ctdb/server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5    Copyright (C) Michael Adam 2010-2013
   6    Copyright (C) Stefan Metzmacher 2010-2011
   7
   8    This program is free software; you can redistribute it and/or modify
   9    it under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3 of the License, or
  11    (at your option) any later version.
  12
  13    This program is distributed in the hope that it will be useful,
  14    but WITHOUT ANY WARRANTY; without even the implied warranty of
  15    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16    GNU General Public License for more details.
  17
  18    You should have received a copy of the GNU General Public License
  19    along with this program; if not, see <http://www.gnu.org/licenses/>.
  20 */
  21
  22 #include "includes.h"
  23 #include "tdb.h"
  24 #include "system/network.h"
  25 #include "system/filesys.h"
  26 #include "system/dir.h"
  27 #include "../include/ctdb_private.h"
  28 #include "db_wrap.h"
  29 #include "lib/util/dlinklist.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34
  35 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  36
  37 struct ctdb_vacuum_child_context {
  38         struct ctdb_vacuum_child_context *next, *prev;
  39         struct ctdb_vacuum_handle *vacuum_handle;
  40         /* fd child writes status to */
  41         int fd[2];
  42         pid_t child_pid;
  43         enum vacuum_child_status status;
  44         struct timeval start_time;
  45 };
  46
  47 struct ctdb_vacuum_handle {
  48         struct ctdb_db_context *ctdb_db;
  49         struct ctdb_vacuum_child_context *child_ctx;
  50         uint32_t fast_path_count;
  51 };
  52
  53
  54 /*  a list of records to possibly delete */
  55 struct vacuum_data {
  56         uint32_t repack_limit;
  57         struct ctdb_context *ctdb;
  58         struct ctdb_db_context *ctdb_db;
  59         struct tdb_context *dest_db;
  60         trbt_tree_t *delete_list;
  61         uint32_t delete_count;
  62         struct ctdb_marshall_buffer **vacuum_fetch_list;
  63         struct timeval start;
  64         bool traverse_error;
  65         bool vacuum;
  66         uint32_t total;
  67         uint32_t vacuumed;
  68         uint32_t copied;
  69         uint32_t fast_added_to_vacuum_fetch_list;
  70         uint32_t fast_added_to_delete_list;
  71         uint32_t fast_deleted;
  72         uint32_t fast_skipped;
  73         uint32_t fast_error;
  74         uint32_t fast_total;
  75         uint32_t full_scheduled;
  76         uint32_t full_skipped;
  77         uint32_t full_error;
  78         uint32_t full_total;
  79         uint32_t delete_left;
  80         uint32_t delete_remote_error;
  81         uint32_t delete_local_error;
  82         uint32_t delete_deleted;
  83         uint32_t delete_skipped;
  84 };
  85
  86 /* this structure contains the information for one record to be deleted */
  87 struct delete_record_data {
  88         struct ctdb_context *ctdb;
  89         struct ctdb_db_context *ctdb_db;
  90         struct ctdb_ltdb_header hdr;
  91         TDB_DATA key;
  92         uint8_t keydata[1];
  93 };
  94
  95 struct delete_records_list {
  96         struct ctdb_marshall_buffer *records;
  97         struct vacuum_data *vdata;
  98 };
  99
 100 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 101                                            const struct ctdb_ltdb_header *hdr,
 102                                            TDB_DATA key);
 103
 104 /**
 105  * Store key and header in a tree, indexed by the key hash.
 106  */
 107 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 108                                                struct ctdb_db_context *ctdb_db,
 109                                                trbt_tree_t *tree,
 110                                                const struct ctdb_ltdb_header *hdr,
 111                                                TDB_DATA key)
 112 {
 113         struct delete_record_data *dd;
 114         uint32_t hash;
 115         size_t len;
 116
 117         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 118
 119         dd = (struct delete_record_data *)talloc_size(tree, len);
 120         if (dd == NULL) {
 121                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 122                 return -1;
 123         }
 124         talloc_set_name_const(dd, "struct delete_record_data");
 125
 126         dd->ctdb      = ctdb;
 127         dd->ctdb_db   = ctdb_db;
 128         dd->key.dsize = key.dsize;
 129         dd->key.dptr  = dd->keydata;
 130         memcpy(dd->keydata, key.dptr, key.dsize);
 131
 132         dd->hdr = *hdr;
 133
 134         hash = ctdb_hash(&key);
 135
 136         trbt_insert32(tree, hash, dd);
 137
 138         return 0;
 139 }
 140
 141 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 142                                      struct ctdb_ltdb_header *hdr)
 143 {
 144         struct ctdb_context *ctdb = vdata->ctdb;
 145         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 146         uint32_t hash;
 147         int ret;
 148
 149         hash = ctdb_hash(&key);
 150
 151         if (trbt_lookup32(vdata->delete_list, hash)) {
 152                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 153                 return 0;
 154         }
 155
 156         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 157                                                   vdata->delete_list,
 158                                                   hdr, key);
 159         if (ret != 0) {
 160                 return -1;
 161         }
 162
 163         vdata->delete_count++;
 164
 165         return 0;
 166 }
 167
 168 /**
 169  * Add a record to the list of records to be sent
 170  * to their lmaster with VACUUM_FETCH.
 171  */
 172 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 173                                            TDB_DATA key)
 174 {
 175         struct ctdb_context *ctdb = vdata->ctdb;
 176         struct ctdb_rec_data *rec;
 177         uint32_t lmaster;
 178         size_t old_size;
 179         struct ctdb_marshall_buffer *vfl;
 180
 181         lmaster = ctdb_lmaster(ctdb, &key);
 182
 183         vfl = vdata->vacuum_fetch_list[lmaster];
 184
 185         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 186         if (rec == NULL) {
 187                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 188                 vdata->traverse_error = true;
 189                 return -1;
 190         }
 191
 192         old_size = talloc_get_size(vfl);
 193         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 194         if (vfl == NULL) {
 195                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 196                 vdata->traverse_error = true;
 197                 return -1;
 198         }
 199         vdata->vacuum_fetch_list[lmaster] = vfl;
 200
 201         vfl->count++;
 202         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 203         talloc_free(rec);
 204
 205         vdata->total++;
 206
 207         return 0;
 208 }
 209
 210
 211 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 212                               struct timeval t, void *private_data);
 213
 214 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 215 {
 216         struct ctdb_ltdb_header *header =
 217                 (struct ctdb_ltdb_header *)private_data;
 218
 219         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 220                 return -1;
 221         }
 222
 223         *header = *(struct ctdb_ltdb_header *)data.dptr;
 224
 225         return 0;
 226 }
 227
 228 /*
 229  * traverse function for gathering the records that can be deleted
 230  */
 231 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 232                            void *private_data)
 233 {
 234         struct vacuum_data *vdata = talloc_get_type(private_data,
 235                                                     struct vacuum_data);
 236         struct ctdb_context *ctdb = vdata->ctdb;
 237         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 238         uint32_t lmaster;
 239         struct ctdb_ltdb_header *hdr;
 240         int res = 0;
 241
 242         vdata->full_total++;
 243
 244         lmaster = ctdb_lmaster(ctdb, &key);
 245         if (lmaster >= ctdb->num_nodes) {
 246                 vdata->full_error++;
 247                 DEBUG(DEBUG_CRIT, (__location__
 248                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 249                                    " with hash[%u]!\n",
 250                                    (unsigned)lmaster,
 251                                    (unsigned)ctdb->num_nodes,
 252                                    (unsigned)ctdb_hash(&key)));
 253                 return -1;
 254         }
 255
 256         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 257                 /* it is not a deleted record */
 258                 vdata->full_skipped++;
 259                 return 0;
 260         }
 261
 262         hdr = (struct ctdb_ltdb_header *)data.dptr;
 263
 264         if (hdr->dmaster != ctdb->pnn) {
 265                 vdata->full_skipped++;
 266                 return 0;
 267         }
 268
 269         /*
 270          * Add the record to this process's delete_queue for processing
 271          * in the subsequent traverse in the fast vacuum run.
 272          */
 273         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 274         if (res != 0) {
 275                 vdata->full_error++;
 276         } else {
 277                 vdata->full_scheduled++;
 278         }
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * traverse the tree of records to delete and marshall them into
 285  * a blob
 286  */
 287 static int delete_marshall_traverse(void *param, void *data)
 288 {
 289         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 290         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 291         struct ctdb_rec_data *rec;
 292         size_t old_size;
 293
 294         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 295         if (rec == NULL) {
 296                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 297                 return 0;
 298         }
 299
 300         old_size = talloc_get_size(recs->records);
 301         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 302         if (recs->records == NULL) {
 303                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 304                 return 0;
 305         }
 306         recs->records->count++;
 307         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 308         return 0;
 309 }
 310
 311 /**
 312  * Variant of delete_marshall_traverse() that bumps the
 313  * RSN of each traversed record in the database.
 314  *
 315  * This is needed to ensure that when rolling out our
 316  * empty record copy before remote deletion, we as the
 317  * record's dmaster keep a higher RSN than the non-dmaster
 318  * nodes. This is needed to prevent old copies from
 319  * resurrection in recoveries.
 320  */
 321 static int delete_marshall_traverse_first(void *param, void *data)
 322 {
 323         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 324         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 325         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 326         struct ctdb_context *ctdb = ctdb_db->ctdb;
 327         struct ctdb_ltdb_header header;
 328         uint32_t lmaster;
 329         uint32_t hash = ctdb_hash(&(dd->key));
 330         int res;
 331
 332         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 333         if (res != 0) {
 334                 DEBUG(DEBUG_ERR,
 335                       (__location__ " Error getting chainlock on record with "
 336                        "key hash [0x%08x] on database db[%s].\n",
 337                        hash, ctdb_db->db_name));
 338                 recs->vdata->delete_skipped++;
 339                 talloc_free(dd);
 340                 return 0;
 341         }
 342
 343         /*
 344          * Verify that the record is still empty, its RSN has not
 345          * changed and that we are still its lmaster and dmaster.
 346          */
 347
 348         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 349                                vacuum_record_parser, &header);
 350         if (res != 0) {
 351                 goto skip;
 352         }
 353
 354         if (header.flags & CTDB_REC_RO_FLAGS) {
 355                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 356                                    "on database db[%s] has read-only flags. "
 357                                    "skipping.\n",
 358                                    hash, ctdb_db->db_name));
 359                 goto skip;
 360         }
 361
 362         if (header.dmaster != ctdb->pnn) {
 363                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 364                                    "on database db[%s] has been migrated away. "
 365                                    "skipping.\n",
 366                                    hash, ctdb_db->db_name));
 367                 goto skip;
 368         }
 369
 370         if (header.rsn != dd->hdr.rsn) {
 371                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 372                                    "on database db[%s] seems to have been "
 373                                    "migrated away and back again (with empty "
 374                                    "data). skipping.\n",
 375                                    hash, ctdb_db->db_name));
 376                 goto skip;
 377         }
 378
 379         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 380
 381         if (lmaster != ctdb->pnn) {
 382                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 383                                    "delete list (key hash [0x%08x], db[%s]). "
 384                                    "Strange! skipping.\n",
 385                                    hash, ctdb_db->db_name));
 386                 goto skip;
 387         }
 388
 389         /*
 390          * Increment the record's RSN to ensure the dmaster (i.e. the current
 391          * node) has the highest RSN of the record in the cluster.
 392          * This is to prevent old record copies from resurrecting in recoveries
 393          * if something should fail during the deletion process.
 394          * Note that ctdb_ltdb_store_server() increments the RSN if called
 395          * on the record's dmaster.
 396          */
 397
 398         res = ctdb_ltdb_store(ctdb_db, dd->key, &header, tdb_null);
 399         if (res != 0) {
 400                 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
 401                                   "key hash [0x%08x] on database db[%s].\n",
 402                                   hash, ctdb_db->db_name));
 403                 goto skip;
 404         }
 405
 406         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 407
 408         goto done;
 409
 410 skip:
 411         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 412
 413         recs->vdata->delete_skipped++;
 414         talloc_free(dd);
 415         dd = NULL;
 416
 417 done:
 418         if (dd == NULL) {
 419                 return 0;
 420         }
 421
 422         return delete_marshall_traverse(param, data);
 423 }
 424
 425 /**
 426  * traverse function for the traversal of the delete_queue,
 427  * the fast-path vacuuming list.
 428  *
 429  *  - If the record has been migrated off the node
 430  *    or has been revived (filled with data) on the node,
 431  *    then skip the record.
 432  *
 433  *  - If the current node is the record's lmaster and it is
 434  *    a record that has never been migrated with data, then
 435  *    delete the record from the local tdb.
 436  *
 437  *  - If the current node is the record's lmaster and it has
 438  *    been migrated with data, then schedule it for the normal
 439  *    vacuuming procedure (i.e. add it to the delete_list).
 440  *
 441  *  - If the current node is NOT the record's lmaster then
 442  *    add it to the list of records that are to be sent to
 443  *    the lmaster with the VACUUM_FETCH message.
 444  */
 445 static int delete_queue_traverse(void *param, void *data)
 446 {
 447         struct delete_record_data *dd =
 448                 talloc_get_type(data, struct delete_record_data);
 449         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 450         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 451         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 452         int res;
 453         struct ctdb_ltdb_header header;
 454         uint32_t lmaster;
 455         uint32_t hash = ctdb_hash(&(dd->key));
 456
 457         vdata->fast_total++;
 458
 459         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 460         if (res != 0) {
 461                 DEBUG(DEBUG_ERR,
 462                       (__location__ " Error getting chainlock on record with "
 463                        "key hash [0x%08x] on database db[%s].\n",
 464                        hash, ctdb_db->db_name));
 465                 vdata->fast_error++;
 466                 return 0;
 467         }
 468
 469         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 470                                vacuum_record_parser, &header);
 471         if (res != 0) {
 472                 goto skipped;
 473         }
 474
 475         if (header.dmaster != ctdb->pnn) {
 476                 /* The record has been migrated off the node. Skip. */
 477                 goto skipped;
 478         }
 479
 480         if (header.rsn != dd->hdr.rsn) {
 481                 /*
 482                  * The record has been migrated off the node and back again.
 483                  * But not requeued for deletion. Skip it.
 484                  */
 485                 goto skipped;
 486         }
 487
 488         /*
 489          * We are dmaster, and the record has no data, and it has
 490          * not been migrated after it has been queued for deletion.
 491          *
 492          * At this stage, the record could still have been revived locally
 493          * and last been written with empty data. This can only be
 494          * fixed with the addition of an active or delete flag. (TODO)
 495          */
 496
 497         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 498
 499         if (lmaster != ctdb->pnn) {
 500                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 501
 502                 if (res != 0) {
 503                         DEBUG(DEBUG_ERR,
 504                               (__location__ " Error adding record to list "
 505                                "of records to send to lmaster.\n"));
 506                         vdata->fast_error++;
 507                 } else {
 508                         vdata->fast_added_to_vacuum_fetch_list++;
 509                 }
 510                 goto done;
 511         }
 512
 513         /* use header->flags or dd->hdr.flags ?? */
 514         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 515                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 516
 517                 if (res != 0) {
 518                         DEBUG(DEBUG_ERR,
 519                               (__location__ " Error adding record to list "
 520                                "of records for deletion on lmaster.\n"));
 521                         vdata->fast_error++;
 522                 } else {
 523                         vdata->fast_added_to_delete_list++;
 524                 }
 525         } else {
 526                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 527
 528                 if (res != 0) {
 529                         DEBUG(DEBUG_ERR,
 530                               (__location__ " Error deleting record with key "
 531                                "hash [0x%08x] from local data base db[%s].\n",
 532                                hash, ctdb_db->db_name));
 533                         vdata->fast_error++;
 534                         goto done;
 535                 }
 536
 537                 DEBUG(DEBUG_DEBUG,
 538                       (__location__ " Deleted record with key hash "
 539                        "[0x%08x] from local data base db[%s].\n",
 540                        hash, ctdb_db->db_name));
 541                 vdata->fast_deleted++;
 542         }
 543
 544         goto done;
 545
 546 skipped:
 547         vdata->fast_skipped++;
 548
 549 done:
 550         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 551
 552         return 0;
 553 }
 554
 555 /**
 556  * Delete the records that we are lmaster and dmaster for and
 557  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 558  * control.
 559  */
 560 static int delete_record_traverse(void *param, void *data)
 561 {
 562         struct delete_record_data *dd =
 563                 talloc_get_type(data, struct delete_record_data);
 564         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 565         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 566         struct ctdb_context *ctdb = ctdb_db->ctdb;
 567         int res;
 568         struct ctdb_ltdb_header header;
 569         uint32_t lmaster;
 570         uint32_t hash = ctdb_hash(&(dd->key));
 571
 572         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 573         if (res != 0) {
 574                 DEBUG(DEBUG_ERR,
 575                       (__location__ " Error getting chainlock on record with "
 576                        "key hash [0x%08x] on database db[%s].\n",
 577                        hash, ctdb_db->db_name));
 578                 vdata->delete_local_error++;
 579                 vdata->delete_left--;
 580                 talloc_free(dd);
 581                 return 0;
 582         }
 583
 584         /*
 585          * Verify that the record is still empty, its RSN has not
 586          * changed and that we are still its lmaster and dmaster.
 587          */
 588
 589         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 590                                vacuum_record_parser, &header);
 591         if (res != 0) {
 592                 goto skip;
 593         }
 594
 595         if (header.flags & CTDB_REC_RO_FLAGS) {
 596                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 597                                    "on database db[%s] has read-only flags. "
 598                                    "skipping.\n",
 599                                    hash, ctdb_db->db_name));
 600                 goto skip;
 601         }
 602
 603         if (header.dmaster != ctdb->pnn) {
 604                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 605                                    "on database db[%s] has been migrated away. "
 606                                    "skipping.\n",
 607                                    hash, ctdb_db->db_name));
 608                 goto skip;
 609         }
 610
 611         if (header.rsn != dd->hdr.rsn + 1) {
 612                 /*
 613                  * The record has been migrated off the node and back again.
 614                  * But not requeued for deletion. Skip it.
 615                  * (Note that the first marshall traverse has bumped the RSN
 616                  *  on disk.)
 617                  */
 618                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 619                                    "on database db[%s] seems to have been "
 620                                    "migrated away and back again (with empty "
 621                                    "data). skipping.\n",
 622                                    hash, ctdb_db->db_name));
 623                 goto skip;
 624         }
 625
 626         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 627
 628         if (lmaster != ctdb->pnn) {
 629                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 630                                    "delete list (key hash [0x%08x], db[%s]). "
 631                                    "Strange! skipping.\n",
 632                                    hash, ctdb_db->db_name));
 633                 goto skip;
 634         }
 635
 636         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 637
 638         if (res != 0) {
 639                 DEBUG(DEBUG_ERR,
 640                       (__location__ " Error deleting record with key hash "
 641                        "[0x%08x] from local data base db[%s].\n",
 642                        hash, ctdb_db->db_name));
 643                 vdata->delete_local_error++;
 644                 goto done;
 645         }
 646
 647         DEBUG(DEBUG_DEBUG,
 648               (__location__ " Deleted record with key hash [0x%08x] from "
 649                "local data base db[%s].\n", hash, ctdb_db->db_name));
 650
 651         vdata->delete_deleted++;
 652         goto done;
 653
 654 skip:
 655         vdata->delete_skipped++;
 656
 657 done:
 658         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 659
 660         talloc_free(dd);
 661         vdata->delete_left--;
 662
 663         return 0;
 664 }
 665
 666 /**
 667  * Traverse the delete_queue.
 668  * Records are either deleted directly or filled
 669  * into the delete list or the vacuum fetch lists
 670  * for further processing.
 671  */
 672 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 673                                       struct vacuum_data *vdata)
 674 {
 675         uint32_t sum;
 676
 677         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 678
 679         sum = vdata->fast_deleted
 680             + vdata->fast_skipped
 681             + vdata->fast_error
 682             + vdata->fast_added_to_delete_list
 683             + vdata->fast_added_to_vacuum_fetch_list;
 684
 685         if (vdata->fast_total != sum) {
 686                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 687                       "counts for db[%s]: total[%u] != sum[%u]\n",
 688                       ctdb_db->db_name, (unsigned)vdata->fast_total,
 689                       (unsigned)sum));
 690         }
 691
 692         if (vdata->fast_total > 0) {
 693                 DEBUG(DEBUG_INFO,
 694                       (__location__
 695                        " fast vacuuming delete_queue traverse statistics: "
 696                        "db[%s] "
 697                        "total[%u] "
 698                        "del[%u] "
 699                        "skp[%u] "
 700                        "err[%u] "
 701                        "adl[%u] "
 702                        "avf[%u]\n",
 703                        ctdb_db->db_name,
 704                        (unsigned)vdata->fast_total,
 705                        (unsigned)vdata->fast_deleted,
 706                        (unsigned)vdata->fast_skipped,
 707                        (unsigned)vdata->fast_error,
 708                        (unsigned)vdata->fast_added_to_delete_list,
 709                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 710         }
 711
 712         return;
 713 }
 714
 715 /**
 716  * read-only traverse of the database, looking for records that
 717  * might be able to be vacuumed.
 718  *
 719  * This is not done each time but only every tunable
 720  * VacuumFastPathCount times.
 721  */
 722 static int ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 723                                    struct vacuum_data *vdata)
 724 {
 725         int ret;
 726
 727         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 728         if (ret == -1 || vdata->traverse_error) {
 729                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 730                                   "'%s'\n", ctdb_db->db_name));
 731                 return -1;
 732         }
 733
 734         if (vdata->full_total > 0) {
 735                 DEBUG(DEBUG_INFO,
 736                       (__location__
 737                        " full vacuuming db traverse statistics: "
 738                        "db[%s] "
 739                        "total[%u] "
 740                        "skp[%u] "
 741                        "err[%u] "
 742                        "sched[%u]\n",
 743                        ctdb_db->db_name,
 744                        (unsigned)vdata->full_total,
 745                        (unsigned)vdata->full_skipped,
 746                        (unsigned)vdata->full_error,
 747                        (unsigned)vdata->full_scheduled));
 748         }
 749
 750         return 0;
 751 }
 752
 753 /**
 754  * Process the vacuum fetch lists:
 755  * For records for which we are not the lmaster, tell the lmaster to
 756  * fetch the record.
 757  */
 758 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 759                                            struct vacuum_data *vdata)
 760 {
 761         int i;
 762         struct ctdb_context *ctdb = ctdb_db->ctdb;
 763
 764         for (i = 0; i < ctdb->num_nodes; i++) {
 765                 TDB_DATA data;
 766                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 767
 768                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 769                         continue;
 770                 }
 771
 772                 if (vfl->count == 0) {
 773                         continue;
 774                 }
 775
 776                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 777                                    vfl->count, ctdb->nodes[i]->pnn,
 778                                    ctdb_db->db_name));
 779
 780                 data.dsize = talloc_get_size(vfl);
 781                 data.dptr  = (void *)vfl;
 782                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 783                                              CTDB_SRVID_VACUUM_FETCH,
 784                                              data) != 0)
 785                 {
 786                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 787                                           "fetch message to %u\n",
 788                                           ctdb->nodes[i]->pnn));
 789                         return -1;
 790                 }
 791         }
 792
 793         return 0;
 794 }
 795
 796 /**
 797  * Process the delete list:
 798  *
 799  * This is the last step of vacuuming that consistently deletes
 800  * those records that have been migrated with data and can hence
 801  * not be deleted when leaving a node.
 802  *
 803  * In this step, the lmaster does the final deletion of those empty
 804  * records that it is also dmaster for. It has ususally received
 805  * at least some of these records previously from the former dmasters
 806  * with the vacuum fetch message.
 807  *
 808  * This last step is implemented as a 3-phase process to protect from
 809  * races leading to data corruption:
 810  *
 811  *  1) Send the lmaster's copy to all other active nodes with the
 812  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 813  *  2) Send the records that could successfully be stored remotely
 814  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 815  *     control. The remote notes delete their local copy.
 816  *  3) The lmaster locally deletes its copies of all records that
 817  *     could successfully be deleted remotely in step #2.
 818  */
 819 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 820                                     struct vacuum_data *vdata)
 821 {
 822         int ret, i;
 823         struct ctdb_context *ctdb = ctdb_db->ctdb;
 824         struct delete_records_list *recs;
 825         TDB_DATA indata;
 826         struct ctdb_node_map *nodemap;
 827         uint32_t *active_nodes;
 828         int num_active_nodes;
 829         TALLOC_CTX *tmp_ctx;
 830         uint32_t sum;
 831
 832         if (vdata->delete_count == 0) {
 833                 return 0;
 834         }
 835
 836         tmp_ctx = talloc_new(vdata);
 837         if (tmp_ctx == NULL) {
 838                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 839                 return 0;
 840         }
 841
 842         vdata->delete_left = vdata->delete_count;
 843
 844         /*
 845          * get the list of currently active nodes
 846          */
 847
 848         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 849                                    CTDB_CURRENT_NODE,
 850                                    tmp_ctx,
 851                                    &nodemap);
 852         if (ret != 0) {
 853                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 854                 ret = -1;
 855                 goto done;
 856         }
 857
 858         active_nodes = list_of_active_nodes(ctdb, nodemap,
 859                                             nodemap, /* talloc context */
 860                                             false /* include self */);
 861         /* yuck! ;-) */
 862         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 863
 864         /*
 865          * Now delete the records all active nodes in a three-phase process:
 866          * 1) send all active remote nodes the current empty copy with this
 867          *    node as DMASTER
 868          * 2) if all nodes could store the new copy,
 869          *    tell all the active remote nodes to delete all their copy
 870          * 3) if all remote nodes deleted their record copy, delete it locally
 871          */
 872
 873         /*
 874          * Step 1:
 875          * Send currently empty record copy to all active nodes for storing.
 876          */
 877
 878         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 879         if (recs == NULL) {
 880                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 881                 ret = -1;
 882                 goto done;
 883         }
 884         recs->records = (struct ctdb_marshall_buffer *)
 885                 talloc_zero_size(recs,
 886                                  offsetof(struct ctdb_marshall_buffer, data));
 887         if (recs->records == NULL) {
 888                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 889                 ret = -1;
 890                 goto done;
 891         }
 892         recs->records->db_id = ctdb_db->db_id;
 893         recs->vdata = vdata;
 894
 895         /*
 896          * traverse the tree of all records we want to delete and
 897          * create a blob we can send to the other nodes.
 898          *
 899          * We call delete_marshall_traverse_first() to bump the
 900          * records' RSNs in the database, to ensure we (as dmaster)
 901          * keep the highest RSN of the records in the cluster.
 902          */
 903         trbt_traversearray32(vdata->delete_list, 1,
 904                              delete_marshall_traverse_first, recs);
 905
 906         indata.dsize = talloc_get_size(recs->records);
 907         indata.dptr  = (void *)recs->records;
 908
 909         for (i = 0; i < num_active_nodes; i++) {
 910                 struct ctdb_marshall_buffer *records;
 911                 struct ctdb_rec_data *rec;
 912                 int32_t res;
 913                 TDB_DATA outdata;
 914
 915                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 916                                 CTDB_CONTROL_RECEIVE_RECORDS, 0,
 917                                 indata, recs, &outdata, &res,
 918                                 NULL, NULL);
 919                 if (ret != 0 || res != 0) {
 920                         DEBUG(DEBUG_ERR, ("Error storing record copies on "
 921                                           "node %u: ret[%d] res[%d]\n",
 922                                           active_nodes[i], ret, res));
 923                         ret = -1;
 924                         goto done;
 925                 }
 926
 927                 /*
 928                  * outdata contains the list of records coming back
 929                  * from the node: These are the records that the
 930                  * remote node could not store. We remove these from
 931                  * the list to process further.
 932                  */
 933                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 934                 rec = (struct ctdb_rec_data *)&records->data[0];
 935                 while (records->count-- > 1) {
 936                         TDB_DATA reckey, recdata;
 937                         struct ctdb_ltdb_header *rechdr;
 938                         struct delete_record_data *dd;
 939
 940                         reckey.dptr = &rec->data[0];
 941                         reckey.dsize = rec->keylen;
 942                         recdata.dptr = &rec->data[reckey.dsize];
 943                         recdata.dsize = rec->datalen;
 944
 945                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 946                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 947                                 ret = -1;
 948                                 goto done;
 949                         }
 950                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 951                         recdata.dptr += sizeof(*rechdr);
 952                         recdata.dsize -= sizeof(*rechdr);
 953
 954                         dd = (struct delete_record_data *)trbt_lookup32(
 955                                         vdata->delete_list,
 956                                         ctdb_hash(&reckey));
 957                         if (dd != NULL) {
 958                                 /*
 959                                  * The other node could not store the record
 960                                  * copy and it is the first node that failed.
 961                                  * So we should remove it from the tree and
 962                                  * update statistics.
 963                                  */
 964                                 talloc_free(dd);
 965                                 vdata->delete_remote_error++;
 966                                 vdata->delete_left--;
 967                         }
 968
 969                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 970                 }
 971         }
 972
 973         if (vdata->delete_left == 0) {
 974                 goto success;
 975         }
 976
 977         /*
 978          * Step 2:
 979          * Send the remaining records to all active nodes for deletion.
 980          *
 981          * The lmaster's (i.e. our) copies of these records have been stored
 982          * successfully on the other nodes.
 983          */
 984
 985         /*
 986          * Create a marshall blob from the remaining list of records to delete.
 987          */
 988
 989         talloc_free(recs->records);
 990
 991         recs->records = (struct ctdb_marshall_buffer *)
 992                 talloc_zero_size(recs,
 993                                  offsetof(struct ctdb_marshall_buffer, data));
 994         if (recs->records == NULL) {
 995                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 996                 ret = -1;
 997                 goto done;
 998         }
 999         recs->records->db_id = ctdb_db->db_id;
1000
1001         trbt_traversearray32(vdata->delete_list, 1,
1002                              delete_marshall_traverse, recs);
1003
1004         indata.dsize = talloc_get_size(recs->records);
1005         indata.dptr  = (void *)recs->records;
1006
1007         for (i = 0; i < num_active_nodes; i++) {
1008                 struct ctdb_marshall_buffer *records;
1009                 struct ctdb_rec_data *rec;
1010                 int32_t res;
1011                 TDB_DATA outdata;
1012
1013                 ret = ctdb_control(ctdb, active_nodes[i], 0,
1014                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1015                                 indata, recs, &outdata, &res,
1016                                 NULL, NULL);
1017                 if (ret != 0 || res != 0) {
1018                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
1019                                           "node %u: ret[%d] res[%d]\n",
1020                                           active_nodes[i], ret, res));
1021                         ret = -1;
1022                         goto done;
1023                 }
1024
1025                 /*
1026                  * outdata contains the list of records coming back
1027                  * from the node: These are the records that the
1028                  * remote node could not delete. We remove these from
1029                  * the list to delete locally.
1030                  */
1031                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1032                 rec = (struct ctdb_rec_data *)&records->data[0];
1033                 while (records->count-- > 1) {
1034                         TDB_DATA reckey, recdata;
1035                         struct ctdb_ltdb_header *rechdr;
1036                         struct delete_record_data *dd;
1037
1038                         reckey.dptr = &rec->data[0];
1039                         reckey.dsize = rec->keylen;
1040                         recdata.dptr = &rec->data[reckey.dsize];
1041                         recdata.dsize = rec->datalen;
1042
1043                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1044                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1045                                 ret = -1;
1046                                 goto done;
1047                         }
1048                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1049                         recdata.dptr += sizeof(*rechdr);
1050                         recdata.dsize -= sizeof(*rechdr);
1051
1052                         dd = (struct delete_record_data *)trbt_lookup32(
1053                                         vdata->delete_list,
1054                                         ctdb_hash(&reckey));
1055                         if (dd != NULL) {
1056                                 /*
1057                                  * The other node could not delete the
1058                                  * record and it is the first node that
1059                                  * failed. So we should remove it from
1060                                  * the tree and update statistics.
1061                                  */
1062                                 talloc_free(dd);
1063                                 vdata->delete_remote_error++;
1064                                 vdata->delete_left--;
1065                         }
1066
1067                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
1068                 }
1069         }
1070
1071         if (vdata->delete_left == 0) {
1072                 goto success;
1073         }
1074
1075         /*
1076          * Step 3:
1077          * Delete the remaining records locally.
1078          *
1079          * These records have successfully been deleted on all
1080          * active remote nodes.
1081          */
1082
1083         trbt_traversearray32(vdata->delete_list, 1,
1084                              delete_record_traverse, vdata);
1085
1086 success:
1087
1088         if (vdata->delete_left != 0) {
1089                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
1090                       "there are %u records left for deletion after "
1091                       "processing delete list\n",
1092                       ctdb_db->db_name,
1093                       (unsigned)vdata->delete_left));
1094         }
1095
1096         sum = vdata->delete_deleted
1097             + vdata->delete_skipped
1098             + vdata->delete_remote_error
1099             + vdata->delete_local_error
1100             + vdata->delete_left;
1101
1102         if (vdata->delete_count != sum) {
1103                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
1104                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1105                       ctdb_db->db_name, (unsigned)vdata->delete_count,
1106                       (unsigned)sum));
1107         }
1108
1109         if (vdata->delete_count > 0) {
1110                 DEBUG(DEBUG_INFO,
1111                       (__location__
1112                        " vacuum delete list statistics: "
1113                        "db[%s] "
1114                        "total[%u] "
1115                        "del[%u] "
1116                        "skip[%u] "
1117                        "rem.err[%u] "
1118                        "loc.err[%u] "
1119                        "left[%u]\n",
1120                        ctdb_db->db_name,
1121                        (unsigned)vdata->delete_count,
1122                        (unsigned)vdata->delete_deleted,
1123                        (unsigned)vdata->delete_skipped,
1124                        (unsigned)vdata->delete_remote_error,
1125                        (unsigned)vdata->delete_local_error,
1126                        (unsigned)vdata->delete_left));
1127         }
1128
1129         ret = 0;
1130
1131 done:
1132         talloc_free(tmp_ctx);
1133
1134         return ret;
1135 }
1136
1137 /**
1138  * initialize the vacuum_data
1139  */
1140 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
1141                                         struct vacuum_data *vdata)
1142 {
1143         int i;
1144         struct ctdb_context *ctdb = ctdb_db->ctdb;
1145
1146         vdata->fast_added_to_delete_list = 0;
1147         vdata->fast_added_to_vacuum_fetch_list = 0;
1148         vdata->fast_deleted = 0;
1149         vdata->fast_skipped = 0;
1150         vdata->fast_error = 0;
1151         vdata->fast_total = 0;
1152         vdata->full_scheduled = 0;
1153         vdata->full_skipped = 0;
1154         vdata->full_error = 0;
1155         vdata->full_total = 0;
1156         vdata->delete_count = 0;
1157         vdata->delete_left = 0;
1158         vdata->delete_remote_error = 0;
1159         vdata->delete_local_error = 0;
1160         vdata->delete_skipped = 0;
1161         vdata->delete_deleted = 0;
1162
1163         /* the list needs to be of length num_nodes */
1164         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1165                                                 struct ctdb_marshall_buffer *,
1166                                                 ctdb->num_nodes);
1167         if (vdata->vacuum_fetch_list == NULL) {
1168                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1169                 return -1;
1170         }
1171         for (i = 0; i < ctdb->num_nodes; i++) {
1172                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1173                         talloc_zero_size(vdata->vacuum_fetch_list,
1174                                          offsetof(struct ctdb_marshall_buffer, data));
1175                 if (vdata->vacuum_fetch_list[i] == NULL) {
1176                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1177                         return -1;
1178                 }
1179                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1180         }
1181
1182         return 0;
1183 }
1184
1185 /**
1186  * Vacuum a DB:
1187  *  - Always do the fast vacuuming run, which traverses
1188  *    the in-memory delete queue: these records have been
1189  *    scheduled for deletion.
1190  *  - Only if explicitly requested, the database is traversed
1191  *    in order to use the traditional heuristics on empty records
1192  *    to trigger deletion.
1193  *    This is done only every VacuumFastPathCount'th vacuuming run.
1194  *
1195  * The traverse runs fill two lists:
1196  *
1197  * - The delete_list:
1198  *   This is the list of empty records the current
1199  *   node is lmaster and dmaster for. These records are later
1200  *   deleted first on other nodes and then locally.
1201  *
1202  *   The fast vacuuming run has a short cut for those records
1203  *   that have never been migrated with data: these records
1204  *   are immediately deleted locally, since they have left
1205  *   no trace on other nodes.
1206  *
1207  * - The vacuum_fetch lists
1208  *   (one for each other lmaster node):
1209  *   The records in this list are sent for deletion to
1210  *   their lmaster in a bulk VACUUM_FETCH message.
1211  *
1212  *   The lmaster then migrates all these records to itelf
1213  *   so that they can be vacuumed there.
1214  *
1215  * This executes in the child context.
1216  */
1217 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1218                           struct vacuum_data *vdata,
1219                           bool full_vacuum_run)
1220 {
1221         struct ctdb_context *ctdb = ctdb_db->ctdb;
1222         int ret, pnn;
1223
1224         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1225                            "%s db_id[0x%08x]\n",
1226                            full_vacuum_run ? "full" : "fast",
1227                            ctdb_db->db_name, ctdb_db->db_id));
1228
1229         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1230         if (ret != 0) {
1231                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1232                 return ret;
1233         }
1234
1235         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1236         if (pnn == -1) {
1237                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1238                 return -1;
1239         }
1240
1241         ctdb->pnn = pnn;
1242
1243         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
1244         if (ret != 0) {
1245                 return ret;
1246         }
1247
1248         if (full_vacuum_run) {
1249                 ret = ctdb_vacuum_traverse_db(ctdb_db, vdata);
1250                 if (ret != 0) {
1251                         return ret;
1252                 }
1253         }
1254
1255         ctdb_process_delete_queue(ctdb_db, vdata);
1256
1257         ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1258         if (ret != 0) {
1259                 return ret;
1260         }
1261
1262         ret = ctdb_process_delete_list(ctdb_db, vdata);
1263         if (ret != 0) {
1264                 return ret;
1265         }
1266
1267         /* this ensures we run our event queue */
1268         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1269
1270         return 0;
1271 }
1272
1273
1274 /*
1275  * traverse function for repacking
1276  */
1277 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
1278                            void *private_data)
1279 {
1280         struct vacuum_data *vdata = (struct vacuum_data *)private_data;
1281
1282         if (vdata->vacuum) {
1283                 uint32_t hash = ctdb_hash(&key);
1284                 struct delete_record_data *kd;
1285                 /*
1286                  * check if we can ignore this record because it's in the delete_list
1287                  */
1288                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1289                 /*
1290                  * there might be hash collisions so we have to compare the keys here to be sure
1291                  */
1292                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1293                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1294                         /*
1295                          * we have to check if the record hasn't changed in the meantime in order to
1296                          * savely remove it from the database
1297                          */
1298                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1299                                 hdr->dmaster == kd->ctdb->pnn &&
1300                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1301                                 kd->hdr.rsn == hdr->rsn) {
1302                                 vdata->vacuumed++;
1303                                 return 0;
1304                         }
1305                 }
1306         }
1307         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1308                 vdata->traverse_error = true;
1309                 return -1;
1310         }
1311         vdata->copied++;
1312         return 0;
1313 }
1314
1315 /*
1316  * repack a tdb
1317  */
1318 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1319 {
1320         struct tdb_context *tmp_db;
1321
1322         if (tdb_transaction_start(tdb) != 0) {
1323                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1324                 return -1;
1325         }
1326
1327         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1328                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1329                           O_RDWR|O_CREAT, 0);
1330         if (tmp_db == NULL) {
1331                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1332                 tdb_transaction_cancel(tdb);
1333                 return -1;
1334         }
1335
1336         vdata->traverse_error = false;
1337         vdata->dest_db = tmp_db;
1338         vdata->vacuum = true;
1339         vdata->vacuumed = 0;
1340         vdata->copied = 0;
1341
1342         /*
1343          * repack and vacuum on-the-fly by not writing the records that are
1344          * no longer needed
1345          */
1346         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1347                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1348                 tdb_transaction_cancel(tdb);
1349                 tdb_close(tmp_db);
1350                 return -1;
1351         }
1352
1353         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1354
1355         if (vdata->traverse_error) {
1356                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1357                 tdb_transaction_cancel(tdb);
1358                 tdb_close(tmp_db);
1359                 return -1;
1360         }
1361
1362         if (tdb_wipe_all(tdb) != 0) {
1363                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1364                 tdb_transaction_cancel(tdb);
1365                 tdb_close(tmp_db);
1366                 return -1;
1367         }
1368
1369         vdata->traverse_error = false;
1370         vdata->dest_db = tdb;
1371         vdata->vacuum = false;
1372         vdata->copied = 0;
1373
1374         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1375                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1376                 tdb_transaction_cancel(tdb);
1377                 tdb_close(tmp_db);
1378                 return -1;
1379         }
1380
1381         if (vdata->traverse_error) {
1382                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1383                 tdb_transaction_cancel(tdb);
1384                 tdb_close(tmp_db);
1385                 return -1;
1386         }
1387
1388         tdb_close(tmp_db);
1389
1390
1391         if (tdb_transaction_commit(tdb) != 0) {
1392                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1393                 return -1;
1394         }
1395         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1396
1397         return 0;
1398 }
1399
1400 /*
1401  * repack and vaccum a db
1402  * called from the child context
1403  */
1404 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1405                                      TALLOC_CTX *mem_ctx,
1406                                      bool full_vacuum_run)
1407 {
1408         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1409         const char *name = ctdb_db->db_name;
1410         int freelist_size = 0;
1411         struct vacuum_data *vdata;
1412
1413         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1414         if (vdata == NULL) {
1415                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1416                 return -1;
1417         }
1418
1419         vdata->ctdb = ctdb_db->ctdb;
1420         vdata->repack_limit = repack_limit;
1421         vdata->delete_list = trbt_create(vdata, 0);
1422         vdata->ctdb_db = ctdb_db;
1423         if (vdata->delete_list == NULL) {
1424                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1425                 talloc_free(vdata);
1426                 return -1;
1427         }
1428
1429         vdata->start = timeval_current();
1430
1431         /*
1432          * gather all records that can be deleted in vdata
1433          */
1434         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1435                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1436         }
1437
1438         if (repack_limit != 0) {
1439                 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1440                 if (freelist_size == -1) {
1441                         DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1442                         talloc_free(vdata);
1443                         return -1;
1444                 }
1445         }
1446
1447         /*
1448          * decide if a repack is necessary
1449          */
1450         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1451         {
1452                 talloc_free(vdata);
1453                 return 0;
1454         }
1455
1456         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1457                         name, freelist_size, vdata->delete_left));
1458
1459         /*
1460          * repack and implicitely get rid of the records we can delete
1461          */
1462         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1463                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1464                 talloc_free(vdata);
1465                 return -1;
1466         }
1467         talloc_free(vdata);
1468
1469         return 0;
1470 }
1471
1472 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1473 {
1474         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1475
1476         return interval;
1477 }
1478
1479 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1480 {
1481         double l = timeval_elapsed(&child_ctx->start_time);
1482         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1483         struct ctdb_context *ctdb = ctdb_db->ctdb;
1484
1485         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1486
1487         if (child_ctx->child_pid != -1) {
1488                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1489         } else {
1490                 /* Bump the number of successful fast-path runs. */
1491                 child_ctx->vacuum_handle->fast_path_count++;
1492         }
1493
1494         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1495
1496         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1497                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1498                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1499
1500         return 0;
1501 }
1502
1503 /*
1504  * this event is generated when a vacuum child process times out
1505  */
1506 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1507                                          struct timeval t, void *private_data)
1508 {
1509         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1510
1511         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1512
1513         child_ctx->status = VACUUM_TIMEOUT;
1514
1515         talloc_free(child_ctx);
1516 }
1517
1518
1519 /*
1520  * this event is generated when a vacuum child process has completed
1521  */
1522 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1523                              uint16_t flags, void *private_data)
1524 {
1525         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1526         char c = 0;
1527         int ret;
1528
1529         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1530         child_ctx->child_pid = -1;
1531
1532         ret = read(child_ctx->fd[0], &c, 1);
1533         if (ret != 1 || c != 0) {
1534                 child_ctx->status = VACUUM_ERROR;
1535                 DEBUG(DEBUG_ERR, ("A vacuum child process failed with an error for database %s. ret=%d c=%d\n", child_ctx->vacuum_handle->ctdb_db->db_name, ret, c));
1536         } else {
1537                 child_ctx->status = VACUUM_OK;
1538         }
1539
1540         talloc_free(child_ctx);
1541 }
1542
1543 /*
1544  * this event is called every time we need to start a new vacuum process
1545  */
1546 static void
1547 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1548                                struct timeval t, void *private_data)
1549 {
1550         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1551         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1552         struct ctdb_context *ctdb = ctdb_db->ctdb;
1553         struct ctdb_vacuum_child_context *child_ctx;
1554         struct tevent_fd *fde;
1555         int ret;
1556
1557         /* we dont vacuum if we are in recovery mode, or db frozen */
1558         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1559             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1560                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1561                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1562                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1563                                    ? "freeze pending"
1564                                    : "frozen"));
1565                 event_add_timed(ctdb->ev, vacuum_handle,
1566                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1567                         ctdb_vacuum_event, vacuum_handle);
1568                 return;
1569         }
1570
1571         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1572         if (child_ctx == NULL) {
1573                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1574                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1575         }
1576
1577
1578         ret = pipe(child_ctx->fd);
1579         if (ret != 0) {
1580                 talloc_free(child_ctx);
1581                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1582                 event_add_timed(ctdb->ev, vacuum_handle,
1583                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1584                         ctdb_vacuum_event, vacuum_handle);
1585                 return;
1586         }
1587
1588         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1589                 vacuum_handle->fast_path_count = 0;
1590         }
1591
1592         child_ctx->child_pid = ctdb_fork(ctdb);
1593         if (child_ctx->child_pid == (pid_t)-1) {
1594                 close(child_ctx->fd[0]);
1595                 close(child_ctx->fd[1]);
1596                 talloc_free(child_ctx);
1597                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1598                 event_add_timed(ctdb->ev, vacuum_handle,
1599                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1600                         ctdb_vacuum_event, vacuum_handle);
1601                 return;
1602         }
1603
1604
1605         if (child_ctx->child_pid == 0) {
1606                 char cc = 0;
1607                 bool full_vacuum_run = false;
1608                 close(child_ctx->fd[0]);
1609
1610                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1611                 ctdb_set_process_name("ctdb_vacuum");
1612                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1613                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1614                         _exit(1);
1615                 }
1616
1617                 /*
1618                  * repack the db
1619                  */
1620                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1621                     (vacuum_handle->fast_path_count == 0))
1622                 {
1623                         full_vacuum_run = true;
1624                 }
1625                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1626                                                full_vacuum_run);
1627
1628                 write(child_ctx->fd[1], &cc, 1);
1629                 _exit(0);
1630         }
1631
1632         set_close_on_exec(child_ctx->fd[0]);
1633         close(child_ctx->fd[1]);
1634
1635         child_ctx->status = VACUUM_RUNNING;
1636         child_ctx->start_time = timeval_current();
1637
1638         DLIST_ADD(ctdb->vacuumers, child_ctx);
1639         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1640
1641         /*
1642          * Clear the fastpath vacuuming list in the parent.
1643          */
1644         talloc_free(ctdb_db->delete_queue);
1645         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1646         if (ctdb_db->delete_queue == NULL) {
1647                 /* fatal here? ... */
1648                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1649                                  "in parent context. Shutting down\n");
1650         }
1651
1652         event_add_timed(ctdb->ev, child_ctx,
1653                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1654                 vacuum_child_timeout, child_ctx);
1655
1656         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1657
1658         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1659                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1660         tevent_fd_set_auto_close(fde);
1661
1662         vacuum_handle->child_ctx = child_ctx;
1663         child_ctx->vacuum_handle = vacuum_handle;
1664 }
1665
1666 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1667 {
1668         /* Simply free them all. */
1669         while (ctdb->vacuumers) {
1670                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1671                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1672                            (int)ctdb->vacuumers->child_pid));
1673                 /* vacuum_child_destructor kills it, removes from list */
1674                 talloc_free(ctdb->vacuumers);
1675         }
1676 }
1677
1678 /* this function initializes the vacuuming context for a database
1679  * starts the vacuuming events
1680  */
1681 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1682 {
1683         if (ctdb_db->persistent != 0) {
1684                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1685                 return 0;
1686         }
1687
1688         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1689         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1690
1691         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1692         ctdb_db->vacuum_handle->fast_path_count = 0;
1693
1694         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1695                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1696                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1697
1698         return 0;
1699 }
1700
1701 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1702                                             const struct ctdb_ltdb_header *hdr,
1703                                             const TDB_DATA key)
1704 {
1705         struct delete_record_data *kd;
1706         uint32_t hash;
1707
1708         hash = (uint32_t)ctdb_hash(&key);
1709
1710         DEBUG(DEBUG_DEBUG, (__location__
1711                             " remove_record_from_delete_queue: "
1712                             "db[%s] "
1713                             "db_id[0x%08x] "
1714                             "key_hash[0x%08x] "
1715                             "lmaster[%u] "
1716                             "migrated_with_data[%s]\n",
1717                              ctdb_db->db_name, ctdb_db->db_id,
1718                              hash,
1719                              ctdb_lmaster(ctdb_db->ctdb, &key),
1720                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1721
1722         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1723         if (kd == NULL) {
1724                 DEBUG(DEBUG_DEBUG, (__location__
1725                                     " remove_record_from_delete_queue: "
1726                                     "record not in queue (hash[0x%08x])\n.",
1727                                     hash));
1728                 return;
1729         }
1730
1731         if ((kd->key.dsize != key.dsize) ||
1732             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1733         {
1734                 DEBUG(DEBUG_DEBUG, (__location__
1735                                     " remove_record_from_delete_queue: "
1736                                     "hash collision for key with hash[0x%08x] "
1737                                     "in db[%s] - skipping\n",
1738                                     hash, ctdb_db->db_name));
1739                 return;
1740         }
1741
1742         DEBUG(DEBUG_DEBUG, (__location__
1743                             " remove_record_from_delete_queue: "
1744                             "removing key with hash[0x%08x]\n",
1745                              hash));
1746
1747         talloc_free(kd);
1748
1749         return;
1750 }
1751
1752 /**
1753  * Insert a record into the ctdb_db context's delete queue,
1754  * handling hash collisions.
1755  */
1756 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1757                                            const struct ctdb_ltdb_header *hdr,
1758                                            TDB_DATA key)
1759 {
1760         struct delete_record_data *kd;
1761         uint32_t hash;
1762         int ret;
1763
1764         hash = (uint32_t)ctdb_hash(&key);
1765
1766         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1767                            "db_id[0x%08x] "
1768                            "key_hash[0x%08x] "
1769                            "lmaster[%u] "
1770                            "migrated_with_data[%s]\n",
1771                             ctdb_db->db_name, ctdb_db->db_id,
1772                             hash,
1773                             ctdb_lmaster(ctdb_db->ctdb, &key),
1774                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1775
1776         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1777         if (kd != NULL) {
1778                 if ((kd->key.dsize != key.dsize) ||
1779                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1780                 {
1781                         DEBUG(DEBUG_INFO,
1782                               (__location__ " schedule for deletion: "
1783                                "hash collision for key hash [0x%08x]. "
1784                                "Skipping the record.\n", hash));
1785                         return 0;
1786                 } else {
1787                         DEBUG(DEBUG_DEBUG,
1788                               (__location__ " schedule for deletion: "
1789                                "updating entry for key with hash [0x%08x].\n",
1790                                hash));
1791                 }
1792         }
1793
1794         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1795                                                   ctdb_db->delete_queue,
1796                                                   hdr, key);
1797         if (ret != 0) {
1798                 DEBUG(DEBUG_INFO,
1799                       (__location__ " schedule for deletion: error "
1800                        "inserting key with hash [0x%08x] into delete queue\n",
1801                        hash));
1802                 return -1;
1803         }
1804
1805         return 0;
1806 }
1807
1808 /**
1809  * Schedule a record for deletetion.
1810  * Called from the parent context.
1811  */
1812 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1813                                            TDB_DATA indata)
1814 {
1815         struct ctdb_control_schedule_for_deletion *dd;
1816         struct ctdb_db_context *ctdb_db;
1817         int ret;
1818         TDB_DATA key;
1819
1820         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1821
1822         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1823         if (ctdb_db == NULL) {
1824                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1825                                   dd->db_id));
1826                 return -1;
1827         }
1828
1829         key.dsize = dd->keylen;
1830         key.dptr = dd->key;
1831
1832         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1833
1834         return ret;
1835 }
1836
1837 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1838                                          const struct ctdb_ltdb_header *hdr,
1839                                          TDB_DATA key)
1840 {
1841         int ret;
1842         struct ctdb_control_schedule_for_deletion *dd;
1843         TDB_DATA indata;
1844         int32_t status;
1845
1846         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1847                 /* main daemon - directly queue */
1848                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1849
1850                 return ret;
1851         }
1852
1853         /* if we dont have a connection to the daemon we can not send
1854            a control. For example sometimes from update_record control child
1855            process.
1856         */
1857         if (!ctdb_db->ctdb->can_send_controls) {
1858                 return -1;
1859         }
1860
1861
1862         /* child process: send the main daemon a control */
1863         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1864         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1865         if (indata.dptr == NULL) {
1866                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1867                 return -1;
1868         }
1869         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1870         dd->db_id = ctdb_db->db_id;
1871         dd->hdr = *hdr;
1872         dd->keylen = key.dsize;
1873         memcpy(dd->key, key.dptr, key.dsize);
1874
1875         ret = ctdb_control(ctdb_db->ctdb,
1876                            CTDB_CURRENT_NODE,
1877                            ctdb_db->db_id,
1878                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1879                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1880                            indata,
1881                            NULL, /* mem_ctx */
1882                            NULL, /* outdata */
1883                            &status,
1884                            NULL, /* timeout : NULL == wait forever */
1885                            NULL); /* error message */
1886
1887         talloc_free(indata.dptr);
1888
1889         if (ret != 0 || status != 0) {
1890                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1891                                   "SCHEDULE_FOR_DELETION "
1892                                   "control.\n"));
1893                 if (status != 0) {
1894                         ret = -1;
1895                 }
1896         }
1897
1898         return ret;
1899 }
1900
1901 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1902                                          const struct ctdb_ltdb_header *hdr,
1903                                          const TDB_DATA key)
1904 {
1905         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1906                 /*
1907                  * Only remove the record from the delete queue if called
1908                  * in the main daemon.
1909                  */
1910                 return;
1911         }
1912
1913         remove_record_from_delete_queue(ctdb_db, hdr, key);
1914
1915         return;
1916 }