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 vacuum_limit;
  57         uint32_t repack_limit;
  58         struct ctdb_context *ctdb;
  59         struct ctdb_db_context *ctdb_db;
  60         struct tdb_context *dest_db;
  61         trbt_tree_t *delete_list;
  62         uint32_t delete_count;
  63         struct ctdb_marshall_buffer **vacuum_fetch_list;
  64         struct timeval start;
  65         bool traverse_error;
  66         bool vacuum;
  67         uint32_t total;
  68         uint32_t vacuumed;
  69         uint32_t copied;
  70         uint32_t fast_added_to_vacuum_fetch_list;
  71         uint32_t fast_added_to_delete_list;
  72         uint32_t fast_deleted;
  73         uint32_t fast_skipped;
  74         uint32_t fast_error;
  75         uint32_t fast_total;
  76         uint32_t full_scheduled;
  77         uint32_t full_skipped;
  78         uint32_t full_error;
  79         uint32_t full_total;
  80         uint32_t delete_left;
  81         uint32_t delete_remote_error;
  82         uint32_t delete_local_error;
  83         uint32_t delete_deleted;
  84         uint32_t delete_skipped;
  85 };
  86
  87 /* this structure contains the information for one record to be deleted */
  88 struct delete_record_data {
  89         struct ctdb_context *ctdb;
  90         struct ctdb_db_context *ctdb_db;
  91         struct ctdb_ltdb_header hdr;
  92         TDB_DATA key;
  93         uint8_t keydata[1];
  94 };
  95
  96 struct delete_records_list {
  97         struct ctdb_marshall_buffer *records;
  98         struct vacuum_data *vdata;
  99 };
 100
 101 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 102                                            const struct ctdb_ltdb_header *hdr,
 103                                            TDB_DATA key);
 104
 105 /**
 106  * Store key and header in a tree, indexed by the key hash.
 107  */
 108 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 109                                                struct ctdb_db_context *ctdb_db,
 110                                                trbt_tree_t *tree,
 111                                                const struct ctdb_ltdb_header *hdr,
 112                                                TDB_DATA key)
 113 {
 114         struct delete_record_data *dd;
 115         uint32_t hash;
 116         size_t len;
 117
 118         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 119
 120         dd = (struct delete_record_data *)talloc_size(tree, len);
 121         if (dd == NULL) {
 122                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 123                 return -1;
 124         }
 125         talloc_set_name_const(dd, "struct delete_record_data");
 126
 127         dd->ctdb      = ctdb;
 128         dd->ctdb_db   = ctdb_db;
 129         dd->key.dsize = key.dsize;
 130         dd->key.dptr  = dd->keydata;
 131         memcpy(dd->keydata, key.dptr, key.dsize);
 132
 133         dd->hdr = *hdr;
 134
 135         hash = ctdb_hash(&key);
 136
 137         trbt_insert32(tree, hash, dd);
 138
 139         return 0;
 140 }
 141
 142 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 143                                      struct ctdb_ltdb_header *hdr)
 144 {
 145         struct ctdb_context *ctdb = vdata->ctdb;
 146         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 147         uint32_t hash;
 148         int ret;
 149
 150         hash = ctdb_hash(&key);
 151
 152         if (trbt_lookup32(vdata->delete_list, hash)) {
 153                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 154                 return 0;
 155         }
 156
 157         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 158                                                   vdata->delete_list,
 159                                                   hdr, key);
 160         if (ret != 0) {
 161                 return -1;
 162         }
 163
 164         vdata->delete_count++;
 165
 166         return 0;
 167 }
 168
 169 /**
 170  * Add a record to the list of records to be sent
 171  * to their lmaster with VACUUM_FETCH.
 172  */
 173 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 174                                            TDB_DATA key)
 175 {
 176         struct ctdb_context *ctdb = vdata->ctdb;
 177         struct ctdb_rec_data *rec;
 178         uint32_t lmaster;
 179         size_t old_size;
 180         struct ctdb_marshall_buffer *vfl;
 181
 182         lmaster = ctdb_lmaster(ctdb, &key);
 183
 184         vfl = vdata->vacuum_fetch_list[lmaster];
 185
 186         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 187         if (rec == NULL) {
 188                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 189                 vdata->traverse_error = true;
 190                 return -1;
 191         }
 192
 193         old_size = talloc_get_size(vfl);
 194         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 195         if (vfl == NULL) {
 196                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 197                 vdata->traverse_error = true;
 198                 return -1;
 199         }
 200         vdata->vacuum_fetch_list[lmaster] = vfl;
 201
 202         vfl->count++;
 203         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 204         talloc_free(rec);
 205
 206         vdata->total++;
 207
 208         return 0;
 209 }
 210
 211
 212 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 213                               struct timeval t, void *private_data);
 214
 215 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 216 {
 217         struct ctdb_ltdb_header *header =
 218                 (struct ctdb_ltdb_header *)private_data;
 219
 220         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 221                 return -1;
 222         }
 223
 224         *header = *(struct ctdb_ltdb_header *)data.dptr;
 225
 226         return 0;
 227 }
 228
 229 /*
 230  * traverse function for gathering the records that can be deleted
 231  */
 232 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 233                            void *private_data)
 234 {
 235         struct vacuum_data *vdata = talloc_get_type(private_data,
 236                                                     struct vacuum_data);
 237         struct ctdb_context *ctdb = vdata->ctdb;
 238         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 239         uint32_t lmaster;
 240         struct ctdb_ltdb_header *hdr;
 241         int res = 0;
 242
 243         vdata->full_total++;
 244
 245         lmaster = ctdb_lmaster(ctdb, &key);
 246         if (lmaster >= ctdb->num_nodes) {
 247                 vdata->full_error++;
 248                 DEBUG(DEBUG_CRIT, (__location__
 249                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 250                                    " with hash[%u]!\n",
 251                                    (unsigned)lmaster,
 252                                    (unsigned)ctdb->num_nodes,
 253                                    (unsigned)ctdb_hash(&key)));
 254                 return -1;
 255         }
 256
 257         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 258                 /* it is not a deleted record */
 259                 vdata->full_skipped++;
 260                 return 0;
 261         }
 262
 263         hdr = (struct ctdb_ltdb_header *)data.dptr;
 264
 265         if (hdr->dmaster != ctdb->pnn) {
 266                 vdata->full_skipped++;
 267                 return 0;
 268         }
 269
 270         /*
 271          * Add the record to this process's delete_queue for processing
 272          * in the subsequent traverse in the fast vacuum run.
 273          */
 274         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 275         if (res != 0) {
 276                 vdata->full_error++;
 277         } else {
 278                 vdata->full_scheduled++;
 279         }
 280
 281         return 0;
 282 }
 283
 284 /*
 285  * traverse the tree of records to delete and marshall them into
 286  * a blob
 287  */
 288 static int delete_marshall_traverse(void *param, void *data)
 289 {
 290         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 291         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 292         struct ctdb_rec_data *rec;
 293         size_t old_size;
 294
 295         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 296         if (rec == NULL) {
 297                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 298                 return 0;
 299         }
 300
 301         old_size = talloc_get_size(recs->records);
 302         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 303         if (recs->records == NULL) {
 304                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 305                 return 0;
 306         }
 307         recs->records->count++;
 308         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 309         return 0;
 310 }
 311
 312 /**
 313  * Variant of delete_marshall_traverse() that bumps the
 314  * RSN of each traversed record in the database.
 315  *
 316  * This is needed to ensure that when rolling out our
 317  * empty record copy before remote deletion, we as the
 318  * record's dmaster keep a higher RSN than the non-dmaster
 319  * nodes. This is needed to prevent old copies from
 320  * resurrection in recoveries.
 321  */
 322 static int delete_marshall_traverse_first(void *param, void *data)
 323 {
 324         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 325         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 326         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 327         struct ctdb_context *ctdb = ctdb_db->ctdb;
 328         struct ctdb_ltdb_header *header;
 329         TDB_DATA tdb_data, ctdb_data;
 330         uint32_t lmaster;
 331         uint32_t hash = ctdb_hash(&(dd->key));
 332         int res;
 333
 334         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 335         if (res != 0) {
 336                 DEBUG(DEBUG_ERR,
 337                       (__location__ " Error getting chainlock on record with "
 338                        "key hash [0x%08x] on database db[%s].\n",
 339                        hash, ctdb_db->db_name));
 340                 recs->vdata->delete_skipped++;
 341                 talloc_free(dd);
 342                 return 0;
 343         }
 344
 345         /*
 346          * Verify that the record is still empty, its RSN has not
 347          * changed and that we are still its lmaster and dmaster.
 348          */
 349
 350         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 351         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 352                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 353                                    "on database db[%s] does not exist or is not"
 354                                    " a ctdb-record.  skipping.\n",
 355                                    hash, ctdb_db->db_name));
 356                 goto skip;
 357         }
 358
 359         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 360                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 361                                    "on database db[%s] has been recycled. "
 362                                    "skipping.\n",
 363                                    hash, ctdb_db->db_name));
 364                 goto skip;
 365         }
 366
 367         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 368
 369         if (header->flags & CTDB_REC_RO_FLAGS) {
 370                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 371                                    "on database db[%s] has read-only flags. "
 372                                    "skipping.\n",
 373                                    hash, ctdb_db->db_name));
 374                 goto skip;
 375         }
 376
 377         if (header->dmaster != ctdb->pnn) {
 378                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 379                                    "on database db[%s] has been migrated away. "
 380                                    "skipping.\n",
 381                                    hash, ctdb_db->db_name));
 382                 goto skip;
 383         }
 384
 385         if (header->rsn != dd->hdr.rsn) {
 386                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 387                                    "on database db[%s] seems to have been "
 388                                    "migrated away and back again (with empty "
 389                                    "data). skipping.\n",
 390                                    hash, ctdb_db->db_name));
 391                 goto skip;
 392         }
 393
 394         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 395
 396         if (lmaster != ctdb->pnn) {
 397                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 398                                    "delete list (key hash [0x%08x], db[%s]). "
 399                                    "Strange! skipping.\n",
 400                                    hash, ctdb_db->db_name));
 401                 goto skip;
 402         }
 403
 404         /*
 405          * Increment the record's RSN to ensure the dmaster (i.e. the current
 406          * node) has the highest RSN of the record in the cluster.
 407          * This is to prevent old record copies from resurrecting in recoveries
 408          * if something should fail during the deletion process.
 409          * Note that ctdb_ltdb_store_server() increments the RSN if called
 410          * on the record's dmaster.
 411          */
 412
 413         ctdb_data.dptr = tdb_data.dptr + sizeof(struct ctdb_ltdb_header);
 414         ctdb_data.dsize = tdb_data.dsize - sizeof(struct ctdb_ltdb_header);
 415
 416         res = ctdb_ltdb_store(ctdb_db, dd->key, header, ctdb_data);
 417         if (res != 0) {
 418                 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
 419                                   "key hash [0x%08x] on database db[%s].\n",
 420                                   hash, ctdb_db->db_name));
 421                 goto skip;
 422         }
 423
 424         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 425
 426         goto done;
 427
 428 skip:
 429         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 430
 431         recs->vdata->delete_skipped++;
 432         talloc_free(dd);
 433         dd = NULL;
 434
 435 done:
 436         if (tdb_data.dptr != NULL) {
 437                 free(tdb_data.dptr);
 438         }
 439
 440         if (dd == NULL) {
 441                 return 0;
 442         }
 443
 444         return delete_marshall_traverse(param, data);
 445 }
 446
 447 /**
 448  * traverse function for the traversal of the delete_queue,
 449  * the fast-path vacuuming list.
 450  *
 451  *  - If the record has been migrated off the node
 452  *    or has been revived (filled with data) on the node,
 453  *    then skip the record.
 454  *
 455  *  - If the current node is the record's lmaster and it is
 456  *    a record that has never been migrated with data, then
 457  *    delete the record from the local tdb.
 458  *
 459  *  - If the current node is the record's lmaster and it has
 460  *    been migrated with data, then schedule it for the normal
 461  *    vacuuming procedure (i.e. add it to the delete_list).
 462  *
 463  *  - If the current node is NOT the record's lmaster then
 464  *    add it to the list of records that are to be sent to
 465  *    the lmaster with the VACUUM_FETCH message.
 466  */
 467 static int delete_queue_traverse(void *param, void *data)
 468 {
 469         struct delete_record_data *dd =
 470                 talloc_get_type(data, struct delete_record_data);
 471         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 472         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 473         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 474         int res;
 475         struct ctdb_ltdb_header header;
 476         uint32_t lmaster;
 477         uint32_t hash = ctdb_hash(&(dd->key));
 478
 479         vdata->fast_total++;
 480
 481         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 482         if (res != 0) {
 483                 DEBUG(DEBUG_ERR,
 484                       (__location__ " Error getting chainlock on record with "
 485                        "key hash [0x%08x] on database db[%s].\n",
 486                        hash, ctdb_db->db_name));
 487                 vdata->fast_error++;
 488                 return 0;
 489         }
 490
 491         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 492                                vacuum_record_parser, &header);
 493         if (res != 0) {
 494                 goto skipped;
 495         }
 496
 497         if (header.dmaster != ctdb->pnn) {
 498                 /* The record has been migrated off the node. Skip. */
 499                 goto skipped;
 500         }
 501
 502         if (header.rsn != dd->hdr.rsn) {
 503                 /*
 504                  * The record has been migrated off the node and back again.
 505                  * But not requeued for deletion. Skip it.
 506                  */
 507                 goto skipped;
 508         }
 509
 510         /*
 511          * We are dmaster, and the record has no data, and it has
 512          * not been migrated after it has been queued for deletion.
 513          *
 514          * At this stage, the record could still have been revived locally
 515          * and last been written with empty data. This can only be
 516          * fixed with the addition of an active or delete flag. (TODO)
 517          */
 518
 519         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 520
 521         if (lmaster != ctdb->pnn) {
 522                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 523
 524                 if (res != 0) {
 525                         DEBUG(DEBUG_ERR,
 526                               (__location__ " Error adding record to list "
 527                                "of records to send to lmaster.\n"));
 528                         vdata->fast_error++;
 529                 } else {
 530                         vdata->fast_added_to_vacuum_fetch_list++;
 531                 }
 532                 goto done;
 533         }
 534
 535         /* use header->flags or dd->hdr.flags ?? */
 536         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 537                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 538
 539                 if (res != 0) {
 540                         DEBUG(DEBUG_ERR,
 541                               (__location__ " Error adding record to list "
 542                                "of records for deletion on lmaster.\n"));
 543                         vdata->fast_error++;
 544                 } else {
 545                         vdata->fast_added_to_delete_list++;
 546                 }
 547         } else {
 548                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 549
 550                 if (res != 0) {
 551                         DEBUG(DEBUG_ERR,
 552                               (__location__ " Error deleting record with key "
 553                                "hash [0x%08x] from local data base db[%s].\n",
 554                                hash, ctdb_db->db_name));
 555                         vdata->fast_error++;
 556                         goto done;
 557                 }
 558
 559                 DEBUG(DEBUG_DEBUG,
 560                       (__location__ " Deleted record with key hash "
 561                        "[0x%08x] from local data base db[%s].\n",
 562                        hash, ctdb_db->db_name));
 563                 vdata->fast_deleted++;
 564         }
 565
 566         goto done;
 567
 568 skipped:
 569         vdata->fast_skipped++;
 570
 571 done:
 572         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 573
 574         return 0;
 575 }
 576
 577 /**
 578  * Delete the records that we are lmaster and dmaster for and
 579  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 580  * control.
 581  */
 582 static int delete_record_traverse(void *param, void *data)
 583 {
 584         struct delete_record_data *dd =
 585                 talloc_get_type(data, struct delete_record_data);
 586         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 587         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 588         struct ctdb_context *ctdb = ctdb_db->ctdb;
 589         int res;
 590         struct ctdb_ltdb_header header;
 591         uint32_t lmaster;
 592         uint32_t hash = ctdb_hash(&(dd->key));
 593
 594         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 595         if (res != 0) {
 596                 DEBUG(DEBUG_ERR,
 597                       (__location__ " Error getting chainlock on record with "
 598                        "key hash [0x%08x] on database db[%s].\n",
 599                        hash, ctdb_db->db_name));
 600                 vdata->delete_local_error++;
 601                 vdata->delete_left--;
 602                 talloc_free(dd);
 603                 return 0;
 604         }
 605
 606         /*
 607          * Verify that the record is still empty, its RSN has not
 608          * changed and that we are still its lmaster and dmaster.
 609          */
 610
 611         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 612                                vacuum_record_parser, &header);
 613         if (res != 0) {
 614                 goto skip;
 615         }
 616
 617         if (header.flags & CTDB_REC_RO_FLAGS) {
 618                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 619                                    "on database db[%s] has read-only flags. "
 620                                    "skipping.\n",
 621                                    hash, ctdb_db->db_name));
 622                 goto skip;
 623         }
 624
 625         if (header.dmaster != ctdb->pnn) {
 626                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 627                                    "on database db[%s] has been migrated away. "
 628                                    "skipping.\n",
 629                                    hash, ctdb_db->db_name));
 630                 goto skip;
 631         }
 632
 633         if (header.rsn != dd->hdr.rsn + 1) {
 634                 /*
 635                  * The record has been migrated off the node and back again.
 636                  * But not requeued for deletion. Skip it.
 637                  * (Note that the first marshall traverse has bumped the RSN
 638                  *  on disk.)
 639                  */
 640                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 641                                    "on database db[%s] seems to have been "
 642                                    "migrated away and back again (with empty "
 643                                    "data). skipping.\n",
 644                                    hash, ctdb_db->db_name));
 645                 goto skip;
 646         }
 647
 648         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 649
 650         if (lmaster != ctdb->pnn) {
 651                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 652                                    "delete list (key hash [0x%08x], db[%s]). "
 653                                    "Strange! skipping.\n",
 654                                    hash, ctdb_db->db_name));
 655                 goto skip;
 656         }
 657
 658         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 659
 660         if (res != 0) {
 661                 DEBUG(DEBUG_ERR,
 662                       (__location__ " Error deleting record with key hash "
 663                        "[0x%08x] from local data base db[%s].\n",
 664                        hash, ctdb_db->db_name));
 665                 vdata->delete_local_error++;
 666                 goto done;
 667         }
 668
 669         DEBUG(DEBUG_DEBUG,
 670               (__location__ " Deleted record with key hash [0x%08x] from "
 671                "local data base db[%s].\n", hash, ctdb_db->db_name));
 672
 673         vdata->delete_deleted++;
 674         goto done;
 675
 676 skip:
 677         vdata->delete_skipped++;
 678
 679 done:
 680         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 681
 682         talloc_free(dd);
 683         vdata->delete_left--;
 684
 685         return 0;
 686 }
 687
 688 /**
 689  * Traverse the delete_queue.
 690  * Records are either deleted directly or filled
 691  * into the delete list or the vacuum fetch lists
 692  * for further processing.
 693  */
 694 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 695                                       struct vacuum_data *vdata)
 696 {
 697         uint32_t sum;
 698
 699         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 700
 701         sum = vdata->fast_deleted
 702             + vdata->fast_skipped
 703             + vdata->fast_error
 704             + vdata->fast_added_to_delete_list
 705             + vdata->fast_added_to_vacuum_fetch_list;
 706
 707         if (vdata->fast_total != sum) {
 708                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 709                       "counts for db[%s]: total[%u] != sum[%u]\n",
 710                       ctdb_db->db_name, (unsigned)vdata->fast_total,
 711                       (unsigned)sum));
 712         }
 713
 714         if (vdata->fast_total > 0) {
 715                 DEBUG(DEBUG_INFO,
 716                       (__location__
 717                        " fast vacuuming delete_queue traverse statistics: "
 718                        "db[%s] "
 719                        "total[%u] "
 720                        "del[%u] "
 721                        "skp[%u] "
 722                        "err[%u] "
 723                        "adl[%u] "
 724                        "avf[%u]\n",
 725                        ctdb_db->db_name,
 726                        (unsigned)vdata->fast_total,
 727                        (unsigned)vdata->fast_deleted,
 728                        (unsigned)vdata->fast_skipped,
 729                        (unsigned)vdata->fast_error,
 730                        (unsigned)vdata->fast_added_to_delete_list,
 731                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 732         }
 733
 734         return;
 735 }
 736
 737 /**
 738  * read-only traverse of the database, looking for records that
 739  * might be able to be vacuumed.
 740  *
 741  * This is not done each time but only every tunable
 742  * VacuumFastPathCount times.
 743  */
 744 static int ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 745                                    struct vacuum_data *vdata)
 746 {
 747         int ret;
 748
 749         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 750         if (ret == -1 || vdata->traverse_error) {
 751                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 752                                   "'%s'\n", ctdb_db->db_name));
 753                 return -1;
 754         }
 755
 756         if (vdata->full_total > 0) {
 757                 DEBUG(DEBUG_INFO,
 758                       (__location__
 759                        " full vacuuming db traverse statistics: "
 760                        "db[%s] "
 761                        "total[%u] "
 762                        "skp[%u] "
 763                        "err[%u] "
 764                        "sched[%u]\n",
 765                        ctdb_db->db_name,
 766                        (unsigned)vdata->full_total,
 767                        (unsigned)vdata->full_skipped,
 768                        (unsigned)vdata->full_error,
 769                        (unsigned)vdata->full_scheduled));
 770         }
 771
 772         return 0;
 773 }
 774
 775 /**
 776  * Process the vacuum fetch lists:
 777  * For records for which we are not the lmaster, tell the lmaster to
 778  * fetch the record.
 779  */
 780 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 781                                            struct vacuum_data *vdata)
 782 {
 783         int i;
 784         struct ctdb_context *ctdb = ctdb_db->ctdb;
 785
 786         for (i = 0; i < ctdb->num_nodes; i++) {
 787                 TDB_DATA data;
 788                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 789
 790                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 791                         continue;
 792                 }
 793
 794                 if (vfl->count == 0) {
 795                         continue;
 796                 }
 797
 798                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 799                                    vfl->count, ctdb->nodes[i]->pnn,
 800                                    ctdb_db->db_name));
 801
 802                 data.dsize = talloc_get_size(vfl);
 803                 data.dptr  = (void *)vfl;
 804                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 805                                              CTDB_SRVID_VACUUM_FETCH,
 806                                              data) != 0)
 807                 {
 808                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 809                                           "fetch message to %u\n",
 810                                           ctdb->nodes[i]->pnn));
 811                         return -1;
 812                 }
 813         }
 814
 815         return 0;
 816 }
 817
 818 /**
 819  * Process the delete list:
 820  *
 821  * This is the last step of vacuuming that consistently deletes
 822  * those records that have been migrated with data and can hence
 823  * not be deleted when leaving a node.
 824  *
 825  * In this step, the lmaster does the final deletion of those empty
 826  * records that it is also dmaster for. It has ususally received
 827  * at least some of these records previously from the former dmasters
 828  * with the vacuum fetch message.
 829  *
 830  * This last step is implemented as a 3-phase process to protect from
 831  * races leading to data corruption:
 832  *
 833  *  1) Send the lmaster's copy to all other active nodes with the
 834  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 835  *  2) Send the records that could successfully be stored remotely
 836  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 837  *     control. The remote notes delete their local copy.
 838  *  3) The lmaster locally deletes its copies of all records that
 839  *     could successfully be deleted remotely in step #2.
 840  */
 841 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 842                                     struct vacuum_data *vdata)
 843 {
 844         int ret, i;
 845         struct ctdb_context *ctdb = ctdb_db->ctdb;
 846         struct delete_records_list *recs;
 847         TDB_DATA indata;
 848         struct ctdb_node_map *nodemap;
 849         uint32_t *active_nodes;
 850         int num_active_nodes;
 851         TALLOC_CTX *tmp_ctx;
 852
 853         if (vdata->delete_count == 0) {
 854                 return 0;
 855         }
 856
 857         tmp_ctx = talloc_new(vdata);
 858         if (tmp_ctx == NULL) {
 859                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 860                 return 0;
 861         }
 862
 863         vdata->delete_left = vdata->delete_count;
 864
 865         /*
 866          * get the list of currently active nodes
 867          */
 868
 869         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 870                                    CTDB_CURRENT_NODE,
 871                                    tmp_ctx,
 872                                    &nodemap);
 873         if (ret != 0) {
 874                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 875                 ret = -1;
 876                 goto done;
 877         }
 878
 879         active_nodes = list_of_active_nodes(ctdb, nodemap,
 880                                             nodemap, /* talloc context */
 881                                             false /* include self */);
 882         /* yuck! ;-) */
 883         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 884
 885         /*
 886          * Now delete the records all active nodes in a three-phase process:
 887          * 1) send all active remote nodes the current empty copy with this
 888          *    node as DMASTER
 889          * 2) if all nodes could store the new copy,
 890          *    tell all the active remote nodes to delete all their copy
 891          * 3) if all remote nodes deleted their record copy, delete it locally
 892          */
 893
 894         /*
 895          * Step 1:
 896          * Send currently empty record copy to all active nodes for storing.
 897          */
 898
 899         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 900         if (recs == NULL) {
 901                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 902                 ret = -1;
 903                 goto done;
 904         }
 905         recs->records = (struct ctdb_marshall_buffer *)
 906                 talloc_zero_size(recs,
 907                                  offsetof(struct ctdb_marshall_buffer, data));
 908         if (recs->records == NULL) {
 909                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 910                 ret = -1;
 911                 goto done;
 912         }
 913         recs->records->db_id = ctdb_db->db_id;
 914         recs->vdata = vdata;
 915
 916         /*
 917          * traverse the tree of all records we want to delete and
 918          * create a blob we can send to the other nodes.
 919          *
 920          * We call delete_marshall_traverse_first() to bump the
 921          * records' RSNs in the database, to ensure we (as dmaster)
 922          * keep the highest RSN of the records in the cluster.
 923          */
 924         trbt_traversearray32(vdata->delete_list, 1,
 925                              delete_marshall_traverse_first, recs);
 926
 927         indata.dsize = talloc_get_size(recs->records);
 928         indata.dptr  = (void *)recs->records;
 929
 930         for (i = 0; i < num_active_nodes; i++) {
 931                 struct ctdb_marshall_buffer *records;
 932                 struct ctdb_rec_data *rec;
 933                 int32_t res;
 934                 TDB_DATA outdata;
 935
 936                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 937                                 CTDB_CONTROL_RECEIVE_RECORDS, 0,
 938                                 indata, recs, &outdata, &res,
 939                                 NULL, NULL);
 940                 if (ret != 0 || res != 0) {
 941                         DEBUG(DEBUG_ERR, ("Error storing record copies on "
 942                                           "node %u: ret[%d] res[%d]\n",
 943                                           active_nodes[i], ret, res));
 944                         ret = -1;
 945                         goto done;
 946                 }
 947
 948                 /*
 949                  * outdata contains the list of records coming back
 950                  * from the node: These are the records that the
 951                  * remote node could not store. We remove these from
 952                  * the list to process further.
 953                  */
 954                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 955                 rec = (struct ctdb_rec_data *)&records->data[0];
 956                 while (records->count-- > 1) {
 957                         TDB_DATA reckey, recdata;
 958                         struct ctdb_ltdb_header *rechdr;
 959                         struct delete_record_data *dd;
 960
 961                         reckey.dptr = &rec->data[0];
 962                         reckey.dsize = rec->keylen;
 963                         recdata.dptr = &rec->data[reckey.dsize];
 964                         recdata.dsize = rec->datalen;
 965
 966                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 967                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 968                                 ret = -1;
 969                                 goto done;
 970                         }
 971                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 972                         recdata.dptr += sizeof(*rechdr);
 973                         recdata.dsize -= sizeof(*rechdr);
 974
 975                         dd = (struct delete_record_data *)trbt_lookup32(
 976                                         vdata->delete_list,
 977                                         ctdb_hash(&reckey));
 978                         if (dd != NULL) {
 979                                 /*
 980                                  * The other node could not store the record
 981                                  * copy and it is the first node that failed.
 982                                  * So we should remove it from the tree and
 983                                  * update statistics.
 984                                  */
 985                                 talloc_free(dd);
 986                                 vdata->delete_remote_error++;
 987                                 vdata->delete_left--;
 988                         }
 989
 990                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 991                 }
 992         }
 993
 994         if (vdata->delete_left == 0) {
 995                 goto success;
 996         }
 997
 998         /*
 999          * Step 2:
1000          * Send the remaining records to all active nodes for deletion.
1001          *
1002          * The lmaster's (i.e. our) copies of these records have been stored
1003          * successfully on the other nodes.
1004          */
1005
1006         /*
1007          * Create a marshall blob from the remaining list of records to delete.
1008          */
1009
1010         talloc_free(recs->records);
1011
1012         recs->records = (struct ctdb_marshall_buffer *)
1013                 talloc_zero_size(recs,
1014                                  offsetof(struct ctdb_marshall_buffer, data));
1015         if (recs->records == NULL) {
1016                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1017                 ret = -1;
1018                 goto done;
1019         }
1020         recs->records->db_id = ctdb_db->db_id;
1021
1022         trbt_traversearray32(vdata->delete_list, 1,
1023                              delete_marshall_traverse, recs);
1024
1025         indata.dsize = talloc_get_size(recs->records);
1026         indata.dptr  = (void *)recs->records;
1027
1028         for (i = 0; i < num_active_nodes; i++) {
1029                 struct ctdb_marshall_buffer *records;
1030                 struct ctdb_rec_data *rec;
1031                 int32_t res;
1032                 TDB_DATA outdata;
1033
1034                 ret = ctdb_control(ctdb, active_nodes[i], 0,
1035                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1036                                 indata, recs, &outdata, &res,
1037                                 NULL, NULL);
1038                 if (ret != 0 || res != 0) {
1039                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
1040                                           "node %u: ret[%d] res[%d]\n",
1041                                           active_nodes[i], ret, res));
1042                         ret = -1;
1043                         goto done;
1044                 }
1045
1046                 /*
1047                  * outdata contains the list of records coming back
1048                  * from the node: These are the records that the
1049                  * remote node could not delete. We remove these from
1050                  * the list to delete locally.
1051                  */
1052                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1053                 rec = (struct ctdb_rec_data *)&records->data[0];
1054                 while (records->count-- > 1) {
1055                         TDB_DATA reckey, recdata;
1056                         struct ctdb_ltdb_header *rechdr;
1057                         struct delete_record_data *dd;
1058
1059                         reckey.dptr = &rec->data[0];
1060                         reckey.dsize = rec->keylen;
1061                         recdata.dptr = &rec->data[reckey.dsize];
1062                         recdata.dsize = rec->datalen;
1063
1064                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1065                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1066                                 ret = -1;
1067                                 goto done;
1068                         }
1069                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1070                         recdata.dptr += sizeof(*rechdr);
1071                         recdata.dsize -= sizeof(*rechdr);
1072
1073                         dd = (struct delete_record_data *)trbt_lookup32(
1074                                         vdata->delete_list,
1075                                         ctdb_hash(&reckey));
1076                         if (dd != NULL) {
1077                                 /*
1078                                  * The other node could not delete the
1079                                  * record and it is the first node that
1080                                  * failed. So we should remove it from
1081                                  * the tree and update statistics.
1082                                  */
1083                                 talloc_free(dd);
1084                                 vdata->delete_remote_error++;
1085                                 vdata->delete_left--;
1086                         }
1087
1088                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
1089                 }
1090         }
1091
1092         if (vdata->delete_left == 0) {
1093                 goto success;
1094         }
1095
1096         /*
1097          * Step 3:
1098          * Delete the remaining records locally.
1099          *
1100          * These records have successfully been deleted on all
1101          * active remote nodes.
1102          */
1103
1104         trbt_traversearray32(vdata->delete_list, 1,
1105                              delete_record_traverse, vdata);
1106
1107 success:
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         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1410         const char *name = ctdb_db->db_name;
1411         int freelist_size = 0;
1412         struct vacuum_data *vdata;
1413
1414         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1415         if (vdata == NULL) {
1416                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1417                 return -1;
1418         }
1419
1420         vdata->ctdb = ctdb_db->ctdb;
1421         vdata->vacuum_limit = vacuum_limit;
1422         vdata->repack_limit = repack_limit;
1423         vdata->delete_list = trbt_create(vdata, 0);
1424         vdata->ctdb_db = ctdb_db;
1425         if (vdata->delete_list == NULL) {
1426                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1427                 talloc_free(vdata);
1428                 return -1;
1429         }
1430
1431         vdata->start = timeval_current();
1432
1433         /*
1434          * gather all records that can be deleted in vdata
1435          */
1436         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1437                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1438         }
1439
1440         if (repack_limit != 0) {
1441                 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1442                 if (freelist_size == -1) {
1443                         DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1444                         talloc_free(vdata);
1445                         return -1;
1446                 }
1447         }
1448
1449         /*
1450          * decide if a repack is necessary
1451          */
1452         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit) &&
1453             vdata->delete_left < vacuum_limit)
1454         {
1455                 talloc_free(vdata);
1456                 return 0;
1457         }
1458
1459         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1460                         name, freelist_size, vdata->delete_left));
1461
1462         /*
1463          * repack and implicitely get rid of the records we can delete
1464          */
1465         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1466                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1467                 talloc_free(vdata);
1468                 return -1;
1469         }
1470         talloc_free(vdata);
1471
1472         return 0;
1473 }
1474
1475 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1476 {
1477         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1478
1479         return interval;
1480 }
1481
1482 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1483 {
1484         double l = timeval_elapsed(&child_ctx->start_time);
1485         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1486         struct ctdb_context *ctdb = ctdb_db->ctdb;
1487
1488         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1489
1490         if (child_ctx->child_pid != -1) {
1491                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1492         } else {
1493                 /* Bump the number of successful fast-path runs. */
1494                 child_ctx->vacuum_handle->fast_path_count++;
1495         }
1496
1497         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1498
1499         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1500                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1501                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1502
1503         return 0;
1504 }
1505
1506 /*
1507  * this event is generated when a vacuum child process times out
1508  */
1509 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1510                                          struct timeval t, void *private_data)
1511 {
1512         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1513
1514         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1515
1516         child_ctx->status = VACUUM_TIMEOUT;
1517
1518         talloc_free(child_ctx);
1519 }
1520
1521
1522 /*
1523  * this event is generated when a vacuum child process has completed
1524  */
1525 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1526                              uint16_t flags, void *private_data)
1527 {
1528         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1529         char c = 0;
1530         int ret;
1531
1532         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1533         child_ctx->child_pid = -1;
1534
1535         ret = read(child_ctx->fd[0], &c, 1);
1536         if (ret != 1 || c != 0) {
1537                 child_ctx->status = VACUUM_ERROR;
1538                 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));
1539         } else {
1540                 child_ctx->status = VACUUM_OK;
1541         }
1542
1543         talloc_free(child_ctx);
1544 }
1545
1546 /*
1547  * this event is called every time we need to start a new vacuum process
1548  */
1549 static void
1550 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1551                                struct timeval t, void *private_data)
1552 {
1553         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1554         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1555         struct ctdb_context *ctdb = ctdb_db->ctdb;
1556         struct ctdb_vacuum_child_context *child_ctx;
1557         struct tevent_fd *fde;
1558         int ret;
1559
1560         /* we dont vacuum if we are in recovery mode, or db frozen */
1561         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1562             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1563                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1564                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1565                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1566                                    ? "freeze pending"
1567                                    : "frozen"));
1568                 event_add_timed(ctdb->ev, vacuum_handle,
1569                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1570                         ctdb_vacuum_event, vacuum_handle);
1571                 return;
1572         }
1573
1574         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1575         if (child_ctx == NULL) {
1576                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1577                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1578         }
1579
1580
1581         ret = pipe(child_ctx->fd);
1582         if (ret != 0) {
1583                 talloc_free(child_ctx);
1584                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1585                 event_add_timed(ctdb->ev, vacuum_handle,
1586                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1587                         ctdb_vacuum_event, vacuum_handle);
1588                 return;
1589         }
1590
1591         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1592                 vacuum_handle->fast_path_count = 0;
1593         }
1594
1595         child_ctx->child_pid = ctdb_fork(ctdb);
1596         if (child_ctx->child_pid == (pid_t)-1) {
1597                 close(child_ctx->fd[0]);
1598                 close(child_ctx->fd[1]);
1599                 talloc_free(child_ctx);
1600                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1601                 event_add_timed(ctdb->ev, vacuum_handle,
1602                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1603                         ctdb_vacuum_event, vacuum_handle);
1604                 return;
1605         }
1606
1607
1608         if (child_ctx->child_pid == 0) {
1609                 char cc = 0;
1610                 bool full_vacuum_run = false;
1611                 close(child_ctx->fd[0]);
1612
1613                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1614                 ctdb_set_process_name("ctdb_vacuum");
1615                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1616                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1617                         _exit(1);
1618                 }
1619
1620                 /*
1621                  * repack the db
1622                  */
1623                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1624                     (vacuum_handle->fast_path_count == 0))
1625                 {
1626                         full_vacuum_run = true;
1627                 }
1628                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1629                                                full_vacuum_run);
1630
1631                 write(child_ctx->fd[1], &cc, 1);
1632                 _exit(0);
1633         }
1634
1635         set_close_on_exec(child_ctx->fd[0]);
1636         close(child_ctx->fd[1]);
1637
1638         child_ctx->status = VACUUM_RUNNING;
1639         child_ctx->start_time = timeval_current();
1640
1641         DLIST_ADD(ctdb->vacuumers, child_ctx);
1642         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1643
1644         /*
1645          * Clear the fastpath vacuuming list in the parent.
1646          */
1647         talloc_free(ctdb_db->delete_queue);
1648         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1649         if (ctdb_db->delete_queue == NULL) {
1650                 /* fatal here? ... */
1651                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1652                                  "in parent context. Shutting down\n");
1653         }
1654
1655         event_add_timed(ctdb->ev, child_ctx,
1656                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1657                 vacuum_child_timeout, child_ctx);
1658
1659         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1660
1661         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1662                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1663         tevent_fd_set_auto_close(fde);
1664
1665         vacuum_handle->child_ctx = child_ctx;
1666         child_ctx->vacuum_handle = vacuum_handle;
1667 }
1668
1669 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1670 {
1671         /* Simply free them all. */
1672         while (ctdb->vacuumers) {
1673                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1674                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1675                            (int)ctdb->vacuumers->child_pid));
1676                 /* vacuum_child_destructor kills it, removes from list */
1677                 talloc_free(ctdb->vacuumers);
1678         }
1679 }
1680
1681 /* this function initializes the vacuuming context for a database
1682  * starts the vacuuming events
1683  */
1684 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1685 {
1686         if (ctdb_db->persistent != 0) {
1687                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1688                 return 0;
1689         }
1690
1691         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1692         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1693
1694         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1695         ctdb_db->vacuum_handle->fast_path_count = 0;
1696
1697         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1698                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1699                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1700
1701         return 0;
1702 }
1703
1704 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1705                                             const struct ctdb_ltdb_header *hdr,
1706                                             const TDB_DATA key)
1707 {
1708         struct delete_record_data *kd;
1709         uint32_t hash;
1710
1711         hash = (uint32_t)ctdb_hash(&key);
1712
1713         DEBUG(DEBUG_DEBUG, (__location__
1714                             " remove_record_from_delete_queue: "
1715                             "db[%s] "
1716                             "db_id[0x%08x] "
1717                             "key_hash[0x%08x] "
1718                             "lmaster[%u] "
1719                             "migrated_with_data[%s]\n",
1720                              ctdb_db->db_name, ctdb_db->db_id,
1721                              hash,
1722                              ctdb_lmaster(ctdb_db->ctdb, &key),
1723                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1724
1725         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1726         if (kd == NULL) {
1727                 DEBUG(DEBUG_DEBUG, (__location__
1728                                     " remove_record_from_delete_queue: "
1729                                     "record not in queue (hash[0x%08x])\n.",
1730                                     hash));
1731                 return;
1732         }
1733
1734         if ((kd->key.dsize != key.dsize) ||
1735             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1736         {
1737                 DEBUG(DEBUG_DEBUG, (__location__
1738                                     " remove_record_from_delete_queue: "
1739                                     "hash collision for key with hash[0x%08x] "
1740                                     "in db[%s] - skipping\n",
1741                                     hash, ctdb_db->db_name));
1742                 return;
1743         }
1744
1745         DEBUG(DEBUG_DEBUG, (__location__
1746                             " remove_record_from_delete_queue: "
1747                             "removing key with hash[0x%08x]\n",
1748                              hash));
1749
1750         talloc_free(kd);
1751
1752         return;
1753 }
1754
1755 /**
1756  * Insert a record into the ctdb_db context's delete queue,
1757  * handling hash collisions.
1758  */
1759 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1760                                            const struct ctdb_ltdb_header *hdr,
1761                                            TDB_DATA key)
1762 {
1763         struct delete_record_data *kd;
1764         uint32_t hash;
1765         int ret;
1766
1767         hash = (uint32_t)ctdb_hash(&key);
1768
1769         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1770                            "db_id[0x%08x] "
1771                            "key_hash[0x%08x] "
1772                            "lmaster[%u] "
1773                            "migrated_with_data[%s]\n",
1774                             ctdb_db->db_name, ctdb_db->db_id,
1775                             hash,
1776                             ctdb_lmaster(ctdb_db->ctdb, &key),
1777                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1778
1779         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1780         if (kd != NULL) {
1781                 if ((kd->key.dsize != key.dsize) ||
1782                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1783                 {
1784                         DEBUG(DEBUG_INFO,
1785                               (__location__ " schedule for deletion: "
1786                                "hash collision for key hash [0x%08x]. "
1787                                "Skipping the record.\n", hash));
1788                         return 0;
1789                 } else {
1790                         DEBUG(DEBUG_DEBUG,
1791                               (__location__ " schedule for deletion: "
1792                                "updating entry for key with hash [0x%08x].\n",
1793                                hash));
1794                 }
1795         }
1796
1797         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1798                                                   ctdb_db->delete_queue,
1799                                                   hdr, key);
1800         if (ret != 0) {
1801                 DEBUG(DEBUG_INFO,
1802                       (__location__ " schedule for deletion: error "
1803                        "inserting key with hash [0x%08x] into delete queue\n",
1804                        hash));
1805                 return -1;
1806         }
1807
1808         return 0;
1809 }
1810
1811 /**
1812  * Schedule a record for deletetion.
1813  * Called from the parent context.
1814  */
1815 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1816                                            TDB_DATA indata)
1817 {
1818         struct ctdb_control_schedule_for_deletion *dd;
1819         struct ctdb_db_context *ctdb_db;
1820         int ret;
1821         TDB_DATA key;
1822
1823         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1824
1825         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1826         if (ctdb_db == NULL) {
1827                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1828                                   dd->db_id));
1829                 return -1;
1830         }
1831
1832         key.dsize = dd->keylen;
1833         key.dptr = dd->key;
1834
1835         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1836
1837         return ret;
1838 }
1839
1840 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1841                                          const struct ctdb_ltdb_header *hdr,
1842                                          TDB_DATA key)
1843 {
1844         int ret;
1845         struct ctdb_control_schedule_for_deletion *dd;
1846         TDB_DATA indata;
1847         int32_t status;
1848
1849         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1850                 /* main daemon - directly queue */
1851                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1852
1853                 return ret;
1854         }
1855
1856         /* if we dont have a connection to the daemon we can not send
1857            a control. For example sometimes from update_record control child
1858            process.
1859         */
1860         if (!ctdb_db->ctdb->can_send_controls) {
1861                 return -1;
1862         }
1863
1864
1865         /* child process: send the main daemon a control */
1866         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1867         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1868         if (indata.dptr == NULL) {
1869                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1870                 return -1;
1871         }
1872         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1873         dd->db_id = ctdb_db->db_id;
1874         dd->hdr = *hdr;
1875         dd->keylen = key.dsize;
1876         memcpy(dd->key, key.dptr, key.dsize);
1877
1878         ret = ctdb_control(ctdb_db->ctdb,
1879                            CTDB_CURRENT_NODE,
1880                            ctdb_db->db_id,
1881                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1882                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1883                            indata,
1884                            NULL, /* mem_ctx */
1885                            NULL, /* outdata */
1886                            &status,
1887                            NULL, /* timeout : NULL == wait forever */
1888                            NULL); /* error message */
1889
1890         talloc_free(indata.dptr);
1891
1892         if (ret != 0 || status != 0) {
1893                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1894                                   "SCHEDULE_FOR_DELETION "
1895                                   "control.\n"));
1896                 if (status != 0) {
1897                         ret = -1;
1898                 }
1899         }
1900
1901         return ret;
1902 }
1903
1904 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1905                                          const struct ctdb_ltdb_header *hdr,
1906                                          const TDB_DATA key)
1907 {
1908         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1909                 /*
1910                  * Only remove the record from the delete queue if called
1911                  * in the main daemon.
1912                  */
1913                 return;
1914         }
1915
1916         remove_record_from_delete_queue(ctdb_db, hdr, key);
1917
1918         return;
1919 }