src/rdb.c

   1 #include "redis.h"
   2 #include "lzf.h"    /* LZF compression library */
   3
   4 #include <math.h>
   5 #include <sys/types.h>
   6 #include <sys/time.h>
   7 #include <sys/resource.h>
   8 #include <sys/wait.h>
   9 #include <arpa/inet.h>
  10 #include <sys/stat.h>
  11
  12 /* Convenience wrapper around fwrite, that returns the number of bytes written
  13  * to the file instead of the number of objects (see fwrite(3)) and -1 in the
  14  * case of an error. It also supports a NULL *fp to skip writing altogether
  15  * instead of writing to /dev/null. */
  16 static int rdbWriteRaw(FILE *fp, void *p, size_t len) {
  17     if (fp != NULL && fwrite(p,len,1,fp) == 0) return -1;
  18     return len;
  19 }
  20
  21 int rdbSaveType(FILE *fp, unsigned char type) {
  22     return rdbWriteRaw(fp,&type,1);
  23 }
  24
  25 int rdbSaveTime(FILE *fp, time_t t) {
  26     int32_t t32 = (int32_t) t;
  27     return rdbWriteRaw(fp,&t32,4);
  28 }
  29
  30 /* check rdbLoadLen() comments for more info */
  31 int rdbSaveLen(FILE *fp, uint32_t len) {
  32     unsigned char buf[2];
  33     int nwritten;
  34
  35     if (len < (1<<6)) {
  36         /* Save a 6 bit len */
  37         buf[0] = (len&0xFF)|(REDIS_RDB_6BITLEN<<6);
  38         if (rdbWriteRaw(fp,buf,1) == -1) return -1;
  39         nwritten = 1;
  40     } else if (len < (1<<14)) {
  41         /* Save a 14 bit len */
  42         buf[0] = ((len>>8)&0xFF)|(REDIS_RDB_14BITLEN<<6);
  43         buf[1] = len&0xFF;
  44         if (rdbWriteRaw(fp,buf,2) == -1) return -1;
  45         nwritten = 2;
  46     } else {
  47         /* Save a 32 bit len */
  48         buf[0] = (REDIS_RDB_32BITLEN<<6);
  49         if (rdbWriteRaw(fp,buf,1) == -1) return -1;
  50         len = htonl(len);
  51         if (rdbWriteRaw(fp,&len,4) == -1) return -1;
  52         nwritten = 1+4;
  53     }
  54     return nwritten;
  55 }
  56
  57 /* Encode 'value' as an integer if possible (if integer will fit the
  58  * supported range). If the function sucessful encoded the integer
  59  * then the (up to 5 bytes) encoded representation is written in the
  60  * string pointed by 'enc' and the length is returned. Otherwise
  61  * 0 is returned. */
  62 int rdbEncodeInteger(long long value, unsigned char *enc) {
  63     /* Finally check if it fits in our ranges */
  64     if (value >= -(1<<7) && value <= (1<<7)-1) {
  65         enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT8;
  66         enc[1] = value&0xFF;
  67         return 2;
  68     } else if (value >= -(1<<15) && value <= (1<<15)-1) {
  69         enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT16;
  70         enc[1] = value&0xFF;
  71         enc[2] = (value>>8)&0xFF;
  72         return 3;
  73     } else if (value >= -((long long)1<<31) && value <= ((long long)1<<31)-1) {
  74         enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT32;
  75         enc[1] = value&0xFF;
  76         enc[2] = (value>>8)&0xFF;
  77         enc[3] = (value>>16)&0xFF;
  78         enc[4] = (value>>24)&0xFF;
  79         return 5;
  80     } else {
  81         return 0;
  82     }
  83 }
  84
  85 /* String objects in the form "2391" "-100" without any space and with a
  86  * range of values that can fit in an 8, 16 or 32 bit signed value can be
  87  * encoded as integers to save space */
  88 int rdbTryIntegerEncoding(char *s, size_t len, unsigned char *enc) {
  89     long long value;
  90     char *endptr, buf[32];
  91
  92     /* Check if it's possible to encode this value as a number */
  93     value = strtoll(s, &endptr, 10);
  94     if (endptr[0] != '\0') return 0;
  95     ll2string(buf,32,value);
  96
  97     /* If the number converted back into a string is not identical
  98      * then it's not possible to encode the string as integer */
  99     if (strlen(buf) != len || memcmp(buf,s,len)) return 0;
 100
 101     return rdbEncodeInteger(value,enc);
 102 }
 103
 104 int rdbSaveLzfStringObject(FILE *fp, unsigned char *s, size_t len) {
 105     size_t comprlen, outlen;
 106     unsigned char byte;
 107     int n, nwritten = 0;
 108     void *out;
 109
 110     /* We require at least four bytes compression for this to be worth it */
 111     if (len <= 4) return 0;
 112     outlen = len-4;
 113     if ((out = zmalloc(outlen+1)) == NULL) return 0;
 114     comprlen = lzf_compress(s, len, out, outlen);
 115     if (comprlen == 0) {
 116         zfree(out);
 117         return 0;
 118     }
 119     /* Data compressed! Let's save it on disk */
 120     byte = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_LZF;
 121     if ((n = rdbWriteRaw(fp,&byte,1)) == -1) goto writeerr;
 122     nwritten += n;
 123
 124     if ((n = rdbSaveLen(fp,comprlen)) == -1) goto writeerr;
 125     nwritten += n;
 126
 127     if ((n = rdbSaveLen(fp,len)) == -1) goto writeerr;
 128     nwritten += n;
 129
 130     if ((n = rdbWriteRaw(fp,out,comprlen)) == -1) goto writeerr;
 131     nwritten += n;
 132
 133     zfree(out);
 134     return nwritten;
 135
 136 writeerr:
 137     zfree(out);
 138     return -1;
 139 }
 140
 141 /* Save a string objet as [len][data] on disk. If the object is a string
 142  * representation of an integer value we try to safe it in a special form */
 143 int rdbSaveRawString(FILE *fp, unsigned char *s, size_t len) {
 144     int enclen;
 145     int n, nwritten = 0;
 146
 147     /* Try integer encoding */
 148     if (len <= 11) {
 149         unsigned char buf[5];
 150         if ((enclen = rdbTryIntegerEncoding((char*)s,len,buf)) > 0) {
 151             if (rdbWriteRaw(fp,buf,enclen) == -1) return -1;
 152             return enclen;
 153         }
 154     }
 155
 156     /* Try LZF compression - under 20 bytes it's unable to compress even
 157      * aaaaaaaaaaaaaaaaaa so skip it */
 158     if (server.rdbcompression && len > 20) {
 159         n = rdbSaveLzfStringObject(fp,s,len);
 160         if (n == -1) return -1;
 161         if (n > 0) return n;
 162         /* Return value of 0 means data can't be compressed, save the old way */
 163     }
 164
 165     /* Store verbatim */
 166     if ((n = rdbSaveLen(fp,len)) == -1) return -1;
 167     nwritten += n;
 168     if (len > 0) {
 169         if (rdbWriteRaw(fp,s,len) == -1) return -1;
 170         nwritten += len;
 171     }
 172     return nwritten;
 173 }
 174
 175 /* Save a long long value as either an encoded string or a string. */
 176 int rdbSaveLongLongAsStringObject(FILE *fp, long long value) {
 177     unsigned char buf[32];
 178     int n, nwritten = 0;
 179     int enclen = rdbEncodeInteger(value,buf);
 180     if (enclen > 0) {
 181         return rdbWriteRaw(fp,buf,enclen);
 182     } else {
 183         /* Encode as string */
 184         enclen = ll2string((char*)buf,32,value);
 185         redisAssert(enclen < 32);
 186         if ((n = rdbSaveLen(fp,enclen)) == -1) return -1;
 187         nwritten += n;
 188         if ((n = rdbWriteRaw(fp,buf,enclen)) == -1) return -1;
 189         nwritten += n;
 190     }
 191     return nwritten;
 192 }
 193
 194 /* Like rdbSaveStringObjectRaw() but handle encoded objects */
 195 int rdbSaveStringObject(FILE *fp, robj *obj) {
 196     /* Avoid to decode the object, then encode it again, if the
 197      * object is alrady integer encoded. */
 198     if (obj->encoding == REDIS_ENCODING_INT) {
 199         return rdbSaveLongLongAsStringObject(fp,(long)obj->ptr);
 200     } else {
 201         redisAssert(obj->encoding == REDIS_ENCODING_RAW);
 202         return rdbSaveRawString(fp,obj->ptr,sdslen(obj->ptr));
 203     }
 204 }
 205
 206 /* Save a double value. Doubles are saved as strings prefixed by an unsigned
 207  * 8 bit integer specifing the length of the representation.
 208  * This 8 bit integer has special values in order to specify the following
 209  * conditions:
 210  * 253: not a number
 211  * 254: + inf
 212  * 255: - inf
 213  */
 214 int rdbSaveDoubleValue(FILE *fp, double val) {
 215     unsigned char buf[128];
 216     int len;
 217
 218     if (isnan(val)) {
 219         buf[0] = 253;
 220         len = 1;
 221     } else if (!isfinite(val)) {
 222         len = 1;
 223         buf[0] = (val < 0) ? 255 : 254;
 224     } else {
 225 #if (DBL_MANT_DIG >= 52) && (LLONG_MAX == 0x7fffffffffffffffLL)
 226         /* Check if the float is in a safe range to be casted into a
 227          * long long. We are assuming that long long is 64 bit here.
 228          * Also we are assuming that there are no implementations around where
 229          * double has precision < 52 bit.
 230          *
 231          * Under this assumptions we test if a double is inside an interval
 232          * where casting to long long is safe. Then using two castings we
 233          * make sure the decimal part is zero. If all this is true we use
 234          * integer printing function that is much faster. */
 235         double min = -4503599627370495; /* (2^52)-1 */
 236         double max = 4503599627370496; /* -(2^52) */
 237         if (val > min && val < max && val == ((double)((long long)val)))
 238             ll2string((char*)buf+1,sizeof(buf),(long long)val);
 239         else
 240 #endif
 241             snprintf((char*)buf+1,sizeof(buf)-1,"%.17g",val);
 242         buf[0] = strlen((char*)buf+1);
 243         len = buf[0]+1;
 244     }
 245     return rdbWriteRaw(fp,buf,len);
 246 }
 247
 248 /* Save a Redis object. */
 249 int rdbSaveObject(FILE *fp, robj *o) {
 250     int n, nwritten = 0;
 251
 252     if (o->type == REDIS_STRING) {
 253         /* Save a string value */
 254         if ((n = rdbSaveStringObject(fp,o)) == -1) return -1;
 255         nwritten += n;
 256     } else if (o->type == REDIS_LIST) {
 257         /* Save a list value */
 258         if (o->encoding == REDIS_ENCODING_ZIPLIST) {
 259             unsigned char *p;
 260             unsigned char *vstr;
 261             unsigned int vlen;
 262             long long vlong;
 263
 264             if ((n = rdbSaveLen(fp,ziplistLen(o->ptr))) == -1) return -1;
 265             nwritten += n;
 266
 267             p = ziplistIndex(o->ptr,0);
 268             while(ziplistGet(p,&vstr,&vlen,&vlong)) {
 269                 if (vstr) {
 270                     if ((n = rdbSaveRawString(fp,vstr,vlen)) == -1)
 271                         return -1;
 272                     nwritten += n;
 273                 } else {
 274                     if ((n = rdbSaveLongLongAsStringObject(fp,vlong)) == -1)
 275                         return -1;
 276                     nwritten += n;
 277                 }
 278                 p = ziplistNext(o->ptr,p);
 279             }
 280         } else if (o->encoding == REDIS_ENCODING_LINKEDLIST) {
 281             list *list = o->ptr;
 282             listIter li;
 283             listNode *ln;
 284
 285             if ((n = rdbSaveLen(fp,listLength(list))) == -1) return -1;
 286             nwritten += n;
 287
 288             listRewind(list,&li);
 289             while((ln = listNext(&li))) {
 290                 robj *eleobj = listNodeValue(ln);
 291                 if ((n = rdbSaveStringObject(fp,eleobj)) == -1) return -1;
 292                 nwritten += n;
 293             }
 294         } else {
 295             redisPanic("Unknown list encoding");
 296         }
 297     } else if (o->type == REDIS_SET) {
 298         /* Save a set value */
 299         if (o->encoding == REDIS_ENCODING_HT) {
 300             dict *set = o->ptr;
 301             dictIterator *di = dictGetIterator(set);
 302             dictEntry *de;
 303
 304             if ((n = rdbSaveLen(fp,dictSize(set))) == -1) return -1;
 305             nwritten += n;
 306
 307             while((de = dictNext(di)) != NULL) {
 308                 robj *eleobj = dictGetEntryKey(de);
 309                 if ((n = rdbSaveStringObject(fp,eleobj)) == -1) return -1;
 310                 nwritten += n;
 311             }
 312             dictReleaseIterator(di);
 313         } else if (o->encoding == REDIS_ENCODING_INTSET) {
 314             intset *is = o->ptr;
 315             int64_t llval;
 316             int i = 0;
 317
 318             if ((n = rdbSaveLen(fp,intsetLen(is))) == -1) return -1;
 319             nwritten += n;
 320
 321             while(intsetGet(is,i++,&llval)) {
 322                 if ((n = rdbSaveLongLongAsStringObject(fp,llval)) == -1) return -1;
 323                 nwritten += n;
 324             }
 325         } else {
 326             redisPanic("Unknown set encoding");
 327         }
 328     } else if (o->type == REDIS_ZSET) {
 329         /* Save a set value */
 330         zset *zs = o->ptr;
 331         dictIterator *di = dictGetIterator(zs->dict);
 332         dictEntry *de;
 333
 334         if ((n = rdbSaveLen(fp,dictSize(zs->dict))) == -1) return -1;
 335         nwritten += n;
 336
 337         while((de = dictNext(di)) != NULL) {
 338             robj *eleobj = dictGetEntryKey(de);
 339             double *score = dictGetEntryVal(de);
 340
 341             if ((n = rdbSaveStringObject(fp,eleobj)) == -1) return -1;
 342             nwritten += n;
 343             if ((n = rdbSaveDoubleValue(fp,*score)) == -1) return -1;
 344             nwritten += n;
 345         }
 346         dictReleaseIterator(di);
 347     } else if (o->type == REDIS_HASH) {
 348         /* Save a hash value */
 349         if (o->encoding == REDIS_ENCODING_ZIPMAP) {
 350             unsigned char *p = zipmapRewind(o->ptr);
 351             unsigned int count = zipmapLen(o->ptr);
 352             unsigned char *key, *val;
 353             unsigned int klen, vlen;
 354
 355             if ((n = rdbSaveLen(fp,count)) == -1) return -1;
 356             nwritten += n;
 357
 358             while((p = zipmapNext(p,&key,&klen,&val,&vlen)) != NULL) {
 359                 if ((n = rdbSaveRawString(fp,key,klen)) == -1) return -1;
 360                 nwritten += n;
 361                 if ((n = rdbSaveRawString(fp,val,vlen)) == -1) return -1;
 362                 nwritten += n;
 363             }
 364         } else {
 365             dictIterator *di = dictGetIterator(o->ptr);
 366             dictEntry *de;
 367
 368             if ((n = rdbSaveLen(fp,dictSize((dict*)o->ptr))) == -1) return -1;
 369             nwritten += n;
 370
 371             while((de = dictNext(di)) != NULL) {
 372                 robj *key = dictGetEntryKey(de);
 373                 robj *val = dictGetEntryVal(de);
 374
 375                 if ((n = rdbSaveStringObject(fp,key)) == -1) return -1;
 376                 nwritten += n;
 377                 if ((n = rdbSaveStringObject(fp,val)) == -1) return -1;
 378                 nwritten += n;
 379             }
 380             dictReleaseIterator(di);
 381         }
 382     } else {
 383         redisPanic("Unknown object type");
 384     }
 385     return nwritten;
 386 }
 387
 388 /* Return the length the object will have on disk if saved with
 389  * the rdbSaveObject() function. Currently we use a trick to get
 390  * this length with very little changes to the code. In the future
 391  * we could switch to a faster solution. */
 392 off_t rdbSavedObjectLen(robj *o) {
 393     int len = rdbSaveObject(NULL,o);
 394     redisAssert(len != -1);
 395     return len;
 396 }
 397
 398 /* Return the number of pages required to save this object in the swap file */
 399 off_t rdbSavedObjectPages(robj *o) {
 400     off_t bytes = rdbSavedObjectLen(o);
 401     return (bytes+(server.vm_page_size-1))/server.vm_page_size;
 402 }
 403
 404 /* Save the DB on disk. Return REDIS_ERR on error, REDIS_OK on success */
 405 int rdbSave(char *filename) {
 406     dictIterator *di = NULL;
 407     dictEntry *de;
 408     FILE *fp;
 409     char tmpfile[256];
 410     int j;
 411     time_t now = time(NULL);
 412
 413     /* Wait for I/O therads to terminate, just in case this is a
 414      * foreground-saving, to avoid seeking the swap file descriptor at the
 415      * same time. */
 416     if (server.vm_enabled)
 417         waitEmptyIOJobsQueue();
 418
 419     snprintf(tmpfile,256,"temp-%d.rdb", (int) getpid());
 420     fp = fopen(tmpfile,"w");
 421     if (!fp) {
 422         redisLog(REDIS_WARNING, "Failed saving the DB: %s", strerror(errno));
 423         return REDIS_ERR;
 424     }
 425     if (fwrite("REDIS0001",9,1,fp) == 0) goto werr;
 426     for (j = 0; j < server.dbnum; j++) {
 427         redisDb *db = server.db+j;
 428         dict *d = db->dict;
 429         if (dictSize(d) == 0) continue;
 430         di = dictGetIterator(d);
 431         if (!di) {
 432             fclose(fp);
 433             return REDIS_ERR;
 434         }
 435
 436         /* Write the SELECT DB opcode */
 437         if (rdbSaveType(fp,REDIS_SELECTDB) == -1) goto werr;
 438         if (rdbSaveLen(fp,j) == -1) goto werr;
 439
 440         /* Iterate this DB writing every entry */
 441         while((de = dictNext(di)) != NULL) {
 442             sds keystr = dictGetEntryKey(de);
 443             robj key, *o = dictGetEntryVal(de);
 444             time_t expiretime;
 445
 446             initStaticStringObject(key,keystr);
 447             expiretime = getExpire(db,&key);
 448
 449             /* Save the expire time */
 450             if (expiretime != -1) {
 451                 /* If this key is already expired skip it */
 452                 if (expiretime < now) continue;
 453                 if (rdbSaveType(fp,REDIS_EXPIRETIME) == -1) goto werr;
 454                 if (rdbSaveTime(fp,expiretime) == -1) goto werr;
 455             }
 456             /* Save the key and associated value. This requires special
 457              * handling if the value is swapped out. */
 458             if (!server.vm_enabled || o->storage == REDIS_VM_MEMORY ||
 459                                       o->storage == REDIS_VM_SWAPPING) {
 460                 /* Save type, key, value */
 461                 if (rdbSaveType(fp,o->type) == -1) goto werr;
 462                 if (rdbSaveStringObject(fp,&key) == -1) goto werr;
 463                 if (rdbSaveObject(fp,o) == -1) goto werr;
 464             } else {
 465                 /* REDIS_VM_SWAPPED or REDIS_VM_LOADING */
 466                 robj *po;
 467                 /* Get a preview of the object in memory */
 468                 po = vmPreviewObject(o);
 469                 /* Save type, key, value */
 470                 if (rdbSaveType(fp,po->type) == -1) goto werr;
 471                 if (rdbSaveStringObject(fp,&key) == -1) goto werr;
 472                 if (rdbSaveObject(fp,po) == -1) goto werr;
 473                 /* Remove the loaded object from memory */
 474                 decrRefCount(po);
 475             }
 476         }
 477         dictReleaseIterator(di);
 478     }
 479     /* EOF opcode */
 480     if (rdbSaveType(fp,REDIS_EOF) == -1) goto werr;
 481
 482     /* Make sure data will not remain on the OS's output buffers */
 483     fflush(fp);
 484     fsync(fileno(fp));
 485     fclose(fp);
 486
 487     /* Use RENAME to make sure the DB file is changed atomically only
 488      * if the generate DB file is ok. */
 489     if (rename(tmpfile,filename) == -1) {
 490         redisLog(REDIS_WARNING,"Error moving temp DB file on the final destination: %s", strerror(errno));
 491         unlink(tmpfile);
 492         return REDIS_ERR;
 493     }
 494     redisLog(REDIS_NOTICE,"DB saved on disk");
 495     server.dirty = 0;
 496     server.lastsave = time(NULL);
 497     return REDIS_OK;
 498
 499 werr:
 500     fclose(fp);
 501     unlink(tmpfile);
 502     redisLog(REDIS_WARNING,"Write error saving DB on disk: %s", strerror(errno));
 503     if (di) dictReleaseIterator(di);
 504     return REDIS_ERR;
 505 }
 506
 507 int rdbSaveBackground(char *filename) {
 508     pid_t childpid;
 509
 510     if (server.bgsavechildpid != -1) return REDIS_ERR;
 511     if (server.vm_enabled) waitEmptyIOJobsQueue();
 512     server.dirty_before_bgsave = server.dirty;
 513     if ((childpid = fork()) == 0) {
 514         /* Child */
 515         if (server.vm_enabled) vmReopenSwapFile();
 516         if (server.ipfd > 0) close(server.ipfd);
 517         if (server.sofd > 0) close(server.sofd);
 518         if (rdbSave(filename) == REDIS_OK) {
 519             _exit(0);
 520         } else {
 521             _exit(1);
 522         }
 523     } else {
 524         /* Parent */
 525         if (childpid == -1) {
 526             redisLog(REDIS_WARNING,"Can't save in background: fork: %s",
 527                 strerror(errno));
 528             return REDIS_ERR;
 529         }
 530         redisLog(REDIS_NOTICE,"Background saving started by pid %d",childpid);
 531         server.bgsavechildpid = childpid;
 532         updateDictResizePolicy();
 533         return REDIS_OK;
 534     }
 535     return REDIS_OK; /* unreached */
 536 }
 537
 538 void rdbRemoveTempFile(pid_t childpid) {
 539     char tmpfile[256];
 540
 541     snprintf(tmpfile,256,"temp-%d.rdb", (int) childpid);
 542     unlink(tmpfile);
 543 }
 544
 545 int rdbLoadType(FILE *fp) {
 546     unsigned char type;
 547     if (fread(&type,1,1,fp) == 0) return -1;
 548     return type;
 549 }
 550
 551 time_t rdbLoadTime(FILE *fp) {
 552     int32_t t32;
 553     if (fread(&t32,4,1,fp) == 0) return -1;
 554     return (time_t) t32;
 555 }
 556
 557 /* Load an encoded length from the DB, see the REDIS_RDB_* defines on the top
 558  * of this file for a description of how this are stored on disk.
 559  *
 560  * isencoded is set to 1 if the readed length is not actually a length but
 561  * an "encoding type", check the above comments for more info */
 562 uint32_t rdbLoadLen(FILE *fp, int *isencoded) {
 563     unsigned char buf[2];
 564     uint32_t len;
 565     int type;
 566
 567     if (isencoded) *isencoded = 0;
 568     if (fread(buf,1,1,fp) == 0) return REDIS_RDB_LENERR;
 569     type = (buf[0]&0xC0)>>6;
 570     if (type == REDIS_RDB_6BITLEN) {
 571         /* Read a 6 bit len */
 572         return buf[0]&0x3F;
 573     } else if (type == REDIS_RDB_ENCVAL) {
 574         /* Read a 6 bit len encoding type */
 575         if (isencoded) *isencoded = 1;
 576         return buf[0]&0x3F;
 577     } else if (type == REDIS_RDB_14BITLEN) {
 578         /* Read a 14 bit len */
 579         if (fread(buf+1,1,1,fp) == 0) return REDIS_RDB_LENERR;
 580         return ((buf[0]&0x3F)<<8)|buf[1];
 581     } else {
 582         /* Read a 32 bit len */
 583         if (fread(&len,4,1,fp) == 0) return REDIS_RDB_LENERR;
 584         return ntohl(len);
 585     }
 586 }
 587
 588 /* Load an integer-encoded object from file 'fp', with the specified
 589  * encoding type 'enctype'. If encode is true the function may return
 590  * an integer-encoded object as reply, otherwise the returned object
 591  * will always be encoded as a raw string. */
 592 robj *rdbLoadIntegerObject(FILE *fp, int enctype, int encode) {
 593     unsigned char enc[4];
 594     long long val;
 595
 596     if (enctype == REDIS_RDB_ENC_INT8) {
 597         if (fread(enc,1,1,fp) == 0) return NULL;
 598         val = (signed char)enc[0];
 599     } else if (enctype == REDIS_RDB_ENC_INT16) {
 600         uint16_t v;
 601         if (fread(enc,2,1,fp) == 0) return NULL;
 602         v = enc[0]|(enc[1]<<8);
 603         val = (int16_t)v;
 604     } else if (enctype == REDIS_RDB_ENC_INT32) {
 605         uint32_t v;
 606         if (fread(enc,4,1,fp) == 0) return NULL;
 607         v = enc[0]|(enc[1]<<8)|(enc[2]<<16)|(enc[3]<<24);
 608         val = (int32_t)v;
 609     } else {
 610         val = 0; /* anti-warning */
 611         redisPanic("Unknown RDB integer encoding type");
 612     }
 613     if (encode)
 614         return createStringObjectFromLongLong(val);
 615     else
 616         return createObject(REDIS_STRING,sdsfromlonglong(val));
 617 }
 618
 619 robj *rdbLoadLzfStringObject(FILE*fp) {
 620     unsigned int len, clen;
 621     unsigned char *c = NULL;
 622     sds val = NULL;
 623
 624     if ((clen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 625     if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 626     if ((c = zmalloc(clen)) == NULL) goto err;
 627     if ((val = sdsnewlen(NULL,len)) == NULL) goto err;
 628     if (fread(c,clen,1,fp) == 0) goto err;
 629     if (lzf_decompress(c,clen,val,len) == 0) goto err;
 630     zfree(c);
 631     return createObject(REDIS_STRING,val);
 632 err:
 633     zfree(c);
 634     sdsfree(val);
 635     return NULL;
 636 }
 637
 638 robj *rdbGenericLoadStringObject(FILE*fp, int encode) {
 639     int isencoded;
 640     uint32_t len;
 641     sds val;
 642
 643     len = rdbLoadLen(fp,&isencoded);
 644     if (isencoded) {
 645         switch(len) {
 646         case REDIS_RDB_ENC_INT8:
 647         case REDIS_RDB_ENC_INT16:
 648         case REDIS_RDB_ENC_INT32:
 649             return rdbLoadIntegerObject(fp,len,encode);
 650         case REDIS_RDB_ENC_LZF:
 651             return rdbLoadLzfStringObject(fp);
 652         default:
 653             redisPanic("Unknown RDB encoding type");
 654         }
 655     }
 656
 657     if (len == REDIS_RDB_LENERR) return NULL;
 658     val = sdsnewlen(NULL,len);
 659     if (len && fread(val,len,1,fp) == 0) {
 660         sdsfree(val);
 661         return NULL;
 662     }
 663     return createObject(REDIS_STRING,val);
 664 }
 665
 666 robj *rdbLoadStringObject(FILE *fp) {
 667     return rdbGenericLoadStringObject(fp,0);
 668 }
 669
 670 robj *rdbLoadEncodedStringObject(FILE *fp) {
 671     return rdbGenericLoadStringObject(fp,1);
 672 }
 673
 674 /* For information about double serialization check rdbSaveDoubleValue() */
 675 int rdbLoadDoubleValue(FILE *fp, double *val) {
 676     char buf[128];
 677     unsigned char len;
 678
 679     if (fread(&len,1,1,fp) == 0) return -1;
 680     switch(len) {
 681     case 255: *val = R_NegInf; return 0;
 682     case 254: *val = R_PosInf; return 0;
 683     case 253: *val = R_Nan; return 0;
 684     default:
 685         if (fread(buf,len,1,fp) == 0) return -1;
 686         buf[len] = '\0';
 687         sscanf(buf, "%lg", val);
 688         return 0;
 689     }
 690 }
 691
 692 /* Load a Redis object of the specified type from the specified file.
 693  * On success a newly allocated object is returned, otherwise NULL. */
 694 robj *rdbLoadObject(int type, FILE *fp) {
 695     robj *o, *ele, *dec;
 696     size_t len;
 697     unsigned int i;
 698
 699     redisLog(REDIS_DEBUG,"LOADING OBJECT %d (at %d)\n",type,ftell(fp));
 700     if (type == REDIS_STRING) {
 701         /* Read string value */
 702         if ((o = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 703         o = tryObjectEncoding(o);
 704     } else if (type == REDIS_LIST) {
 705         /* Read list value */
 706         if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 707
 708         /* Use a real list when there are too many entries */
 709         if (len > server.list_max_ziplist_entries) {
 710             o = createListObject();
 711         } else {
 712             o = createZiplistObject();
 713         }
 714
 715         /* Load every single element of the list */
 716         while(len--) {
 717             if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 718
 719             /* If we are using a ziplist and the value is too big, convert
 720              * the object to a real list. */
 721             if (o->encoding == REDIS_ENCODING_ZIPLIST &&
 722                 ele->encoding == REDIS_ENCODING_RAW &&
 723                 sdslen(ele->ptr) > server.list_max_ziplist_value)
 724                     listTypeConvert(o,REDIS_ENCODING_LINKEDLIST);
 725
 726             if (o->encoding == REDIS_ENCODING_ZIPLIST) {
 727                 dec = getDecodedObject(ele);
 728                 o->ptr = ziplistPush(o->ptr,dec->ptr,sdslen(dec->ptr),REDIS_TAIL);
 729                 decrRefCount(dec);
 730                 decrRefCount(ele);
 731             } else {
 732                 ele = tryObjectEncoding(ele);
 733                 listAddNodeTail(o->ptr,ele);
 734             }
 735         }
 736     } else if (type == REDIS_SET) {
 737         /* Read list/set value */
 738         if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 739
 740         /* Use a regular set when there are too many entries. */
 741         if (len > server.set_max_intset_entries) {
 742             o = createSetObject();
 743             /* It's faster to expand the dict to the right size asap in order
 744              * to avoid rehashing */
 745             if (len > DICT_HT_INITIAL_SIZE)
 746                 dictExpand(o->ptr,len);
 747         } else {
 748             o = createIntsetObject();
 749         }
 750
 751         /* Load every single element of the list/set */
 752         for (i = 0; i < len; i++) {
 753             long long llval;
 754             if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 755             ele = tryObjectEncoding(ele);
 756
 757             if (o->encoding == REDIS_ENCODING_INTSET) {
 758                 /* Fetch integer value from element */
 759                 if (isObjectRepresentableAsLongLong(ele,&llval) == REDIS_OK) {
 760                     o->ptr = intsetAdd(o->ptr,llval,NULL);
 761                 } else {
 762                     setTypeConvert(o,REDIS_ENCODING_HT);
 763                     dictExpand(o->ptr,len);
 764                 }
 765             }
 766
 767             /* This will also be called when the set was just converted
 768              * to regular hashtable encoded set */
 769             if (o->encoding == REDIS_ENCODING_HT) {
 770                 dictAdd((dict*)o->ptr,ele,NULL);
 771             } else {
 772                 decrRefCount(ele);
 773             }
 774         }
 775     } else if (type == REDIS_ZSET) {
 776         /* Read list/set value */
 777         size_t zsetlen;
 778         zset *zs;
 779
 780         if ((zsetlen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 781         o = createZsetObject();
 782         zs = o->ptr;
 783         /* Load every single element of the list/set */
 784         while(zsetlen--) {
 785             robj *ele;
 786             double score;
 787             zskiplistNode *znode;
 788
 789             if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 790             ele = tryObjectEncoding(ele);
 791             if (rdbLoadDoubleValue(fp,&score) == -1) return NULL;
 792             znode = zslInsert(zs->zsl,score,ele);
 793             dictAdd(zs->dict,ele,&znode->score);
 794             incrRefCount(ele); /* added to skiplist */
 795         }
 796     } else if (type == REDIS_HASH) {
 797         size_t hashlen;
 798
 799         if ((hashlen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL;
 800         o = createHashObject();
 801         /* Too many entries? Use an hash table. */
 802         if (hashlen > server.hash_max_zipmap_entries)
 803             convertToRealHash(o);
 804         /* Load every key/value, then set it into the zipmap or hash
 805          * table, as needed. */
 806         while(hashlen--) {
 807             robj *key, *val;
 808
 809             if ((key = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 810             if ((val = rdbLoadEncodedStringObject(fp)) == NULL) return NULL;
 811             /* If we are using a zipmap and there are too big values
 812              * the object is converted to real hash table encoding. */
 813             if (o->encoding != REDIS_ENCODING_HT &&
 814                ((key->encoding == REDIS_ENCODING_RAW &&
 815                 sdslen(key->ptr) > server.hash_max_zipmap_value) ||
 816                 (val->encoding == REDIS_ENCODING_RAW &&
 817                 sdslen(val->ptr) > server.hash_max_zipmap_value)))
 818             {
 819                     convertToRealHash(o);
 820             }
 821
 822             if (o->encoding == REDIS_ENCODING_ZIPMAP) {
 823                 unsigned char *zm = o->ptr;
 824                 robj *deckey, *decval;
 825
 826                 /* We need raw string objects to add them to the zipmap */
 827                 deckey = getDecodedObject(key);
 828                 decval = getDecodedObject(val);
 829                 zm = zipmapSet(zm,deckey->ptr,sdslen(deckey->ptr),
 830                                   decval->ptr,sdslen(decval->ptr),NULL);
 831                 o->ptr = zm;
 832                 decrRefCount(deckey);
 833                 decrRefCount(decval);
 834                 decrRefCount(key);
 835                 decrRefCount(val);
 836             } else {
 837                 key = tryObjectEncoding(key);
 838                 val = tryObjectEncoding(val);
 839                 dictAdd((dict*)o->ptr,key,val);
 840             }
 841         }
 842     } else {
 843         redisPanic("Unknown object type");
 844     }
 845     return o;
 846 }
 847
 848 /* Mark that we are loading in the global state and setup the fields
 849  * needed to provide loading stats. */
 850 void startLoading(FILE *fp) {
 851     struct stat sb;
 852
 853     /* Load the DB */
 854     server.loading = 1;
 855     server.loading_start_time = time(NULL);
 856     if (fstat(fileno(fp), &sb) == -1) {
 857         server.loading_total_bytes = 1; /* just to avoid division by zero */
 858     } else {
 859         server.loading_total_bytes = sb.st_size;
 860     }
 861 }
 862
 863 /* Refresh the loading progress info */
 864 void loadingProgress(off_t pos) {
 865     server.loading_loaded_bytes = pos;
 866 }
 867
 868 /* Loading finished */
 869 void stopLoading(void) {
 870     server.loading = 0;
 871 }
 872
 873 int rdbLoad(char *filename) {
 874     FILE *fp;
 875     uint32_t dbid;
 876     int type, retval, rdbver;
 877     int swap_all_values = 0;
 878     redisDb *db = server.db+0;
 879     char buf[1024];
 880     time_t expiretime, now = time(NULL);
 881     long loops = 0;
 882
 883     fp = fopen(filename,"r");
 884     if (!fp) return REDIS_ERR;
 885     if (fread(buf,9,1,fp) == 0) goto eoferr;
 886     buf[9] = '\0';
 887     if (memcmp(buf,"REDIS",5) != 0) {
 888         fclose(fp);
 889         redisLog(REDIS_WARNING,"Wrong signature trying to load DB from file");
 890         return REDIS_ERR;
 891     }
 892     rdbver = atoi(buf+5);
 893     if (rdbver != 1) {
 894         fclose(fp);
 895         redisLog(REDIS_WARNING,"Can't handle RDB format version %d",rdbver);
 896         return REDIS_ERR;
 897     }
 898
 899     startLoading(fp);
 900     while(1) {
 901         robj *key, *val;
 902         int force_swapout;
 903
 904         expiretime = -1;
 905
 906         /* Serve the clients from time to time */
 907         if (!(loops++ % 1000)) {
 908             loadingProgress(ftello(fp));
 909             aeProcessEvents(server.el, AE_FILE_EVENTS|AE_DONT_WAIT);
 910         }
 911
 912         /* Read type. */
 913         if ((type = rdbLoadType(fp)) == -1) goto eoferr;
 914         if (type == REDIS_EXPIRETIME) {
 915             if ((expiretime = rdbLoadTime(fp)) == -1) goto eoferr;
 916             /* We read the time so we need to read the object type again */
 917             if ((type = rdbLoadType(fp)) == -1) goto eoferr;
 918         }
 919         if (type == REDIS_EOF) break;
 920         /* Handle SELECT DB opcode as a special case */
 921         if (type == REDIS_SELECTDB) {
 922             if ((dbid = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR)
 923                 goto eoferr;
 924             if (dbid >= (unsigned)server.dbnum) {
 925                 redisLog(REDIS_WARNING,"FATAL: Data file was created with a Redis server configured to handle more than %d databases. Exiting\n", server.dbnum);
 926                 exit(1);
 927             }
 928             db = server.db+dbid;
 929             continue;
 930         }
 931         /* Read key */
 932         if ((key = rdbLoadStringObject(fp)) == NULL) goto eoferr;
 933         /* Read value */
 934         if ((val = rdbLoadObject(type,fp)) == NULL) goto eoferr;
 935         /* Check if the key already expired */
 936         if (expiretime != -1 && expiretime < now) {
 937             decrRefCount(key);
 938             decrRefCount(val);
 939             continue;
 940         }
 941         /* Add the new object in the hash table */
 942         retval = dbAdd(db,key,val);
 943         if (retval == REDIS_ERR) {
 944             redisLog(REDIS_WARNING,"Loading DB, duplicated key (%s) found! Unrecoverable error, exiting now.", key->ptr);
 945             exit(1);
 946         }
 947         /* Set the expire time if needed */
 948         if (expiretime != -1) setExpire(db,key,expiretime);
 949
 950         /* Handle swapping while loading big datasets when VM is on */
 951
 952         /* If we detecter we are hopeless about fitting something in memory
 953          * we just swap every new key on disk. Directly...
 954          * Note that's important to check for this condition before resorting
 955          * to random sampling, otherwise we may try to swap already
 956          * swapped keys. */
 957         if (swap_all_values) {
 958             dictEntry *de = dictFind(db->dict,key->ptr);
 959
 960             /* de may be NULL since the key already expired */
 961             if (de) {
 962                 vmpointer *vp;
 963                 val = dictGetEntryVal(de);
 964
 965                 if (val->refcount == 1 &&
 966                     (vp = vmSwapObjectBlocking(val)) != NULL)
 967                     dictGetEntryVal(de) = vp;
 968             }
 969             decrRefCount(key);
 970             continue;
 971         }
 972         decrRefCount(key);
 973
 974         /* Flush data on disk once 32 MB of additional RAM are used... */
 975         force_swapout = 0;
 976         if ((zmalloc_used_memory() - server.vm_max_memory) > 1024*1024*32)
 977             force_swapout = 1;
 978
 979         /* If we have still some hope of having some value fitting memory
 980          * then we try random sampling. */
 981         if (!swap_all_values && server.vm_enabled && force_swapout) {
 982             while (zmalloc_used_memory() > server.vm_max_memory) {
 983                 if (vmSwapOneObjectBlocking() == REDIS_ERR) break;
 984             }
 985             if (zmalloc_used_memory() > server.vm_max_memory)
 986                 swap_all_values = 1; /* We are already using too much mem */
 987         }
 988     }
 989     fclose(fp);
 990     stopLoading();
 991     return REDIS_OK;
 992
 993 eoferr: /* unexpected end of file is handled here with a fatal exit */
 994     redisLog(REDIS_WARNING,"Short read or OOM loading DB. Unrecoverable error, aborting now.");
 995     exit(1);
 996     return REDIS_ERR; /* Just to avoid warning */
 997 }
 998
 999 /* A background saving child (BGSAVE) terminated its work. Handle this. */
1000 void backgroundSaveDoneHandler(int statloc) {
1001     int exitcode = WEXITSTATUS(statloc);
1002     int bysignal = WIFSIGNALED(statloc);
1003
1004     if (!bysignal && exitcode == 0) {
1005         redisLog(REDIS_NOTICE,
1006             "Background saving terminated with success");
1007         server.dirty = server.dirty - server.dirty_before_bgsave;
1008         server.lastsave = time(NULL);
1009     } else if (!bysignal && exitcode != 0) {
1010         redisLog(REDIS_WARNING, "Background saving error");
1011     } else {
1012         redisLog(REDIS_WARNING,
1013             "Background saving terminated by signal %d", WTERMSIG(statloc));
1014         rdbRemoveTempFile(server.bgsavechildpid);
1015     }
1016     server.bgsavechildpid = -1;
1017     /* Possibly there are slaves waiting for a BGSAVE in order to be served
1018      * (the first stage of SYNC is a bulk transfer of dump.rdb) */
1019     updateSlavesWaitingBgsave(exitcode == 0 ? REDIS_OK : REDIS_ERR);
1020 }