]>
Commit | Line | Data |
---|---|---|
e2641e09 | 1 | #include "redis.h" |
2 | #include "lzf.h" /* LZF compression library */ | |
3 | ||
4 | #include <math.h> | |
3688d7f3 | 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> | |
e2641e09 | 10 | |
11 | int rdbSaveType(FILE *fp, unsigned char type) { | |
12 | if (fwrite(&type,1,1,fp) == 0) return -1; | |
13 | return 0; | |
14 | } | |
15 | ||
16 | int rdbSaveTime(FILE *fp, time_t t) { | |
17 | int32_t t32 = (int32_t) t; | |
18 | if (fwrite(&t32,4,1,fp) == 0) return -1; | |
19 | return 0; | |
20 | } | |
21 | ||
22 | /* check rdbLoadLen() comments for more info */ | |
23 | int rdbSaveLen(FILE *fp, uint32_t len) { | |
24 | unsigned char buf[2]; | |
25 | ||
26 | if (len < (1<<6)) { | |
27 | /* Save a 6 bit len */ | |
28 | buf[0] = (len&0xFF)|(REDIS_RDB_6BITLEN<<6); | |
29 | if (fwrite(buf,1,1,fp) == 0) return -1; | |
30 | } else if (len < (1<<14)) { | |
31 | /* Save a 14 bit len */ | |
32 | buf[0] = ((len>>8)&0xFF)|(REDIS_RDB_14BITLEN<<6); | |
33 | buf[1] = len&0xFF; | |
34 | if (fwrite(buf,2,1,fp) == 0) return -1; | |
35 | } else { | |
36 | /* Save a 32 bit len */ | |
37 | buf[0] = (REDIS_RDB_32BITLEN<<6); | |
38 | if (fwrite(buf,1,1,fp) == 0) return -1; | |
39 | len = htonl(len); | |
40 | if (fwrite(&len,4,1,fp) == 0) return -1; | |
41 | } | |
42 | return 0; | |
43 | } | |
44 | ||
45 | /* Encode 'value' as an integer if possible (if integer will fit the | |
46 | * supported range). If the function sucessful encoded the integer | |
47 | * then the (up to 5 bytes) encoded representation is written in the | |
48 | * string pointed by 'enc' and the length is returned. Otherwise | |
49 | * 0 is returned. */ | |
50 | int rdbEncodeInteger(long long value, unsigned char *enc) { | |
51 | /* Finally check if it fits in our ranges */ | |
52 | if (value >= -(1<<7) && value <= (1<<7)-1) { | |
53 | enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT8; | |
54 | enc[1] = value&0xFF; | |
55 | return 2; | |
56 | } else if (value >= -(1<<15) && value <= (1<<15)-1) { | |
57 | enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT16; | |
58 | enc[1] = value&0xFF; | |
59 | enc[2] = (value>>8)&0xFF; | |
60 | return 3; | |
61 | } else if (value >= -((long long)1<<31) && value <= ((long long)1<<31)-1) { | |
62 | enc[0] = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_INT32; | |
63 | enc[1] = value&0xFF; | |
64 | enc[2] = (value>>8)&0xFF; | |
65 | enc[3] = (value>>16)&0xFF; | |
66 | enc[4] = (value>>24)&0xFF; | |
67 | return 5; | |
68 | } else { | |
69 | return 0; | |
70 | } | |
71 | } | |
72 | ||
73 | /* String objects in the form "2391" "-100" without any space and with a | |
74 | * range of values that can fit in an 8, 16 or 32 bit signed value can be | |
75 | * encoded as integers to save space */ | |
76 | int rdbTryIntegerEncoding(char *s, size_t len, unsigned char *enc) { | |
77 | long long value; | |
78 | char *endptr, buf[32]; | |
79 | ||
80 | /* Check if it's possible to encode this value as a number */ | |
81 | value = strtoll(s, &endptr, 10); | |
82 | if (endptr[0] != '\0') return 0; | |
83 | ll2string(buf,32,value); | |
84 | ||
85 | /* If the number converted back into a string is not identical | |
86 | * then it's not possible to encode the string as integer */ | |
87 | if (strlen(buf) != len || memcmp(buf,s,len)) return 0; | |
88 | ||
89 | return rdbEncodeInteger(value,enc); | |
90 | } | |
91 | ||
92 | int rdbSaveLzfStringObject(FILE *fp, unsigned char *s, size_t len) { | |
93 | size_t comprlen, outlen; | |
94 | unsigned char byte; | |
95 | void *out; | |
96 | ||
97 | /* We require at least four bytes compression for this to be worth it */ | |
98 | if (len <= 4) return 0; | |
99 | outlen = len-4; | |
100 | if ((out = zmalloc(outlen+1)) == NULL) return 0; | |
101 | comprlen = lzf_compress(s, len, out, outlen); | |
102 | if (comprlen == 0) { | |
103 | zfree(out); | |
104 | return 0; | |
105 | } | |
106 | /* Data compressed! Let's save it on disk */ | |
107 | byte = (REDIS_RDB_ENCVAL<<6)|REDIS_RDB_ENC_LZF; | |
108 | if (fwrite(&byte,1,1,fp) == 0) goto writeerr; | |
109 | if (rdbSaveLen(fp,comprlen) == -1) goto writeerr; | |
110 | if (rdbSaveLen(fp,len) == -1) goto writeerr; | |
111 | if (fwrite(out,comprlen,1,fp) == 0) goto writeerr; | |
112 | zfree(out); | |
113 | return comprlen; | |
114 | ||
115 | writeerr: | |
116 | zfree(out); | |
117 | return -1; | |
118 | } | |
119 | ||
120 | /* Save a string objet as [len][data] on disk. If the object is a string | |
121 | * representation of an integer value we try to safe it in a special form */ | |
122 | int rdbSaveRawString(FILE *fp, unsigned char *s, size_t len) { | |
123 | int enclen; | |
124 | ||
125 | /* Try integer encoding */ | |
126 | if (len <= 11) { | |
127 | unsigned char buf[5]; | |
128 | if ((enclen = rdbTryIntegerEncoding((char*)s,len,buf)) > 0) { | |
129 | if (fwrite(buf,enclen,1,fp) == 0) return -1; | |
130 | return 0; | |
131 | } | |
132 | } | |
133 | ||
134 | /* Try LZF compression - under 20 bytes it's unable to compress even | |
135 | * aaaaaaaaaaaaaaaaaa so skip it */ | |
136 | if (server.rdbcompression && len > 20) { | |
137 | int retval; | |
138 | ||
139 | retval = rdbSaveLzfStringObject(fp,s,len); | |
140 | if (retval == -1) return -1; | |
141 | if (retval > 0) return 0; | |
142 | /* retval == 0 means data can't be compressed, save the old way */ | |
143 | } | |
144 | ||
145 | /* Store verbatim */ | |
146 | if (rdbSaveLen(fp,len) == -1) return -1; | |
147 | if (len && fwrite(s,len,1,fp) == 0) return -1; | |
148 | return 0; | |
149 | } | |
150 | ||
151 | /* Save a long long value as either an encoded string or a string. */ | |
152 | int rdbSaveLongLongAsStringObject(FILE *fp, long long value) { | |
153 | unsigned char buf[32]; | |
154 | int enclen = rdbEncodeInteger(value,buf); | |
155 | if (enclen > 0) { | |
156 | if (fwrite(buf,enclen,1,fp) == 0) return -1; | |
157 | } else { | |
158 | /* Encode as string */ | |
159 | enclen = ll2string((char*)buf,32,value); | |
160 | redisAssert(enclen < 32); | |
161 | if (rdbSaveLen(fp,enclen) == -1) return -1; | |
162 | if (fwrite(buf,enclen,1,fp) == 0) return -1; | |
163 | } | |
164 | return 0; | |
165 | } | |
166 | ||
167 | /* Like rdbSaveStringObjectRaw() but handle encoded objects */ | |
168 | int rdbSaveStringObject(FILE *fp, robj *obj) { | |
169 | /* Avoid to decode the object, then encode it again, if the | |
170 | * object is alrady integer encoded. */ | |
171 | if (obj->encoding == REDIS_ENCODING_INT) { | |
172 | return rdbSaveLongLongAsStringObject(fp,(long)obj->ptr); | |
173 | } else { | |
174 | redisAssert(obj->encoding == REDIS_ENCODING_RAW); | |
175 | return rdbSaveRawString(fp,obj->ptr,sdslen(obj->ptr)); | |
176 | } | |
177 | } | |
178 | ||
179 | /* Save a double value. Doubles are saved as strings prefixed by an unsigned | |
180 | * 8 bit integer specifing the length of the representation. | |
181 | * This 8 bit integer has special values in order to specify the following | |
182 | * conditions: | |
183 | * 253: not a number | |
184 | * 254: + inf | |
185 | * 255: - inf | |
186 | */ | |
187 | int rdbSaveDoubleValue(FILE *fp, double val) { | |
188 | unsigned char buf[128]; | |
189 | int len; | |
190 | ||
191 | if (isnan(val)) { | |
192 | buf[0] = 253; | |
193 | len = 1; | |
194 | } else if (!isfinite(val)) { | |
195 | len = 1; | |
196 | buf[0] = (val < 0) ? 255 : 254; | |
197 | } else { | |
198 | #if (DBL_MANT_DIG >= 52) && (LLONG_MAX == 0x7fffffffffffffffLL) | |
199 | /* Check if the float is in a safe range to be casted into a | |
200 | * long long. We are assuming that long long is 64 bit here. | |
201 | * Also we are assuming that there are no implementations around where | |
202 | * double has precision < 52 bit. | |
203 | * | |
204 | * Under this assumptions we test if a double is inside an interval | |
205 | * where casting to long long is safe. Then using two castings we | |
206 | * make sure the decimal part is zero. If all this is true we use | |
207 | * integer printing function that is much faster. */ | |
208 | double min = -4503599627370495; /* (2^52)-1 */ | |
209 | double max = 4503599627370496; /* -(2^52) */ | |
210 | if (val > min && val < max && val == ((double)((long long)val))) | |
211 | ll2string((char*)buf+1,sizeof(buf),(long long)val); | |
212 | else | |
213 | #endif | |
214 | snprintf((char*)buf+1,sizeof(buf)-1,"%.17g",val); | |
215 | buf[0] = strlen((char*)buf+1); | |
216 | len = buf[0]+1; | |
217 | } | |
218 | if (fwrite(buf,len,1,fp) == 0) return -1; | |
219 | return 0; | |
220 | } | |
221 | ||
222 | /* Save a Redis object. */ | |
223 | int rdbSaveObject(FILE *fp, robj *o) { | |
224 | if (o->type == REDIS_STRING) { | |
225 | /* Save a string value */ | |
226 | if (rdbSaveStringObject(fp,o) == -1) return -1; | |
227 | } else if (o->type == REDIS_LIST) { | |
228 | /* Save a list value */ | |
229 | if (o->encoding == REDIS_ENCODING_ZIPLIST) { | |
230 | unsigned char *p; | |
231 | unsigned char *vstr; | |
232 | unsigned int vlen; | |
233 | long long vlong; | |
234 | ||
235 | if (rdbSaveLen(fp,ziplistLen(o->ptr)) == -1) return -1; | |
236 | p = ziplistIndex(o->ptr,0); | |
237 | while(ziplistGet(p,&vstr,&vlen,&vlong)) { | |
238 | if (vstr) { | |
239 | if (rdbSaveRawString(fp,vstr,vlen) == -1) | |
240 | return -1; | |
241 | } else { | |
242 | if (rdbSaveLongLongAsStringObject(fp,vlong) == -1) | |
243 | return -1; | |
244 | } | |
245 | p = ziplistNext(o->ptr,p); | |
246 | } | |
247 | } else if (o->encoding == REDIS_ENCODING_LINKEDLIST) { | |
248 | list *list = o->ptr; | |
249 | listIter li; | |
250 | listNode *ln; | |
251 | ||
252 | if (rdbSaveLen(fp,listLength(list)) == -1) return -1; | |
253 | listRewind(list,&li); | |
254 | while((ln = listNext(&li))) { | |
255 | robj *eleobj = listNodeValue(ln); | |
256 | if (rdbSaveStringObject(fp,eleobj) == -1) return -1; | |
257 | } | |
258 | } else { | |
259 | redisPanic("Unknown list encoding"); | |
260 | } | |
261 | } else if (o->type == REDIS_SET) { | |
262 | /* Save a set value */ | |
263 | dict *set = o->ptr; | |
264 | dictIterator *di = dictGetIterator(set); | |
265 | dictEntry *de; | |
266 | ||
267 | if (rdbSaveLen(fp,dictSize(set)) == -1) return -1; | |
268 | while((de = dictNext(di)) != NULL) { | |
269 | robj *eleobj = dictGetEntryKey(de); | |
270 | ||
271 | if (rdbSaveStringObject(fp,eleobj) == -1) return -1; | |
272 | } | |
273 | dictReleaseIterator(di); | |
274 | } else if (o->type == REDIS_ZSET) { | |
275 | /* Save a set value */ | |
276 | zset *zs = o->ptr; | |
277 | dictIterator *di = dictGetIterator(zs->dict); | |
278 | dictEntry *de; | |
279 | ||
280 | if (rdbSaveLen(fp,dictSize(zs->dict)) == -1) return -1; | |
281 | while((de = dictNext(di)) != NULL) { | |
282 | robj *eleobj = dictGetEntryKey(de); | |
283 | double *score = dictGetEntryVal(de); | |
284 | ||
285 | if (rdbSaveStringObject(fp,eleobj) == -1) return -1; | |
286 | if (rdbSaveDoubleValue(fp,*score) == -1) return -1; | |
287 | } | |
288 | dictReleaseIterator(di); | |
289 | } else if (o->type == REDIS_HASH) { | |
290 | /* Save a hash value */ | |
291 | if (o->encoding == REDIS_ENCODING_ZIPMAP) { | |
292 | unsigned char *p = zipmapRewind(o->ptr); | |
293 | unsigned int count = zipmapLen(o->ptr); | |
294 | unsigned char *key, *val; | |
295 | unsigned int klen, vlen; | |
296 | ||
297 | if (rdbSaveLen(fp,count) == -1) return -1; | |
298 | while((p = zipmapNext(p,&key,&klen,&val,&vlen)) != NULL) { | |
299 | if (rdbSaveRawString(fp,key,klen) == -1) return -1; | |
300 | if (rdbSaveRawString(fp,val,vlen) == -1) return -1; | |
301 | } | |
302 | } else { | |
303 | dictIterator *di = dictGetIterator(o->ptr); | |
304 | dictEntry *de; | |
305 | ||
306 | if (rdbSaveLen(fp,dictSize((dict*)o->ptr)) == -1) return -1; | |
307 | while((de = dictNext(di)) != NULL) { | |
308 | robj *key = dictGetEntryKey(de); | |
309 | robj *val = dictGetEntryVal(de); | |
310 | ||
311 | if (rdbSaveStringObject(fp,key) == -1) return -1; | |
312 | if (rdbSaveStringObject(fp,val) == -1) return -1; | |
313 | } | |
314 | dictReleaseIterator(di); | |
315 | } | |
316 | } else { | |
317 | redisPanic("Unknown object type"); | |
318 | } | |
319 | return 0; | |
320 | } | |
321 | ||
322 | /* Return the length the object will have on disk if saved with | |
323 | * the rdbSaveObject() function. Currently we use a trick to get | |
324 | * this length with very little changes to the code. In the future | |
325 | * we could switch to a faster solution. */ | |
326 | off_t rdbSavedObjectLen(robj *o, FILE *fp) { | |
327 | if (fp == NULL) fp = server.devnull; | |
328 | rewind(fp); | |
329 | redisAssert(rdbSaveObject(fp,o) != 1); | |
330 | return ftello(fp); | |
331 | } | |
332 | ||
333 | /* Return the number of pages required to save this object in the swap file */ | |
334 | off_t rdbSavedObjectPages(robj *o, FILE *fp) { | |
335 | off_t bytes = rdbSavedObjectLen(o,fp); | |
336 | ||
337 | return (bytes+(server.vm_page_size-1))/server.vm_page_size; | |
338 | } | |
339 | ||
340 | /* Save the DB on disk. Return REDIS_ERR on error, REDIS_OK on success */ | |
341 | int rdbSave(char *filename) { | |
342 | dictIterator *di = NULL; | |
343 | dictEntry *de; | |
344 | FILE *fp; | |
345 | char tmpfile[256]; | |
346 | int j; | |
347 | time_t now = time(NULL); | |
348 | ||
349 | /* Wait for I/O therads to terminate, just in case this is a | |
350 | * foreground-saving, to avoid seeking the swap file descriptor at the | |
351 | * same time. */ | |
352 | if (server.vm_enabled) | |
353 | waitEmptyIOJobsQueue(); | |
354 | ||
355 | snprintf(tmpfile,256,"temp-%d.rdb", (int) getpid()); | |
356 | fp = fopen(tmpfile,"w"); | |
357 | if (!fp) { | |
358 | redisLog(REDIS_WARNING, "Failed saving the DB: %s", strerror(errno)); | |
359 | return REDIS_ERR; | |
360 | } | |
361 | if (fwrite("REDIS0001",9,1,fp) == 0) goto werr; | |
362 | for (j = 0; j < server.dbnum; j++) { | |
363 | redisDb *db = server.db+j; | |
364 | dict *d = db->dict; | |
365 | if (dictSize(d) == 0) continue; | |
366 | di = dictGetIterator(d); | |
367 | if (!di) { | |
368 | fclose(fp); | |
369 | return REDIS_ERR; | |
370 | } | |
371 | ||
372 | /* Write the SELECT DB opcode */ | |
373 | if (rdbSaveType(fp,REDIS_SELECTDB) == -1) goto werr; | |
374 | if (rdbSaveLen(fp,j) == -1) goto werr; | |
375 | ||
376 | /* Iterate this DB writing every entry */ | |
377 | while((de = dictNext(di)) != NULL) { | |
378 | sds keystr = dictGetEntryKey(de); | |
379 | robj key, *o = dictGetEntryVal(de); | |
380 | time_t expiretime; | |
381 | ||
382 | initStaticStringObject(key,keystr); | |
383 | expiretime = getExpire(db,&key); | |
384 | ||
385 | /* Save the expire time */ | |
386 | if (expiretime != -1) { | |
387 | /* If this key is already expired skip it */ | |
388 | if (expiretime < now) continue; | |
389 | if (rdbSaveType(fp,REDIS_EXPIRETIME) == -1) goto werr; | |
390 | if (rdbSaveTime(fp,expiretime) == -1) goto werr; | |
391 | } | |
392 | /* Save the key and associated value. This requires special | |
393 | * handling if the value is swapped out. */ | |
394 | if (!server.vm_enabled || o->storage == REDIS_VM_MEMORY || | |
395 | o->storage == REDIS_VM_SWAPPING) { | |
396 | /* Save type, key, value */ | |
397 | if (rdbSaveType(fp,o->type) == -1) goto werr; | |
398 | if (rdbSaveStringObject(fp,&key) == -1) goto werr; | |
399 | if (rdbSaveObject(fp,o) == -1) goto werr; | |
400 | } else { | |
401 | /* REDIS_VM_SWAPPED or REDIS_VM_LOADING */ | |
402 | robj *po; | |
403 | /* Get a preview of the object in memory */ | |
404 | po = vmPreviewObject(o); | |
405 | /* Save type, key, value */ | |
406 | if (rdbSaveType(fp,po->type) == -1) goto werr; | |
407 | if (rdbSaveStringObject(fp,&key) == -1) goto werr; | |
408 | if (rdbSaveObject(fp,po) == -1) goto werr; | |
409 | /* Remove the loaded object from memory */ | |
410 | decrRefCount(po); | |
411 | } | |
412 | } | |
413 | dictReleaseIterator(di); | |
414 | } | |
415 | /* EOF opcode */ | |
416 | if (rdbSaveType(fp,REDIS_EOF) == -1) goto werr; | |
417 | ||
418 | /* Make sure data will not remain on the OS's output buffers */ | |
419 | fflush(fp); | |
420 | fsync(fileno(fp)); | |
421 | fclose(fp); | |
422 | ||
423 | /* Use RENAME to make sure the DB file is changed atomically only | |
424 | * if the generate DB file is ok. */ | |
425 | if (rename(tmpfile,filename) == -1) { | |
426 | redisLog(REDIS_WARNING,"Error moving temp DB file on the final destination: %s", strerror(errno)); | |
427 | unlink(tmpfile); | |
428 | return REDIS_ERR; | |
429 | } | |
430 | redisLog(REDIS_NOTICE,"DB saved on disk"); | |
431 | server.dirty = 0; | |
432 | server.lastsave = time(NULL); | |
433 | return REDIS_OK; | |
434 | ||
435 | werr: | |
436 | fclose(fp); | |
437 | unlink(tmpfile); | |
438 | redisLog(REDIS_WARNING,"Write error saving DB on disk: %s", strerror(errno)); | |
439 | if (di) dictReleaseIterator(di); | |
440 | return REDIS_ERR; | |
441 | } | |
442 | ||
443 | int rdbSaveBackground(char *filename) { | |
444 | pid_t childpid; | |
445 | ||
446 | if (server.bgsavechildpid != -1) return REDIS_ERR; | |
447 | if (server.vm_enabled) waitEmptyIOJobsQueue(); | |
448 | if ((childpid = fork()) == 0) { | |
449 | /* Child */ | |
450 | if (server.vm_enabled) vmReopenSwapFile(); | |
451 | close(server.fd); | |
452 | if (rdbSave(filename) == REDIS_OK) { | |
453 | _exit(0); | |
454 | } else { | |
455 | _exit(1); | |
456 | } | |
457 | } else { | |
458 | /* Parent */ | |
459 | if (childpid == -1) { | |
460 | redisLog(REDIS_WARNING,"Can't save in background: fork: %s", | |
461 | strerror(errno)); | |
462 | return REDIS_ERR; | |
463 | } | |
464 | redisLog(REDIS_NOTICE,"Background saving started by pid %d",childpid); | |
465 | server.bgsavechildpid = childpid; | |
466 | updateDictResizePolicy(); | |
467 | return REDIS_OK; | |
468 | } | |
469 | return REDIS_OK; /* unreached */ | |
470 | } | |
471 | ||
472 | void rdbRemoveTempFile(pid_t childpid) { | |
473 | char tmpfile[256]; | |
474 | ||
475 | snprintf(tmpfile,256,"temp-%d.rdb", (int) childpid); | |
476 | unlink(tmpfile); | |
477 | } | |
478 | ||
479 | int rdbLoadType(FILE *fp) { | |
480 | unsigned char type; | |
481 | if (fread(&type,1,1,fp) == 0) return -1; | |
482 | return type; | |
483 | } | |
484 | ||
485 | time_t rdbLoadTime(FILE *fp) { | |
486 | int32_t t32; | |
487 | if (fread(&t32,4,1,fp) == 0) return -1; | |
488 | return (time_t) t32; | |
489 | } | |
490 | ||
491 | /* Load an encoded length from the DB, see the REDIS_RDB_* defines on the top | |
492 | * of this file for a description of how this are stored on disk. | |
493 | * | |
494 | * isencoded is set to 1 if the readed length is not actually a length but | |
495 | * an "encoding type", check the above comments for more info */ | |
496 | uint32_t rdbLoadLen(FILE *fp, int *isencoded) { | |
497 | unsigned char buf[2]; | |
498 | uint32_t len; | |
499 | int type; | |
500 | ||
501 | if (isencoded) *isencoded = 0; | |
502 | if (fread(buf,1,1,fp) == 0) return REDIS_RDB_LENERR; | |
503 | type = (buf[0]&0xC0)>>6; | |
504 | if (type == REDIS_RDB_6BITLEN) { | |
505 | /* Read a 6 bit len */ | |
506 | return buf[0]&0x3F; | |
507 | } else if (type == REDIS_RDB_ENCVAL) { | |
508 | /* Read a 6 bit len encoding type */ | |
509 | if (isencoded) *isencoded = 1; | |
510 | return buf[0]&0x3F; | |
511 | } else if (type == REDIS_RDB_14BITLEN) { | |
512 | /* Read a 14 bit len */ | |
513 | if (fread(buf+1,1,1,fp) == 0) return REDIS_RDB_LENERR; | |
514 | return ((buf[0]&0x3F)<<8)|buf[1]; | |
515 | } else { | |
516 | /* Read a 32 bit len */ | |
517 | if (fread(&len,4,1,fp) == 0) return REDIS_RDB_LENERR; | |
518 | return ntohl(len); | |
519 | } | |
520 | } | |
521 | ||
522 | /* Load an integer-encoded object from file 'fp', with the specified | |
523 | * encoding type 'enctype'. If encode is true the function may return | |
524 | * an integer-encoded object as reply, otherwise the returned object | |
525 | * will always be encoded as a raw string. */ | |
526 | robj *rdbLoadIntegerObject(FILE *fp, int enctype, int encode) { | |
527 | unsigned char enc[4]; | |
528 | long long val; | |
529 | ||
530 | if (enctype == REDIS_RDB_ENC_INT8) { | |
531 | if (fread(enc,1,1,fp) == 0) return NULL; | |
532 | val = (signed char)enc[0]; | |
533 | } else if (enctype == REDIS_RDB_ENC_INT16) { | |
534 | uint16_t v; | |
535 | if (fread(enc,2,1,fp) == 0) return NULL; | |
536 | v = enc[0]|(enc[1]<<8); | |
537 | val = (int16_t)v; | |
538 | } else if (enctype == REDIS_RDB_ENC_INT32) { | |
539 | uint32_t v; | |
540 | if (fread(enc,4,1,fp) == 0) return NULL; | |
541 | v = enc[0]|(enc[1]<<8)|(enc[2]<<16)|(enc[3]<<24); | |
542 | val = (int32_t)v; | |
543 | } else { | |
544 | val = 0; /* anti-warning */ | |
545 | redisPanic("Unknown RDB integer encoding type"); | |
546 | } | |
547 | if (encode) | |
548 | return createStringObjectFromLongLong(val); | |
549 | else | |
550 | return createObject(REDIS_STRING,sdsfromlonglong(val)); | |
551 | } | |
552 | ||
553 | robj *rdbLoadLzfStringObject(FILE*fp) { | |
554 | unsigned int len, clen; | |
555 | unsigned char *c = NULL; | |
556 | sds val = NULL; | |
557 | ||
558 | if ((clen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL; | |
559 | if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL; | |
560 | if ((c = zmalloc(clen)) == NULL) goto err; | |
561 | if ((val = sdsnewlen(NULL,len)) == NULL) goto err; | |
562 | if (fread(c,clen,1,fp) == 0) goto err; | |
563 | if (lzf_decompress(c,clen,val,len) == 0) goto err; | |
564 | zfree(c); | |
565 | return createObject(REDIS_STRING,val); | |
566 | err: | |
567 | zfree(c); | |
568 | sdsfree(val); | |
569 | return NULL; | |
570 | } | |
571 | ||
572 | robj *rdbGenericLoadStringObject(FILE*fp, int encode) { | |
573 | int isencoded; | |
574 | uint32_t len; | |
575 | sds val; | |
576 | ||
577 | len = rdbLoadLen(fp,&isencoded); | |
578 | if (isencoded) { | |
579 | switch(len) { | |
580 | case REDIS_RDB_ENC_INT8: | |
581 | case REDIS_RDB_ENC_INT16: | |
582 | case REDIS_RDB_ENC_INT32: | |
583 | return rdbLoadIntegerObject(fp,len,encode); | |
584 | case REDIS_RDB_ENC_LZF: | |
585 | return rdbLoadLzfStringObject(fp); | |
586 | default: | |
587 | redisPanic("Unknown RDB encoding type"); | |
588 | } | |
589 | } | |
590 | ||
591 | if (len == REDIS_RDB_LENERR) return NULL; | |
592 | val = sdsnewlen(NULL,len); | |
593 | if (len && fread(val,len,1,fp) == 0) { | |
594 | sdsfree(val); | |
595 | return NULL; | |
596 | } | |
597 | return createObject(REDIS_STRING,val); | |
598 | } | |
599 | ||
600 | robj *rdbLoadStringObject(FILE *fp) { | |
601 | return rdbGenericLoadStringObject(fp,0); | |
602 | } | |
603 | ||
604 | robj *rdbLoadEncodedStringObject(FILE *fp) { | |
605 | return rdbGenericLoadStringObject(fp,1); | |
606 | } | |
607 | ||
608 | /* For information about double serialization check rdbSaveDoubleValue() */ | |
609 | int rdbLoadDoubleValue(FILE *fp, double *val) { | |
610 | char buf[128]; | |
611 | unsigned char len; | |
612 | ||
613 | if (fread(&len,1,1,fp) == 0) return -1; | |
614 | switch(len) { | |
615 | case 255: *val = R_NegInf; return 0; | |
616 | case 254: *val = R_PosInf; return 0; | |
617 | case 253: *val = R_Nan; return 0; | |
618 | default: | |
619 | if (fread(buf,len,1,fp) == 0) return -1; | |
620 | buf[len] = '\0'; | |
621 | sscanf(buf, "%lg", val); | |
622 | return 0; | |
623 | } | |
624 | } | |
625 | ||
626 | /* Load a Redis object of the specified type from the specified file. | |
627 | * On success a newly allocated object is returned, otherwise NULL. */ | |
628 | robj *rdbLoadObject(int type, FILE *fp) { | |
629 | robj *o, *ele, *dec; | |
630 | size_t len; | |
631 | ||
632 | redisLog(REDIS_DEBUG,"LOADING OBJECT %d (at %d)\n",type,ftell(fp)); | |
633 | if (type == REDIS_STRING) { | |
634 | /* Read string value */ | |
635 | if ((o = rdbLoadEncodedStringObject(fp)) == NULL) return NULL; | |
636 | o = tryObjectEncoding(o); | |
637 | } else if (type == REDIS_LIST) { | |
638 | /* Read list value */ | |
639 | if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL; | |
640 | ||
641 | /* Use a real list when there are too many entries */ | |
642 | if (len > server.list_max_ziplist_entries) { | |
643 | o = createListObject(); | |
644 | } else { | |
645 | o = createZiplistObject(); | |
646 | } | |
647 | ||
648 | /* Load every single element of the list */ | |
649 | while(len--) { | |
650 | if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL; | |
651 | ||
652 | /* If we are using a ziplist and the value is too big, convert | |
653 | * the object to a real list. */ | |
654 | if (o->encoding == REDIS_ENCODING_ZIPLIST && | |
655 | ele->encoding == REDIS_ENCODING_RAW && | |
656 | sdslen(ele->ptr) > server.list_max_ziplist_value) | |
657 | listTypeConvert(o,REDIS_ENCODING_LINKEDLIST); | |
658 | ||
659 | if (o->encoding == REDIS_ENCODING_ZIPLIST) { | |
660 | dec = getDecodedObject(ele); | |
661 | o->ptr = ziplistPush(o->ptr,dec->ptr,sdslen(dec->ptr),REDIS_TAIL); | |
662 | decrRefCount(dec); | |
663 | decrRefCount(ele); | |
664 | } else { | |
665 | ele = tryObjectEncoding(ele); | |
666 | listAddNodeTail(o->ptr,ele); | |
667 | } | |
668 | } | |
669 | } else if (type == REDIS_SET) { | |
670 | /* Read list/set value */ | |
671 | if ((len = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL; | |
672 | o = createSetObject(); | |
673 | /* It's faster to expand the dict to the right size asap in order | |
674 | * to avoid rehashing */ | |
675 | if (len > DICT_HT_INITIAL_SIZE) | |
676 | dictExpand(o->ptr,len); | |
677 | /* Load every single element of the list/set */ | |
678 | while(len--) { | |
679 | if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL; | |
680 | ele = tryObjectEncoding(ele); | |
681 | dictAdd((dict*)o->ptr,ele,NULL); | |
682 | } | |
683 | } else if (type == REDIS_ZSET) { | |
684 | /* Read list/set value */ | |
685 | size_t zsetlen; | |
686 | zset *zs; | |
687 | ||
688 | if ((zsetlen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL; | |
689 | o = createZsetObject(); | |
690 | zs = o->ptr; | |
691 | /* Load every single element of the list/set */ | |
692 | while(zsetlen--) { | |
693 | robj *ele; | |
694 | double *score = zmalloc(sizeof(double)); | |
695 | ||
696 | if ((ele = rdbLoadEncodedStringObject(fp)) == NULL) return NULL; | |
697 | ele = tryObjectEncoding(ele); | |
698 | if (rdbLoadDoubleValue(fp,score) == -1) return NULL; | |
699 | dictAdd(zs->dict,ele,score); | |
700 | zslInsert(zs->zsl,*score,ele); | |
701 | incrRefCount(ele); /* added to skiplist */ | |
702 | } | |
703 | } else if (type == REDIS_HASH) { | |
704 | size_t hashlen; | |
705 | ||
706 | if ((hashlen = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) return NULL; | |
707 | o = createHashObject(); | |
708 | /* Too many entries? Use an hash table. */ | |
709 | if (hashlen > server.hash_max_zipmap_entries) | |
710 | convertToRealHash(o); | |
711 | /* Load every key/value, then set it into the zipmap or hash | |
712 | * table, as needed. */ | |
713 | while(hashlen--) { | |
714 | robj *key, *val; | |
715 | ||
716 | if ((key = rdbLoadEncodedStringObject(fp)) == NULL) return NULL; | |
717 | if ((val = rdbLoadEncodedStringObject(fp)) == NULL) return NULL; | |
718 | /* If we are using a zipmap and there are too big values | |
719 | * the object is converted to real hash table encoding. */ | |
720 | if (o->encoding != REDIS_ENCODING_HT && | |
721 | ((key->encoding == REDIS_ENCODING_RAW && | |
722 | sdslen(key->ptr) > server.hash_max_zipmap_value) || | |
723 | (val->encoding == REDIS_ENCODING_RAW && | |
724 | sdslen(val->ptr) > server.hash_max_zipmap_value))) | |
725 | { | |
726 | convertToRealHash(o); | |
727 | } | |
728 | ||
729 | if (o->encoding == REDIS_ENCODING_ZIPMAP) { | |
730 | unsigned char *zm = o->ptr; | |
731 | robj *deckey, *decval; | |
732 | ||
733 | /* We need raw string objects to add them to the zipmap */ | |
734 | deckey = getDecodedObject(key); | |
735 | decval = getDecodedObject(val); | |
736 | zm = zipmapSet(zm,deckey->ptr,sdslen(deckey->ptr), | |
737 | decval->ptr,sdslen(decval->ptr),NULL); | |
738 | o->ptr = zm; | |
739 | decrRefCount(deckey); | |
740 | decrRefCount(decval); | |
741 | decrRefCount(key); | |
742 | decrRefCount(val); | |
743 | } else { | |
744 | key = tryObjectEncoding(key); | |
745 | val = tryObjectEncoding(val); | |
746 | dictAdd((dict*)o->ptr,key,val); | |
747 | } | |
748 | } | |
749 | } else { | |
750 | redisPanic("Unknown object type"); | |
751 | } | |
752 | return o; | |
753 | } | |
754 | ||
755 | int rdbLoad(char *filename) { | |
756 | FILE *fp; | |
757 | uint32_t dbid; | |
758 | int type, retval, rdbver; | |
759 | int swap_all_values = 0; | |
760 | redisDb *db = server.db+0; | |
761 | char buf[1024]; | |
762 | time_t expiretime, now = time(NULL); | |
763 | ||
764 | fp = fopen(filename,"r"); | |
765 | if (!fp) return REDIS_ERR; | |
766 | if (fread(buf,9,1,fp) == 0) goto eoferr; | |
767 | buf[9] = '\0'; | |
768 | if (memcmp(buf,"REDIS",5) != 0) { | |
769 | fclose(fp); | |
770 | redisLog(REDIS_WARNING,"Wrong signature trying to load DB from file"); | |
771 | return REDIS_ERR; | |
772 | } | |
773 | rdbver = atoi(buf+5); | |
774 | if (rdbver != 1) { | |
775 | fclose(fp); | |
776 | redisLog(REDIS_WARNING,"Can't handle RDB format version %d",rdbver); | |
777 | return REDIS_ERR; | |
778 | } | |
779 | while(1) { | |
780 | robj *key, *val; | |
781 | int force_swapout; | |
782 | ||
783 | expiretime = -1; | |
784 | /* Read type. */ | |
785 | if ((type = rdbLoadType(fp)) == -1) goto eoferr; | |
786 | if (type == REDIS_EXPIRETIME) { | |
787 | if ((expiretime = rdbLoadTime(fp)) == -1) goto eoferr; | |
788 | /* We read the time so we need to read the object type again */ | |
789 | if ((type = rdbLoadType(fp)) == -1) goto eoferr; | |
790 | } | |
791 | if (type == REDIS_EOF) break; | |
792 | /* Handle SELECT DB opcode as a special case */ | |
793 | if (type == REDIS_SELECTDB) { | |
794 | if ((dbid = rdbLoadLen(fp,NULL)) == REDIS_RDB_LENERR) | |
795 | goto eoferr; | |
796 | if (dbid >= (unsigned)server.dbnum) { | |
797 | redisLog(REDIS_WARNING,"FATAL: Data file was created with a Redis server configured to handle more than %d databases. Exiting\n", server.dbnum); | |
798 | exit(1); | |
799 | } | |
800 | db = server.db+dbid; | |
801 | continue; | |
802 | } | |
803 | /* Read key */ | |
804 | if ((key = rdbLoadStringObject(fp)) == NULL) goto eoferr; | |
805 | /* Read value */ | |
806 | if ((val = rdbLoadObject(type,fp)) == NULL) goto eoferr; | |
807 | /* Check if the key already expired */ | |
808 | if (expiretime != -1 && expiretime < now) { | |
809 | decrRefCount(key); | |
810 | decrRefCount(val); | |
811 | continue; | |
812 | } | |
813 | /* Add the new object in the hash table */ | |
814 | retval = dbAdd(db,key,val); | |
815 | if (retval == REDIS_ERR) { | |
816 | redisLog(REDIS_WARNING,"Loading DB, duplicated key (%s) found! Unrecoverable error, exiting now.", key->ptr); | |
817 | exit(1); | |
818 | } | |
819 | /* Set the expire time if needed */ | |
820 | if (expiretime != -1) setExpire(db,key,expiretime); | |
821 | ||
822 | /* Handle swapping while loading big datasets when VM is on */ | |
823 | ||
824 | /* If we detecter we are hopeless about fitting something in memory | |
825 | * we just swap every new key on disk. Directly... | |
826 | * Note that's important to check for this condition before resorting | |
827 | * to random sampling, otherwise we may try to swap already | |
828 | * swapped keys. */ | |
829 | if (swap_all_values) { | |
830 | dictEntry *de = dictFind(db->dict,key->ptr); | |
831 | ||
832 | /* de may be NULL since the key already expired */ | |
833 | if (de) { | |
834 | vmpointer *vp; | |
835 | val = dictGetEntryVal(de); | |
836 | ||
837 | if (val->refcount == 1 && | |
838 | (vp = vmSwapObjectBlocking(val)) != NULL) | |
839 | dictGetEntryVal(de) = vp; | |
840 | } | |
841 | decrRefCount(key); | |
842 | continue; | |
843 | } | |
844 | decrRefCount(key); | |
845 | ||
846 | /* Flush data on disk once 32 MB of additional RAM are used... */ | |
847 | force_swapout = 0; | |
848 | if ((zmalloc_used_memory() - server.vm_max_memory) > 1024*1024*32) | |
849 | force_swapout = 1; | |
850 | ||
851 | /* If we have still some hope of having some value fitting memory | |
852 | * then we try random sampling. */ | |
853 | if (!swap_all_values && server.vm_enabled && force_swapout) { | |
854 | while (zmalloc_used_memory() > server.vm_max_memory) { | |
855 | if (vmSwapOneObjectBlocking() == REDIS_ERR) break; | |
856 | } | |
857 | if (zmalloc_used_memory() > server.vm_max_memory) | |
858 | swap_all_values = 1; /* We are already using too much mem */ | |
859 | } | |
860 | } | |
861 | fclose(fp); | |
862 | return REDIS_OK; | |
863 | ||
864 | eoferr: /* unexpected end of file is handled here with a fatal exit */ | |
865 | redisLog(REDIS_WARNING,"Short read or OOM loading DB. Unrecoverable error, aborting now."); | |
866 | exit(1); | |
867 | return REDIS_ERR; /* Just to avoid warning */ | |
868 | } | |
869 | ||
870 | /* A background saving child (BGSAVE) terminated its work. Handle this. */ | |
871 | void backgroundSaveDoneHandler(int statloc) { | |
872 | int exitcode = WEXITSTATUS(statloc); | |
873 | int bysignal = WIFSIGNALED(statloc); | |
874 | ||
875 | if (!bysignal && exitcode == 0) { | |
876 | redisLog(REDIS_NOTICE, | |
877 | "Background saving terminated with success"); | |
878 | server.dirty = 0; | |
879 | server.lastsave = time(NULL); | |
880 | } else if (!bysignal && exitcode != 0) { | |
881 | redisLog(REDIS_WARNING, "Background saving error"); | |
882 | } else { | |
883 | redisLog(REDIS_WARNING, | |
884 | "Background saving terminated by signal %d", WTERMSIG(statloc)); | |
885 | rdbRemoveTempFile(server.bgsavechildpid); | |
886 | } | |
887 | server.bgsavechildpid = -1; | |
888 | /* Possibly there are slaves waiting for a BGSAVE in order to be served | |
889 | * (the first stage of SYNC is a bulk transfer of dump.rdb) */ | |
890 | updateSlavesWaitingBgsave(exitcode == 0 ? REDIS_OK : REDIS_ERR); | |
891 | } |