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         int ret;
 677
 678         ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
 679                                    delete_queue_traverse, vdata);
 680
 681         if (ret != 0) {
 682                 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
 683                       "the delete queue.\n"));
 684         }
 685
 686         sum = vdata->fast_deleted
 687             + vdata->fast_skipped
 688             + vdata->fast_error
 689             + vdata->fast_added_to_delete_list
 690             + vdata->fast_added_to_vacuum_fetch_list;
 691
 692         if (vdata->fast_total != sum) {
 693                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 694                       "counts for db[%s]: total[%u] != sum[%u]\n",
 695                       ctdb_db->db_name, (unsigned)vdata->fast_total,
 696                       (unsigned)sum));
 697         }
 698
 699         if (vdata->fast_total > 0) {
 700                 DEBUG(DEBUG_INFO,
 701                       (__location__
 702                        " fast vacuuming delete_queue traverse statistics: "
 703                        "db[%s] "
 704                        "total[%u] "
 705                        "del[%u] "
 706                        "skp[%u] "
 707                        "err[%u] "
 708                        "adl[%u] "
 709                        "avf[%u]\n",
 710                        ctdb_db->db_name,
 711                        (unsigned)vdata->fast_total,
 712                        (unsigned)vdata->fast_deleted,
 713                        (unsigned)vdata->fast_skipped,
 714                        (unsigned)vdata->fast_error,
 715                        (unsigned)vdata->fast_added_to_delete_list,
 716                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 717         }
 718
 719         return;
 720 }
 721
 722 /**
 723  * read-only traverse of the database, looking for records that
 724  * might be able to be vacuumed.
 725  *
 726  * This is not done each time but only every tunable
 727  * VacuumFastPathCount times.
 728  */
 729 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 730                                     struct vacuum_data *vdata)
 731 {
 732         int ret;
 733
 734         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 735         if (ret == -1 || vdata->traverse_error) {
 736                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 737                                   "'%s'\n", ctdb_db->db_name));
 738         }
 739
 740         if (vdata->full_total > 0) {
 741                 DEBUG(DEBUG_INFO,
 742                       (__location__
 743                        " full vacuuming db traverse statistics: "
 744                        "db[%s] "
 745                        "total[%u] "
 746                        "skp[%u] "
 747                        "err[%u] "
 748                        "sched[%u]\n",
 749                        ctdb_db->db_name,
 750                        (unsigned)vdata->full_total,
 751                        (unsigned)vdata->full_skipped,
 752                        (unsigned)vdata->full_error,
 753                        (unsigned)vdata->full_scheduled));
 754         }
 755
 756         return;
 757 }
 758
 759 /**
 760  * Process the vacuum fetch lists:
 761  * For records for which we are not the lmaster, tell the lmaster to
 762  * fetch the record.
 763  */
 764 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 765                                             struct vacuum_data *vdata)
 766 {
 767         int i;
 768         struct ctdb_context *ctdb = ctdb_db->ctdb;
 769
 770         for (i = 0; i < ctdb->num_nodes; i++) {
 771                 TDB_DATA data;
 772                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 773
 774                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 775                         continue;
 776                 }
 777
 778                 if (vfl->count == 0) {
 779                         continue;
 780                 }
 781
 782                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 783                                    vfl->count, ctdb->nodes[i]->pnn,
 784                                    ctdb_db->db_name));
 785
 786                 data.dsize = talloc_get_size(vfl);
 787                 data.dptr  = (void *)vfl;
 788                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 789                                              CTDB_SRVID_VACUUM_FETCH,
 790                                              data) != 0)
 791                 {
 792                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 793                                           "fetch message to %u\n",
 794                                           ctdb->nodes[i]->pnn));
 795                 }
 796         }
 797
 798         return;
 799 }
 800
 801 /**
 802  * Process the delete list:
 803  *
 804  * This is the last step of vacuuming that consistently deletes
 805  * those records that have been migrated with data and can hence
 806  * not be deleted when leaving a node.
 807  *
 808  * In this step, the lmaster does the final deletion of those empty
 809  * records that it is also dmaster for. It has ususally received
 810  * at least some of these records previously from the former dmasters
 811  * with the vacuum fetch message.
 812  *
 813  * This last step is implemented as a 3-phase process to protect from
 814  * races leading to data corruption:
 815  *
 816  *  1) Send the lmaster's copy to all other active nodes with the
 817  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 818  *  2) Send the records that could successfully be stored remotely
 819  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 820  *     control. The remote notes delete their local copy.
 821  *  3) The lmaster locally deletes its copies of all records that
 822  *     could successfully be deleted remotely in step #2.
 823  */
 824 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 825                                      struct vacuum_data *vdata)
 826 {
 827         int ret, i;
 828         struct ctdb_context *ctdb = ctdb_db->ctdb;
 829         struct delete_records_list *recs;
 830         TDB_DATA indata;
 831         struct ctdb_node_map *nodemap;
 832         uint32_t *active_nodes;
 833         int num_active_nodes;
 834         TALLOC_CTX *tmp_ctx;
 835         uint32_t sum;
 836
 837         if (vdata->delete_count == 0) {
 838                 return;
 839         }
 840
 841         tmp_ctx = talloc_new(vdata);
 842         if (tmp_ctx == NULL) {
 843                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 844                 return;
 845         }
 846
 847         vdata->delete_left = vdata->delete_count;
 848
 849         /*
 850          * get the list of currently active nodes
 851          */
 852
 853         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 854                                    CTDB_CURRENT_NODE,
 855                                    tmp_ctx,
 856                                    &nodemap);
 857         if (ret != 0) {
 858                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 859                 goto done;
 860         }
 861
 862         active_nodes = list_of_active_nodes(ctdb, nodemap,
 863                                             nodemap, /* talloc context */
 864                                             false /* include self */);
 865         /* yuck! ;-) */
 866         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 867
 868         /*
 869          * Now delete the records all active nodes in a three-phase process:
 870          * 1) send all active remote nodes the current empty copy with this
 871          *    node as DMASTER
 872          * 2) if all nodes could store the new copy,
 873          *    tell all the active remote nodes to delete all their copy
 874          * 3) if all remote nodes deleted their record copy, delete it locally
 875          */
 876
 877         /*
 878          * Step 1:
 879          * Send currently empty record copy to all active nodes for storing.
 880          */
 881
 882         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 883         if (recs == NULL) {
 884                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 885                 goto done;
 886         }
 887         recs->records = (struct ctdb_marshall_buffer *)
 888                 talloc_zero_size(recs,
 889                                  offsetof(struct ctdb_marshall_buffer, data));
 890         if (recs->records == NULL) {
 891                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 892                 goto done;
 893         }
 894         recs->records->db_id = ctdb_db->db_id;
 895         recs->vdata = vdata;
 896
 897         /*
 898          * traverse the tree of all records we want to delete and
 899          * create a blob we can send to the other nodes.
 900          *
 901          * We call delete_marshall_traverse_first() to bump the
 902          * records' RSNs in the database, to ensure we (as dmaster)
 903          * keep the highest RSN of the records in the cluster.
 904          */
 905         ret = trbt_traversearray32(vdata->delete_list, 1,
 906                                    delete_marshall_traverse_first, recs);
 907         if (ret != 0) {
 908                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
 909                       "delete list for first marshalling.\n"));
 910         }
 911
 912         indata.dsize = talloc_get_size(recs->records);
 913         indata.dptr  = (void *)recs->records;
 914
 915         for (i = 0; i < num_active_nodes; i++) {
 916                 struct ctdb_marshall_buffer *records;
 917                 struct ctdb_rec_data *rec;
 918                 int32_t res;
 919                 TDB_DATA outdata;
 920
 921                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 922                                 CTDB_CONTROL_RECEIVE_RECORDS, 0,
 923                                 indata, recs, &outdata, &res,
 924                                 NULL, NULL);
 925                 if (ret != 0 || res != 0) {
 926                         DEBUG(DEBUG_ERR, ("Error storing record copies on "
 927                                           "node %u: ret[%d] res[%d]\n",
 928                                           active_nodes[i], ret, res));
 929                         goto done;
 930                 }
 931
 932                 /*
 933                  * outdata contains the list of records coming back
 934                  * from the node: These are the records that the
 935                  * remote node could not store. We remove these from
 936                  * the list to process further.
 937                  */
 938                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 939                 rec = (struct ctdb_rec_data *)&records->data[0];
 940                 while (records->count-- > 1) {
 941                         TDB_DATA reckey, recdata;
 942                         struct ctdb_ltdb_header *rechdr;
 943                         struct delete_record_data *dd;
 944
 945                         reckey.dptr = &rec->data[0];
 946                         reckey.dsize = rec->keylen;
 947                         recdata.dptr = &rec->data[reckey.dsize];
 948                         recdata.dsize = rec->datalen;
 949
 950                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 951                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 952                                 goto done;
 953                         }
 954                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 955                         recdata.dptr += sizeof(*rechdr);
 956                         recdata.dsize -= sizeof(*rechdr);
 957
 958                         dd = (struct delete_record_data *)trbt_lookup32(
 959                                         vdata->delete_list,
 960                                         ctdb_hash(&reckey));
 961                         if (dd != NULL) {
 962                                 /*
 963                                  * The other node could not store the record
 964                                  * copy and it is the first node that failed.
 965                                  * So we should remove it from the tree and
 966                                  * update statistics.
 967                                  */
 968                                 talloc_free(dd);
 969                                 vdata->delete_remote_error++;
 970                                 vdata->delete_left--;
 971                         }
 972
 973                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 974                 }
 975         }
 976
 977         if (vdata->delete_left == 0) {
 978                 goto success;
 979         }
 980
 981         /*
 982          * Step 2:
 983          * Send the remaining records to all active nodes for deletion.
 984          *
 985          * The lmaster's (i.e. our) copies of these records have been stored
 986          * successfully on the other nodes.
 987          */
 988
 989         /*
 990          * Create a marshall blob from the remaining list of records to delete.
 991          */
 992
 993         talloc_free(recs->records);
 994
 995         recs->records = (struct ctdb_marshall_buffer *)
 996                 talloc_zero_size(recs,
 997                                  offsetof(struct ctdb_marshall_buffer, data));
 998         if (recs->records == NULL) {
 999                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1000                 goto done;
1001         }
1002         recs->records->db_id = ctdb_db->db_id;
1003
1004         ret = trbt_traversearray32(vdata->delete_list, 1,
1005                                    delete_marshall_traverse, recs);
1006         if (ret != 0) {
1007                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1008                       "delete list for second marshalling.\n"));
1009         }
1010
1011         indata.dsize = talloc_get_size(recs->records);
1012         indata.dptr  = (void *)recs->records;
1013
1014         for (i = 0; i < num_active_nodes; i++) {
1015                 struct ctdb_marshall_buffer *records;
1016                 struct ctdb_rec_data *rec;
1017                 int32_t res;
1018                 TDB_DATA outdata;
1019
1020                 ret = ctdb_control(ctdb, active_nodes[i], 0,
1021                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1022                                 indata, recs, &outdata, &res,
1023                                 NULL, NULL);
1024                 if (ret != 0 || res != 0) {
1025                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
1026                                           "node %u: ret[%d] res[%d]\n",
1027                                           active_nodes[i], ret, res));
1028                         goto done;
1029                 }
1030
1031                 /*
1032                  * outdata contains the list of records coming back
1033                  * from the node: These are the records that the
1034                  * remote node could not delete. We remove these from
1035                  * the list to delete locally.
1036                  */
1037                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1038                 rec = (struct ctdb_rec_data *)&records->data[0];
1039                 while (records->count-- > 1) {
1040                         TDB_DATA reckey, recdata;
1041                         struct ctdb_ltdb_header *rechdr;
1042                         struct delete_record_data *dd;
1043
1044                         reckey.dptr = &rec->data[0];
1045                         reckey.dsize = rec->keylen;
1046                         recdata.dptr = &rec->data[reckey.dsize];
1047                         recdata.dsize = rec->datalen;
1048
1049                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1050                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1051                                 goto done;
1052                         }
1053                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1054                         recdata.dptr += sizeof(*rechdr);
1055                         recdata.dsize -= sizeof(*rechdr);
1056
1057                         dd = (struct delete_record_data *)trbt_lookup32(
1058                                         vdata->delete_list,
1059                                         ctdb_hash(&reckey));
1060                         if (dd != NULL) {
1061                                 /*
1062                                  * The other node could not delete the
1063                                  * record and it is the first node that
1064                                  * failed. So we should remove it from
1065                                  * the tree and update statistics.
1066                                  */
1067                                 talloc_free(dd);
1068                                 vdata->delete_remote_error++;
1069                                 vdata->delete_left--;
1070                         }
1071
1072                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
1073                 }
1074         }
1075
1076         if (vdata->delete_left == 0) {
1077                 goto success;
1078         }
1079
1080         /*
1081          * Step 3:
1082          * Delete the remaining records locally.
1083          *
1084          * These records have successfully been deleted on all
1085          * active remote nodes.
1086          */
1087
1088         ret = trbt_traversearray32(vdata->delete_list, 1,
1089                                    delete_record_traverse, vdata);
1090         if (ret != 0) {
1091                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1092                       "delete list for deletion.\n"));
1093         }
1094
1095 success:
1096
1097         if (vdata->delete_left != 0) {
1098                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
1099                       "there are %u records left for deletion after "
1100                       "processing delete list\n",
1101                       ctdb_db->db_name,
1102                       (unsigned)vdata->delete_left));
1103         }
1104
1105         sum = vdata->delete_deleted
1106             + vdata->delete_skipped
1107             + vdata->delete_remote_error
1108             + vdata->delete_local_error
1109             + vdata->delete_left;
1110
1111         if (vdata->delete_count != sum) {
1112                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
1113                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1114                       ctdb_db->db_name, (unsigned)vdata->delete_count,
1115                       (unsigned)sum));
1116         }
1117
1118         if (vdata->delete_count > 0) {
1119                 DEBUG(DEBUG_INFO,
1120                       (__location__
1121                        " vacuum delete list statistics: "
1122                        "db[%s] "
1123                        "total[%u] "
1124                        "del[%u] "
1125                        "skip[%u] "
1126                        "rem.err[%u] "
1127                        "loc.err[%u] "
1128                        "left[%u]\n",
1129                        ctdb_db->db_name,
1130                        (unsigned)vdata->delete_count,
1131                        (unsigned)vdata->delete_deleted,
1132                        (unsigned)vdata->delete_skipped,
1133                        (unsigned)vdata->delete_remote_error,
1134                        (unsigned)vdata->delete_local_error,
1135                        (unsigned)vdata->delete_left));
1136         }
1137
1138 done:
1139         talloc_free(tmp_ctx);
1140
1141         return;
1142 }
1143
1144 /**
1145  * initialize the vacuum_data
1146  */
1147 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
1148                                         struct vacuum_data *vdata)
1149 {
1150         int i;
1151         struct ctdb_context *ctdb = ctdb_db->ctdb;
1152
1153         vdata->fast_added_to_delete_list = 0;
1154         vdata->fast_added_to_vacuum_fetch_list = 0;
1155         vdata->fast_deleted = 0;
1156         vdata->fast_skipped = 0;
1157         vdata->fast_error = 0;
1158         vdata->fast_total = 0;
1159         vdata->full_scheduled = 0;
1160         vdata->full_skipped = 0;
1161         vdata->full_error = 0;
1162         vdata->full_total = 0;
1163         vdata->delete_count = 0;
1164         vdata->delete_left = 0;
1165         vdata->delete_remote_error = 0;
1166         vdata->delete_local_error = 0;
1167         vdata->delete_skipped = 0;
1168         vdata->delete_deleted = 0;
1169
1170         /* the list needs to be of length num_nodes */
1171         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1172                                                 struct ctdb_marshall_buffer *,
1173                                                 ctdb->num_nodes);
1174         if (vdata->vacuum_fetch_list == NULL) {
1175                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1176                 return -1;
1177         }
1178         for (i = 0; i < ctdb->num_nodes; i++) {
1179                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1180                         talloc_zero_size(vdata->vacuum_fetch_list,
1181                                          offsetof(struct ctdb_marshall_buffer, data));
1182                 if (vdata->vacuum_fetch_list[i] == NULL) {
1183                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1184                         return -1;
1185                 }
1186                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1187         }
1188
1189         return 0;
1190 }
1191
1192 /**
1193  * Vacuum a DB:
1194  *  - Always do the fast vacuuming run, which traverses
1195  *    the in-memory delete queue: these records have been
1196  *    scheduled for deletion.
1197  *  - Only if explicitly requested, the database is traversed
1198  *    in order to use the traditional heuristics on empty records
1199  *    to trigger deletion.
1200  *    This is done only every VacuumFastPathCount'th vacuuming run.
1201  *
1202  * The traverse runs fill two lists:
1203  *
1204  * - The delete_list:
1205  *   This is the list of empty records the current
1206  *   node is lmaster and dmaster for. These records are later
1207  *   deleted first on other nodes and then locally.
1208  *
1209  *   The fast vacuuming run has a short cut for those records
1210  *   that have never been migrated with data: these records
1211  *   are immediately deleted locally, since they have left
1212  *   no trace on other nodes.
1213  *
1214  * - The vacuum_fetch lists
1215  *   (one for each other lmaster node):
1216  *   The records in this list are sent for deletion to
1217  *   their lmaster in a bulk VACUUM_FETCH message.
1218  *
1219  *   The lmaster then migrates all these records to itelf
1220  *   so that they can be vacuumed there.
1221  *
1222  * This executes in the child context.
1223  */
1224 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1225                           struct vacuum_data *vdata,
1226                           bool full_vacuum_run)
1227 {
1228         struct ctdb_context *ctdb = ctdb_db->ctdb;
1229         int ret, pnn;
1230
1231         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1232                            "%s db_id[0x%08x]\n",
1233                            full_vacuum_run ? "full" : "fast",
1234                            ctdb_db->db_name, ctdb_db->db_id));
1235
1236         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1237         if (ret != 0) {
1238                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1239                 return ret;
1240         }
1241
1242         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1243         if (pnn == -1) {
1244                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1245                 return -1;
1246         }
1247
1248         ctdb->pnn = pnn;
1249
1250         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
1251         if (ret != 0) {
1252                 return ret;
1253         }
1254
1255         if (full_vacuum_run) {
1256                 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1257         }
1258
1259         ctdb_process_delete_queue(ctdb_db, vdata);
1260
1261         ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1262
1263         ctdb_process_delete_list(ctdb_db, vdata);
1264
1265         /* this ensures we run our event queue */
1266         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1267
1268         return 0;
1269 }
1270
1271
1272 /*
1273  * traverse function for repacking
1274  */
1275 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
1276                            void *private_data)
1277 {
1278         struct vacuum_data *vdata = (struct vacuum_data *)private_data;
1279
1280         if (vdata->vacuum) {
1281                 uint32_t hash = ctdb_hash(&key);
1282                 struct delete_record_data *kd;
1283                 /*
1284                  * check if we can ignore this record because it's in the delete_list
1285                  */
1286                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1287                 /*
1288                  * there might be hash collisions so we have to compare the keys here to be sure
1289                  */
1290                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1291                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1292                         /*
1293                          * we have to check if the record hasn't changed in the meantime in order to
1294                          * savely remove it from the database
1295                          */
1296                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1297                                 hdr->dmaster == kd->ctdb->pnn &&
1298                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1299                                 kd->hdr.rsn == hdr->rsn) {
1300                                 vdata->vacuumed++;
1301                                 return 0;
1302                         }
1303                 }
1304         }
1305         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1306                 vdata->traverse_error = true;
1307                 return -1;
1308         }
1309         vdata->copied++;
1310         return 0;
1311 }
1312
1313 /*
1314  * repack a tdb
1315  */
1316 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1317 {
1318         struct tdb_context *tmp_db;
1319
1320         if (tdb_transaction_start(tdb) != 0) {
1321                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1322                 return -1;
1323         }
1324
1325         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1326                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1327                           O_RDWR|O_CREAT, 0);
1328         if (tmp_db == NULL) {
1329                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1330                 tdb_transaction_cancel(tdb);
1331                 return -1;
1332         }
1333
1334         vdata->traverse_error = false;
1335         vdata->dest_db = tmp_db;
1336         vdata->vacuum = true;
1337         vdata->vacuumed = 0;
1338         vdata->copied = 0;
1339
1340         /*
1341          * repack and vacuum on-the-fly by not writing the records that are
1342          * no longer needed
1343          */
1344         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1345                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1346                 tdb_transaction_cancel(tdb);
1347                 tdb_close(tmp_db);
1348                 return -1;
1349         }
1350
1351         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1352
1353         if (vdata->traverse_error) {
1354                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1355                 tdb_transaction_cancel(tdb);
1356                 tdb_close(tmp_db);
1357                 return -1;
1358         }
1359
1360         if (tdb_wipe_all(tdb) != 0) {
1361                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1362                 tdb_transaction_cancel(tdb);
1363                 tdb_close(tmp_db);
1364                 return -1;
1365         }
1366
1367         vdata->traverse_error = false;
1368         vdata->dest_db = tdb;
1369         vdata->vacuum = false;
1370         vdata->copied = 0;
1371
1372         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1373                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1374                 tdb_transaction_cancel(tdb);
1375                 tdb_close(tmp_db);
1376                 return -1;
1377         }
1378
1379         if (vdata->traverse_error) {
1380                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1381                 tdb_transaction_cancel(tdb);
1382                 tdb_close(tmp_db);
1383                 return -1;
1384         }
1385
1386         tdb_close(tmp_db);
1387
1388
1389         if (tdb_transaction_commit(tdb) != 0) {
1390                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1391                 return -1;
1392         }
1393         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1394
1395         return 0;
1396 }
1397
1398 /*
1399  * repack and vaccum a db
1400  * called from the child context
1401  */
1402 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1403                                      TALLOC_CTX *mem_ctx,
1404                                      bool full_vacuum_run)
1405 {
1406         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1407         const char *name = ctdb_db->db_name;
1408         int freelist_size = 0;
1409         struct vacuum_data *vdata;
1410
1411         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1412         if (vdata == NULL) {
1413                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1414                 return -1;
1415         }
1416
1417         vdata->ctdb = ctdb_db->ctdb;
1418         vdata->repack_limit = repack_limit;
1419         vdata->delete_list = trbt_create(vdata, 0);
1420         vdata->ctdb_db = ctdb_db;
1421         if (vdata->delete_list == NULL) {
1422                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1423                 talloc_free(vdata);
1424                 return -1;
1425         }
1426
1427         vdata->start = timeval_current();
1428
1429         /*
1430          * gather all records that can be deleted in vdata
1431          */
1432         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1433                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1434         }
1435
1436         if (repack_limit != 0) {
1437                 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1438                 if (freelist_size == -1) {
1439                         DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1440                         talloc_free(vdata);
1441                         return -1;
1442                 }
1443         }
1444
1445         /*
1446          * decide if a repack is necessary
1447          */
1448         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1449         {
1450                 talloc_free(vdata);
1451                 return 0;
1452         }
1453
1454         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1455                         name, freelist_size, vdata->delete_left));
1456
1457         /*
1458          * repack and implicitely get rid of the records we can delete
1459          */
1460         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1461                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1462                 talloc_free(vdata);
1463                 return -1;
1464         }
1465         talloc_free(vdata);
1466
1467         return 0;
1468 }
1469
1470 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1471 {
1472         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1473
1474         return interval;
1475 }
1476
1477 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1478 {
1479         double l = timeval_elapsed(&child_ctx->start_time);
1480         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1481         struct ctdb_context *ctdb = ctdb_db->ctdb;
1482
1483         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1484
1485         if (child_ctx->child_pid != -1) {
1486                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1487         } else {
1488                 /* Bump the number of successful fast-path runs. */
1489                 child_ctx->vacuum_handle->fast_path_count++;
1490         }
1491
1492         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1493
1494         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1495                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1496                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1497
1498         return 0;
1499 }
1500
1501 /*
1502  * this event is generated when a vacuum child process times out
1503  */
1504 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1505                                          struct timeval t, void *private_data)
1506 {
1507         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1508
1509         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1510
1511         child_ctx->status = VACUUM_TIMEOUT;
1512
1513         talloc_free(child_ctx);
1514 }
1515
1516
1517 /*
1518  * this event is generated when a vacuum child process has completed
1519  */
1520 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1521                              uint16_t flags, void *private_data)
1522 {
1523         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1524         char c = 0;
1525         int ret;
1526
1527         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1528         child_ctx->child_pid = -1;
1529
1530         ret = read(child_ctx->fd[0], &c, 1);
1531         if (ret != 1 || c != 0) {
1532                 child_ctx->status = VACUUM_ERROR;
1533                 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));
1534         } else {
1535                 child_ctx->status = VACUUM_OK;
1536         }
1537
1538         talloc_free(child_ctx);
1539 }
1540
1541 /*
1542  * this event is called every time we need to start a new vacuum process
1543  */
1544 static void
1545 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1546                                struct timeval t, void *private_data)
1547 {
1548         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1549         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1550         struct ctdb_context *ctdb = ctdb_db->ctdb;
1551         struct ctdb_vacuum_child_context *child_ctx;
1552         struct tevent_fd *fde;
1553         int ret;
1554
1555         /* we dont vacuum if we are in recovery mode, or db frozen */
1556         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1557             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1558                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1559                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1560                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1561                                    ? "freeze pending"
1562                                    : "frozen"));
1563                 event_add_timed(ctdb->ev, vacuum_handle,
1564                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1565                         ctdb_vacuum_event, vacuum_handle);
1566                 return;
1567         }
1568
1569         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1570         if (child_ctx == NULL) {
1571                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1572                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1573         }
1574
1575
1576         ret = pipe(child_ctx->fd);
1577         if (ret != 0) {
1578                 talloc_free(child_ctx);
1579                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1580                 event_add_timed(ctdb->ev, vacuum_handle,
1581                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1582                         ctdb_vacuum_event, vacuum_handle);
1583                 return;
1584         }
1585
1586         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1587                 vacuum_handle->fast_path_count = 0;
1588         }
1589
1590         child_ctx->child_pid = ctdb_fork(ctdb);
1591         if (child_ctx->child_pid == (pid_t)-1) {
1592                 close(child_ctx->fd[0]);
1593                 close(child_ctx->fd[1]);
1594                 talloc_free(child_ctx);
1595                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1596                 event_add_timed(ctdb->ev, vacuum_handle,
1597                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1598                         ctdb_vacuum_event, vacuum_handle);
1599                 return;
1600         }
1601
1602
1603         if (child_ctx->child_pid == 0) {
1604                 char cc = 0;
1605                 bool full_vacuum_run = false;
1606                 close(child_ctx->fd[0]);
1607
1608                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1609                 ctdb_set_process_name("ctdb_vacuum");
1610                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1611                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1612                         _exit(1);
1613                 }
1614
1615                 /*
1616                  * repack the db
1617                  */
1618                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1619                     (vacuum_handle->fast_path_count == 0))
1620                 {
1621                         full_vacuum_run = true;
1622                 }
1623                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1624                                                full_vacuum_run);
1625
1626                 write(child_ctx->fd[1], &cc, 1);
1627                 _exit(0);
1628         }
1629
1630         set_close_on_exec(child_ctx->fd[0]);
1631         close(child_ctx->fd[1]);
1632
1633         child_ctx->status = VACUUM_RUNNING;
1634         child_ctx->start_time = timeval_current();
1635
1636         DLIST_ADD(ctdb->vacuumers, child_ctx);
1637         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1638
1639         /*
1640          * Clear the fastpath vacuuming list in the parent.
1641          */
1642         talloc_free(ctdb_db->delete_queue);
1643         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1644         if (ctdb_db->delete_queue == NULL) {
1645                 /* fatal here? ... */
1646                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1647                                  "in parent context. Shutting down\n");
1648         }
1649
1650         event_add_timed(ctdb->ev, child_ctx,
1651                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1652                 vacuum_child_timeout, child_ctx);
1653
1654         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1655
1656         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1657                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1658         tevent_fd_set_auto_close(fde);
1659
1660         vacuum_handle->child_ctx = child_ctx;
1661         child_ctx->vacuum_handle = vacuum_handle;
1662 }
1663
1664 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1665 {
1666         /* Simply free them all. */
1667         while (ctdb->vacuumers) {
1668                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1669                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1670                            (int)ctdb->vacuumers->child_pid));
1671                 /* vacuum_child_destructor kills it, removes from list */
1672                 talloc_free(ctdb->vacuumers);
1673         }
1674 }
1675
1676 /* this function initializes the vacuuming context for a database
1677  * starts the vacuuming events
1678  */
1679 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1680 {
1681         if (ctdb_db->persistent != 0) {
1682                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1683                 return 0;
1684         }
1685
1686         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1687         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1688
1689         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1690         ctdb_db->vacuum_handle->fast_path_count = 0;
1691
1692         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1693                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1694                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1695
1696         return 0;
1697 }
1698
1699 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1700                                             const struct ctdb_ltdb_header *hdr,
1701                                             const TDB_DATA key)
1702 {
1703         struct delete_record_data *kd;
1704         uint32_t hash;
1705
1706         hash = (uint32_t)ctdb_hash(&key);
1707
1708         DEBUG(DEBUG_DEBUG, (__location__
1709                             " remove_record_from_delete_queue: "
1710                             "db[%s] "
1711                             "db_id[0x%08x] "
1712                             "key_hash[0x%08x] "
1713                             "lmaster[%u] "
1714                             "migrated_with_data[%s]\n",
1715                              ctdb_db->db_name, ctdb_db->db_id,
1716                              hash,
1717                              ctdb_lmaster(ctdb_db->ctdb, &key),
1718                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1719
1720         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1721         if (kd == NULL) {
1722                 DEBUG(DEBUG_DEBUG, (__location__
1723                                     " remove_record_from_delete_queue: "
1724                                     "record not in queue (hash[0x%08x])\n.",
1725                                     hash));
1726                 return;
1727         }
1728
1729         if ((kd->key.dsize != key.dsize) ||
1730             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1731         {
1732                 DEBUG(DEBUG_DEBUG, (__location__
1733                                     " remove_record_from_delete_queue: "
1734                                     "hash collision for key with hash[0x%08x] "
1735                                     "in db[%s] - skipping\n",
1736                                     hash, ctdb_db->db_name));
1737                 return;
1738         }
1739
1740         DEBUG(DEBUG_DEBUG, (__location__
1741                             " remove_record_from_delete_queue: "
1742                             "removing key with hash[0x%08x]\n",
1743                              hash));
1744
1745         talloc_free(kd);
1746
1747         return;
1748 }
1749
1750 /**
1751  * Insert a record into the ctdb_db context's delete queue,
1752  * handling hash collisions.
1753  */
1754 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1755                                            const struct ctdb_ltdb_header *hdr,
1756                                            TDB_DATA key)
1757 {
1758         struct delete_record_data *kd;
1759         uint32_t hash;
1760         int ret;
1761
1762         hash = (uint32_t)ctdb_hash(&key);
1763
1764         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1765                            "db_id[0x%08x] "
1766                            "key_hash[0x%08x] "
1767                            "lmaster[%u] "
1768                            "migrated_with_data[%s]\n",
1769                             ctdb_db->db_name, ctdb_db->db_id,
1770                             hash,
1771                             ctdb_lmaster(ctdb_db->ctdb, &key),
1772                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1773
1774         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1775         if (kd != NULL) {
1776                 if ((kd->key.dsize != key.dsize) ||
1777                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1778                 {
1779                         DEBUG(DEBUG_INFO,
1780                               (__location__ " schedule for deletion: "
1781                                "hash collision for key hash [0x%08x]. "
1782                                "Skipping the record.\n", hash));
1783                         return 0;
1784                 } else {
1785                         DEBUG(DEBUG_DEBUG,
1786                               (__location__ " schedule for deletion: "
1787                                "updating entry for key with hash [0x%08x].\n",
1788                                hash));
1789                 }
1790         }
1791
1792         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1793                                                   ctdb_db->delete_queue,
1794                                                   hdr, key);
1795         if (ret != 0) {
1796                 DEBUG(DEBUG_INFO,
1797                       (__location__ " schedule for deletion: error "
1798                        "inserting key with hash [0x%08x] into delete queue\n",
1799                        hash));
1800                 return -1;
1801         }
1802
1803         return 0;
1804 }
1805
1806 /**
1807  * Schedule a record for deletetion.
1808  * Called from the parent context.
1809  */
1810 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1811                                            TDB_DATA indata)
1812 {
1813         struct ctdb_control_schedule_for_deletion *dd;
1814         struct ctdb_db_context *ctdb_db;
1815         int ret;
1816         TDB_DATA key;
1817
1818         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1819
1820         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1821         if (ctdb_db == NULL) {
1822                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1823                                   dd->db_id));
1824                 return -1;
1825         }
1826
1827         key.dsize = dd->keylen;
1828         key.dptr = dd->key;
1829
1830         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1831
1832         return ret;
1833 }
1834
1835 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1836                                          const struct ctdb_ltdb_header *hdr,
1837                                          TDB_DATA key)
1838 {
1839         int ret;
1840         struct ctdb_control_schedule_for_deletion *dd;
1841         TDB_DATA indata;
1842         int32_t status;
1843
1844         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1845                 /* main daemon - directly queue */
1846                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1847
1848                 return ret;
1849         }
1850
1851         /* if we dont have a connection to the daemon we can not send
1852            a control. For example sometimes from update_record control child
1853            process.
1854         */
1855         if (!ctdb_db->ctdb->can_send_controls) {
1856                 return -1;
1857         }
1858
1859
1860         /* child process: send the main daemon a control */
1861         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1862         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1863         if (indata.dptr == NULL) {
1864                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1865                 return -1;
1866         }
1867         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1868         dd->db_id = ctdb_db->db_id;
1869         dd->hdr = *hdr;
1870         dd->keylen = key.dsize;
1871         memcpy(dd->key, key.dptr, key.dsize);
1872
1873         ret = ctdb_control(ctdb_db->ctdb,
1874                            CTDB_CURRENT_NODE,
1875                            ctdb_db->db_id,
1876                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1877                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1878                            indata,
1879                            NULL, /* mem_ctx */
1880                            NULL, /* outdata */
1881                            &status,
1882                            NULL, /* timeout : NULL == wait forever */
1883                            NULL); /* error message */
1884
1885         talloc_free(indata.dptr);
1886
1887         if (ret != 0 || status != 0) {
1888                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1889                                   "SCHEDULE_FOR_DELETION "
1890                                   "control.\n"));
1891                 if (status != 0) {
1892                         ret = -1;
1893                 }
1894         }
1895
1896         return ret;
1897 }
1898
1899 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1900                                          const struct ctdb_ltdb_header *hdr,
1901                                          const TDB_DATA key)
1902 {
1903         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1904                 /*
1905                  * Only remove the record from the delete queue if called
1906                  * in the main daemon.
1907                  */
1908                 return;
1909         }
1910
1911         remove_record_from_delete_queue(ctdb_db, hdr, key);
1912
1913         return;
1914 }