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_added_to_vacuum_fetch_list;
  77         uint32_t full_added_to_delete_list;
  78         uint32_t full_skipped;
  79         uint32_t full_error;
  80         uint32_t full_total;
  81         uint32_t delete_left;
  82         uint32_t delete_remote_error;
  83         uint32_t delete_local_error;
  84         uint32_t delete_deleted;
  85         uint32_t delete_skipped;
  86 };
  87
  88 /* this structure contains the information for one record to be deleted */
  89 struct delete_record_data {
  90         struct ctdb_context *ctdb;
  91         struct ctdb_db_context *ctdb_db;
  92         struct ctdb_ltdb_header hdr;
  93         TDB_DATA key;
  94         uint8_t keydata[1];
  95 };
  96
  97 struct delete_records_list {
  98         struct ctdb_marshall_buffer *records;
  99         struct vacuum_data *vdata;
 100 };
 101
 102 /**
 103  * Store key and header in a tree, indexed by the key hash.
 104  */
 105 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 106                                                struct ctdb_db_context *ctdb_db,
 107                                                trbt_tree_t *tree,
 108                                                const struct ctdb_ltdb_header *hdr,
 109                                                TDB_DATA key)
 110 {
 111         struct delete_record_data *dd;
 112         uint32_t hash;
 113         size_t len;
 114
 115         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 116
 117         dd = (struct delete_record_data *)talloc_size(tree, len);
 118         if (dd == NULL) {
 119                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 120                 return -1;
 121         }
 122         talloc_set_name_const(dd, "struct delete_record_data");
 123
 124         dd->ctdb      = ctdb;
 125         dd->ctdb_db   = ctdb_db;
 126         dd->key.dsize = key.dsize;
 127         dd->key.dptr  = dd->keydata;
 128         memcpy(dd->keydata, key.dptr, key.dsize);
 129
 130         dd->hdr = *hdr;
 131
 132         hash = ctdb_hash(&key);
 133
 134         trbt_insert32(tree, hash, dd);
 135
 136         return 0;
 137 }
 138
 139 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 140                                      struct ctdb_ltdb_header *hdr)
 141 {
 142         struct ctdb_context *ctdb = vdata->ctdb;
 143         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 144         uint32_t hash;
 145         int ret;
 146
 147         hash = ctdb_hash(&key);
 148
 149         if (trbt_lookup32(vdata->delete_list, hash)) {
 150                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 151                 return 0;
 152         }
 153
 154         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 155                                                   vdata->delete_list,
 156                                                   hdr, key);
 157         if (ret != 0) {
 158                 return -1;
 159         }
 160
 161         vdata->delete_count++;
 162
 163         return 0;
 164 }
 165
 166 /**
 167  * Add a record to the list of records to be sent
 168  * to their lmaster with VACUUM_FETCH.
 169  */
 170 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 171                                            TDB_DATA key)
 172 {
 173         struct ctdb_context *ctdb = vdata->ctdb;
 174         struct ctdb_rec_data *rec;
 175         uint32_t lmaster;
 176         size_t old_size;
 177         struct ctdb_marshall_buffer *vfl;
 178
 179         lmaster = ctdb_lmaster(ctdb, &key);
 180
 181         vfl = vdata->vacuum_fetch_list[lmaster];
 182
 183         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 184         if (rec == NULL) {
 185                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 186                 vdata->traverse_error = true;
 187                 return -1;
 188         }
 189
 190         old_size = talloc_get_size(vfl);
 191         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 192         if (vfl == NULL) {
 193                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 194                 vdata->traverse_error = true;
 195                 return -1;
 196         }
 197         vdata->vacuum_fetch_list[lmaster] = vfl;
 198
 199         vfl->count++;
 200         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 201         talloc_free(rec);
 202
 203         vdata->total++;
 204
 205         return 0;
 206 }
 207
 208
 209 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 210                               struct timeval t, void *private_data);
 211
 212
 213 /*
 214  * traverse function for gathering the records that can be deleted
 215  */
 216 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 217 {
 218         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 219         struct ctdb_context *ctdb = vdata->ctdb;
 220         uint32_t lmaster;
 221         struct ctdb_ltdb_header *hdr;
 222         int res = 0;
 223
 224         vdata->full_total++;
 225
 226         lmaster = ctdb_lmaster(ctdb, &key);
 227         if (lmaster >= ctdb->num_nodes) {
 228                 vdata->full_error++;
 229                 DEBUG(DEBUG_CRIT, (__location__
 230                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 231                                    " with hash[%u]!\n",
 232                                    (unsigned)lmaster,
 233                                    (unsigned)ctdb->num_nodes,
 234                                    (unsigned)ctdb_hash(&key)));
 235                 return -1;
 236         }
 237
 238         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 239                 /* it is not a deleted record */
 240                 vdata->full_skipped++;
 241                 return 0;
 242         }
 243
 244         hdr = (struct ctdb_ltdb_header *)data.dptr;
 245
 246         if (hdr->dmaster != ctdb->pnn) {
 247                 vdata->full_skipped++;
 248                 return 0;
 249         }
 250
 251         if (lmaster == ctdb->pnn) {
 252                 /*
 253                  * We are both lmaster and dmaster, and the record is empty.
 254                  * So we should be able to delete it.
 255                  */
 256                 res = add_record_to_delete_list(vdata, key, hdr);
 257                 if (res != 0) {
 258                         vdata->full_error++;
 259                 } else {
 260                         vdata->full_added_to_delete_list++;
 261                 }
 262         } else {
 263                 /*
 264                  * We are not lmaster.
 265                  * Add the record to the blob ready to send to the nodes.
 266                  */
 267                 res = add_record_to_vacuum_fetch_list(vdata, key);
 268                 if (res != 0) {
 269                         vdata->full_error++;
 270                 } else {
 271                         vdata->full_added_to_vacuum_fetch_list++;
 272                 }
 273         }
 274
 275         return res;
 276 }
 277
 278 /*
 279  * traverse the tree of records to delete and marshall them into
 280  * a blob
 281  */
 282 static int delete_marshall_traverse(void *param, void *data)
 283 {
 284         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 285         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 286         struct ctdb_rec_data *rec;
 287         size_t old_size;
 288
 289         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 290         if (rec == NULL) {
 291                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 292                 return 0;
 293         }
 294
 295         old_size = talloc_get_size(recs->records);
 296         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 297         if (recs->records == NULL) {
 298                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 299                 return 0;
 300         }
 301         recs->records->count++;
 302         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 303         return 0;
 304 }
 305
 306 /**
 307  * Variant of delete_marshall_traverse() that bumps the
 308  * RSN of each traversed record in the database.
 309  *
 310  * This is needed to ensure that when rolling out our
 311  * empty record copy before remote deletion, we as the
 312  * record's dmaster keep a higher RSN than the non-dmaster
 313  * nodes. This is needed to prevent old copies from
 314  * resurrection in recoveries.
 315  */
 316 static int delete_marshall_traverse_first(void *param, void *data)
 317 {
 318         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 319         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 320         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 321         struct ctdb_context *ctdb = ctdb_db->ctdb;
 322         struct ctdb_ltdb_header *header;
 323         TDB_DATA tdb_data, ctdb_data;
 324         uint32_t lmaster;
 325         uint32_t hash = ctdb_hash(&(dd->key));
 326         int res;
 327
 328         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 329         if (res != 0) {
 330                 DEBUG(DEBUG_ERR,
 331                       (__location__ " Error getting chainlock on record with "
 332                        "key hash [0x%08x] on database db[%s].\n",
 333                        hash, ctdb_db->db_name));
 334                 recs->vdata->delete_skipped++;
 335                 talloc_free(dd);
 336                 return 0;
 337         }
 338
 339         /*
 340          * Verify that the record is still empty, its RSN has not
 341          * changed and that we are still its lmaster and dmaster.
 342          */
 343
 344         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 345         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 346                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 347                                    "on database db[%s] does not exist or is not"
 348                                    " a ctdb-record.  skipping.\n",
 349                                    hash, ctdb_db->db_name));
 350                 goto skip;
 351         }
 352
 353         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 354                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 355                                    "on database db[%s] has been recycled. "
 356                                    "skipping.\n",
 357                                    hash, ctdb_db->db_name));
 358                 goto skip;
 359         }
 360
 361         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 362
 363         if (header->flags & CTDB_REC_RO_FLAGS) {
 364                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 365                                    "on database db[%s] has read-only flags. "
 366                                    "skipping.\n",
 367                                    hash, ctdb_db->db_name));
 368                 goto skip;
 369         }
 370
 371         if (header->dmaster != ctdb->pnn) {
 372                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 373                                    "on database db[%s] has been migrated away. "
 374                                    "skipping.\n",
 375                                    hash, ctdb_db->db_name));
 376                 goto skip;
 377         }
 378
 379         if (header->rsn != dd->hdr.rsn) {
 380                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 381                                    "on database db[%s] seems to have been "
 382                                    "migrated away and back again (with empty "
 383                                    "data). skipping.\n",
 384                                    hash, ctdb_db->db_name));
 385                 goto skip;
 386         }
 387
 388         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 389
 390         if (lmaster != ctdb->pnn) {
 391                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 392                                    "delete list (key hash [0x%08x], db[%s]). "
 393                                    "Strange! skipping.\n",
 394                                    hash, ctdb_db->db_name));
 395                 goto skip;
 396         }
 397
 398         /*
 399          * Increment the record's RSN to ensure the dmaster (i.e. the current
 400          * node) has the highest RSN of the record in the cluster.
 401          * This is to prevent old record copies from resurrecting in recoveries
 402          * if something should fail during the deletion process.
 403          * Note that ctdb_ltdb_store_server() increments the RSN if called
 404          * on the record's dmaster.
 405          */
 406
 407         ctdb_data.dptr = tdb_data.dptr + sizeof(struct ctdb_ltdb_header);
 408         ctdb_data.dsize = tdb_data.dsize - sizeof(struct ctdb_ltdb_header);
 409
 410         res = ctdb_ltdb_store(ctdb_db, dd->key, header, ctdb_data);
 411         if (res != 0) {
 412                 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
 413                                   "key hash [0x%08x] on database db[%s].\n",
 414                                   hash, ctdb_db->db_name));
 415                 goto skip;
 416         }
 417
 418         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 419
 420         goto done;
 421
 422 skip:
 423         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 424
 425         recs->vdata->delete_skipped++;
 426         talloc_free(dd);
 427         dd = NULL;
 428
 429 done:
 430         if (tdb_data.dptr != NULL) {
 431                 free(tdb_data.dptr);
 432         }
 433
 434         if (dd == NULL) {
 435                 return 0;
 436         }
 437
 438         return delete_marshall_traverse(param, data);
 439 }
 440
 441 /**
 442  * traverse function for the traversal of the delete_queue,
 443  * the fast-path vacuuming list.
 444  *
 445  *  - If the record has been migrated off the node
 446  *    or has been revived (filled with data) on the node,
 447  *    then skip the record.
 448  *
 449  *  - If the current node is the record's lmaster and it is
 450  *    a record that has never been migrated with data, then
 451  *    delete the record from the local tdb.
 452  *
 453  *  - If the current node is the record's lmaster and it has
 454  *    been migrated with data, then schedule it for the normal
 455  *    vacuuming procedure (i.e. add it to the delete_list).
 456  *
 457  *  - If the current node is NOT the record's lmaster then
 458  *    add it to the list of records that are to be sent to
 459  *    the lmaster with the VACUUM_FETCH message.
 460  */
 461 static int delete_queue_traverse(void *param, void *data)
 462 {
 463         struct delete_record_data *dd =
 464                 talloc_get_type(data, struct delete_record_data);
 465         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 466         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 467         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 468         int res;
 469         struct ctdb_ltdb_header *header;
 470         TDB_DATA tdb_data;
 471         uint32_t lmaster;
 472         uint32_t hash = ctdb_hash(&(dd->key));
 473
 474         vdata->fast_total++;
 475
 476         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 477         if (res != 0) {
 478                 DEBUG(DEBUG_ERR,
 479                       (__location__ " Error getting chainlock on record with "
 480                        "key hash [0x%08x] on database db[%s].\n",
 481                        hash, ctdb_db->db_name));
 482                 vdata->fast_error++;
 483                 return 0;
 484         }
 485
 486         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 487         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 488                 /* Does not exist or not a ctdb record. Skip. */
 489                 goto skipped;
 490         }
 491
 492         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 493                 /* The record has been recycled (filled with data). Skip. */
 494                 goto skipped;
 495         }
 496
 497         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 498
 499         if (header->dmaster != ctdb->pnn) {
 500                 /* The record has been migrated off the node. Skip. */
 501                 goto skipped;
 502         }
 503
 504         if (header->rsn != dd->hdr.rsn) {
 505                 /*
 506                  * The record has been migrated off the node and back again.
 507                  * But not requeued for deletion. Skip it.
 508                  */
 509                 goto skipped;
 510         }
 511
 512         /*
 513          * We are dmaster, and the record has no data, and it has
 514          * not been migrated after it has been queued for deletion.
 515          *
 516          * At this stage, the record could still have been revived locally
 517          * and last been written with empty data. This can only be
 518          * fixed with the addition of an active or delete flag. (TODO)
 519          */
 520
 521         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 522
 523         if (lmaster != ctdb->pnn) {
 524                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 525
 526                 if (res != 0) {
 527                         DEBUG(DEBUG_ERR,
 528                               (__location__ " Error adding record to list "
 529                                "of records to send to lmaster.\n"));
 530                         vdata->fast_error++;
 531                 } else {
 532                         vdata->fast_added_to_vacuum_fetch_list++;
 533                 }
 534                 goto done;
 535         }
 536
 537         /* use header->flags or dd->hdr.flags ?? */
 538         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 539                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 540
 541                 if (res != 0) {
 542                         DEBUG(DEBUG_ERR,
 543                               (__location__ " Error adding record to list "
 544                                "of records for deletion on lmaster.\n"));
 545                         vdata->fast_error++;
 546                 } else {
 547                         vdata->fast_added_to_delete_list++;
 548                 }
 549         } else {
 550                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 551
 552                 if (res != 0) {
 553                         DEBUG(DEBUG_ERR,
 554                               (__location__ " Error deleting record with key "
 555                                "hash [0x%08x] from local data base db[%s].\n",
 556                                hash, ctdb_db->db_name));
 557                         vdata->fast_error++;
 558                 } else {
 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
 567         goto done;
 568
 569 skipped:
 570         vdata->fast_skipped++;
 571
 572 done:
 573         if (tdb_data.dptr != NULL) {
 574                 free(tdb_data.dptr);
 575         }
 576         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 577
 578         return 0;
 579 }
 580
 581 /**
 582  * Delete the records that we are lmaster and dmaster for and
 583  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 584  * control.
 585  */
 586 static int delete_record_traverse(void *param, void *data)
 587 {
 588         struct delete_record_data *dd =
 589                 talloc_get_type(data, struct delete_record_data);
 590         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 591         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 592         struct ctdb_context *ctdb = ctdb_db->ctdb;
 593         int res;
 594         struct ctdb_ltdb_header *header;
 595         TDB_DATA tdb_data;
 596         uint32_t lmaster;
 597         bool deleted = false;
 598         uint32_t hash = ctdb_hash(&(dd->key));
 599
 600         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 601         if (res != 0) {
 602                 DEBUG(DEBUG_ERR,
 603                       (__location__ " Error getting chainlock on record with "
 604                        "key hash [0x%08x] on database db[%s].\n",
 605                        hash, ctdb_db->db_name));
 606                 vdata->delete_local_error++;
 607                 return 0;
 608         }
 609
 610         /*
 611          * Verify that the record is still empty, its RSN has not
 612          * changed and that we are still its lmaster and dmaster.
 613          */
 614
 615         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 616         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 617                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 618                                    "on database db[%s] does not exist or is not"
 619                                    " a ctdb-record.  skipping.\n",
 620                                    hash, ctdb_db->db_name));
 621                 vdata->delete_skipped++;
 622                 goto done;
 623         }
 624
 625         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 626                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 627                                    "on database db[%s] has been recycled. "
 628                                    "skipping.\n",
 629                                    hash, ctdb_db->db_name));
 630                 vdata->delete_skipped++;
 631                 goto done;
 632         }
 633
 634         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 635
 636         if (header->flags & CTDB_REC_RO_FLAGS) {
 637                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 638                                    "on database db[%s] has read-only flags. "
 639                                    "skipping.\n",
 640                                    hash, ctdb_db->db_name));
 641                 vdata->delete_skipped++;
 642                 goto done;
 643         }
 644
 645         if (header->dmaster != ctdb->pnn) {
 646                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 647                                    "on database db[%s] has been migrated away. "
 648                                    "skipping.\n",
 649                                    hash, ctdb_db->db_name));
 650                 vdata->delete_skipped++;
 651                 goto done;
 652         }
 653
 654         if (header->rsn != dd->hdr.rsn + 1) {
 655                 /*
 656                  * The record has been migrated off the node and back again.
 657                  * But not requeued for deletion. Skip it.
 658                  * (Note that the first marshall traverse has bumped the RSN
 659                  *  on disk.)
 660                  */
 661                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 662                                    "on database db[%s] seems to have been "
 663                                    "migrated away and back again (with empty "
 664                                    "data). skipping.\n",
 665                                    hash, ctdb_db->db_name));
 666                 vdata->delete_skipped++;
 667                 goto done;
 668         }
 669
 670         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 671
 672         if (lmaster != ctdb->pnn) {
 673                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 674                                    "delete list (key hash [0x%08x], db[%s]). "
 675                                    "Strange! skipping.\n",
 676                                    hash, ctdb_db->db_name));
 677                 vdata->delete_skipped++;
 678                 goto done;
 679         }
 680
 681         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 682
 683         if (res != 0) {
 684                 DEBUG(DEBUG_ERR,
 685                       (__location__ " Error deleting record with key hash "
 686                        "[0x%08x] from local data base db[%s].\n",
 687                        hash, ctdb_db->db_name));
 688                 vdata->delete_local_error++;
 689                 goto done;
 690         }
 691
 692         deleted = true;
 693
 694         DEBUG(DEBUG_DEBUG,
 695               (__location__ " Deleted record with key hash [0x%08x] from "
 696                "local data base db[%s].\n", hash, ctdb_db->db_name));
 697
 698 done:
 699         if (tdb_data.dptr != NULL) {
 700                 free(tdb_data.dptr);
 701         }
 702
 703         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 704
 705         if (deleted) {
 706                 /*
 707                  * successfully deleted the record locally.
 708                  * remove it from the list and update statistics.
 709                  */
 710                 talloc_free(dd);
 711                 vdata->delete_deleted++;
 712                 vdata->delete_left--;
 713         }
 714
 715         return 0;
 716 }
 717
 718 /**
 719  * Fast vacuuming run:
 720  * Traverse the delete_queue.
 721  * This fills the same lists as the database traverse.
 722  */
 723 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
 724                                 struct vacuum_data *vdata)
 725 {
 726         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 727
 728         if (vdata->fast_total > 0) {
 729                 DEBUG(DEBUG_INFO,
 730                       (__location__
 731                        " fast vacuuming delete_queue traverse statistics: "
 732                        "db[%s] "
 733                        "total[%u] "
 734                        "del[%u] "
 735                        "skp[%u] "
 736                        "err[%u] "
 737                        "adl[%u] "
 738                        "avf[%u]\n",
 739                        ctdb_db->db_name,
 740                        (unsigned)vdata->fast_total,
 741                        (unsigned)vdata->fast_deleted,
 742                        (unsigned)vdata->fast_skipped,
 743                        (unsigned)vdata->fast_error,
 744                        (unsigned)vdata->fast_added_to_delete_list,
 745                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 746         }
 747
 748         return;
 749 }
 750
 751 /**
 752  * Full vacuum run:
 753  * read-only traverse of the database, looking for records that
 754  * might be able to be vacuumed.
 755  *
 756  * This is not done each time but only every tunable
 757  * VacuumFastPathCount times.
 758  */
 759 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
 760                                struct vacuum_data *vdata,
 761                                bool full_vacuum_run)
 762 {
 763         int ret;
 764
 765         if (!full_vacuum_run) {
 766                 return 0;
 767         }
 768
 769         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 770         if (ret == -1 || vdata->traverse_error) {
 771                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 772                                   "'%s'\n", ctdb_db->db_name));
 773                 return -1;
 774         }
 775
 776         if (vdata->full_total > 0) {
 777                 DEBUG(DEBUG_INFO,
 778                       (__location__
 779                        " full vacuuming db traverse statistics: "
 780                        "db[%s] "
 781                        "total[%u] "
 782                        "skp[%u] "
 783                        "err[%u] "
 784                        "adl[%u] "
 785                        "avf[%u]\n",
 786                        ctdb_db->db_name,
 787                        (unsigned)vdata->full_total,
 788                        (unsigned)vdata->full_skipped,
 789                        (unsigned)vdata->full_error,
 790                        (unsigned)vdata->full_added_to_delete_list,
 791                        (unsigned)vdata->full_added_to_vacuum_fetch_list));
 792         }
 793
 794         return 0;
 795 }
 796
 797 /**
 798  * Process the vacuum fetch lists:
 799  * For records for which we are not the lmaster, tell the lmaster to
 800  * fetch the record.
 801  */
 802 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 803                                            struct vacuum_data *vdata)
 804 {
 805         int i;
 806         struct ctdb_context *ctdb = ctdb_db->ctdb;
 807
 808         for (i = 0; i < ctdb->num_nodes; i++) {
 809                 TDB_DATA data;
 810                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 811
 812                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 813                         continue;
 814                 }
 815
 816                 if (vfl->count == 0) {
 817                         continue;
 818                 }
 819
 820                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 821                                    vfl->count, ctdb->nodes[i]->pnn,
 822                                    ctdb_db->db_name));
 823
 824                 data.dsize = talloc_get_size(vfl);
 825                 data.dptr  = (void *)vfl;
 826                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 827                                              CTDB_SRVID_VACUUM_FETCH,
 828                                              data) != 0)
 829                 {
 830                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 831                                           "fetch message to %u\n",
 832                                           ctdb->nodes[i]->pnn));
 833                         return -1;
 834                 }
 835         }
 836
 837         return 0;
 838 }
 839
 840 /**
 841  * Process the delete list:
 842  *
 843  * This is the last step of vacuuming that consistently deletes
 844  * those records that have been migrated with data and can hence
 845  * not be deleted when leaving a node.
 846  *
 847  * In this step, the lmaster does the final deletion of those empty
 848  * records that it is also dmaster for. It has ususally received
 849  * at least some of these records previously from the former dmasters
 850  * with the vacuum fetch message.
 851  *
 852  * This last step is implemented as a 3-phase process to protect from
 853  * races leading to data corruption:
 854  *
 855  *  1) Send the lmaster's copy to all other active nodes with the
 856  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 857  *  2) Send the records that could successfully be stored remotely
 858  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 859  *     control. The remote notes delete their local copy.
 860  *  3) The lmaster locally deletes its copies of all records that
 861  *     could successfully be deleted remotely in step #2.
 862  */
 863 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 864                                     struct vacuum_data *vdata)
 865 {
 866         int ret, i;
 867         struct ctdb_context *ctdb = ctdb_db->ctdb;
 868         struct delete_records_list *recs;
 869         TDB_DATA indata;
 870         struct ctdb_node_map *nodemap;
 871         uint32_t *active_nodes;
 872         int num_active_nodes;
 873         TALLOC_CTX *tmp_ctx;
 874
 875         if (vdata->delete_count == 0) {
 876                 return 0;
 877         }
 878
 879         tmp_ctx = talloc_new(vdata);
 880         if (tmp_ctx == NULL) {
 881                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 882                 return 0;
 883         }
 884
 885         vdata->delete_left = vdata->delete_count;
 886
 887         /*
 888          * get the list of currently active nodes
 889          */
 890
 891         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 892                                    CTDB_CURRENT_NODE,
 893                                    tmp_ctx,
 894                                    &nodemap);
 895         if (ret != 0) {
 896                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 897                 ret = -1;
 898                 goto done;
 899         }
 900
 901         active_nodes = list_of_active_nodes(ctdb, nodemap,
 902                                             nodemap, /* talloc context */
 903                                             false /* include self */);
 904         /* yuck! ;-) */
 905         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 906
 907         /*
 908          * Now delete the records all active nodes in a three-phase process:
 909          * 1) send all active remote nodes the current empty copy with this
 910          *    node as DMASTER
 911          * 2) if all nodes could store the new copy,
 912          *    tell all the active remote nodes to delete all their copy
 913          * 3) if all remote nodes deleted their record copy, delete it locally
 914          */
 915
 916         /*
 917          * Step 1:
 918          * Send currently empty record copy to all active nodes for storing.
 919          */
 920
 921         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 922         if (recs == NULL) {
 923                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 924                 ret = -1;
 925                 goto done;
 926         }
 927         recs->records = (struct ctdb_marshall_buffer *)
 928                 talloc_zero_size(recs,
 929                                  offsetof(struct ctdb_marshall_buffer, data));
 930         if (recs->records == NULL) {
 931                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 932                 ret = -1;
 933                 goto done;
 934         }
 935         recs->records->db_id = ctdb_db->db_id;
 936         recs->vdata = vdata;
 937
 938         /*
 939          * traverse the tree of all records we want to delete and
 940          * create a blob we can send to the other nodes.
 941          *
 942          * We call delete_marshall_traverse_first() to bump the
 943          * records' RSNs in the database, to ensure we (as dmaster)
 944          * keep the highest RSN of the records in the cluster.
 945          */
 946         trbt_traversearray32(vdata->delete_list, 1,
 947                              delete_marshall_traverse_first, recs);
 948
 949         indata.dsize = talloc_get_size(recs->records);
 950         indata.dptr  = (void *)recs->records;
 951
 952         for (i = 0; i < num_active_nodes; i++) {
 953                 struct ctdb_marshall_buffer *records;
 954                 struct ctdb_rec_data *rec;
 955                 int32_t res;
 956                 TDB_DATA outdata;
 957
 958                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 959                                 CTDB_CONTROL_RECEIVE_RECORDS, 0,
 960                                 indata, recs, &outdata, &res,
 961                                 NULL, NULL);
 962                 if (ret != 0 || res != 0) {
 963                         DEBUG(DEBUG_ERR, ("Error storing record copies on "
 964                                           "node %u: ret[%d] res[%d]\n",
 965                                           active_nodes[i], ret, res));
 966                         ret = -1;
 967                         goto done;
 968                 }
 969
 970                 /*
 971                  * outdata contains the list of records coming back
 972                  * from the node: These are the records that the
 973                  * remote node could not store. We remove these from
 974                  * the list to process further.
 975                  */
 976                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 977                 rec = (struct ctdb_rec_data *)&records->data[0];
 978                 while (records->count-- > 1) {
 979                         TDB_DATA reckey, recdata;
 980                         struct ctdb_ltdb_header *rechdr;
 981                         struct delete_record_data *dd;
 982
 983                         reckey.dptr = &rec->data[0];
 984                         reckey.dsize = rec->keylen;
 985                         recdata.dptr = &rec->data[reckey.dsize];
 986                         recdata.dsize = rec->datalen;
 987
 988                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 989                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 990                                 ret = -1;
 991                                 goto done;
 992                         }
 993                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 994                         recdata.dptr += sizeof(*rechdr);
 995                         recdata.dsize -= sizeof(*rechdr);
 996
 997                         dd = (struct delete_record_data *)trbt_lookup32(
 998                                         vdata->delete_list,
 999                                         ctdb_hash(&reckey));
1000                         if (dd != NULL) {
1001                                 /*
1002                                  * The other node could not store the record
1003                                  * copy and it is the first node that failed.
1004                                  * So we should remove it from the tree and
1005                                  * update statistics.
1006                                  */
1007                                 talloc_free(dd);
1008                                 vdata->delete_remote_error++;
1009                                 vdata->delete_left--;
1010                         }
1011
1012                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
1013                 }
1014         }
1015
1016         if (vdata->delete_left == 0) {
1017                 goto success;
1018         }
1019
1020         /*
1021          * Step 2:
1022          * Send the remaining records to all active nodes for deletion.
1023          *
1024          * The lmaster's (i.e. our) copies of these records have been stored
1025          * successfully on the other nodes.
1026          */
1027
1028         /*
1029          * Create a marshall blob from the remaining list of records to delete.
1030          */
1031
1032         talloc_free(recs->records);
1033
1034         recs->records = (struct ctdb_marshall_buffer *)
1035                 talloc_zero_size(recs,
1036                                  offsetof(struct ctdb_marshall_buffer, data));
1037         if (recs->records == NULL) {
1038                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1039                 ret = -1;
1040                 goto done;
1041         }
1042         recs->records->db_id = ctdb_db->db_id;
1043
1044         trbt_traversearray32(vdata->delete_list, 1,
1045                              delete_marshall_traverse, recs);
1046
1047         indata.dsize = talloc_get_size(recs->records);
1048         indata.dptr  = (void *)recs->records;
1049
1050         for (i = 0; i < num_active_nodes; i++) {
1051                 struct ctdb_marshall_buffer *records;
1052                 struct ctdb_rec_data *rec;
1053                 int32_t res;
1054                 TDB_DATA outdata;
1055
1056                 ret = ctdb_control(ctdb, active_nodes[i], 0,
1057                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1058                                 indata, recs, &outdata, &res,
1059                                 NULL, NULL);
1060                 if (ret != 0 || res != 0) {
1061                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
1062                                           "node %u: ret[%d] res[%d]\n",
1063                                           active_nodes[i], ret, res));
1064                         ret = -1;
1065                         goto done;
1066                 }
1067
1068                 /*
1069                  * outdata contains the list of records coming back
1070                  * from the node: These are the records that the
1071                  * remote node could not delete. We remove these from
1072                  * the list to delete locally.
1073                  */
1074                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1075                 rec = (struct ctdb_rec_data *)&records->data[0];
1076                 while (records->count-- > 1) {
1077                         TDB_DATA reckey, recdata;
1078                         struct ctdb_ltdb_header *rechdr;
1079                         struct delete_record_data *dd;
1080
1081                         reckey.dptr = &rec->data[0];
1082                         reckey.dsize = rec->keylen;
1083                         recdata.dptr = &rec->data[reckey.dsize];
1084                         recdata.dsize = rec->datalen;
1085
1086                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1087                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1088                                 ret = -1;
1089                                 goto done;
1090                         }
1091                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1092                         recdata.dptr += sizeof(*rechdr);
1093                         recdata.dsize -= sizeof(*rechdr);
1094
1095                         dd = (struct delete_record_data *)trbt_lookup32(
1096                                         vdata->delete_list,
1097                                         ctdb_hash(&reckey));
1098                         if (dd != NULL) {
1099                                 /*
1100                                  * The other node could not delete the
1101                                  * record and it is the first node that
1102                                  * failed. So we should remove it from
1103                                  * the tree and update statistics.
1104                                  */
1105                                 talloc_free(dd);
1106                                 vdata->delete_remote_error++;
1107                                 vdata->delete_left--;
1108                         }
1109
1110                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
1111                 }
1112         }
1113
1114         if (vdata->delete_left == 0) {
1115                 goto success;
1116         }
1117
1118         /*
1119          * Step 3:
1120          * Delete the remaining records locally.
1121          *
1122          * These records have successfully been deleted on all
1123          * active remote nodes.
1124          */
1125
1126         trbt_traversearray32(vdata->delete_list, 1,
1127                              delete_record_traverse, vdata);
1128
1129 success:
1130
1131         if (vdata->delete_count > 0) {
1132                 DEBUG(DEBUG_INFO,
1133                       (__location__
1134                        " vacuum delete list statistics: "
1135                        "db[%s] "
1136                        "coll[%u] "
1137                        "rem.err[%u] "
1138                        "loc.err[%u] "
1139                        "skip[%u] "
1140                        "del[%u] "
1141                        "left[%u]\n",
1142                        ctdb_db->db_name,
1143                        (unsigned)vdata->delete_count,
1144                        (unsigned)vdata->delete_remote_error,
1145                        (unsigned)vdata->delete_local_error,
1146                        (unsigned)vdata->delete_skipped,
1147                        (unsigned)vdata->delete_deleted,
1148                        (unsigned)vdata->delete_left));
1149         }
1150
1151         ret = 0;
1152
1153 done:
1154         talloc_free(tmp_ctx);
1155
1156         return ret;
1157 }
1158
1159 /**
1160  * initialize the vacuum_data
1161  */
1162 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
1163                                         struct vacuum_data *vdata)
1164 {
1165         int i;
1166         struct ctdb_context *ctdb = ctdb_db->ctdb;
1167
1168         vdata->fast_added_to_delete_list = 0;
1169         vdata->fast_added_to_vacuum_fetch_list = 0;
1170         vdata->fast_deleted = 0;
1171         vdata->fast_skipped = 0;
1172         vdata->fast_error = 0;
1173         vdata->fast_total = 0;
1174         vdata->full_added_to_delete_list = 0;
1175         vdata->full_added_to_vacuum_fetch_list = 0;
1176         vdata->full_skipped = 0;
1177         vdata->full_error = 0;
1178         vdata->full_total = 0;
1179         vdata->delete_count = 0;
1180         vdata->delete_left = 0;
1181         vdata->delete_remote_error = 0;
1182         vdata->delete_local_error = 0;
1183         vdata->delete_skipped = 0;
1184         vdata->delete_deleted = 0;
1185
1186         /* the list needs to be of length num_nodes */
1187         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1188                                                 struct ctdb_marshall_buffer *,
1189                                                 ctdb->num_nodes);
1190         if (vdata->vacuum_fetch_list == NULL) {
1191                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1192                 return -1;
1193         }
1194         for (i = 0; i < ctdb->num_nodes; i++) {
1195                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1196                         talloc_zero_size(vdata->vacuum_fetch_list,
1197                                          offsetof(struct ctdb_marshall_buffer, data));
1198                 if (vdata->vacuum_fetch_list[i] == NULL) {
1199                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1200                         return -1;
1201                 }
1202                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1203         }
1204
1205         return 0;
1206 }
1207
1208 /**
1209  * Vacuum a DB:
1210  *  - Always do the fast vacuuming run, which traverses
1211  *    the in-memory delete queue: these records have been
1212  *    scheduled for deletion.
1213  *  - Only if explicitly requested, the database is traversed
1214  *    in order to use the traditional heuristics on empty records
1215  *    to trigger deletion.
1216  *    This is done only every VacuumFastPathCount'th vacuuming run.
1217  *
1218  * The traverse runs fill two lists:
1219  *
1220  * - The delete_list:
1221  *   This is the list of empty records the current
1222  *   node is lmaster and dmaster for. These records are later
1223  *   deleted first on other nodes and then locally.
1224  *
1225  *   The fast vacuuming run has a short cut for those records
1226  *   that have never been migrated with data: these records
1227  *   are immediately deleted locally, since they have left
1228  *   no trace on other nodes.
1229  *
1230  * - The vacuum_fetch lists
1231  *   (one for each other lmaster node):
1232  *   The records in this list are sent for deletion to
1233  *   their lmaster in a bulk VACUUM_FETCH message.
1234  *
1235  *   The lmaster then migrates all these records to itelf
1236  *   so that they can be vacuumed there.
1237  *
1238  * This executes in the child context.
1239  */
1240 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1241                           struct vacuum_data *vdata,
1242                           bool full_vacuum_run)
1243 {
1244         struct ctdb_context *ctdb = ctdb_db->ctdb;
1245         int ret, pnn;
1246
1247         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1248                            "%s db_id[0x%08x]\n",
1249                            full_vacuum_run ? "full" : "fast",
1250                            ctdb_db->db_name, ctdb_db->db_id));
1251
1252         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1253         if (ret != 0) {
1254                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1255                 return ret;
1256         }
1257
1258         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1259         if (pnn == -1) {
1260                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1261                 return -1;
1262         }
1263
1264         ctdb->pnn = pnn;
1265
1266         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
1267         if (ret != 0) {
1268                 return ret;
1269         }
1270
1271         ctdb_vacuum_db_fast(ctdb_db, vdata);
1272
1273         ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
1274         if (ret != 0) {
1275                 return ret;
1276         }
1277
1278         ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1279         if (ret != 0) {
1280                 return ret;
1281         }
1282
1283         ret = ctdb_process_delete_list(ctdb_db, vdata);
1284         if (ret != 0) {
1285                 return ret;
1286         }
1287
1288         /* this ensures we run our event queue */
1289         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1290
1291         return 0;
1292 }
1293
1294
1295 /*
1296  * traverse function for repacking
1297  */
1298 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
1299 {
1300         struct vacuum_data *vdata = (struct vacuum_data *)private;
1301
1302         if (vdata->vacuum) {
1303                 uint32_t hash = ctdb_hash(&key);
1304                 struct delete_record_data *kd;
1305                 /*
1306                  * check if we can ignore this record because it's in the delete_list
1307                  */
1308                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1309                 /*
1310                  * there might be hash collisions so we have to compare the keys here to be sure
1311                  */
1312                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1313                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1314                         /*
1315                          * we have to check if the record hasn't changed in the meantime in order to
1316                          * savely remove it from the database
1317                          */
1318                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1319                                 hdr->dmaster == kd->ctdb->pnn &&
1320                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1321                                 kd->hdr.rsn == hdr->rsn) {
1322                                 vdata->vacuumed++;
1323                                 return 0;
1324                         }
1325                 }
1326         }
1327         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1328                 vdata->traverse_error = true;
1329                 return -1;
1330         }
1331         vdata->copied++;
1332         return 0;
1333 }
1334
1335 /*
1336  * repack a tdb
1337  */
1338 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1339 {
1340         struct tdb_context *tmp_db;
1341
1342         if (tdb_transaction_start(tdb) != 0) {
1343                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1344                 return -1;
1345         }
1346
1347         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1348                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1349                           O_RDWR|O_CREAT, 0);
1350         if (tmp_db == NULL) {
1351                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1352                 tdb_transaction_cancel(tdb);
1353                 return -1;
1354         }
1355
1356         vdata->traverse_error = false;
1357         vdata->dest_db = tmp_db;
1358         vdata->vacuum = true;
1359         vdata->vacuumed = 0;
1360         vdata->copied = 0;
1361
1362         /*
1363          * repack and vacuum on-the-fly by not writing the records that are
1364          * no longer needed
1365          */
1366         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1367                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1368                 tdb_transaction_cancel(tdb);
1369                 tdb_close(tmp_db);
1370                 return -1;
1371         }
1372
1373         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1374
1375         if (vdata->traverse_error) {
1376                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1377                 tdb_transaction_cancel(tdb);
1378                 tdb_close(tmp_db);
1379                 return -1;
1380         }
1381
1382         if (tdb_wipe_all(tdb) != 0) {
1383                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1384                 tdb_transaction_cancel(tdb);
1385                 tdb_close(tmp_db);
1386                 return -1;
1387         }
1388
1389         vdata->traverse_error = false;
1390         vdata->dest_db = tdb;
1391         vdata->vacuum = false;
1392         vdata->copied = 0;
1393
1394         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1395                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1396                 tdb_transaction_cancel(tdb);
1397                 tdb_close(tmp_db);
1398                 return -1;
1399         }
1400
1401         if (vdata->traverse_error) {
1402                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1403                 tdb_transaction_cancel(tdb);
1404                 tdb_close(tmp_db);
1405                 return -1;
1406         }
1407
1408         tdb_close(tmp_db);
1409
1410
1411         if (tdb_transaction_commit(tdb) != 0) {
1412                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1413                 return -1;
1414         }
1415         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1416
1417         return 0;
1418 }
1419
1420 /*
1421  * repack and vaccum a db
1422  * called from the child context
1423  */
1424 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1425                                      TALLOC_CTX *mem_ctx,
1426                                      bool full_vacuum_run)
1427 {
1428         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1429         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1430         const char *name = ctdb_db->db_name;
1431         int freelist_size;
1432         struct vacuum_data *vdata;
1433
1434         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1435         if (freelist_size == -1) {
1436                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1437                 return -1;
1438         }
1439
1440         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1441         if (vdata == NULL) {
1442                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1443                 return -1;
1444         }
1445
1446         vdata->ctdb = ctdb_db->ctdb;
1447         vdata->vacuum_limit = vacuum_limit;
1448         vdata->repack_limit = repack_limit;
1449         vdata->delete_list = trbt_create(vdata, 0);
1450         vdata->ctdb_db = ctdb_db;
1451         if (vdata->delete_list == NULL) {
1452                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1453                 talloc_free(vdata);
1454                 return -1;
1455         }
1456
1457         vdata->start = timeval_current();
1458
1459         /*
1460          * gather all records that can be deleted in vdata
1461          */
1462         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1463                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1464         }
1465
1466         /*
1467          * decide if a repack is necessary
1468          */
1469         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1470         {
1471                 talloc_free(vdata);
1472                 return 0;
1473         }
1474
1475         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1476                         name, freelist_size, vdata->delete_left));
1477
1478         /*
1479          * repack and implicitely get rid of the records we can delete
1480          */
1481         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1482                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1483                 talloc_free(vdata);
1484                 return -1;
1485         }
1486         talloc_free(vdata);
1487
1488         return 0;
1489 }
1490
1491 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1492 {
1493         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1494
1495         return interval;
1496 }
1497
1498 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1499 {
1500         double l = timeval_elapsed(&child_ctx->start_time);
1501         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1502         struct ctdb_context *ctdb = ctdb_db->ctdb;
1503
1504         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1505
1506         if (child_ctx->child_pid != -1) {
1507                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1508         } else {
1509                 /* Bump the number of successful fast-path runs. */
1510                 child_ctx->vacuum_handle->fast_path_count++;
1511         }
1512
1513         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1514
1515         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1516                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1517                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1518
1519         return 0;
1520 }
1521
1522 /*
1523  * this event is generated when a vacuum child process times out
1524  */
1525 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1526                                          struct timeval t, void *private_data)
1527 {
1528         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1529
1530         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1531
1532         child_ctx->status = VACUUM_TIMEOUT;
1533
1534         talloc_free(child_ctx);
1535 }
1536
1537
1538 /*
1539  * this event is generated when a vacuum child process has completed
1540  */
1541 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1542                              uint16_t flags, void *private_data)
1543 {
1544         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1545         char c = 0;
1546         int ret;
1547
1548         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1549         child_ctx->child_pid = -1;
1550
1551         ret = read(child_ctx->fd[0], &c, 1);
1552         if (ret != 1 || c != 0) {
1553                 child_ctx->status = VACUUM_ERROR;
1554                 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));
1555         } else {
1556                 child_ctx->status = VACUUM_OK;
1557         }
1558
1559         talloc_free(child_ctx);
1560 }
1561
1562 /*
1563  * this event is called every time we need to start a new vacuum process
1564  */
1565 static void
1566 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1567                                struct timeval t, void *private_data)
1568 {
1569         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1570         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1571         struct ctdb_context *ctdb = ctdb_db->ctdb;
1572         struct ctdb_vacuum_child_context *child_ctx;
1573         struct tevent_fd *fde;
1574         int ret;
1575
1576         /* we dont vacuum if we are in recovery mode, or db frozen */
1577         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1578             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1579                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1580                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1581                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1582                                    ? "freeze pending"
1583                                    : "frozen"));
1584                 event_add_timed(ctdb->ev, vacuum_handle,
1585                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1586                         ctdb_vacuum_event, vacuum_handle);
1587                 return;
1588         }
1589
1590         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1591         if (child_ctx == NULL) {
1592                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1593                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1594         }
1595
1596
1597         ret = pipe(child_ctx->fd);
1598         if (ret != 0) {
1599                 talloc_free(child_ctx);
1600                 DEBUG(DEBUG_ERR, ("Failed to create pipe for 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         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1608                 vacuum_handle->fast_path_count = 0;
1609         }
1610
1611         child_ctx->child_pid = ctdb_fork(ctdb);
1612         if (child_ctx->child_pid == (pid_t)-1) {
1613                 close(child_ctx->fd[0]);
1614                 close(child_ctx->fd[1]);
1615                 talloc_free(child_ctx);
1616                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1617                 event_add_timed(ctdb->ev, vacuum_handle,
1618                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1619                         ctdb_vacuum_event, vacuum_handle);
1620                 return;
1621         }
1622
1623
1624         if (child_ctx->child_pid == 0) {
1625                 char cc = 0;
1626                 bool full_vacuum_run = false;
1627                 close(child_ctx->fd[0]);
1628
1629                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1630                 ctdb_set_process_name("ctdb_vacuum");
1631                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1632                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1633                         _exit(1);
1634                 }
1635
1636                 /*
1637                  * repack the db
1638                  */
1639                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1640                     (vacuum_handle->fast_path_count == 0))
1641                 {
1642                         full_vacuum_run = true;
1643                 }
1644                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1645                                                full_vacuum_run);
1646
1647                 write(child_ctx->fd[1], &cc, 1);
1648                 _exit(0);
1649         }
1650
1651         set_close_on_exec(child_ctx->fd[0]);
1652         close(child_ctx->fd[1]);
1653
1654         child_ctx->status = VACUUM_RUNNING;
1655         child_ctx->start_time = timeval_current();
1656
1657         DLIST_ADD(ctdb->vacuumers, child_ctx);
1658         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1659
1660         /*
1661          * Clear the fastpath vacuuming list in the parent.
1662          */
1663         talloc_free(ctdb_db->delete_queue);
1664         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1665         if (ctdb_db->delete_queue == NULL) {
1666                 /* fatal here? ... */
1667                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1668                                  "in parent context. Shutting down\n");
1669         }
1670
1671         event_add_timed(ctdb->ev, child_ctx,
1672                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1673                 vacuum_child_timeout, child_ctx);
1674
1675         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1676
1677         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1678                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1679         tevent_fd_set_auto_close(fde);
1680
1681         vacuum_handle->child_ctx = child_ctx;
1682         child_ctx->vacuum_handle = vacuum_handle;
1683 }
1684
1685 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1686 {
1687         /* Simply free them all. */
1688         while (ctdb->vacuumers) {
1689                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1690                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1691                            (int)ctdb->vacuumers->child_pid));
1692                 /* vacuum_child_destructor kills it, removes from list */
1693                 talloc_free(ctdb->vacuumers);
1694         }
1695 }
1696
1697 /* this function initializes the vacuuming context for a database
1698  * starts the vacuuming events
1699  */
1700 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1701 {
1702         if (ctdb_db->persistent != 0) {
1703                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1704                 return 0;
1705         }
1706
1707         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1708         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1709
1710         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1711         ctdb_db->vacuum_handle->fast_path_count = 0;
1712
1713         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1714                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1715                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1716
1717         return 0;
1718 }
1719
1720 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1721                                             const struct ctdb_ltdb_header *hdr,
1722                                             const TDB_DATA key)
1723 {
1724         struct delete_record_data *kd;
1725         uint32_t hash;
1726
1727         hash = (uint32_t)ctdb_hash(&key);
1728
1729         DEBUG(DEBUG_DEBUG, (__location__
1730                             " remove_record_from_delete_queue: "
1731                             "db[%s] "
1732                             "db_id[0x%08x] "
1733                             "key_hash[0x%08x] "
1734                             "lmaster[%u] "
1735                             "migrated_with_data[%s]\n",
1736                              ctdb_db->db_name, ctdb_db->db_id,
1737                              hash,
1738                              ctdb_lmaster(ctdb_db->ctdb, &key),
1739                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1740
1741         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1742         if (kd == NULL) {
1743                 DEBUG(DEBUG_DEBUG, (__location__
1744                                     " remove_record_from_delete_queue: "
1745                                     "record not in queue (hash[0x%08x])\n.",
1746                                     hash));
1747                 return;
1748         }
1749
1750         if ((kd->key.dsize != key.dsize) ||
1751             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1752         {
1753                 DEBUG(DEBUG_DEBUG, (__location__
1754                                     " remove_record_from_delete_queue: "
1755                                     "hash collision for key with hash[0x%08x] "
1756                                     "in db[%s] - skipping\n",
1757                                     hash, ctdb_db->db_name));
1758                 return;
1759         }
1760
1761         DEBUG(DEBUG_DEBUG, (__location__
1762                             " remove_record_from_delete_queue: "
1763                             "removing key with hash[0x%08x]\n",
1764                              hash));
1765
1766         talloc_free(kd);
1767
1768         return;
1769 }
1770
1771 /**
1772  * Insert a record into the ctdb_db context's delete queue,
1773  * handling hash collisions.
1774  */
1775 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1776                                            const struct ctdb_ltdb_header *hdr,
1777                                            TDB_DATA key)
1778 {
1779         struct delete_record_data *kd;
1780         uint32_t hash;
1781         int ret;
1782
1783         hash = (uint32_t)ctdb_hash(&key);
1784
1785         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1786                            "db_id[0x%08x] "
1787                            "key_hash[0x%08x] "
1788                            "lmaster[%u] "
1789                            "migrated_with_data[%s]\n",
1790                             ctdb_db->db_name, ctdb_db->db_id,
1791                             hash,
1792                             ctdb_lmaster(ctdb_db->ctdb, &key),
1793                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1794
1795         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1796         if (kd != NULL) {
1797                 if ((kd->key.dsize != key.dsize) ||
1798                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1799                 {
1800                         DEBUG(DEBUG_INFO,
1801                               (__location__ " schedule for deletion: "
1802                                "hash collision for key hash [0x%08x]. "
1803                                "Skipping the record.\n", hash));
1804                         return 0;
1805                 } else {
1806                         DEBUG(DEBUG_DEBUG,
1807                               (__location__ " schedule for deletion: "
1808                                "updating entry for key with hash [0x%08x].\n",
1809                                hash));
1810                 }
1811         }
1812
1813         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1814                                                   ctdb_db->delete_queue,
1815                                                   hdr, key);
1816         if (ret != 0) {
1817                 DEBUG(DEBUG_INFO,
1818                       (__location__ " schedule for deletion: error "
1819                        "inserting key with hash [0x%08x] into delete queue\n",
1820                        hash));
1821                 return -1;
1822         }
1823
1824         return 0;
1825 }
1826
1827 /**
1828  * Schedule a record for deletetion.
1829  * Called from the parent context.
1830  */
1831 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1832                                            TDB_DATA indata)
1833 {
1834         struct ctdb_control_schedule_for_deletion *dd;
1835         struct ctdb_db_context *ctdb_db;
1836         int ret;
1837         TDB_DATA key;
1838
1839         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1840
1841         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1842         if (ctdb_db == NULL) {
1843                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1844                                   dd->db_id));
1845                 return -1;
1846         }
1847
1848         key.dsize = dd->keylen;
1849         key.dptr = dd->key;
1850
1851         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1852
1853         return ret;
1854 }
1855
1856 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1857                                          const struct ctdb_ltdb_header *hdr,
1858                                          TDB_DATA key)
1859 {
1860         int ret;
1861         struct ctdb_control_schedule_for_deletion *dd;
1862         TDB_DATA indata;
1863         int32_t status;
1864
1865         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1866                 /* main daemon - directly queue */
1867                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1868
1869                 return ret;
1870         }
1871
1872         /* if we dont have a connection to the daemon we can not send
1873            a control. For example sometimes from update_record control child
1874            process.
1875         */
1876         if (!ctdb_db->ctdb->can_send_controls) {
1877                 return -1;
1878         }
1879
1880
1881         /* child process: send the main daemon a control */
1882         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1883         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1884         if (indata.dptr == NULL) {
1885                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1886                 return -1;
1887         }
1888         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1889         dd->db_id = ctdb_db->db_id;
1890         dd->hdr = *hdr;
1891         dd->keylen = key.dsize;
1892         memcpy(dd->key, key.dptr, key.dsize);
1893
1894         ret = ctdb_control(ctdb_db->ctdb,
1895                            CTDB_CURRENT_NODE,
1896                            ctdb_db->db_id,
1897                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1898                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1899                            indata,
1900                            NULL, /* mem_ctx */
1901                            NULL, /* outdata */
1902                            &status,
1903                            NULL, /* timeout : NULL == wait forever */
1904                            NULL); /* error message */
1905
1906         talloc_free(indata.dptr);
1907
1908         if (ret != 0 || status != 0) {
1909                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1910                                   "SCHEDULE_FOR_DELETION "
1911                                   "control.\n"));
1912                 if (status != 0) {
1913                         ret = -1;
1914                 }
1915         }
1916
1917         return ret;
1918 }
1919
1920 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1921                                          const struct ctdb_ltdb_header *hdr,
1922                                          const TDB_DATA key)
1923 {
1924         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1925                 /*
1926                  * Only remove the record from the delete queue if called
1927                  * in the main daemon.
1928                  */
1929                 return;
1930         }
1931
1932         remove_record_from_delete_queue(ctdb_db, hdr, key);
1933
1934         return;
1935 }