source3/lib/dbwrap_ctdb.c

   1 /*
   2    Unix SMB/CIFS implementation.
   3    Database interface wrapper around ctdbd
   4    Copyright (C) Volker Lendecke 2007-2009
   5    Copyright (C) Michael Adam 2009
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program.  If not, see <http://www.gnu.org/licenses/>.
  19 */
  20
  21 #include "includes.h"
  22 #ifdef CLUSTER_SUPPORT
  23 #include "ctdb.h"
  24 #include "ctdb_private.h"
  25 #include "ctdbd_conn.h"
  26 #include "g_lock.h"
  27
  28 struct db_ctdb_transaction_handle {
  29         struct db_ctdb_ctx *ctx;
  30         /*
  31          * we store the reads and writes done under a transaction:
  32          * - one list stores both reads and writes (m_all),
  33          * - the other just writes (m_write)
  34          */
  35         struct ctdb_marshall_buffer *m_all;
  36         struct ctdb_marshall_buffer *m_write;
  37         uint32_t nesting;
  38         bool nested_cancel;
  39         char *lock_name;
  40 };
  41
  42 struct db_ctdb_ctx {
  43         struct db_context *db;
  44         struct tdb_wrap *wtdb;
  45         uint32 db_id;
  46         struct db_ctdb_transaction_handle *transaction;
  47         struct g_lock_ctx *lock_ctx;
  48 };
  49
  50 struct db_ctdb_rec {
  51         struct db_ctdb_ctx *ctdb_ctx;
  52         struct ctdb_ltdb_header header;
  53 };
  54
  55 static struct db_record *fetch_locked_internal(struct db_ctdb_ctx *ctx,
  56                                                TALLOC_CTX *mem_ctx,
  57                                                TDB_DATA key,
  58                                                bool persistent);
  59
  60 static NTSTATUS tdb_error_to_ntstatus(struct tdb_context *tdb)
  61 {
  62         NTSTATUS status;
  63         enum TDB_ERROR tret = tdb_error(tdb);
  64
  65         switch (tret) {
  66         case TDB_ERR_EXISTS:
  67                 status = NT_STATUS_OBJECT_NAME_COLLISION;
  68                 break;
  69         case TDB_ERR_NOEXIST:
  70                 status = NT_STATUS_OBJECT_NAME_NOT_FOUND;
  71                 break;
  72         default:
  73                 status = NT_STATUS_INTERNAL_DB_CORRUPTION;
  74                 break;
  75         }
  76
  77         return status;
  78 }
  79
  80
  81 /**
  82  * fetch a record from the tdb, separating out the header
  83  * information and returning the body of the record.
  84  */
  85 static NTSTATUS db_ctdb_ltdb_fetch(struct db_ctdb_ctx *db,
  86                                    TDB_DATA key,
  87                                    struct ctdb_ltdb_header *header,
  88                                    TALLOC_CTX *mem_ctx,
  89                                    TDB_DATA *data)
  90 {
  91         TDB_DATA rec;
  92         NTSTATUS status;
  93
  94         rec = tdb_fetch(db->wtdb->tdb, key);
  95         if (rec.dsize < sizeof(struct ctdb_ltdb_header)) {
  96                 status = NT_STATUS_NOT_FOUND;
  97                 if (data) {
  98                         ZERO_STRUCTP(data);
  99                 }
 100                 if (header) {
 101                         header->dmaster = (uint32_t)-1;
 102                         header->rsn = 0;
 103                 }
 104                 goto done;
 105         }
 106
 107         if (header) {
 108                 *header = *(struct ctdb_ltdb_header *)rec.dptr;
 109         }
 110
 111         if (data) {
 112                 data->dsize = rec.dsize - sizeof(struct ctdb_ltdb_header);
 113                 if (data->dsize == 0) {
 114                         data->dptr = NULL;
 115                 } else {
 116                         data->dptr = (unsigned char *)talloc_memdup(mem_ctx,
 117                                         rec.dptr
 118                                          + sizeof(struct ctdb_ltdb_header),
 119                                         data->dsize);
 120                         if (data->dptr == NULL) {
 121                                 status = NT_STATUS_NO_MEMORY;
 122                                 goto done;
 123                         }
 124                 }
 125         }
 126
 127         status = NT_STATUS_OK;
 128
 129 done:
 130         SAFE_FREE(rec.dptr);
 131         return status;
 132 }
 133
 134 /*
 135  * Store a record together with the ctdb record header
 136  * in the local copy of the database.
 137  */
 138 static NTSTATUS db_ctdb_ltdb_store(struct db_ctdb_ctx *db,
 139                                    TDB_DATA key,
 140                                    struct ctdb_ltdb_header *header,
 141                                    TDB_DATA data)
 142 {
 143         TALLOC_CTX *tmp_ctx = talloc_stackframe();
 144         TDB_DATA rec;
 145         int ret;
 146
 147         rec.dsize = data.dsize + sizeof(struct ctdb_ltdb_header);
 148         rec.dptr = (uint8_t *)talloc_size(tmp_ctx, rec.dsize);
 149
 150         if (rec.dptr == NULL) {
 151                 talloc_free(tmp_ctx);
 152                 return NT_STATUS_NO_MEMORY;
 153         }
 154
 155         memcpy(rec.dptr, header, sizeof(struct ctdb_ltdb_header));
 156         memcpy(sizeof(struct ctdb_ltdb_header) + (uint8_t *)rec.dptr, data.dptr, data.dsize);
 157
 158         ret = tdb_store(db->wtdb->tdb, key, rec, TDB_REPLACE);
 159
 160         talloc_free(tmp_ctx);
 161
 162         return (ret == 0) ? NT_STATUS_OK
 163                           : tdb_error_to_ntstatus(db->wtdb->tdb);
 164
 165 }
 166
 167 /*
 168   form a ctdb_rec_data record from a key/data pair
 169
 170   note that header may be NULL. If not NULL then it is included in the data portion
 171   of the record
 172  */
 173 static struct ctdb_rec_data *db_ctdb_marshall_record(TALLOC_CTX *mem_ctx, uint32_t reqid,
 174                                                   TDB_DATA key,
 175                                                   struct ctdb_ltdb_header *header,
 176                                                   TDB_DATA data)
 177 {
 178         size_t length;
 179         struct ctdb_rec_data *d;
 180
 181         length = offsetof(struct ctdb_rec_data, data) + key.dsize +
 182                 data.dsize + (header?sizeof(*header):0);
 183         d = (struct ctdb_rec_data *)talloc_size(mem_ctx, length);
 184         if (d == NULL) {
 185                 return NULL;
 186         }
 187         d->length = length;
 188         d->reqid = reqid;
 189         d->keylen = key.dsize;
 190         memcpy(&d->data[0], key.dptr, key.dsize);
 191         if (header) {
 192                 d->datalen = data.dsize + sizeof(*header);
 193                 memcpy(&d->data[key.dsize], header, sizeof(*header));
 194                 memcpy(&d->data[key.dsize+sizeof(*header)], data.dptr, data.dsize);
 195         } else {
 196                 d->datalen = data.dsize;
 197                 memcpy(&d->data[key.dsize], data.dptr, data.dsize);
 198         }
 199         return d;
 200 }
 201
 202
 203 /* helper function for marshalling multiple records */
 204 static struct ctdb_marshall_buffer *db_ctdb_marshall_add(TALLOC_CTX *mem_ctx,
 205                                                struct ctdb_marshall_buffer *m,
 206                                                uint64_t db_id,
 207                                                uint32_t reqid,
 208                                                TDB_DATA key,
 209                                                struct ctdb_ltdb_header *header,
 210                                                TDB_DATA data)
 211 {
 212         struct ctdb_rec_data *r;
 213         size_t m_size, r_size;
 214         struct ctdb_marshall_buffer *m2 = NULL;
 215
 216         r = db_ctdb_marshall_record(talloc_tos(), reqid, key, header, data);
 217         if (r == NULL) {
 218                 talloc_free(m);
 219                 return NULL;
 220         }
 221
 222         if (m == NULL) {
 223                 m = (struct ctdb_marshall_buffer *)talloc_zero_size(
 224                         mem_ctx, offsetof(struct ctdb_marshall_buffer, data));
 225                 if (m == NULL) {
 226                         goto done;
 227                 }
 228                 m->db_id = db_id;
 229         }
 230
 231         m_size = talloc_get_size(m);
 232         r_size = talloc_get_size(r);
 233
 234         m2 = (struct ctdb_marshall_buffer *)talloc_realloc_size(
 235                 mem_ctx, m,  m_size + r_size);
 236         if (m2 == NULL) {
 237                 talloc_free(m);
 238                 goto done;
 239         }
 240
 241         memcpy(m_size + (uint8_t *)m2, r, r_size);
 242
 243         m2->count++;
 244
 245 done:
 246         talloc_free(r);
 247         return m2;
 248 }
 249
 250 /* we've finished marshalling, return a data blob with the marshalled records */
 251 static TDB_DATA db_ctdb_marshall_finish(struct ctdb_marshall_buffer *m)
 252 {
 253         TDB_DATA data;
 254         data.dptr = (uint8_t *)m;
 255         data.dsize = talloc_get_size(m);
 256         return data;
 257 }
 258
 259 /*
 260    loop over a marshalling buffer
 261
 262      - pass r==NULL to start
 263      - loop the number of times indicated by m->count
 264 */
 265 static struct ctdb_rec_data *db_ctdb_marshall_loop_next(struct ctdb_marshall_buffer *m, struct ctdb_rec_data *r,
 266                                                      uint32_t *reqid,
 267                                                      struct ctdb_ltdb_header *header,
 268                                                      TDB_DATA *key, TDB_DATA *data)
 269 {
 270         if (r == NULL) {
 271                 r = (struct ctdb_rec_data *)&m->data[0];
 272         } else {
 273                 r = (struct ctdb_rec_data *)(r->length + (uint8_t *)r);
 274         }
 275
 276         if (reqid != NULL) {
 277                 *reqid = r->reqid;
 278         }
 279
 280         if (key != NULL) {
 281                 key->dptr   = &r->data[0];
 282                 key->dsize  = r->keylen;
 283         }
 284         if (data != NULL) {
 285                 data->dptr  = &r->data[r->keylen];
 286                 data->dsize = r->datalen;
 287                 if (header != NULL) {
 288                         data->dptr += sizeof(*header);
 289                         data->dsize -= sizeof(*header);
 290                 }
 291         }
 292
 293         if (header != NULL) {
 294                 if (r->datalen < sizeof(*header)) {
 295                         return NULL;
 296                 }
 297                 *header = *(struct ctdb_ltdb_header *)&r->data[r->keylen];
 298         }
 299
 300         return r;
 301 }
 302
 303 /**
 304  * CTDB transaction destructor
 305  */
 306 static int db_ctdb_transaction_destructor(struct db_ctdb_transaction_handle *h)
 307 {
 308         NTSTATUS status;
 309
 310         status = g_lock_unlock(h->ctx->lock_ctx, h->lock_name);
 311         if (!NT_STATUS_IS_OK(status)) {
 312                 DEBUG(0, ("g_lock_unlock failed: %s\n", nt_errstr(status)));
 313                 return -1;
 314         }
 315         return 0;
 316 }
 317
 318 /**
 319  * CTDB dbwrap API: transaction_start function
 320  * starts a transaction on a persistent database
 321  */
 322 static int db_ctdb_transaction_start(struct db_context *db)
 323 {
 324         struct db_ctdb_transaction_handle *h;
 325         NTSTATUS status;
 326         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
 327                                                         struct db_ctdb_ctx);
 328
 329         if (!db->persistent) {
 330                 DEBUG(0,("transactions not supported on non-persistent database 0x%08x\n",
 331                          ctx->db_id));
 332                 return -1;
 333         }
 334
 335         if (ctx->transaction) {
 336                 ctx->transaction->nesting++;
 337                 return 0;
 338         }
 339
 340         h = talloc_zero(db, struct db_ctdb_transaction_handle);
 341         if (h == NULL) {
 342                 DEBUG(0,(__location__ " oom for transaction handle\n"));
 343                 return -1;
 344         }
 345
 346         h->ctx = ctx;
 347
 348         h->lock_name = talloc_asprintf(h, "transaction_db_0x%08x",
 349                                        (unsigned int)ctx->db_id);
 350         if (h->lock_name == NULL) {
 351                 DEBUG(0, ("talloc_asprintf failed\n"));
 352                 TALLOC_FREE(h);
 353                 return -1;
 354         }
 355
 356         /*
 357          * Wait a day, i.e. forever...
 358          */
 359         status = g_lock_lock(ctx->lock_ctx, h->lock_name, G_LOCK_WRITE,
 360                              timeval_set(86400, 0));
 361         if (!NT_STATUS_IS_OK(status)) {
 362                 DEBUG(0, ("g_lock_lock failed: %s\n", nt_errstr(status)));
 363                 TALLOC_FREE(h);
 364                 return -1;
 365         }
 366
 367         talloc_set_destructor(h, db_ctdb_transaction_destructor);
 368
 369         ctx->transaction = h;
 370
 371         DEBUG(5,(__location__ " Started transaction on db 0x%08x\n", ctx->db_id));
 372
 373         return 0;
 374 }
 375
 376 static bool pull_newest_from_marshall_buffer(struct ctdb_marshall_buffer *buf,
 377                                              TDB_DATA key,
 378                                              struct ctdb_ltdb_header *pheader,
 379                                              TALLOC_CTX *mem_ctx,
 380                                              TDB_DATA *pdata)
 381 {
 382         struct ctdb_rec_data *rec = NULL;
 383         struct ctdb_ltdb_header h;
 384         bool found;
 385         TDB_DATA data;
 386         int i;
 387
 388         if (buf == NULL) {
 389                 return false;
 390         }
 391
 392         /*
 393          * Walk the list of records written during this
 394          * transaction. If we want to read one we have already
 395          * written, return the last written sample. Thus we do not do
 396          * a "break;" for the first hit, this record might have been
 397          * overwritten later.
 398          */
 399
 400         for (i=0; i<buf->count; i++) {
 401                 TDB_DATA tkey, tdata;
 402                 uint32_t reqid;
 403
 404                 rec = db_ctdb_marshall_loop_next(buf, rec, &reqid, &h, &tkey,
 405                                                  &tdata);
 406                 if (rec == NULL) {
 407                         return false;
 408                 }
 409
 410                 if (tdb_data_equal(key, tkey)) {
 411                         found = true;
 412                         data = tdata;
 413                 }
 414         }
 415
 416         if (!found) {
 417                 return false;
 418         }
 419
 420         if (pdata != NULL) {
 421                 data.dptr = (uint8_t *)talloc_memdup(mem_ctx, data.dptr,
 422                                                      data.dsize);
 423                 if ((data.dsize != 0) && (data.dptr == NULL)) {
 424                         return false;
 425                 }
 426                 *pdata = data;
 427         }
 428
 429         if (pheader != NULL) {
 430                 *pheader = h;
 431         }
 432
 433         return true;
 434 }
 435
 436 /*
 437   fetch a record inside a transaction
 438  */
 439 static int db_ctdb_transaction_fetch(struct db_ctdb_ctx *db,
 440                                      TALLOC_CTX *mem_ctx,
 441                                      TDB_DATA key, TDB_DATA *data)
 442 {
 443         struct db_ctdb_transaction_handle *h = db->transaction;
 444         NTSTATUS status;
 445         bool found;
 446
 447         found = pull_newest_from_marshall_buffer(h->m_write, key, NULL,
 448                                                  mem_ctx, data);
 449         if (found) {
 450                 return 0;
 451         }
 452
 453         status = db_ctdb_ltdb_fetch(h->ctx, key, NULL, mem_ctx, data);
 454
 455         if (NT_STATUS_EQUAL(status, NT_STATUS_NOT_FOUND)) {
 456                 *data = tdb_null;
 457         } else if (!NT_STATUS_IS_OK(status)) {
 458                 return -1;
 459         }
 460
 461         h->m_all = db_ctdb_marshall_add(h, h->m_all, h->ctx->db_id, 1, key,
 462                                         NULL, *data);
 463         if (h->m_all == NULL) {
 464                 DEBUG(0,(__location__ " Failed to add to marshalling "
 465                          "record\n"));
 466                 data->dsize = 0;
 467                 talloc_free(data->dptr);
 468                 return -1;
 469         }
 470
 471         return 0;
 472 }
 473
 474
 475 static NTSTATUS db_ctdb_store_transaction(struct db_record *rec, TDB_DATA data, int flag);
 476 static NTSTATUS db_ctdb_delete_transaction(struct db_record *rec);
 477
 478 static struct db_record *db_ctdb_fetch_locked_transaction(struct db_ctdb_ctx *ctx,
 479                                                           TALLOC_CTX *mem_ctx,
 480                                                           TDB_DATA key)
 481 {
 482         struct db_record *result;
 483         TDB_DATA ctdb_data;
 484
 485         if (!(result = talloc(mem_ctx, struct db_record))) {
 486                 DEBUG(0, ("talloc failed\n"));
 487                 return NULL;
 488         }
 489
 490         result->private_data = ctx->transaction;
 491
 492         result->key.dsize = key.dsize;
 493         result->key.dptr = (uint8 *)talloc_memdup(result, key.dptr, key.dsize);
 494         if (result->key.dptr == NULL) {
 495                 DEBUG(0, ("talloc failed\n"));
 496                 TALLOC_FREE(result);
 497                 return NULL;
 498         }
 499
 500         result->store = db_ctdb_store_transaction;
 501         result->delete_rec = db_ctdb_delete_transaction;
 502
 503         if (pull_newest_from_marshall_buffer(ctx->transaction->m_write, key,
 504                                              NULL, result, &result->value)) {
 505                 return result;
 506         }
 507
 508         ctdb_data = tdb_fetch(ctx->wtdb->tdb, key);
 509         if (ctdb_data.dptr == NULL) {
 510                 /* create the record */
 511                 result->value = tdb_null;
 512                 return result;
 513         }
 514
 515         result->value.dsize = ctdb_data.dsize - sizeof(struct ctdb_ltdb_header);
 516         result->value.dptr = NULL;
 517
 518         if ((result->value.dsize != 0)
 519             && !(result->value.dptr = (uint8 *)talloc_memdup(
 520                          result, ctdb_data.dptr + sizeof(struct ctdb_ltdb_header),
 521                          result->value.dsize))) {
 522                 DEBUG(0, ("talloc failed\n"));
 523                 TALLOC_FREE(result);
 524         }
 525
 526         SAFE_FREE(ctdb_data.dptr);
 527
 528         return result;
 529 }
 530
 531 static int db_ctdb_record_destructor(struct db_record **recp)
 532 {
 533         struct db_record *rec = talloc_get_type_abort(*recp, struct db_record);
 534         struct db_ctdb_transaction_handle *h = talloc_get_type_abort(
 535                 rec->private_data, struct db_ctdb_transaction_handle);
 536         int ret = h->ctx->db->transaction_commit(h->ctx->db);
 537         if (ret != 0) {
 538                 DEBUG(0,(__location__ " transaction_commit failed\n"));
 539         }
 540         return 0;
 541 }
 542
 543 /*
 544   auto-create a transaction for persistent databases
 545  */
 546 static struct db_record *db_ctdb_fetch_locked_persistent(struct db_ctdb_ctx *ctx,
 547                                                          TALLOC_CTX *mem_ctx,
 548                                                          TDB_DATA key)
 549 {
 550         int res;
 551         struct db_record *rec, **recp;
 552
 553         res = db_ctdb_transaction_start(ctx->db);
 554         if (res == -1) {
 555                 return NULL;
 556         }
 557
 558         rec = db_ctdb_fetch_locked_transaction(ctx, mem_ctx, key);
 559         if (rec == NULL) {
 560                 ctx->db->transaction_cancel(ctx->db);
 561                 return NULL;
 562         }
 563
 564         /* destroy this transaction when we release the lock */
 565         recp = talloc(rec, struct db_record *);
 566         if (recp == NULL) {
 567                 ctx->db->transaction_cancel(ctx->db);
 568                 talloc_free(rec);
 569                 return NULL;
 570         }
 571         *recp = rec;
 572         talloc_set_destructor(recp, db_ctdb_record_destructor);
 573         return rec;
 574 }
 575
 576
 577 /*
 578   stores a record inside a transaction
 579  */
 580 static int db_ctdb_transaction_store(struct db_ctdb_transaction_handle *h,
 581                                      TDB_DATA key, TDB_DATA data)
 582 {
 583         TALLOC_CTX *tmp_ctx = talloc_new(h);
 584         TDB_DATA rec;
 585         struct ctdb_ltdb_header header;
 586
 587         ZERO_STRUCT(header);
 588
 589         /* we need the header so we can update the RSN */
 590
 591         if (!pull_newest_from_marshall_buffer(h->m_write, key, &header,
 592                                               NULL, NULL)) {
 593
 594                 rec = tdb_fetch(h->ctx->wtdb->tdb, key);
 595
 596                 if (rec.dptr != NULL) {
 597                         memcpy(&header, rec.dptr,
 598                                sizeof(struct ctdb_ltdb_header));
 599                         rec.dsize -= sizeof(struct ctdb_ltdb_header);
 600
 601                         /*
 602                          * a special case, we are writing the same
 603                          * data that is there now
 604                          */
 605                         if (data.dsize == rec.dsize &&
 606                             memcmp(data.dptr,
 607                                    rec.dptr + sizeof(struct ctdb_ltdb_header),
 608                                    data.dsize) == 0) {
 609                                 SAFE_FREE(rec.dptr);
 610                                 talloc_free(tmp_ctx);
 611                                 return 0;
 612                         }
 613                 }
 614                 SAFE_FREE(rec.dptr);
 615         }
 616
 617         header.dmaster = get_my_vnn();
 618         header.rsn++;
 619
 620         h->m_all = db_ctdb_marshall_add(h, h->m_all, h->ctx->db_id, 0, key,
 621                                         NULL, data);
 622         if (h->m_all == NULL) {
 623                 DEBUG(0,(__location__ " Failed to add to marshalling "
 624                          "record\n"));
 625                 talloc_free(tmp_ctx);
 626                 return -1;
 627         }
 628
 629         h->m_write = db_ctdb_marshall_add(h, h->m_write, h->ctx->db_id, 0, key, &header, data);
 630         if (h->m_write == NULL) {
 631                 DEBUG(0,(__location__ " Failed to add to marshalling record\n"));
 632                 talloc_free(tmp_ctx);
 633                 return -1;
 634         }
 635
 636         talloc_free(tmp_ctx);
 637         return 0;
 638 }
 639
 640
 641 /*
 642    a record store inside a transaction
 643  */
 644 static NTSTATUS db_ctdb_store_transaction(struct db_record *rec, TDB_DATA data, int flag)
 645 {
 646         struct db_ctdb_transaction_handle *h = talloc_get_type_abort(
 647                 rec->private_data, struct db_ctdb_transaction_handle);
 648         int ret;
 649
 650         ret = db_ctdb_transaction_store(h, rec->key, data);
 651         if (ret != 0) {
 652                 return tdb_error_to_ntstatus(h->ctx->wtdb->tdb);
 653         }
 654         return NT_STATUS_OK;
 655 }
 656
 657 /*
 658    a record delete inside a transaction
 659  */
 660 static NTSTATUS db_ctdb_delete_transaction(struct db_record *rec)
 661 {
 662         struct db_ctdb_transaction_handle *h = talloc_get_type_abort(
 663                 rec->private_data, struct db_ctdb_transaction_handle);
 664         int ret;
 665
 666         ret = db_ctdb_transaction_store(h, rec->key, tdb_null);
 667         if (ret != 0) {
 668                 return tdb_error_to_ntstatus(h->ctx->wtdb->tdb);
 669         }
 670         return NT_STATUS_OK;
 671 }
 672
 673 /*
 674   commit a transaction
 675  */
 676 static int db_ctdb_transaction_commit(struct db_context *db)
 677 {
 678         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
 679                                                         struct db_ctdb_ctx);
 680         NTSTATUS rets;
 681         int status;
 682         struct db_ctdb_transaction_handle *h = ctx->transaction;
 683
 684         if (h == NULL) {
 685                 DEBUG(0,(__location__ " transaction commit with no open transaction on db 0x%08x\n", ctx->db_id));
 686                 return -1;
 687         }
 688
 689         if (h->nested_cancel) {
 690                 db->transaction_cancel(db);
 691                 DEBUG(5,(__location__ " Failed transaction commit after nested cancel\n"));
 692                 return -1;
 693         }
 694
 695         if (h->nesting != 0) {
 696                 h->nesting--;
 697                 return 0;
 698         }
 699
 700         DEBUG(5,(__location__ " Commit transaction on db 0x%08x\n", ctx->db_id));
 701
 702 again:
 703         if (h->m_write == NULL) {
 704                 /* no changes were made, potentially after a retry */
 705                 goto done;
 706         }
 707
 708         /* tell ctdbd to commit to the other nodes */
 709         rets = ctdbd_control_local(messaging_ctdbd_connection(),
 710                                    CTDB_CONTROL_TRANS3_COMMIT,
 711                                    h->ctx->db_id, 0,
 712                                    db_ctdb_marshall_finish(h->m_write),
 713                                    NULL, NULL, &status);
 714         if (!NT_STATUS_IS_OK(rets) || status != 0) {
 715                 /*
 716                  * TODO:
 717                  * check the database sequence number and
 718                  * compare it to the seqnum after applying the
 719                  * marshall buffer. If it is the same: return success.
 720                  */
 721                 goto again;
 722         }
 723
 724 done:
 725         h->ctx->transaction = NULL;
 726         talloc_free(h);
 727         return 0;
 728 }
 729
 730
 731 /*
 732   cancel a transaction
 733  */
 734 static int db_ctdb_transaction_cancel(struct db_context *db)
 735 {
 736         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
 737                                                         struct db_ctdb_ctx);
 738         struct db_ctdb_transaction_handle *h = ctx->transaction;
 739
 740         if (h == NULL) {
 741                 DEBUG(0,(__location__ " transaction cancel with no open transaction on db 0x%08x\n", ctx->db_id));
 742                 return -1;
 743         }
 744
 745         if (h->nesting != 0) {
 746                 h->nesting--;
 747                 h->nested_cancel = true;
 748                 return 0;
 749         }
 750
 751         DEBUG(5,(__location__ " Cancel transaction on db 0x%08x\n", ctx->db_id));
 752
 753         ctx->transaction = NULL;
 754         talloc_free(h);
 755         return 0;
 756 }
 757
 758
 759 static NTSTATUS db_ctdb_store(struct db_record *rec, TDB_DATA data, int flag)
 760 {
 761         struct db_ctdb_rec *crec = talloc_get_type_abort(
 762                 rec->private_data, struct db_ctdb_rec);
 763
 764         return db_ctdb_ltdb_store(crec->ctdb_ctx, rec->key, &(crec->header), data);
 765 }
 766
 767
 768
 769 static NTSTATUS db_ctdb_delete(struct db_record *rec)
 770 {
 771         TDB_DATA data;
 772
 773         /*
 774          * We have to store the header with empty data. TODO: Fix the
 775          * tdb-level cleanup
 776          */
 777
 778         ZERO_STRUCT(data);
 779
 780         return db_ctdb_store(rec, data, 0);
 781
 782 }
 783
 784 static int db_ctdb_record_destr(struct db_record* data)
 785 {
 786         struct db_ctdb_rec *crec = talloc_get_type_abort(
 787                 data->private_data, struct db_ctdb_rec);
 788
 789         DEBUG(10, (DEBUGLEVEL > 10
 790                    ? "Unlocking db %u key %s\n"
 791                    : "Unlocking db %u key %.20s\n",
 792                    (int)crec->ctdb_ctx->db_id,
 793                    hex_encode_talloc(data, (unsigned char *)data->key.dptr,
 794                               data->key.dsize)));
 795
 796         if (tdb_chainunlock(crec->ctdb_ctx->wtdb->tdb, data->key) != 0) {
 797                 DEBUG(0, ("tdb_chainunlock failed\n"));
 798                 return -1;
 799         }
 800
 801         return 0;
 802 }
 803
 804 static struct db_record *fetch_locked_internal(struct db_ctdb_ctx *ctx,
 805                                                TALLOC_CTX *mem_ctx,
 806                                                TDB_DATA key,
 807                                                bool persistent)
 808 {
 809         struct db_record *result;
 810         struct db_ctdb_rec *crec;
 811         NTSTATUS status;
 812         TDB_DATA ctdb_data;
 813         int migrate_attempts = 0;
 814
 815         if (!(result = talloc(mem_ctx, struct db_record))) {
 816                 DEBUG(0, ("talloc failed\n"));
 817                 return NULL;
 818         }
 819
 820         if (!(crec = TALLOC_ZERO_P(result, struct db_ctdb_rec))) {
 821                 DEBUG(0, ("talloc failed\n"));
 822                 TALLOC_FREE(result);
 823                 return NULL;
 824         }
 825
 826         result->private_data = (void *)crec;
 827         crec->ctdb_ctx = ctx;
 828
 829         result->key.dsize = key.dsize;
 830         result->key.dptr = (uint8 *)talloc_memdup(result, key.dptr, key.dsize);
 831         if (result->key.dptr == NULL) {
 832                 DEBUG(0, ("talloc failed\n"));
 833                 TALLOC_FREE(result);
 834                 return NULL;
 835         }
 836
 837         /*
 838          * Do a blocking lock on the record
 839          */
 840 again:
 841
 842         if (DEBUGLEVEL >= 10) {
 843                 char *keystr = hex_encode_talloc(result, key.dptr, key.dsize);
 844                 DEBUG(10, (DEBUGLEVEL > 10
 845                            ? "Locking db %u key %s\n"
 846                            : "Locking db %u key %.20s\n",
 847                            (int)crec->ctdb_ctx->db_id, keystr));
 848                 TALLOC_FREE(keystr);
 849         }
 850
 851         if (tdb_chainlock(ctx->wtdb->tdb, key) != 0) {
 852                 DEBUG(3, ("tdb_chainlock failed\n"));
 853                 TALLOC_FREE(result);
 854                 return NULL;
 855         }
 856
 857         result->store = db_ctdb_store;
 858         result->delete_rec = db_ctdb_delete;
 859         talloc_set_destructor(result, db_ctdb_record_destr);
 860
 861         ctdb_data = tdb_fetch(ctx->wtdb->tdb, key);
 862
 863         /*
 864          * See if we have a valid record and we are the dmaster. If so, we can
 865          * take the shortcut and just return it.
 866          */
 867
 868         if ((ctdb_data.dptr == NULL) ||
 869             (ctdb_data.dsize < sizeof(struct ctdb_ltdb_header)) ||
 870             ((struct ctdb_ltdb_header *)ctdb_data.dptr)->dmaster != get_my_vnn()
 871 #if 0
 872             || (random() % 2 != 0)
 873 #endif
 874 ) {
 875                 SAFE_FREE(ctdb_data.dptr);
 876                 tdb_chainunlock(ctx->wtdb->tdb, key);
 877                 talloc_set_destructor(result, NULL);
 878
 879                 migrate_attempts += 1;
 880
 881                 DEBUG(10, ("ctdb_data.dptr = %p, dmaster = %u (%u)\n",
 882                            ctdb_data.dptr, ctdb_data.dptr ?
 883                            ((struct ctdb_ltdb_header *)ctdb_data.dptr)->dmaster : -1,
 884                            get_my_vnn()));
 885
 886                 status = ctdbd_migrate(messaging_ctdbd_connection(),ctx->db_id, key);
 887                 if (!NT_STATUS_IS_OK(status)) {
 888                         DEBUG(5, ("ctdb_migrate failed: %s\n",
 889                                   nt_errstr(status)));
 890                         TALLOC_FREE(result);
 891                         return NULL;
 892                 }
 893                 /* now its migrated, try again */
 894                 goto again;
 895         }
 896
 897         if (migrate_attempts > 10) {
 898                 DEBUG(0, ("db_ctdb_fetch_locked needed %d attempts\n",
 899                           migrate_attempts));
 900         }
 901
 902         memcpy(&crec->header, ctdb_data.dptr, sizeof(crec->header));
 903
 904         result->value.dsize = ctdb_data.dsize - sizeof(crec->header);
 905         result->value.dptr = NULL;
 906
 907         if ((result->value.dsize != 0)
 908             && !(result->value.dptr = (uint8 *)talloc_memdup(
 909                          result, ctdb_data.dptr + sizeof(crec->header),
 910                          result->value.dsize))) {
 911                 DEBUG(0, ("talloc failed\n"));
 912                 TALLOC_FREE(result);
 913         }
 914
 915         SAFE_FREE(ctdb_data.dptr);
 916
 917         return result;
 918 }
 919
 920 static struct db_record *db_ctdb_fetch_locked(struct db_context *db,
 921                                               TALLOC_CTX *mem_ctx,
 922                                               TDB_DATA key)
 923 {
 924         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
 925                                                         struct db_ctdb_ctx);
 926
 927         if (ctx->transaction != NULL) {
 928                 return db_ctdb_fetch_locked_transaction(ctx, mem_ctx, key);
 929         }
 930
 931         if (db->persistent) {
 932                 return db_ctdb_fetch_locked_persistent(ctx, mem_ctx, key);
 933         }
 934
 935         return fetch_locked_internal(ctx, mem_ctx, key, db->persistent);
 936 }
 937
 938 /*
 939   fetch (unlocked, no migration) operation on ctdb
 940  */
 941 static int db_ctdb_fetch(struct db_context *db, TALLOC_CTX *mem_ctx,
 942                          TDB_DATA key, TDB_DATA *data)
 943 {
 944         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
 945                                                         struct db_ctdb_ctx);
 946         NTSTATUS status;
 947         TDB_DATA ctdb_data;
 948
 949         if (ctx->transaction) {
 950                 return db_ctdb_transaction_fetch(ctx, mem_ctx, key, data);
 951         }
 952
 953         /* try a direct fetch */
 954         ctdb_data = tdb_fetch(ctx->wtdb->tdb, key);
 955
 956         /*
 957          * See if we have a valid record and we are the dmaster. If so, we can
 958          * take the shortcut and just return it.
 959          * we bypass the dmaster check for persistent databases
 960          */
 961         if ((ctdb_data.dptr != NULL) &&
 962             (ctdb_data.dsize >= sizeof(struct ctdb_ltdb_header)) &&
 963             (db->persistent ||
 964              ((struct ctdb_ltdb_header *)ctdb_data.dptr)->dmaster == get_my_vnn())) {
 965                 /* we are the dmaster - avoid the ctdb protocol op */
 966
 967                 data->dsize = ctdb_data.dsize - sizeof(struct ctdb_ltdb_header);
 968                 if (data->dsize == 0) {
 969                         SAFE_FREE(ctdb_data.dptr);
 970                         data->dptr = NULL;
 971                         return 0;
 972                 }
 973
 974                 data->dptr = (uint8 *)talloc_memdup(
 975                         mem_ctx, ctdb_data.dptr+sizeof(struct ctdb_ltdb_header),
 976                         data->dsize);
 977
 978                 SAFE_FREE(ctdb_data.dptr);
 979
 980                 if (data->dptr == NULL) {
 981                         return -1;
 982                 }
 983                 return 0;
 984         }
 985
 986         SAFE_FREE(ctdb_data.dptr);
 987
 988         /* we weren't able to get it locally - ask ctdb to fetch it for us */
 989         status = ctdbd_fetch(messaging_ctdbd_connection(),ctx->db_id, key, mem_ctx, data);
 990         if (!NT_STATUS_IS_OK(status)) {
 991                 DEBUG(5, ("ctdbd_fetch failed: %s\n", nt_errstr(status)));
 992                 return -1;
 993         }
 994
 995         return 0;
 996 }
 997
 998 struct traverse_state {
 999         struct db_context *db;
1000         int (*fn)(struct db_record *rec, void *private_data);
1001         void *private_data;
1002 };
1003
1004 static void traverse_callback(TDB_DATA key, TDB_DATA data, void *private_data)
1005 {
1006         struct traverse_state *state = (struct traverse_state *)private_data;
1007         struct db_record *rec;
1008         TALLOC_CTX *tmp_ctx = talloc_new(state->db);
1009         /* we have to give them a locked record to prevent races */
1010         rec = db_ctdb_fetch_locked(state->db, tmp_ctx, key);
1011         if (rec && rec->value.dsize > 0) {
1012                 state->fn(rec, state->private_data);
1013         }
1014         talloc_free(tmp_ctx);
1015 }
1016
1017 static int traverse_persistent_callback(TDB_CONTEXT *tdb, TDB_DATA kbuf, TDB_DATA dbuf,
1018                                         void *private_data)
1019 {
1020         struct traverse_state *state = (struct traverse_state *)private_data;
1021         struct db_record *rec;
1022         TALLOC_CTX *tmp_ctx = talloc_new(state->db);
1023         int ret = 0;
1024         /* we have to give them a locked record to prevent races */
1025         rec = db_ctdb_fetch_locked(state->db, tmp_ctx, kbuf);
1026         if (rec && rec->value.dsize > 0) {
1027                 ret = state->fn(rec, state->private_data);
1028         }
1029         talloc_free(tmp_ctx);
1030         return ret;
1031 }
1032
1033 static int db_ctdb_traverse(struct db_context *db,
1034                             int (*fn)(struct db_record *rec,
1035                                       void *private_data),
1036                             void *private_data)
1037 {
1038         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
1039                                                         struct db_ctdb_ctx);
1040         struct traverse_state state;
1041
1042         state.db = db;
1043         state.fn = fn;
1044         state.private_data = private_data;
1045
1046         if (db->persistent) {
1047                 /* for persistent databases we don't need to do a ctdb traverse,
1048                    we can do a faster local traverse */
1049                 return tdb_traverse(ctx->wtdb->tdb, traverse_persistent_callback, &state);
1050         }
1051
1052
1053         ctdbd_traverse(ctx->db_id, traverse_callback, &state);
1054         return 0;
1055 }
1056
1057 static NTSTATUS db_ctdb_store_deny(struct db_record *rec, TDB_DATA data, int flag)
1058 {
1059         return NT_STATUS_MEDIA_WRITE_PROTECTED;
1060 }
1061
1062 static NTSTATUS db_ctdb_delete_deny(struct db_record *rec)
1063 {
1064         return NT_STATUS_MEDIA_WRITE_PROTECTED;
1065 }
1066
1067 static void traverse_read_callback(TDB_DATA key, TDB_DATA data, void *private_data)
1068 {
1069         struct traverse_state *state = (struct traverse_state *)private_data;
1070         struct db_record rec;
1071         rec.key = key;
1072         rec.value = data;
1073         rec.store = db_ctdb_store_deny;
1074         rec.delete_rec = db_ctdb_delete_deny;
1075         rec.private_data = state->db;
1076         state->fn(&rec, state->private_data);
1077 }
1078
1079 static int traverse_persistent_callback_read(TDB_CONTEXT *tdb, TDB_DATA kbuf, TDB_DATA dbuf,
1080                                         void *private_data)
1081 {
1082         struct traverse_state *state = (struct traverse_state *)private_data;
1083         struct db_record rec;
1084         rec.key = kbuf;
1085         rec.value = dbuf;
1086         rec.store = db_ctdb_store_deny;
1087         rec.delete_rec = db_ctdb_delete_deny;
1088         rec.private_data = state->db;
1089
1090         if (rec.value.dsize <= sizeof(struct ctdb_ltdb_header)) {
1091                 /* a deleted record */
1092                 return 0;
1093         }
1094         rec.value.dsize -= sizeof(struct ctdb_ltdb_header);
1095         rec.value.dptr += sizeof(struct ctdb_ltdb_header);
1096
1097         return state->fn(&rec, state->private_data);
1098 }
1099
1100 static int db_ctdb_traverse_read(struct db_context *db,
1101                                  int (*fn)(struct db_record *rec,
1102                                            void *private_data),
1103                                  void *private_data)
1104 {
1105         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
1106                                                         struct db_ctdb_ctx);
1107         struct traverse_state state;
1108
1109         state.db = db;
1110         state.fn = fn;
1111         state.private_data = private_data;
1112
1113         if (db->persistent) {
1114                 /* for persistent databases we don't need to do a ctdb traverse,
1115                    we can do a faster local traverse */
1116                 return tdb_traverse_read(ctx->wtdb->tdb, traverse_persistent_callback_read, &state);
1117         }
1118
1119         ctdbd_traverse(ctx->db_id, traverse_read_callback, &state);
1120         return 0;
1121 }
1122
1123 static int db_ctdb_get_seqnum(struct db_context *db)
1124 {
1125         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
1126                                                         struct db_ctdb_ctx);
1127         return tdb_get_seqnum(ctx->wtdb->tdb);
1128 }
1129
1130 static int db_ctdb_get_flags(struct db_context *db)
1131 {
1132         struct db_ctdb_ctx *ctx = talloc_get_type_abort(db->private_data,
1133                                                         struct db_ctdb_ctx);
1134         return tdb_get_flags(ctx->wtdb->tdb);
1135 }
1136
1137 struct db_context *db_open_ctdb(TALLOC_CTX *mem_ctx,
1138                                 const char *name,
1139                                 int hash_size, int tdb_flags,
1140                                 int open_flags, mode_t mode)
1141 {
1142         struct db_context *result;
1143         struct db_ctdb_ctx *db_ctdb;
1144         char *db_path;
1145         struct ctdbd_connection *conn;
1146
1147         if (!lp_clustering()) {
1148                 DEBUG(10, ("Clustering disabled -- no ctdb\n"));
1149                 return NULL;
1150         }
1151
1152         if (!(result = TALLOC_ZERO_P(mem_ctx, struct db_context))) {
1153                 DEBUG(0, ("talloc failed\n"));
1154                 TALLOC_FREE(result);
1155                 return NULL;
1156         }
1157
1158         if (!(db_ctdb = TALLOC_P(result, struct db_ctdb_ctx))) {
1159                 DEBUG(0, ("talloc failed\n"));
1160                 TALLOC_FREE(result);
1161                 return NULL;
1162         }
1163
1164         db_ctdb->transaction = NULL;
1165         db_ctdb->db = result;
1166
1167         conn = messaging_ctdbd_connection();
1168
1169         if (!NT_STATUS_IS_OK(ctdbd_db_attach(conn, name, &db_ctdb->db_id, tdb_flags))) {
1170                 DEBUG(0, ("ctdbd_db_attach failed for %s\n", name));
1171                 TALLOC_FREE(result);
1172                 return NULL;
1173         }
1174
1175         db_path = ctdbd_dbpath(conn, db_ctdb, db_ctdb->db_id);
1176
1177         result->persistent = ((tdb_flags & TDB_CLEAR_IF_FIRST) == 0);
1178
1179         /* only pass through specific flags */
1180         tdb_flags &= TDB_SEQNUM;
1181
1182         /* honor permissions if user has specified O_CREAT */
1183         if (open_flags & O_CREAT) {
1184                 chmod(db_path, mode);
1185         }
1186
1187         db_ctdb->wtdb = tdb_wrap_open(db_ctdb, db_path, hash_size, tdb_flags, O_RDWR, 0);
1188         if (db_ctdb->wtdb == NULL) {
1189                 DEBUG(0, ("Could not open tdb %s: %s\n", db_path, strerror(errno)));
1190                 TALLOC_FREE(result);
1191                 return NULL;
1192         }
1193         talloc_free(db_path);
1194
1195         if (result->persistent) {
1196                 db_ctdb->lock_ctx = g_lock_ctx_init(db_ctdb,
1197                                                     ctdb_conn_msg_ctx(conn));
1198                 if (db_ctdb->lock_ctx == NULL) {
1199                         DEBUG(0, ("g_lock_ctx_init failed\n"));
1200                         TALLOC_FREE(result);
1201                         return NULL;
1202                 }
1203         }
1204
1205         result->private_data = (void *)db_ctdb;
1206         result->fetch_locked = db_ctdb_fetch_locked;
1207         result->fetch = db_ctdb_fetch;
1208         result->traverse = db_ctdb_traverse;
1209         result->traverse_read = db_ctdb_traverse_read;
1210         result->get_seqnum = db_ctdb_get_seqnum;
1211         result->get_flags = db_ctdb_get_flags;
1212         result->transaction_start = db_ctdb_transaction_start;
1213         result->transaction_commit = db_ctdb_transaction_commit;
1214         result->transaction_cancel = db_ctdb_transaction_cancel;
1215
1216         DEBUG(3,("db_open_ctdb: opened database '%s' with dbid 0x%x\n",
1217                  name, db_ctdb->db_id));
1218
1219         return result;
1220 }
1221 #endif