src/migrate.c

   1 #include "redis.h"
   2 #include "endianconv.h"
   3
   4 #include <sys/stat.h>
   5 #include <lmdb.h>
   6
   7 /* -----------------------------------------------------------------------------
   8  * DUMP, RESTORE and MIGRATE commands
   9  * -------------------------------------------------------------------------- */
  10
  11 MDB_env *env;
  12 MDB_dbi dbi;
  13
  14 /* Generates a DUMP-format representation of the object 'o', adding it to the
  15  * io stream pointed by 'rio'. This function can't fail. */
  16 void createDumpPayload(rio *payload, robj *o) {
  17     unsigned char buf[2];
  18     uint64_t crc;
  19
  20     /* Serialize the object in a RDB-like format. It consist of an object type
  21      * byte followed by the serialized object. This is understood by RESTORE. */
  22     rioInitWithBuffer(payload,sdsempty());
  23     redisAssert(rdbSaveObjectType(payload,o));
  24     redisAssert(rdbSaveObject(payload,o));
  25
  26     /* Write the footer, this is how it looks like:
  27      * ----------------+---------------------+---------------+
  28      * ... RDB payload | 2 bytes RDB version | 8 bytes CRC64 |
  29      * ----------------+---------------------+---------------+
  30      * RDB version and CRC are both in little endian.
  31      */
  32
  33     /* RDB version */
  34     buf[0] = REDIS_RDB_VERSION & 0xff;
  35     buf[1] = (REDIS_RDB_VERSION >> 8) & 0xff;
  36     payload->io.buffer.ptr = sdscatlen(payload->io.buffer.ptr,buf,2);
  37
  38     /* CRC64 */
  39     crc = crc64(0,(unsigned char*)payload->io.buffer.ptr,
  40                 sdslen(payload->io.buffer.ptr));
  41     memrev64ifbe(&crc);
  42     payload->io.buffer.ptr = sdscatlen(payload->io.buffer.ptr,&crc,8);
  43 }
  44
  45 /* Verify that the RDB version of the dump payload matches the one of this Redis
  46  * instance and that the checksum is ok.
  47  * If the DUMP payload looks valid REDIS_OK is returned, otherwise REDIS_ERR
  48  * is returned. */
  49 int verifyDumpPayload(unsigned char *p, size_t len) {
  50     unsigned char *footer;
  51     uint16_t rdbver;
  52     uint64_t crc;
  53
  54     /* At least 2 bytes of RDB version and 8 of CRC64 should be present. */
  55     if (len < 10) return REDIS_ERR;
  56     footer = p+(len-10);
  57
  58     /* Verify RDB version */
  59     rdbver = (footer[1] << 8) | footer[0];
  60     if (rdbver != REDIS_RDB_VERSION) return REDIS_ERR;
  61
  62     /* Verify CRC64 */
  63     crc = crc64(0,p,len-8);
  64     memrev64ifbe(&crc);
  65     return (memcmp(&crc,footer+2,8) == 0) ? REDIS_OK : REDIS_ERR;
  66 }
  67
  68 /* DUMP keyname
  69  * DUMP is actually not used by Redis Cluster but it is the obvious
  70  * complement of RESTORE and can be useful for different applications. */
  71 void dumpCommand(redisClient *c) {
  72     robj *o, *dumpobj;
  73     rio payload;
  74
  75     /* Check if the key is here. */
  76     if ((o = lookupKeyRead(c->db,c->argv[1])) == NULL) {
  77         addReply(c,shared.nullbulk);
  78         return;
  79     }
  80
  81     /* Create the DUMP encoded representation. */
  82     createDumpPayload(&payload,o);
  83
  84     /* Transfer to the client */
  85     dumpobj = createObject(REDIS_STRING,payload.io.buffer.ptr);
  86     addReplyBulk(c,dumpobj);
  87     decrRefCount(dumpobj);
  88     return;
  89 }
  90
  91 /* RESTORE key ttl serialized-value */
  92 void restoreCommand(redisClient *c) {
  93     long ttl;
  94     rio payload;
  95     int type;
  96     robj *obj;
  97
  98     /* Make sure this key does not already exist here... */
  99     if (lookupKeyWrite(c->db,c->argv[1]) != NULL) {
 100         addReplyError(c,"Target key name is busy.");
 101         return;
 102     }
 103
 104     /* Check if the TTL value makes sense */
 105     if (getLongFromObjectOrReply(c,c->argv[2],&ttl,NULL) != REDIS_OK) {
 106         return;
 107     } else if (ttl < 0) {
 108         addReplyError(c,"Invalid TTL value, must be >= 0");
 109         return;
 110     }
 111
 112     /* Verify RDB version and data checksum. */
 113     if (verifyDumpPayload(c->argv[3]->ptr,sdslen(c->argv[3]->ptr)) == REDIS_ERR) {
 114         addReplyError(c,"DUMP payload version or checksum are wrong");
 115         return;
 116     }
 117
 118     rioInitWithBuffer(&payload,c->argv[3]->ptr);
 119     if (((type = rdbLoadObjectType(&payload)) == -1) ||
 120         ((obj = rdbLoadObject(type,&payload)) == NULL))
 121     {
 122         addReplyError(c,"Bad data format");
 123         return;
 124     }
 125
 126     /* Create the key and set the TTL if any */
 127     dbAdd(c->db,c->argv[1],obj);
 128     if (ttl) setExpire(c->db,c->argv[1],mstime()+ttl);
 129     signalModifiedKey(c->db,c->argv[1]);
 130     addReply(c,shared.ok);
 131     server.dirty++;
 132 }
 133
 134 /* MIGRATE host port key dbid timeout */
 135 void migrateCommand(redisClient *c) {
 136     int fd;
 137     long timeout;
 138     long dbid;
 139     long long ttl = 0, expireat;
 140     robj *o;
 141     rio cmd, payload;
 142
 143     /* Sanity check */
 144     if (getLongFromObjectOrReply(c,c->argv[5],&timeout,NULL) != REDIS_OK)
 145         return;
 146     if (getLongFromObjectOrReply(c,c->argv[4],&dbid,NULL) != REDIS_OK)
 147         return;
 148     if (timeout <= 0) timeout = 1000;
 149
 150     /* Check if the key is here. If not we reply with success as there is
 151      * nothing to migrate (for instance the key expired in the meantime), but
 152      * we include such information in the reply string. */
 153     if ((o = lookupKeyRead(c->db,c->argv[3])) == NULL) {
 154         addReplySds(c,sdsnew("+NOKEY\r\n"));
 155         return;
 156     }
 157
 158     /* Connect */
 159     fd = anetTcpNonBlockConnect(server.neterr,c->argv[1]->ptr,
 160                 atoi(c->argv[2]->ptr));
 161     if (fd == -1) {
 162         addReplyErrorFormat(c,"Can't connect to target node: %s",
 163             server.neterr);
 164         return;
 165     }
 166     if ((aeWait(fd,AE_WRITABLE,timeout) & AE_WRITABLE) == 0) {
 167         close(fd);
 168         addReplySds(c,sdsnew("-IOERR error or timeout connecting to the client\r\n"));
 169         return;
 170     }
 171
 172     /* Create RESTORE payload and generate the protocol to call the command. */
 173     rioInitWithBuffer(&cmd,sdsempty());
 174     redisAssertWithInfo(c,NULL,rioWriteBulkCount(&cmd,'*',2));
 175     redisAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,"SELECT",6));
 176     redisAssertWithInfo(c,NULL,rioWriteBulkLongLong(&cmd,dbid));
 177
 178     expireat = getExpire(c->db,c->argv[3]);
 179     if (expireat != -1) {
 180         ttl = expireat-mstime();
 181         if (ttl < 1) ttl = 1;
 182     }
 183     redisAssertWithInfo(c,NULL,rioWriteBulkCount(&cmd,'*',4));
 184     redisAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,"RESTORE",7));
 185     redisAssertWithInfo(c,NULL,c->argv[3]->encoding == REDIS_ENCODING_RAW);
 186     redisAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,c->argv[3]->ptr,sdslen(c->argv[3]->ptr)));
 187     redisAssertWithInfo(c,NULL,rioWriteBulkLongLong(&cmd,ttl));
 188
 189     /* Finally the last argument that is the serailized object payload
 190      * in the DUMP format. */
 191     createDumpPayload(&payload,o);
 192     redisAssertWithInfo(c,NULL,rioWriteBulkString(&cmd,payload.io.buffer.ptr,
 193                                 sdslen(payload.io.buffer.ptr)));
 194     sdsfree(payload.io.buffer.ptr);
 195
 196     /* Tranfer the query to the other node in 64K chunks. */
 197     {
 198         sds buf = cmd.io.buffer.ptr;
 199         size_t pos = 0, towrite;
 200         int nwritten = 0;
 201
 202         while ((towrite = sdslen(buf)-pos) > 0) {
 203             towrite = (towrite > (64*1024) ? (64*1024) : towrite);
 204             nwritten = syncWrite(fd,buf+pos,towrite,timeout);
 205             if (nwritten != (signed)towrite) goto socket_wr_err;
 206             pos += nwritten;
 207         }
 208     }
 209
 210     /* Read back the reply. */
 211     {
 212         char buf1[1024];
 213         char buf2[1024];
 214
 215         /* Read the two replies */
 216         if (syncReadLine(fd, buf1, sizeof(buf1), timeout) <= 0)
 217             goto socket_rd_err;
 218         if (syncReadLine(fd, buf2, sizeof(buf2), timeout) <= 0)
 219             goto socket_rd_err;
 220         if (buf1[0] == '-' || buf2[0] == '-') {
 221             addReplyErrorFormat(c,"Target instance replied with error: %s",
 222                 (buf1[0] == '-') ? buf1+1 : buf2+1);
 223         } else {
 224             robj *aux;
 225
 226             dbDelete(c->db,c->argv[3]);
 227             signalModifiedKey(c->db,c->argv[3]);
 228             addReply(c,shared.ok);
 229             server.dirty++;
 230
 231             /* Translate MIGRATE as DEL for replication/AOF. */
 232             aux = createStringObject("DEL",3);
 233             rewriteClientCommandVector(c,2,aux,c->argv[3]);
 234             decrRefCount(aux);
 235         }
 236     }
 237
 238     sdsfree(cmd.io.buffer.ptr);
 239     close(fd);
 240     return;
 241
 242 socket_wr_err:
 243     addReplySds(c,sdsnew("-IOERR error or timeout writing to target instance\r\n"));
 244     sdsfree(cmd.io.buffer.ptr);
 245     close(fd);
 246     return;
 247
 248 socket_rd_err:
 249     addReplySds(c,sdsnew("-IOERR error or timeout reading from target node\r\n"));
 250     sdsfree(cmd.io.buffer.ptr);
 251     close(fd);
 252     return;
 253 }
 254
 255 void stopKeyArchive(void) {
 256     redisAssert(env != NULL);
 257
 258     mdb_dbi_close(env, dbi);
 259     mdb_env_close(env);
 260     env = NULL;
 261
 262     server.mdb_state = REDIS_MDB_OFF;
 263 }
 264
 265 int startKeyArchive(void) {
 266     redisAssert(env == NULL);
 267
 268     int ret;
 269
 270     ret = mdb_env_create(&env);
 271     if (ret != 0) return ret;
 272
 273     ret = mdb_env_set_mapsize(env, server.mdb_mapsize);
 274     if (ret != 0) return ret;
 275
 276     ret = mdb_env_set_maxdbs(env, 1);
 277     if (ret != 0) return ret;
 278
 279     mkdir(server.mdb_environment, 0644);
 280
 281     ret = mdb_env_open(env, server.mdb_environment, MDB_FIXEDMAP | MDB_NOSYNC, 0664);
 282     if (ret != 0) return ret;
 283
 284     MDB_txn *txn;
 285     ret = mdb_txn_begin(env, NULL, 0, &txn);
 286     if (ret != 0) return ret;
 287
 288     ret = mdb_open(txn, NULL, 0, &dbi);
 289     if (ret != 0) return ret;
 290
 291     mdb_txn_commit(txn);
 292
 293     server.mdb_state = REDIS_MDB_ON;
 294     return 0;
 295 }
 296
 297 int archive(redisDb *db, robj *key) {
 298     if (server.mdb_state == REDIS_MDB_OFF)
 299         return 1;
 300     redisAssert(env != NULL);
 301
 302     MDB_val kval;
 303     kval.mv_data = key->ptr;
 304     kval.mv_size = sdslen((sds)key->ptr);
 305
 306     robj *object;
 307     object = lookupKey(db, key);
 308     if (object == NULL)
 309         return 0;
 310
 311     if (object->archived != 0)
 312         return 1;
 313
 314     rio payload;
 315     createDumpPayload(&payload, object);
 316
 317     MDB_val dval;
 318     dval.mv_size = sdslen(payload.io.buffer.ptr);
 319     dval.mv_data = payload.io.buffer.ptr;
 320
 321     int ret;
 322
 323     MDB_txn *txn;
 324     ret = mdb_txn_begin(env, NULL, 0, &txn);
 325     if (ret != 0)
 326         goto archive_err;
 327
 328     ret = mdb_put(txn, dbi, &kval, &dval, 0);
 329     if (ret != 0) {
 330         mdb_txn_abort(txn);
 331         goto archive_err;
 332     }
 333
 334     mdb_txn_commit(txn);
 335     sdsfree(payload.io.buffer.ptr);
 336     return 1;
 337
 338 archive_err:
 339     sdsfree(payload.io.buffer.ptr);
 340     redisAssert(0);
 341     return 0;
 342 }
 343
 344 robj *recover(redisDb *db, robj *key) {
 345     if (server.mdb_state == REDIS_MDB_OFF)
 346         return NULL;
 347
 348     int ret;
 349
 350     MDB_val kval;
 351     kval.mv_data = key->ptr;
 352     kval.mv_size = sdslen((sds)key->ptr);
 353
 354     MDB_txn *txn;
 355     ret = mdb_txn_begin(env, NULL, 0, &txn);
 356     if (ret != 0)
 357         return NULL;
 358
 359     MDB_cursor *cursor;
 360     ret = mdb_cursor_open(txn, dbi, &cursor);
 361     if (ret != 0) {
 362         mdb_txn_abort(txn);
 363         return NULL;
 364     }
 365
 366     MDB_val pval;
 367     ret = mdb_cursor_get(cursor, &kval, &pval, MDB_SET);
 368     if (ret != 0) {
 369         mdb_txn_abort(txn);
 370         return NULL;
 371     }
 372
 373     sds sval = sdsnewlen(pval.mv_data, pval.mv_size);
 374     mdb_cursor_close(cursor);
 375     mdb_txn_abort(txn);
 376
 377     rio payload;
 378     rioInitWithBuffer(&payload, sval);
 379
 380     int type = rdbLoadObjectType(&payload);
 381     if (type == -1)
 382         goto recover_err;
 383
 384     robj *object = rdbLoadObject(type, &payload);
 385     if (object == NULL)
 386         goto recover_err;
 387
 388     object->archived = 1;
 389
 390     dbAdd(db, key, object);
 391     signalModifiedKey(db, key);
 392     server.dirty++;
 393
 394     sdsfree(sval);
 395     return object;
 396
 397 recover_err:
 398     sdsfree(sval);
 399     return NULL;
 400 }