys/resource.h>
   9 #include <sys/wait.h>
  10
  11 /* Called when the user switches from "appendonly yes" to "appendonly no"
  12  * at runtime using the CONFIG command. */
  13 void stopAppendOnly(void) {
  14     flushAppendOnlyFile();
  15     aof_fsync(server.appendfd);
  16     close(server.appendfd);
  17
  18     server.appendfd = -1;
  19     server.appendseldb = -1;
  20     server.appendonly = 0;
  21     /* rewrite operation in progress? kill it, wait child exit */
  22     if (server.bgsavechildpid != -1) {
  23         int statloc;
  24
  25         if (kill(server.bgsavechildpid,SIGKILL) != -1)
  26             wait3(&statloc,0,NULL);
  27         /* reset the buffer accumulating changes while the child saves */
  28         sdsfree(server.bgrewritebuf);
  29         server.bgrewritebuf = sdsempty();
  30         server.bgsavechildpid = -1;
  31     }
  32 }
  33
  34 /* Called when the user switches from "appendonly no" to "appendonly yes"
  35  * at runtime using the CONFIG command. */
  36 int startAppendOnly(void) {
  37     server.appendonly = 1;
  38     server.lastfsync = time(NULL);
  39     server.appendfd = open(server.appendfilename,O_WRONLY|O_APPEND|O_CREAT,0644);
  40     if (server.appendfd == -1) {
  41         redisLog(REDIS_WARNING,"Used tried to switch on AOF via CONFIG, but I can't open the AOF file: %s",strerror(errno));
  42         return REDIS_ERR;
  43     }
  44     if (rewriteAppendOnlyFileBackground() == REDIS_ERR) {
  45         server.appendonly = 0;
  46         close(server.appendfd);
  47         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));
  48         return REDIS_ERR;
  49     }
  50     return REDIS_OK;
  51 }
  52
  53 /* Write the append only file buffer on disk.
  54  *
  55  * Since we are required to write the AOF before replying to the client,
  56  * and the only way the client socket can get a write is entering when the
  57  * the event loop, we accumulate all the AOF writes in a memory
  58  * buffer and write it on disk using this function just before entering
  59  * the event loop again. */
  60 void flushAppendOnlyFile(void) {
  61     time_t now;
  62     ssize_t nwritten;
  63
  64     if (sdslen(server.aofbuf) == 0) return;
  65
  66     /* We want to perform a single write. This should be guaranteed atomic
  67      * at least if the filesystem we are writing is a real physical one.
  68      * While this will save us against the server being killed I don't think
  69      * there is much to do about the whole server stopping for power problems
  70      * or alike */
  71      nwritten = write(server.appendfd,server.aofbuf,sdslen(server.aofbuf));
  72      if (nwritten != (signed)sdslen(server.aofbuf)) {
  73         /* Ooops, we are in troubles. The best thing to do for now is
  74          * aborting instead of giving the illusion that everything is
  75          * working as expected. */
  76          if (nwritten == -1) {
  77             redisLog(REDIS_WARNING,"Exiting on error writing to the append-only file: %s",strerror(errno));
  78          } else {
  79             redisLog(REDIS_WARNING,"Exiting on short write while writing to the append-only file: %s",strerror(errno));
  80          }
  81          exit(1);
  82     }
  83     sdsfree(server.aofbuf);
  84     server.aofbuf = sdsempty();
  85
  86     /* Don't Fsync if no-appendfsync-on-rewrite is set to yes and we have
  87      * childs performing heavy I/O on disk. */
  88     if (server.no_appendfsync_on_rewrite &&
  89         (server.bgrewritechildpid != -1 || server.bgsavechildpid != -1))
  90             return;
  91     /* Fsync if needed */
  92     now = time(NULL);
  93     if (server.appendfsync == APPENDFSYNC_ALWAYS ||
  94         (server.appendfsync == APPENDFSYNC_EVERYSEC &&
  95          now-server.lastfsync > 1))
  96     {
  97         /* aof_fsync is defined as fdatasync() for Linux in order to avoid
  98          * flushing metadata. */
  99         aof_fsync(server.appendfd); /* Let's try to get this data on the disk */
 100         server.lastfsync = now;
 101     }
 102 }
 103
 104 sds catAppendOnlyGenericCommand(sds buf, int argc, robj **argv) {
 105     int j;
 106     buf = sdscatprintf(buf,"*%d\r\n",argc);
 107     for (j = 0; j < argc; j++) {
 108         robj *o = getDecodedObject(argv[j]);
 109         buf = sdscatprintf(buf,"$%lu\r\n",(unsigned long)sdslen(o->ptr));
 110         buf = sdscatlen(buf,o->ptr,sdslen(o->ptr));
 111         buf = sdscatlen(buf,"\r\n",2);
 112         decrRefCount(o);
 113     }
 114     return buf;
 115 }
 116
 117 sds catAppendOnlyExpireAtCommand(sds buf, robj *key, robj *seconds) {
 118     int argc = 3;
 119     long when;
 120     robj *argv[3];
 121
 122     /* Make sure we can use strtol */
 123     seconds = getDecodedObject(seconds);
 124     when = time(NULL)+strtol(seconds->ptr,NULL,10);
 125     decrRefCount(seconds);
 126
 127     argv[0] = createStringObject("EXPIREAT",8);
 128     argv[1] = key;
 129     argv[2] = createObject(REDIS_STRING,
 130         sdscatprintf(sdsempty(),"%ld",when));
 131     buf = catAppendOnlyGenericCommand(buf, argc, argv);
 132     decrRefCount(argv[0]);
 133     decrRefCount(argv[2]);
 134     return buf;
 135 }
 136
 137 void feedAppendOnlyFile(struct redisCommand *cmd, int dictid, robj **argv, int argc) {
 138     sds buf = sdsempty();
 139     robj *tmpargv[3];
 140
 141     /* The DB this command was targetting is not the same as the last command
 142      * we appendend. To issue a SELECT command is needed. */
 143     if (dictid != server.appendseldb) {
 144         char seldb[64];
 145
 146         snprintf(seldb,sizeof(seldb),"%d",dictid);
 147         buf = sdscatprintf(buf,"*2\r\n$6\r\nSELECT\r\n$%lu\r\n%s\r\n",
 148             (unsigned long)strlen(seldb),seldb);
 149         server.appendseldb = dictid;
 150     }
 151
 152     if (cmd->proc == expireCommand) {
 153         /* Translate EXPIRE into EXPIREAT */
 154         buf = catAppendOnlyExpireAtCommand(buf,argv[1],argv[2]);
 155     } else if (cmd->proc == setexCommand) {
 156         /* Translate SETEX to SET and EXPIREAT */
 157         tmpargv[0] = createStringObject("SET",3);
 158         tmpargv[1] = argv[1];
 159         tmpargv[2] = argv[3];
 160         buf = catAppendOnlyGenericCommand(buf,3,tmpargv);
 161         decrRefCount(tmpargv[0]);
 162         buf = catAppendOnlyExpireAtCommand(buf,argv[1],argv[2]);
 163     } else {
 164         buf = catAppendOnlyGenericCommand(buf,argc,argv);
 165     }
 166
 167     /* Append to the AOF buffer. This will be flushed on disk just before
 168      * of re-entering the event loop, so before the client will get a
 169      * positive reply about the operation performed. */
 170     server.aofbuf = sdscatlen(server.aofbuf,buf,sdslen(buf));
 171
 172     /* If a background append only file rewriting is in progress we want to
 173      * accumulate the differences between the child DB and the current one
 174      * in a buffer, so that when the child process will do its work we
 175      * can append the differences to the new append only file. */
 176     if (server.bgrewritechildpid != -1)
 177         server.bgrewritebuf = sdscatlen(server.bgrewritebuf,buf,sdslen(buf));
 178
 179     sdsfree(buf);
 180 }
 181
 182 /* In Redis commands are always executed in the context of a client, so in
 183  * order to load the append only file we need to create a fake client. */
 184 struct redisClient *createFakeClient(void) {
 185     struct redisClient *c = zmalloc(sizeof(*c));
 186
 187     selectDb(c,0);
 188     c->fd = -1;
 189     c->querybuf = sdsempty();
 190     c->argc = 0;
 191     c->argv = NULL;
 192     c->flags = 0;
 193     /* We set the fake client as a slave waiting for the synchronization
 194      * so that Redis will not try to send replies to this client. */
 195     c->replstate = REDIS_REPL_WAIT_BGSAVE_START;
 196     c->reply = listCreate();
 197     c->watched_keys = listCreate();
 198     listSetFreeMethod(c->reply,decrRefCount);
 199     listSetDupMethod(c->reply,dupClientReplyValue);
 200     initClientMultiState(c);
 201     return c;
 202 }
 203
 204 void freeFakeClient(struct redisClient *c) {
 205     sdsfree(c->querybuf);
 206     listRelease(c->reply);
 207     listRelease(c->watched_keys);
 208     freeClientMultiState(c);
 209     zfree(c);
 210 }
 211
 212 /* Replay the append log file. On error REDIS_OK is returned. On non fatal
 213  * error (the append only file is zero-length) REDIS_ERR is returned. On
 214  * fatal error an error message is logged and the program exists. */
 215 int loadAppendOnlyFile(char *filename) {
 216     struct redisClient *fakeClient;
 217     FILE *fp = fopen(filename,"r");
 218     struct redis_stat sb;
 219     int appendonly = server.appendonly;
 220
 221     if (redis_fstat(fileno(fp),&sb) != -1 && sb.st_size == 0)
 222         return REDIS_ERR;
 223
 224     if (fp == NULL) {
 225         redisLog(REDIS_WARNING,"Fatal error: can't open the append log file for reading: %s",strerror(errno));
 226         exit(1);
 227     }
 228
 229     /* Temporarily disable AOF, to prevent EXEC from feeding a MULTI
 230      * to the same file we're about to read. */
 231     server.appendonly = 0;
 232
 233     fakeClient = createFakeClient();
 234     while(1) {
 235         int argc, j;
 236         unsigned long len;
 237         robj **argv;
 238         char buf[128];
 239         sds argsds;
 240         struct redisCommand *cmd;
 241         int force_swapout;
 242
 243         if (fgets(buf,sizeof(buf),fp) == NULL) {
 244             if (feof(fp))
 245                 break;
 246             else
 247                 goto readerr;
 248         }
 249         if (buf[0] != '*') goto fmterr;
 250         argc = atoi(buf+1);
 251         argv = zmalloc(sizeof(robj*)*argc);
 252         for (j = 0; j < argc; j++) {
 253             if (fgets(buf,sizeof(buf),fp) == NULL) goto readerr;
 254             if (buf[0] != '$') goto fmterr;
 255             len = strtol(buf+1,NULL,10);
 256             argsds = sdsnewlen(NULL,len);
 257             if (len && fread(argsds,len,1,fp) == 0) goto fmterr;
 258             argv[j] = createObject(REDIS_STRING,argsds);
 259             if (fread(buf,2,1,fp) == 0) goto fmterr; /* discard CRLF */
 260         }
 261
 262         /* Command lookup */
 263         cmd = lookupCommand(argv[0]->ptr);
 264         if (!cmd) {
 265             redisLog(REDIS_WARNING,"Unknown command '%s' reading the append only file", argv[0]->ptr);
 266             exit(1);
 267         }
 268         /* Try object encoding */
 269         if (cmd->flags & REDIS_CMD_BULK)
 270             argv[argc-1] = tryObjectEncoding(argv[argc-1]);
 271         /* Run the command in the context of a fake client */
 272         fakeClient->argc = argc;
 273         fakeClient->argv = argv;
 274         cmd->proc(fakeClient);
 275         /* Discard the reply objects list from the fake client */
 276         while(listLength(fakeClient->reply))
 277             listDelNode(fakeClient->reply,listFirst(fakeClient->reply));
 278         /* Clean up, ready for the next command */
 279         for (j = 0; j < argc; j++) decrRefCount(argv[j]);
 280         zfree(argv);
 281         /* Handle swapping while loading big datasets when VM is on */
 282         force_swapout = 0;
 283         if ((zmalloc_used_memory() - server.vm_max_memory) > 1024*1024*32)
 284             force_swapout = 1;
 285
 286         if (server.vm_enabled && force_swapout) {
 287             while (zmalloc_used_memory() > server.vm_max_memory) {
 288                 if (vmSwapOneObjectBlocking() == REDIS_ERR) break;
 289             }
 290         }
 291     }
 292
 293     /* This point can only be reached when EOF is reached without errors.
 294      * If the client is in the middle of a MULTI/EXEC, log error and quit. */
 295     if (fakeClient->flags & REDIS_MULTI) goto readerr;
 296
 297     fclose(fp);
 298     freeFakeClient(fakeClient);
 299     server.appendonly = appendonly;
 300     return REDIS_OK;
 301
 302 readerr:
 303     if (feof(fp)) {
 304         redisLog(REDIS_WARNING,"Unexpected end of file reading the append only file");
 305     } else {
 306         redisLog(REDIS_WARNING,"Unrecoverable error reading the append only file: %s", strerror(errno));
 307     }
 308     exit(1);
 309 fmterr:
 310     redisLog(REDIS_WARNING,"Bad file format reading the append only file");
 311     exit(1);
 312 }
 313
 314 /* Write binary-safe string into a file in the bulkformat
 315  * $<count>\r\n<payload>\r\n */
 316 int fwriteBulkString(FILE *fp, char *s, unsigned long len) {
 317     char cbuf[128];
 318     int clen;
 319     cbuf[0] = '$';
 320     clen = 1+ll2string(cbuf+1,sizeof(cbuf)-1,len);
 321     cbuf[clen++] = '\r';
 322     cbuf[clen++] = '\n';
 323     if (fwrite(cbuf,clen,1,fp) == 0) return 0;
 324     if (len > 0 && fwrite(s,len,1,fp) == 0) return 0;
 325     if (fwrite("\r\n",2,1,fp) == 0) return 0;
 326     return 1;
 327 }
 328
 329 /* Write a double value in bulk format $<count>\r\n<payload>\r\n */
 330 int fwriteBulkDouble(FILE *fp, double d) {
 331     char buf[128], dbuf[128];
 332
 333     snprintf(dbuf,sizeof(dbuf),"%.17g\r\n",d);
 334     snprintf(buf,sizeof(buf),"$%lu\r\n",(unsigned long)strlen(dbuf)-2);
 335     if (fwrite(buf,strlen(buf),1,fp) == 0) return 0;
 336     if (fwrite(dbuf,strlen(dbuf),1,fp) == 0) return 0;
 337     return 1;
 338 }
 339
 340 /* Write a long value in bulk format $<count>\r\n<payload>\r\n */
 341 int fwriteBulkLongLong(FILE *fp, long long l) {
 342     char bbuf[128], lbuf[128];
 343     unsigned int blen, llen;
 344     llen = ll2string(lbuf,32,l);
 345     blen = snprintf(bbuf,sizeof(bbuf),"$%u\r\n%s\r\n",llen,lbuf);
 346     if (fwrite(bbuf,blen,1,fp) == 0) return 0;
 347     return 1;
 348 }
 349
 350 /* Delegate writing an object to writing a bulk string or bulk long long. */
 351 int fwriteBulkObject(FILE *fp, robj *obj) {
 352     /* Avoid using getDecodedObject to help copy-on-write (we are often
 353      * in a child process when this function is called). */
 354     if (obj->encoding == REDIS_ENCODING_INT) {
 355         return fwriteBulkLongLong(fp,(long)obj->ptr);
 356     } else if (obj->encoding == REDIS_ENCODING_RAW) {
 357         return fwriteBulkString(fp,obj->ptr,sdslen(obj->ptr));
 358     } else {
 359         redisPanic("Unknown string encoding");
 360     }
 361 }
 362
 363 /* Write a sequence of commands able to fully rebuild the dataset into
 364  * "filename". Used both by REWRITEAOF and BGREWRITEAOF. */
 365 int rewriteAppendOnlyFile(char *filename) {
 366     dictIterator *di = NULL;
 367     dictEntry *de;
 368     FILE *fp;
 369     char tmpfile[256];
 370     int j;
 371     time_t now = time(NULL);
 372
 373     /* Note that we have to use a different temp name here compared to the
 374      * one used by rewriteAppendOnlyFileBackground() function. */
 375     snprintf(tmpfile,256,"temp-rewriteaof-%d.aof", (int) getpid());
 376     fp = fopen(tmpfile,"w");
 377     if (!fp) {
 378         redisLog(REDIS_WARNING, "Failed rewriting the append only file: %s", strerror(errno));
 379         return REDIS_ERR;
 380     }
 381     for (j = 0; j < server.dbnum; j++) {
 382         char selectcmd[] = "*2\r\n$6\r\nSELECT\r\n";
 383         redisDb *db = server.db+j;
 384         dict *d = db->dict;
 385         if (dictSize(d) == 0) continue;
 386         di = dictGetIterator(d);
 387         if (!di) {
 388             fclose(fp);
 389             return REDIS_ERR;
 390         }
 391
 392         /* SELECT the new DB */
 393         if (fwrite(selectcmd,sizeof(selectcmd)-1,1,fp) == 0) goto werr;
 394         if (fwriteBulkLongLong(fp,j) == 0) goto werr;
 395
 396         /* Iterate this DB writing every entry */
 397         while((de = dictNext(di)) != NULL) {
 398             sds keystr = dictGetEntryKey(de);
 399             robj key, *o;
 400             time_t expiretime;
 401             int swapped;
 402
 403             keystr = dictGetEntryKey(de);
 404             o = dictGetEntryVal(de);
 405             initStaticStringObject(key,keystr);
 406             /* If the value for this key is swapped, load a preview in memory.
 407              * We use a "swapped" flag to remember if we need to free the
 408              * value object instead to just increment the ref count anyway
 409              * in order to avoid copy-on-write of pages if we are forked() */
 410             if (!server.vm_enabled || o->storage == REDIS_VM_MEMORY ||
 411                 o->storage == REDIS_VM_SWAPPING) {
 412                 swapped = 0;
 413             } else {
 414                 o = vmPreviewObject(o);
 415                 swapped = 1;
 416             }
 417             expiretime = getExpire(db,&key);
 418
 419             /* Save the key and associated value */
 420             if (o->type == REDIS_STRING) {
 421                 /* Emit a SET command */
 422                 char cmd[]="*3\r\n$3\r\nSET\r\n";
 423                 if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 424                 /* Key and value */
 425                 if (fwriteBulkObject(fp,&key) == 0) goto werr;
 426                 if (fwriteBulkObject(fp,o) == 0) goto werr;
 427             } else if (o->type == REDIS_LIST) {
 428                 /* Emit the RPUSHes needed to rebuild the list */
 429                 char cmd[]="*3\r\n$5\r\nRPUSH\r\n";
 430                 if (o->encoding == REDIS_ENCODING_ZIPLIST) {
 431                     unsigned char *zl = o->ptr;
 432                     unsigned char *p = ziplistIndex(zl,0);
 433                     unsigned char *vstr;
 434                     unsigned int vlen;
 435                     long long vlong;
 436
 437                     while(ziplistGet(p,&vstr,&vlen,&vlong)) {
 438                         if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 439                         if (fwriteBulkObject(fp,&key) == 0) goto werr;
 440                         if (vstr) {
 441                             if (fwriteBulkString(fp,(char*)vstr,vlen) == 0)
 442                                 goto werr;
 443                         } else {
 444                             if (fwriteBulkLongLong(fp,vlong) == 0)
 445                                 goto werr;
 446                         }
 447                         p = ziplistNext(zl,p);
 448                     }
 449                 } else if (o->encoding == REDIS_ENCODING_LINKEDLIST) {
 450                     list *list = o->ptr;
 451                     listNode *ln;
 452                     listIter li;
 453
 454                     listRewind(list,&li);
 455                     while((ln = listNext(&li))) {
 456                         robj *eleobj = listNodeValue(ln);
 457
 458                         if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 459                         if (fwriteBulkObject(fp,&key) == 0) goto werr;
 460                         if (fwriteBulkObject(fp,eleobj) == 0) goto werr;
 461                     }
 462                 } else {
 463                     redisPanic("Unknown list encoding");
 464                 }
 465             } else if (o->type == REDIS_SET) {
 466                 char cmd[]="*3\r\n$4\r\nSADD\r\n";
 467
 468                 /* Emit the SADDs needed to rebuild the set */
 469                 if (o->encoding == REDIS_ENCODING_INTSET) {
 470                     int ii = 0;
 471                     long long llval;
 472                     while(intsetGet(o->ptr,ii++,&llval)) {
 473                         if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 474                         if (fwriteBulkObject(fp,&key) == 0) goto werr;
 475                         if (fwriteBulkLongLong(fp,llval) == 0) goto werr;
 476                     }
 477                 } else if (o->encoding == REDIS_ENCODING_HT) {
 478                     dictIterator *di = dictGetIterator(o->ptr);
 479                     dictEntry *de;
 480                     while((de = dictNext(di)) != NULL) {
 481                         robj *eleobj = dictGetEntryKey(de);
 482                         if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 483                         if (fwriteBulkObject(fp,&key) == 0) goto werr;
 484                         if (fwriteBulkObject(fp,eleobj) == 0) goto werr;
 485                     }
 486                     dictReleaseIterator(di);
 487                 } else {
 488                     redisPanic("Unknown set encoding");
 489                 }
 490             } else if (o->type == REDIS_ZSET) {
 491                 /* Emit the ZADDs needed to rebuild the sorted set */
 492                 zset *zs = o->ptr;
 493                 dictIterator *di = dictGetIterator(zs->dict);
 494                 dictEntry *de;
 495
 496                 while((de = dictNext(di)) != NULL) {
 497                     char cmd[]="*4\r\n$4\r\nZADD\r\n";
 498                     robj *eleobj = dictGetEntryKey(de);
 499                     double *score = dictGetEntryVal(de);
 500
 501                     if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 502                     if (fwriteBulkObject(fp,&key) == 0) goto werr;
 503                     if (fwriteBulkDouble(fp,*score) == 0) goto werr;
 504                     if (fwriteBulkObject(fp,eleobj) == 0) goto werr;
 505                 }
 506                 dictReleaseIterator(di);
 507             } else if (o->type == REDIS_HASH) {
 508                 char cmd[]="*4\r\n$4\r\nHSET\r\n";
 509
 510                 /* Emit the HSETs needed to rebuild the hash */
 511                 if (o->encoding == REDIS_ENCODING_ZIPMAP) {
 512                     unsigned char *p = zipmapRewind(o->ptr);
 513                     unsigned char *field, *val;
 514                     unsigned int flen, vlen;
 515
 516                     while((p = zipmapNext(p,&field,&flen,&val,&vlen)) != NULL) {
 517                         if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 518                         if (fwriteBulkObject(fp,&key) == 0) goto werr;
 519                         if (fwriteBulkString(fp,(char*)field,flen) == 0)
 520                             goto werr;
 521                         if (fwriteBulkString(fp,(char*)val,vlen) == 0)
 522                             goto werr;
 523                     }
 524                 } else {
 525                     dictIterator *di = dictGetIterator(o->ptr);
 526                     dictEntry *de;
 527
 528                     while((de = dictNext(di)) != NULL) {
 529                         robj *field = dictGetEntryKey(de);
 530                         robj *val = dictGetEntryVal(de);
 531
 532                         if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 533                         if (fwriteBulkObject(fp,&key) == 0) goto werr;
 534                         if (fwriteBulkObject(fp,field) == 0) goto werr;
 535                         if (fwriteBulkObject(fp,val) == 0) goto werr;
 536                     }
 537                     dictReleaseIterator(di);
 538                 }
 539             } else {
 540                 redisPanic("Unknown object type");
 541             }
 542             /* Save the expire time */
 543             if (expiretime != -1) {
 544                 char cmd[]="*3\r\n$8\r\nEXPIREAT\r\n";
 545                 /* If this key is already expired skip it */
 546                 if (expiretime < now) continue;
 547                 if (fwrite(cmd,sizeof(cmd)-1,1,fp) == 0) goto werr;
 548                 if (fwriteBulkObject(fp,&key) == 0) goto werr;
 549                 if (fwriteBulkLongLong(fp,expiretime) == 0) goto werr;
 550             }
 551             if (swapped) decrRefCount(o);
 552         }
 553         dictReleaseIterator(di);
 554     }
 555
 556     /* Make sure data will not remain on the OS's output buffers */
 557     fflush(fp);
 558     aof_fsync(fileno(fp));
 559     fclose(fp);
 560
 561     /* Use RENAME to make sure the DB file is changed atomically only
 562      * if the generate DB file is ok. */
 563     if (rename(tmpfile,filename) == -1) {
 564         redisLog(REDIS_WARNING,"Error moving temp append only file on the final destination: %s", strerror(errno));
 565         unlink(tmpfile);
 566         return REDIS_ERR;
 567     }
 568     redisLog(REDIS_NOTICE,"SYNC append only file rewrite performed");
 569     return REDIS_OK;
 570
 571 werr:
 572     fclose(fp);
 573     unlink(tmpfile);
 574     redisLog(REDIS_WARNING,"Write error writing append only file on disk: %s", strerror(errno));
 575     if (di) dictReleaseIterator(di);
 576     return REDIS_ERR;
 577 }
 578
 579 /* This is how rewriting of the append only file in background works:
 580  *
 581  * 1) The user calls BGREWRITEAOF
 582  * 2) Redis calls this function, that forks():
 583  *    2a) the child rewrite the append only file in a temp file.
 584  *    2b) the parent accumulates differences in server.bgrewritebuf.
 585  * 3) When the child finished '2a' exists.
 586  * 4) The parent will trap the exit code, if it's OK, will append the
 587  *    data accumulated into server.bgrewritebuf into the temp file, and
 588  *    finally will rename(2) the temp file in the actual file name.
 589  *    The the new file is reopened as the new append only file. Profit!
 590  */
 591 int rewriteAppendOnlyFileBackground(void) {
 592     pid_t childpid;
 593
 594     if (server.bgrewritechildpid != -1) return REDIS_ERR;
 595     if (server.vm_enabled) waitEmptyIOJobsQueue();
 596     if ((childpid = fork()) == 0) {
 597         /* Child */
 598         char tmpfile[256];
 599
 600         if (server.vm_enabled) vmReopenSwapFile();
 601         close(server.fd);
 602         snprintf(tmpfile,256,"temp-rewriteaof-bg-%d.aof", (int) getpid());
 603         if (rewriteAppendOnlyFile(tmpfile) == REDIS_OK) {
 604             _exit(0);
 605         } else {
 606             _exit(1);
 607         }
 608     } else {
 609         /* Parent */
 610         if (childpid == -1) {
 611             redisLog(REDIS_WARNING,
 612                 "Can't rewrite append only file in background: fork: %s",
 613                 strerror(errno));
 614             return REDIS_ERR;
 615         }
 616         redisLog(REDIS_NOTICE,
 617             "Background append only file rewriting started by pid %d",childpid);
 618         server.bgrewritechildpid = childpid;
 619         updateDictResizePolicy();
 620         /* We set appendseldb to -1 in order to force the next call to the
 621          * feedAppendOnlyFile() to issue a SELECT command, so the differences
 622          * accumulated by the parent into server.bgrewritebuf will start
 623          * with a SELECT statement and it will be safe to merge. */
 624         server.appendseldb = -1;
 625         return REDIS_OK;
 626     }
 627     return REDIS_OK; /* unreached */
 628 }
 629
 630 void bgrewriteaofCommand(redisClient *c) {
 631     if (server.bgrewritechildpid != -1) {
 632         addReplySds(c,sdsnew("-ERR background append only file rewriting already in progress\r\n"));
 633         return;
 634     }
 635     if (rewriteAppendOnlyFileBackground() == REDIS_OK) {
 636         char *status = "+Background append only file rewriting started\r\n";
 637         addReplySds(c,sdsnew(status));
 638     } else {
 639         addReply(c,shared.err);
 640     }
 641 }
 642
 643 void aofRemoveTempFile(pid_t childpid) {
 644     char tmpfile[256];
 645
 646     snprintf(tmpfile,256,"temp-rewriteaof-bg-%d.aof", (int) childpid);
 647     unlink(tmpfile);
 648 }
 649
 650 /* A background append only file rewriting (BGREWRITEAOF) terminated its work.
 651  * Handle this. */
 652 void backgroundRewriteDoneHandler(int statloc) {
 653     int exitcode = WEXITSTATUS(statloc);
 654     int bysignal = WIFSIGNALED(statloc);
 655
 656     if (!bysignal && exitcode == 0) {
 657         int fd;
 658         char tmpfile[256];
 659
 660         redisLog(REDIS_NOTICE,
 661             "Background append only file rewriting terminated with success");
 662         /* Now it's time to flush the differences accumulated by the parent */
 663         snprintf(tmpfile,256,"temp-rewriteaof-bg-%d.aof", (int) server.bgrewritechildpid);
 664         fd = open(tmpfile,O_WRONLY|O_APPEND);
 665         if (fd == -1) {
 666             redisLog(REDIS_WARNING, "Not able to open the temp append only file produced by the child: %s", strerror(errno));
 667             goto cleanup;
 668         }
 669         /* Flush our data... */
 670         if (write(fd,server.bgrewritebuf,sdslen(server.bgrewritebuf)) !=
 671                 (signed) sdslen(server.bgrewritebuf)) {
 672             redisLog(REDIS_WARNING, "Error or short write trying to flush the parent diff of the append log file in the child temp file: %s", strerror(errno));
 673             close(fd);
 674             goto cleanup;
 675         }
 676         redisLog(REDIS_NOTICE,"Parent diff flushed into the new append log file with success (%lu bytes)",sdslen(server.bgrewritebuf));
 677         /* Now our work is to rename the temp file into the stable file. And
 678          * switch the file descriptor used by the server for append only. */
 679         if (rename(tmpfile,server.appendfilename) == -1) {
 680             redisLog(REDIS_WARNING,"Can't rename the temp append only file into the stable one: %s", strerror(errno));
 681             close(fd);
 682             goto cleanup;
 683         }
 684         /* Mission completed... almost */
 685         redisLog(REDIS_NOTICE,"Append only file successfully rewritten.");
 686         if (server.appendfd != -1) {
 687             /* If append only is actually enabled... */
 688             close(server.appendfd);
 689             server.appendfd = fd;
 690             if (server.appendfsync != APPENDFSYNC_NO) aof_fsync(fd);
 691             server.appendseldb = -1; /* Make sure it will issue SELECT */
 692             redisLog(REDIS_NOTICE,"The new append only file was selected for future appends.");
 693         } else {
 694             /* If append only is disabled we just generate a dump in this
 695              * format. Why not? */
 696             close(fd);
 697         }
 698     } else if (!bysignal && exitcode != 0) {
 699         redisLog(REDIS_WARNING, "Background append only file rewriting error");
 700     } else {
 701         redisLog(REDIS_WARNING,
 702             "Background append only file rewriting terminated by signal %d",
 703             WTERMSIG(statloc));
 704     }
 705 cleanup:
 706     sdsfree(server.bgrewritebuf);
 707     server.bgrewritebuf = sdsempty();
 708     aofRemoveTempFile(server.bgrewritechildpid);
 709     server.bgrewritechildpid = -1;
 710 }