ctdb/server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5    Copyright (C) Michael Adam 2010-2013
   6    Copyright (C) Stefan Metzmacher 2010-2011
   7
   8    This program is free software; you can redistribute it and/or modify
   9    it under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3 of the License, or
  11    (at your option) any later version.
  12
  13    This program is distributed in the hope that it will be useful,
  14    but WITHOUT ANY WARRANTY; without even the implied warranty of
  15    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16    GNU General Public License for more details.
  17
  18    You should have received a copy of the GNU General Public License
  19    along with this program; if not, see <http://www.gnu.org/licenses/>.
  20 */
  21
  22 #include "includes.h"
  23 #include "tdb.h"
  24 #include "system/network.h"
  25 #include "system/filesys.h"
  26 #include "system/dir.h"
  27 #include "../include/ctdb_private.h"
  28 #include "db_wrap.h"
  29 #include "lib/util/dlinklist.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34
  35 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  36
  37 struct ctdb_vacuum_child_context {
  38         struct ctdb_vacuum_child_context *next, *prev;
  39         struct ctdb_vacuum_handle *vacuum_handle;
  40         /* fd child writes status to */
  41         int fd[2];
  42         pid_t child_pid;
  43         enum vacuum_child_status status;
  44         struct timeval start_time;
  45 };
  46
  47 struct ctdb_vacuum_handle {
  48         struct ctdb_db_context *ctdb_db;
  49         struct ctdb_vacuum_child_context *child_ctx;
  50         uint32_t fast_path_count;
  51 };
  52
  53
  54 /*  a list of records to possibly delete */
  55 struct vacuum_data {
  56         uint32_t repack_limit;
  57         struct ctdb_context *ctdb;
  58         struct ctdb_db_context *ctdb_db;
  59         struct tdb_context *dest_db;
  60         trbt_tree_t *delete_list;
  61         uint32_t delete_count;
  62         struct ctdb_marshall_buffer **vacuum_fetch_list;
  63         struct timeval start;
  64         bool traverse_error;
  65         bool vacuum;
  66         uint32_t total;
  67         uint32_t vacuumed;
  68         uint32_t copied;
  69         uint32_t fast_added_to_vacuum_fetch_list;
  70         uint32_t fast_added_to_delete_list;
  71         uint32_t fast_deleted;
  72         uint32_t fast_skipped;
  73         uint32_t fast_error;
  74         uint32_t fast_total;
  75         uint32_t full_scheduled;
  76         uint32_t full_skipped;
  77         uint32_t full_error;
  78         uint32_t full_total;
  79         uint32_t delete_left;
  80         uint32_t delete_remote_error;
  81         uint32_t delete_local_error;
  82         uint32_t delete_deleted;
  83         uint32_t delete_skipped;
  84 };
  85
  86 /* this structure contains the information for one record to be deleted */
  87 struct delete_record_data {
  88         struct ctdb_context *ctdb;
  89         struct ctdb_db_context *ctdb_db;
  90         struct ctdb_ltdb_header hdr;
  91         TDB_DATA key;
  92         uint8_t keydata[1];
  93 };
  94
  95 struct delete_records_list {
  96         struct ctdb_marshall_buffer *records;
  97         struct vacuum_data *vdata;
  98 };
  99
 100 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 101                                            const struct ctdb_ltdb_header *hdr,
 102                                            TDB_DATA key);
 103
 104 /**
 105  * Store key and header in a tree, indexed by the key hash.
 106  */
 107 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 108                                                struct ctdb_db_context *ctdb_db,
 109                                                trbt_tree_t *tree,
 110                                                const struct ctdb_ltdb_header *hdr,
 111                                                TDB_DATA key)
 112 {
 113         struct delete_record_data *dd;
 114         uint32_t hash;
 115         size_t len;
 116
 117         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 118
 119         dd = (struct delete_record_data *)talloc_size(tree, len);
 120         if (dd == NULL) {
 121                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 122                 return -1;
 123         }
 124         talloc_set_name_const(dd, "struct delete_record_data");
 125
 126         dd->ctdb      = ctdb;
 127         dd->ctdb_db   = ctdb_db;
 128         dd->key.dsize = key.dsize;
 129         dd->key.dptr  = dd->keydata;
 130         memcpy(dd->keydata, key.dptr, key.dsize);
 131
 132         dd->hdr = *hdr;
 133
 134         hash = ctdb_hash(&key);
 135
 136         trbt_insert32(tree, hash, dd);
 137
 138         return 0;
 139 }
 140
 141 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 142                                      struct ctdb_ltdb_header *hdr)
 143 {
 144         struct ctdb_context *ctdb = vdata->ctdb;
 145         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 146         uint32_t hash;
 147         int ret;
 148
 149         hash = ctdb_hash(&key);
 150
 151         if (trbt_lookup32(vdata->delete_list, hash)) {
 152                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 153                 return 0;
 154         }
 155
 156         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 157                                                   vdata->delete_list,
 158                                                   hdr, key);
 159         if (ret != 0) {
 160                 return -1;
 161         }
 162
 163         vdata->delete_count++;
 164
 165         return 0;
 166 }
 167
 168 /**
 169  * Add a record to the list of records to be sent
 170  * to their lmaster with VACUUM_FETCH.
 171  */
 172 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 173                                            TDB_DATA key)
 174 {
 175         struct ctdb_context *ctdb = vdata->ctdb;
 176         struct ctdb_rec_data *rec;
 177         uint32_t lmaster;
 178         size_t old_size;
 179         struct ctdb_marshall_buffer *vfl;
 180
 181         lmaster = ctdb_lmaster(ctdb, &key);
 182
 183         vfl = vdata->vacuum_fetch_list[lmaster];
 184
 185         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 186         if (rec == NULL) {
 187                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 188                 vdata->traverse_error = true;
 189                 return -1;
 190         }
 191
 192         old_size = talloc_get_size(vfl);
 193         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 194         if (vfl == NULL) {
 195                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 196                 vdata->traverse_error = true;
 197                 return -1;
 198         }
 199         vdata->vacuum_fetch_list[lmaster] = vfl;
 200
 201         vfl->count++;
 202         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 203         talloc_free(rec);
 204
 205         vdata->total++;
 206
 207         return 0;
 208 }
 209
 210
 211 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 212                               struct timeval t, void *private_data);
 213
 214 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 215 {
 216         struct ctdb_ltdb_header *header =
 217                 (struct ctdb_ltdb_header *)private_data;
 218
 219         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 220                 return -1;
 221         }
 222
 223         *header = *(struct ctdb_ltdb_header *)data.dptr;
 224
 225         return 0;
 226 }
 227
 228 /*
 229  * traverse function for gathering the records that can be deleted
 230  */
 231 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 232                            void *private_data)
 233 {
 234         struct vacuum_data *vdata = talloc_get_type(private_data,
 235                                                     struct vacuum_data);
 236         struct ctdb_context *ctdb = vdata->ctdb;
 237         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 238         uint32_t lmaster;
 239         struct ctdb_ltdb_header *hdr;
 240         int res = 0;
 241
 242         vdata->full_total++;
 243
 244         lmaster = ctdb_lmaster(ctdb, &key);
 245         if (lmaster >= ctdb->num_nodes) {
 246                 vdata->full_error++;
 247                 DEBUG(DEBUG_CRIT, (__location__
 248                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 249                                    " with hash[%u]!\n",
 250                                    (unsigned)lmaster,
 251                                    (unsigned)ctdb->num_nodes,
 252                                    (unsigned)ctdb_hash(&key)));
 253                 return -1;
 254         }
 255
 256         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 257                 /* it is not a deleted record */
 258                 vdata->full_skipped++;
 259                 return 0;
 260         }
 261
 262         hdr = (struct ctdb_ltdb_header *)data.dptr;
 263
 264         if (hdr->dmaster != ctdb->pnn) {
 265                 vdata->full_skipped++;
 266                 return 0;
 267         }
 268
 269         /*
 270          * Add the record to this process's delete_queue for processing
 271          * in the subsequent traverse in the fast vacuum run.
 272          */
 273         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 274         if (res != 0) {
 275                 vdata->full_error++;
 276         } else {
 277                 vdata->full_scheduled++;
 278         }
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * traverse the tree of records to delete and marshall them into
 285  * a blob
 286  */
 287 static int delete_marshall_traverse(void *param, void *data)
 288 {
 289         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 290         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 291         struct ctdb_rec_data *rec;
 292         size_t old_size;
 293
 294         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 295         if (rec == NULL) {
 296                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 297                 return 0;
 298         }
 299
 300         old_size = talloc_get_size(recs->records);
 301         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 302         if (recs->records == NULL) {
 303                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 304                 return 0;
 305         }
 306         recs->records->count++;
 307         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 308         return 0;
 309 }
 310
 311 /**
 312  * Variant of delete_marshall_traverse() that bumps the
 313  * RSN of each traversed record in the database.
 314  *
 315  * This is needed to ensure that when rolling out our
 316  * empty record copy before remote deletion, we as the
 317  * record's dmaster keep a higher RSN than the non-dmaster
 318  * nodes. This is needed to prevent old copies from
 319  * resurrection in recoveries.
 320  */
 321 static int delete_marshall_traverse_first(void *param, void *data)
 322 {
 323         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 324         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 325         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 326         struct ctdb_context *ctdb = ctdb_db->ctdb;
 327         struct ctdb_ltdb_header *header;
 328         TDB_DATA tdb_data, ctdb_data;
 329         uint32_t lmaster;
 330         uint32_t hash = ctdb_hash(&(dd->key));
 331         int res;
 332
 333         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 334         if (res != 0) {
 335                 DEBUG(DEBUG_ERR,
 336                       (__location__ " Error getting chainlock on record with "
 337                        "key hash [0x%08x] on database db[%s].\n",
 338                        hash, ctdb_db->db_name));
 339                 recs->vdata->delete_skipped++;
 340                 talloc_free(dd);
 341                 return 0;
 342         }
 343
 344         /*
 345          * Verify that the record is still empty, its RSN has not
 346          * changed and that we are still its lmaster and dmaster.
 347          */
 348
 349         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 350         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 351                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 352                                    "on database db[%s] does not exist or is not"
 353                                    " a ctdb-record.  skipping.\n",
 354                                    hash, ctdb_db->db_name));
 355                 goto skip;
 356         }
 357
 358         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 359                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 360                                    "on database db[%s] has been recycled. "
 361                                    "skipping.\n",
 362                                    hash, ctdb_db->db_name));
 363                 goto skip;
 364         }
 365
 366         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 367
 368         if (header->flags & CTDB_REC_RO_FLAGS) {
 369                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 370                                    "on database db[%s] has read-only flags. "
 371                                    "skipping.\n",
 372                                    hash, ctdb_db->db_name));
 373                 goto skip;
 374         }
 375
 376         if (header->dmaster != ctdb->pnn) {
 377                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 378                                    "on database db[%s] has been migrated away. "
 379                                    "skipping.\n",
 380                                    hash, ctdb_db->db_name));
 381                 goto skip;
 382         }
 383
 384         if (header->rsn != dd->hdr.rsn) {
 385                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 386                                    "on database db[%s] seems to have been "
 387                                    "migrated away and back again (with empty "
 388                                    "data). skipping.\n",
 389                                    hash, ctdb_db->db_name));
 390                 goto skip;
 391         }
 392
 393         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 394
 395         if (lmaster != ctdb->pnn) {
 396                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 397                                    "delete list (key hash [0x%08x], db[%s]). "
 398                                    "Strange! skipping.\n",
 399                                    hash, ctdb_db->db_name));
 400                 goto skip;
 401         }
 402
 403         /*
 404          * Increment the record's RSN to ensure the dmaster (i.e. the current
 405          * node) has the highest RSN of the record in the cluster.
 406          * This is to prevent old record copies from resurrecting in recoveries
 407          * if something should fail during the deletion process.
 408          * Note that ctdb_ltdb_store_server() increments the RSN if called
 409          * on the record's dmaster.
 410          */
 411
 412         ctdb_data.dptr = tdb_data.dptr + sizeof(struct ctdb_ltdb_header);
 413         ctdb_data.dsize = tdb_data.dsize - sizeof(struct ctdb_ltdb_header);
 414
 415         res = ctdb_ltdb_store(ctdb_db, dd->key, header, ctdb_data);
 416         if (res != 0) {
 417                 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
 418                                   "key hash [0x%08x] on database db[%s].\n",
 419                                   hash, ctdb_db->db_name));
 420                 goto skip;
 421         }
 422
 423         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 424
 425         goto done;
 426
 427 skip:
 428         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 429
 430         recs->vdata->delete_skipped++;
 431         talloc_free(dd);
 432         dd = NULL;
 433
 434 done:
 435         if (tdb_data.dptr != NULL) {
 436                 free(tdb_data.dptr);
 437         }
 438
 439         if (dd == NULL) {
 440                 return 0;
 441         }
 442
 443         return delete_marshall_traverse(param, data);
 444 }
 445
 446 /**
 447  * traverse function for the traversal of the delete_queue,
 448  * the fast-path vacuuming list.
 449  *
 450  *  - If the record has been migrated off the node
 451  *    or has been revived (filled with data) on the node,
 452  *    then skip the record.
 453  *
 454  *  - If the current node is the record's lmaster and it is
 455  *    a record that has never been migrated with data, then
 456  *    delete the record from the local tdb.
 457  *
 458  *  - If the current node is the record's lmaster and it has
 459  *    been migrated with data, then schedule it for the normal
 460  *    vacuuming procedure (i.e. add it to the delete_list).
 461  *
 462  *  - If the current node is NOT the record's lmaster then
 463  *    add it to the list of records that are to be sent to
 464  *    the lmaster with the VACUUM_FETCH message.
 465  */
 466 static int delete_queue_traverse(void *param, void *data)
 467 {
 468         struct delete_record_data *dd =
 469                 talloc_get_type(data, struct delete_record_data);
 470         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 471         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 472         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 473         int res;
 474         struct ctdb_ltdb_header header;
 475         uint32_t lmaster;
 476         uint32_t hash = ctdb_hash(&(dd->key));
 477
 478         vdata->fast_total++;
 479
 480         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 481         if (res != 0) {
 482                 DEBUG(DEBUG_ERR,
 483                       (__location__ " Error getting chainlock on record with "
 484                        "key hash [0x%08x] on database db[%s].\n",
 485                        hash, ctdb_db->db_name));
 486                 vdata->fast_error++;
 487                 return 0;
 488         }
 489
 490         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 491                                vacuum_record_parser, &header);
 492         if (res != 0) {
 493                 goto skipped;
 494         }
 495
 496         if (header.dmaster != ctdb->pnn) {
 497                 /* The record has been migrated off the node. Skip. */
 498                 goto skipped;
 499         }
 500
 501         if (header.rsn != dd->hdr.rsn) {
 502                 /*
 503                  * The record has been migrated off the node and back again.
 504                  * But not requeued for deletion. Skip it.
 505                  */
 506                 goto skipped;
 507         }
 508
 509         /*
 510          * We are dmaster, and the record has no data, and it has
 511          * not been migrated after it has been queued for deletion.
 512          *
 513          * At this stage, the record could still have been revived locally
 514          * and last been written with empty data. This can only be
 515          * fixed with the addition of an active or delete flag. (TODO)
 516          */
 517
 518         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 519
 520         if (lmaster != ctdb->pnn) {
 521                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 522
 523                 if (res != 0) {
 524                         DEBUG(DEBUG_ERR,
 525                               (__location__ " Error adding record to list "
 526                                "of records to send to lmaster.\n"));
 527                         vdata->fast_error++;
 528                 } else {
 529                         vdata->fast_added_to_vacuum_fetch_list++;
 530                 }
 531                 goto done;
 532         }
 533
 534         /* use header->flags or dd->hdr.flags ?? */
 535         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 536                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 537
 538                 if (res != 0) {
 539                         DEBUG(DEBUG_ERR,
 540                               (__location__ " Error adding record to list "
 541                                "of records for deletion on lmaster.\n"));
 542                         vdata->fast_error++;
 543                 } else {
 544                         vdata->fast_added_to_delete_list++;
 545                 }
 546         } else {
 547                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 548
 549                 if (res != 0) {
 550                         DEBUG(DEBUG_ERR,
 551                               (__location__ " Error deleting record with key "
 552                                "hash [0x%08x] from local data base db[%s].\n",
 553                                hash, ctdb_db->db_name));
 554                         vdata->fast_error++;
 555                         goto done;
 556                 }
 557
 558                 DEBUG(DEBUG_DEBUG,
 559                       (__location__ " Deleted record with key hash "
 560                        "[0x%08x] from local data base db[%s].\n",
 561                        hash, ctdb_db->db_name));
 562                 vdata->fast_deleted++;
 563         }
 564
 565         goto done;
 566
 567 skipped:
 568         vdata->fast_skipped++;
 569
 570 done:
 571         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 572
 573         return 0;
 574 }
 575
 576 /**
 577  * Delete the records that we are lmaster and dmaster for and
 578  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 579  * control.
 580  */
 581 static int delete_record_traverse(void *param, void *data)
 582 {
 583         struct delete_record_data *dd =
 584                 talloc_get_type(data, struct delete_record_data);
 585         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 586         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 587         struct ctdb_context *ctdb = ctdb_db->ctdb;
 588         int res;
 589         struct ctdb_ltdb_header header;
 590         uint32_t lmaster;
 591         uint32_t hash = ctdb_hash(&(dd->key));
 592
 593         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 594         if (res != 0) {
 595                 DEBUG(DEBUG_ERR,
 596                       (__location__ " Error getting chainlock on record with "
 597                        "key hash [0x%08x] on database db[%s].\n",
 598                        hash, ctdb_db->db_name));
 599                 vdata->delete_local_error++;
 600                 vdata->delete_left--;
 601                 talloc_free(dd);
 602                 return 0;
 603         }
 604
 605         /*
 606          * Verify that the record is still empty, its RSN has not
 607          * changed and that we are still its lmaster and dmaster.
 608          */
 609
 610         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 611                                vacuum_record_parser, &header);
 612         if (res != 0) {
 613                 goto skip;
 614         }
 615
 616         if (header.flags & CTDB_REC_RO_FLAGS) {
 617                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 618                                    "on database db[%s] has read-only flags. "
 619                                    "skipping.\n",
 620                                    hash, ctdb_db->db_name));
 621                 goto skip;
 622         }
 623
 624         if (header.dmaster != ctdb->pnn) {
 625                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 626                                    "on database db[%s] has been migrated away. "
 627                                    "skipping.\n",
 628                                    hash, ctdb_db->db_name));
 629                 goto skip;
 630         }
 631
 632         if (header.rsn != dd->hdr.rsn + 1) {
 633                 /*
 634                  * The record has been migrated off the node and back again.
 635                  * But not requeued for deletion. Skip it.
 636                  * (Note that the first marshall traverse has bumped the RSN
 637                  *  on disk.)
 638                  */
 639                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 640                                    "on database db[%s] seems to have been "
 641                                    "migrated away and back again (with empty "
 642                                    "data). skipping.\n",
 643                                    hash, ctdb_db->db_name));
 644                 goto skip;
 645         }
 646
 647         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 648
 649         if (lmaster != ctdb->pnn) {
 650                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 651                                    "delete list (key hash [0x%08x], db[%s]). "
 652                                    "Strange! skipping.\n",
 653                                    hash, ctdb_db->db_name));
 654                 goto skip;
 655         }
 656
 657         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 658
 659         if (res != 0) {
 660                 DEBUG(DEBUG_ERR,
 661                       (__location__ " Error deleting record with key hash "
 662                        "[0x%08x] from local data base db[%s].\n",
 663                        hash, ctdb_db->db_name));
 664                 vdata->delete_local_error++;
 665                 goto done;
 666         }
 667
 668         DEBUG(DEBUG_DEBUG,
 669               (__location__ " Deleted record with key hash [0x%08x] from "
 670                "local data base db[%s].\n", hash, ctdb_db->db_name));
 671
 672         vdata->delete_deleted++;
 673         goto done;
 674
 675 skip:
 676         vdata->delete_skipped++;
 677
 678 done:
 679         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 680
 681         talloc_free(dd);
 682         vdata->delete_left--;
 683
 684         return 0;
 685 }
 686
 687 /**
 688  * Traverse the delete_queue.
 689  * Records are either deleted directly or filled
 690  * into the delete list or the vacuum fetch lists
 691  * for further processing.
 692  */
 693 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 694                                       struct vacuum_data *vdata)
 695 {
 696         uint32_t sum;
 697
 698         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 699
 700         sum = vdata->fast_deleted
 701             + vdata->fast_skipped
 702             + vdata->fast_error
 703             + vdata->fast_added_to_delete_list
 704             + vdata->fast_added_to_vacuum_fetch_list;
 705
 706         if (vdata->fast_total != sum) {
 707                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 708                       "counts for db[%s]: total[%u] != sum[%u]\n",
 709                       ctdb_db->db_name, (unsigned)vdata->fast_total,
 710                       (unsigned)sum));
 711         }
 712
 713         if (vdata->fast_total > 0) {
 714                 DEBUG(DEBUG_INFO,
 715                       (__location__
 716                        " fast vacuuming delete_queue traverse statistics: "
 717                        "db[%s] "
 718                        "total[%u] "
 719                        "del[%u] "
 720                        "skp[%u] "
 721                        "err[%u] "
 722                        "adl[%u] "
 723                        "avf[%u]\n",
 724                        ctdb_db->db_name,
 725                        (unsigned)vdata->fast_total,
 726                        (unsigned)vdata->fast_deleted,
 727                        (unsigned)vdata->fast_skipped,
 728                        (unsigned)vdata->fast_error,
 729                        (unsigned)vdata->fast_added_to_delete_list,
 730                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 731         }
 732
 733         return;
 734 }
 735
 736 /**
 737  * read-only traverse of the database, looking for records that
 738  * might be able to be vacuumed.
 739  *
 740  * This is not done each time but only every tunable
 741  * VacuumFastPathCount times.
 742  */
 743 static int ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 744                                    struct vacuum_data *vdata)
 745 {
 746         int ret;
 747
 748         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 749         if (ret == -1 || vdata->traverse_error) {
 750                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 751                                   "'%s'\n", ctdb_db->db_name));
 752                 return -1;
 753         }
 754
 755         if (vdata->full_total > 0) {
 756                 DEBUG(DEBUG_INFO,
 757                       (__location__
 758                        " full vacuuming db traverse statistics: "
 759                        "db[%s] "
 760                        "total[%u] "
 761                        "skp[%u] "
 762                        "err[%u] "
 763                        "sched[%u]\n",
 764                        ctdb_db->db_name,
 765                        (unsigned)vdata->full_total,
 766                        (unsigned)vdata->full_skipped,
 767                        (unsigned)vdata->full_error,
 768                        (unsigned)vdata->full_scheduled));
 769         }
 770
 771         return 0;
 772 }
 773
 774 /**
 775  * Process the vacuum fetch lists:
 776  * For records for which we are not the lmaster, tell the lmaster to
 777  * fetch the record.
 778  */
 779 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 780                                            struct vacuum_data *vdata)
 781 {
 782         int i;
 783         struct ctdb_context *ctdb = ctdb_db->ctdb;
 784
 785         for (i = 0; i < ctdb->num_nodes; i++) {
 786                 TDB_DATA data;
 787                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 788
 789                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 790                         continue;
 791                 }
 792
 793                 if (vfl->count == 0) {
 794                         continue;
 795                 }
 796
 797                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 798                                    vfl->count, ctdb->nodes[i]->pnn,
 799                                    ctdb_db->db_name));
 800
 801                 data.dsize = talloc_get_size(vfl);
 802                 data.dptr  = (void *)vfl;
 803                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 804                                              CTDB_SRVID_VACUUM_FETCH,
 805                                              data) != 0)
 806                 {
 807                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 808                                           "fetch message to %u\n",
 809                                           ctdb->nodes[i]->pnn));
 810                         return -1;
 811                 }
 812         }
 813
 814         return 0;
 815 }
 816
 817 /**
 818  * Process the delete list:
 819  *
 820  * This is the last step of vacuuming that consistently deletes
 821  * those records that have been migrated with data and can hence
 822  * not be deleted when leaving a node.
 823  *
 824  * In this step, the lmaster does the final deletion of those empty
 825  * records that it is also dmaster for. It has ususally received
 826  * at least some of these records previously from the former dmasters
 827  * with the vacuum fetch message.
 828  *
 829  * This last step is implemented as a 3-phase process to protect from
 830  * races leading to data corruption:
 831  *
 832  *  1) Send the lmaster's copy to all other active nodes with the
 833  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 834  *  2) Send the records that could successfully be stored remotely
 835  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 836  *     control. The remote notes delete their local copy.
 837  *  3) The lmaster locally deletes its copies of all records that
 838  *     could successfully be deleted remotely in step #2.
 839  */
 840 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 841                                     struct vacuum_data *vdata)
 842 {
 843         int ret, i;
 844         struct ctdb_context *ctdb = ctdb_db->ctdb;
 845         struct delete_records_list *recs;
 846         TDB_DATA indata;
 847         struct ctdb_node_map *nodemap;
 848         uint32_t *active_nodes;
 849         int num_active_nodes;
 850         TALLOC_CTX *tmp_ctx;
 851         uint32_t sum;
 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_left != 0) {
1110                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
1111                       "there are %u records left for deletion after "
1112                       "processing delete list\n",
1113                       ctdb_db->db_name,
1114                       (unsigned)vdata->delete_left));
1115         }
1116
1117         sum = vdata->delete_deleted
1118             + vdata->delete_skipped
1119             + vdata->delete_remote_error
1120             + vdata->delete_local_error
1121             + vdata->delete_left;
1122
1123         if (vdata->delete_count != sum) {
1124                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
1125                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1126                       ctdb_db->db_name, (unsigned)vdata->delete_count,
1127                       (unsigned)sum));
1128         }
1129
1130         if (vdata->delete_count > 0) {
1131                 DEBUG(DEBUG_INFO,
1132                       (__location__
1133                        " vacuum delete list statistics: "
1134                        "db[%s] "
1135                        "total[%u] "
1136                        "del[%u] "
1137                        "skip[%u] "
1138                        "rem.err[%u] "
1139                        "loc.err[%u] "
1140                        "left[%u]\n",
1141                        ctdb_db->db_name,
1142                        (unsigned)vdata->delete_count,
1143                        (unsigned)vdata->delete_deleted,
1144                        (unsigned)vdata->delete_skipped,
1145                        (unsigned)vdata->delete_remote_error,
1146                        (unsigned)vdata->delete_local_error,
1147                        (unsigned)vdata->delete_left));
1148         }
1149
1150         ret = 0;
1151
1152 done:
1153         talloc_free(tmp_ctx);
1154
1155         return ret;
1156 }
1157
1158 /**
1159  * initialize the vacuum_data
1160  */
1161 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
1162                                         struct vacuum_data *vdata)
1163 {
1164         int i;
1165         struct ctdb_context *ctdb = ctdb_db->ctdb;
1166
1167         vdata->fast_added_to_delete_list = 0;
1168         vdata->fast_added_to_vacuum_fetch_list = 0;
1169         vdata->fast_deleted = 0;
1170         vdata->fast_skipped = 0;
1171         vdata->fast_error = 0;
1172         vdata->fast_total = 0;
1173         vdata->full_scheduled = 0;
1174         vdata->full_skipped = 0;
1175         vdata->full_error = 0;
1176         vdata->full_total = 0;
1177         vdata->delete_count = 0;
1178         vdata->delete_left = 0;
1179         vdata->delete_remote_error = 0;
1180         vdata->delete_local_error = 0;
1181         vdata->delete_skipped = 0;
1182         vdata->delete_deleted = 0;
1183
1184         /* the list needs to be of length num_nodes */
1185         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1186                                                 struct ctdb_marshall_buffer *,
1187                                                 ctdb->num_nodes);
1188         if (vdata->vacuum_fetch_list == NULL) {
1189                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1190                 return -1;
1191         }
1192         for (i = 0; i < ctdb->num_nodes; i++) {
1193                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1194                         talloc_zero_size(vdata->vacuum_fetch_list,
1195                                          offsetof(struct ctdb_marshall_buffer, data));
1196                 if (vdata->vacuum_fetch_list[i] == NULL) {
1197                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1198                         return -1;
1199                 }
1200                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1201         }
1202
1203         return 0;
1204 }
1205
1206 /**
1207  * Vacuum a DB:
1208  *  - Always do the fast vacuuming run, which traverses
1209  *    the in-memory delete queue: these records have been
1210  *    scheduled for deletion.
1211  *  - Only if explicitly requested, the database is traversed
1212  *    in order to use the traditional heuristics on empty records
1213  *    to trigger deletion.
1214  *    This is done only every VacuumFastPathCount'th vacuuming run.
1215  *
1216  * The traverse runs fill two lists:
1217  *
1218  * - The delete_list:
1219  *   This is the list of empty records the current
1220  *   node is lmaster and dmaster for. These records are later
1221  *   deleted first on other nodes and then locally.
1222  *
1223  *   The fast vacuuming run has a short cut for those records
1224  *   that have never been migrated with data: these records
1225  *   are immediately deleted locally, since they have left
1226  *   no trace on other nodes.
1227  *
1228  * - The vacuum_fetch lists
1229  *   (one for each other lmaster node):
1230  *   The records in this list are sent for deletion to
1231  *   their lmaster in a bulk VACUUM_FETCH message.
1232  *
1233  *   The lmaster then migrates all these records to itelf
1234  *   so that they can be vacuumed there.
1235  *
1236  * This executes in the child context.
1237  */
1238 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1239                           struct vacuum_data *vdata,
1240                           bool full_vacuum_run)
1241 {
1242         struct ctdb_context *ctdb = ctdb_db->ctdb;
1243         int ret, pnn;
1244
1245         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1246                            "%s db_id[0x%08x]\n",
1247                            full_vacuum_run ? "full" : "fast",
1248                            ctdb_db->db_name, ctdb_db->db_id));
1249
1250         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1251         if (ret != 0) {
1252                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1253                 return ret;
1254         }
1255
1256         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1257         if (pnn == -1) {
1258                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1259                 return -1;
1260         }
1261
1262         ctdb->pnn = pnn;
1263
1264         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
1265         if (ret != 0) {
1266                 return ret;
1267         }
1268
1269         if (full_vacuum_run) {
1270                 ret = ctdb_vacuum_traverse_db(ctdb_db, vdata);
1271                 if (ret != 0) {
1272                         return ret;
1273                 }
1274         }
1275
1276         ctdb_process_delete_queue(ctdb_db, vdata);
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,
1299                            void *private_data)
1300 {
1301         struct vacuum_data *vdata = (struct vacuum_data *)private_data;
1302
1303         if (vdata->vacuum) {
1304                 uint32_t hash = ctdb_hash(&key);
1305                 struct delete_record_data *kd;
1306                 /*
1307                  * check if we can ignore this record because it's in the delete_list
1308                  */
1309                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1310                 /*
1311                  * there might be hash collisions so we have to compare the keys here to be sure
1312                  */
1313                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1314                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1315                         /*
1316                          * we have to check if the record hasn't changed in the meantime in order to
1317                          * savely remove it from the database
1318                          */
1319                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1320                                 hdr->dmaster == kd->ctdb->pnn &&
1321                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1322                                 kd->hdr.rsn == hdr->rsn) {
1323                                 vdata->vacuumed++;
1324                                 return 0;
1325                         }
1326                 }
1327         }
1328         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1329                 vdata->traverse_error = true;
1330                 return -1;
1331         }
1332         vdata->copied++;
1333         return 0;
1334 }
1335
1336 /*
1337  * repack a tdb
1338  */
1339 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1340 {
1341         struct tdb_context *tmp_db;
1342
1343         if (tdb_transaction_start(tdb) != 0) {
1344                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1345                 return -1;
1346         }
1347
1348         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1349                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1350                           O_RDWR|O_CREAT, 0);
1351         if (tmp_db == NULL) {
1352                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1353                 tdb_transaction_cancel(tdb);
1354                 return -1;
1355         }
1356
1357         vdata->traverse_error = false;
1358         vdata->dest_db = tmp_db;
1359         vdata->vacuum = true;
1360         vdata->vacuumed = 0;
1361         vdata->copied = 0;
1362
1363         /*
1364          * repack and vacuum on-the-fly by not writing the records that are
1365          * no longer needed
1366          */
1367         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1368                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1369                 tdb_transaction_cancel(tdb);
1370                 tdb_close(tmp_db);
1371                 return -1;
1372         }
1373
1374         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1375
1376         if (vdata->traverse_error) {
1377                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1378                 tdb_transaction_cancel(tdb);
1379                 tdb_close(tmp_db);
1380                 return -1;
1381         }
1382
1383         if (tdb_wipe_all(tdb) != 0) {
1384                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1385                 tdb_transaction_cancel(tdb);
1386                 tdb_close(tmp_db);
1387                 return -1;
1388         }
1389
1390         vdata->traverse_error = false;
1391         vdata->dest_db = tdb;
1392         vdata->vacuum = false;
1393         vdata->copied = 0;
1394
1395         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1396                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1397                 tdb_transaction_cancel(tdb);
1398                 tdb_close(tmp_db);
1399                 return -1;
1400         }
1401
1402         if (vdata->traverse_error) {
1403                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1404                 tdb_transaction_cancel(tdb);
1405                 tdb_close(tmp_db);
1406                 return -1;
1407         }
1408
1409         tdb_close(tmp_db);
1410
1411
1412         if (tdb_transaction_commit(tdb) != 0) {
1413                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1414                 return -1;
1415         }
1416         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1417
1418         return 0;
1419 }
1420
1421 /*
1422  * repack and vaccum a db
1423  * called from the child context
1424  */
1425 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1426                                      TALLOC_CTX *mem_ctx,
1427                                      bool full_vacuum_run)
1428 {
1429         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1430         const char *name = ctdb_db->db_name;
1431         int freelist_size = 0;
1432         struct vacuum_data *vdata;
1433
1434         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1435         if (vdata == NULL) {
1436                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1437                 return -1;
1438         }
1439
1440         vdata->ctdb = ctdb_db->ctdb;
1441         vdata->repack_limit = repack_limit;
1442         vdata->delete_list = trbt_create(vdata, 0);
1443         vdata->ctdb_db = ctdb_db;
1444         if (vdata->delete_list == NULL) {
1445                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1446                 talloc_free(vdata);
1447                 return -1;
1448         }
1449
1450         vdata->start = timeval_current();
1451
1452         /*
1453          * gather all records that can be deleted in vdata
1454          */
1455         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1456                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1457         }
1458
1459         if (repack_limit != 0) {
1460                 freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1461                 if (freelist_size == -1) {
1462                         DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1463                         talloc_free(vdata);
1464                         return -1;
1465                 }
1466         }
1467
1468         /*
1469          * decide if a repack is necessary
1470          */
1471         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1472         {
1473                 talloc_free(vdata);
1474                 return 0;
1475         }
1476
1477         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1478                         name, freelist_size, vdata->delete_left));
1479
1480         /*
1481          * repack and implicitely get rid of the records we can delete
1482          */
1483         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1484                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1485                 talloc_free(vdata);
1486                 return -1;
1487         }
1488         talloc_free(vdata);
1489
1490         return 0;
1491 }
1492
1493 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1494 {
1495         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1496
1497         return interval;
1498 }
1499
1500 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1501 {
1502         double l = timeval_elapsed(&child_ctx->start_time);
1503         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1504         struct ctdb_context *ctdb = ctdb_db->ctdb;
1505
1506         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1507
1508         if (child_ctx->child_pid != -1) {
1509                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1510         } else {
1511                 /* Bump the number of successful fast-path runs. */
1512                 child_ctx->vacuum_handle->fast_path_count++;
1513         }
1514
1515         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1516
1517         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1518                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1519                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1520
1521         return 0;
1522 }
1523
1524 /*
1525  * this event is generated when a vacuum child process times out
1526  */
1527 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1528                                          struct timeval t, void *private_data)
1529 {
1530         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1531
1532         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1533
1534         child_ctx->status = VACUUM_TIMEOUT;
1535
1536         talloc_free(child_ctx);
1537 }
1538
1539
1540 /*
1541  * this event is generated when a vacuum child process has completed
1542  */
1543 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1544                              uint16_t flags, void *private_data)
1545 {
1546         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1547         char c = 0;
1548         int ret;
1549
1550         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1551         child_ctx->child_pid = -1;
1552
1553         ret = read(child_ctx->fd[0], &c, 1);
1554         if (ret != 1 || c != 0) {
1555                 child_ctx->status = VACUUM_ERROR;
1556                 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));
1557         } else {
1558                 child_ctx->status = VACUUM_OK;
1559         }
1560
1561         talloc_free(child_ctx);
1562 }
1563
1564 /*
1565  * this event is called every time we need to start a new vacuum process
1566  */
1567 static void
1568 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1569                                struct timeval t, void *private_data)
1570 {
1571         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1572         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1573         struct ctdb_context *ctdb = ctdb_db->ctdb;
1574         struct ctdb_vacuum_child_context *child_ctx;
1575         struct tevent_fd *fde;
1576         int ret;
1577
1578         /* we dont vacuum if we are in recovery mode, or db frozen */
1579         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1580             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1581                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1582                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1583                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1584                                    ? "freeze pending"
1585                                    : "frozen"));
1586                 event_add_timed(ctdb->ev, vacuum_handle,
1587                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1588                         ctdb_vacuum_event, vacuum_handle);
1589                 return;
1590         }
1591
1592         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1593         if (child_ctx == NULL) {
1594                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1595                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1596         }
1597
1598
1599         ret = pipe(child_ctx->fd);
1600         if (ret != 0) {
1601                 talloc_free(child_ctx);
1602                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1603                 event_add_timed(ctdb->ev, vacuum_handle,
1604                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1605                         ctdb_vacuum_event, vacuum_handle);
1606                 return;
1607         }
1608
1609         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1610                 vacuum_handle->fast_path_count = 0;
1611         }
1612
1613         child_ctx->child_pid = ctdb_fork(ctdb);
1614         if (child_ctx->child_pid == (pid_t)-1) {
1615                 close(child_ctx->fd[0]);
1616                 close(child_ctx->fd[1]);
1617                 talloc_free(child_ctx);
1618                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1619                 event_add_timed(ctdb->ev, vacuum_handle,
1620                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1621                         ctdb_vacuum_event, vacuum_handle);
1622                 return;
1623         }
1624
1625
1626         if (child_ctx->child_pid == 0) {
1627                 char cc = 0;
1628                 bool full_vacuum_run = false;
1629                 close(child_ctx->fd[0]);
1630
1631                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1632                 ctdb_set_process_name("ctdb_vacuum");
1633                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1634                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1635                         _exit(1);
1636                 }
1637
1638                 /*
1639                  * repack the db
1640                  */
1641                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1642                     (vacuum_handle->fast_path_count == 0))
1643                 {
1644                         full_vacuum_run = true;
1645                 }
1646                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1647                                                full_vacuum_run);
1648
1649                 write(child_ctx->fd[1], &cc, 1);
1650                 _exit(0);
1651         }
1652
1653         set_close_on_exec(child_ctx->fd[0]);
1654         close(child_ctx->fd[1]);
1655
1656         child_ctx->status = VACUUM_RUNNING;
1657         child_ctx->start_time = timeval_current();
1658
1659         DLIST_ADD(ctdb->vacuumers, child_ctx);
1660         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1661
1662         /*
1663          * Clear the fastpath vacuuming list in the parent.
1664          */
1665         talloc_free(ctdb_db->delete_queue);
1666         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1667         if (ctdb_db->delete_queue == NULL) {
1668                 /* fatal here? ... */
1669                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1670                                  "in parent context. Shutting down\n");
1671         }
1672
1673         event_add_timed(ctdb->ev, child_ctx,
1674                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1675                 vacuum_child_timeout, child_ctx);
1676
1677         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1678
1679         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1680                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1681         tevent_fd_set_auto_close(fde);
1682
1683         vacuum_handle->child_ctx = child_ctx;
1684         child_ctx->vacuum_handle = vacuum_handle;
1685 }
1686
1687 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1688 {
1689         /* Simply free them all. */
1690         while (ctdb->vacuumers) {
1691                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1692                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1693                            (int)ctdb->vacuumers->child_pid));
1694                 /* vacuum_child_destructor kills it, removes from list */
1695                 talloc_free(ctdb->vacuumers);
1696         }
1697 }
1698
1699 /* this function initializes the vacuuming context for a database
1700  * starts the vacuuming events
1701  */
1702 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1703 {
1704         if (ctdb_db->persistent != 0) {
1705                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1706                 return 0;
1707         }
1708
1709         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1710         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1711
1712         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1713         ctdb_db->vacuum_handle->fast_path_count = 0;
1714
1715         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1716                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1717                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1718
1719         return 0;
1720 }
1721
1722 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1723                                             const struct ctdb_ltdb_header *hdr,
1724                                             const TDB_DATA key)
1725 {
1726         struct delete_record_data *kd;
1727         uint32_t hash;
1728
1729         hash = (uint32_t)ctdb_hash(&key);
1730
1731         DEBUG(DEBUG_DEBUG, (__location__
1732                             " remove_record_from_delete_queue: "
1733                             "db[%s] "
1734                             "db_id[0x%08x] "
1735                             "key_hash[0x%08x] "
1736                             "lmaster[%u] "
1737                             "migrated_with_data[%s]\n",
1738                              ctdb_db->db_name, ctdb_db->db_id,
1739                              hash,
1740                              ctdb_lmaster(ctdb_db->ctdb, &key),
1741                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1742
1743         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1744         if (kd == NULL) {
1745                 DEBUG(DEBUG_DEBUG, (__location__
1746                                     " remove_record_from_delete_queue: "
1747                                     "record not in queue (hash[0x%08x])\n.",
1748                                     hash));
1749                 return;
1750         }
1751
1752         if ((kd->key.dsize != key.dsize) ||
1753             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1754         {
1755                 DEBUG(DEBUG_DEBUG, (__location__
1756                                     " remove_record_from_delete_queue: "
1757                                     "hash collision for key with hash[0x%08x] "
1758                                     "in db[%s] - skipping\n",
1759                                     hash, ctdb_db->db_name));
1760                 return;
1761         }
1762
1763         DEBUG(DEBUG_DEBUG, (__location__
1764                             " remove_record_from_delete_queue: "
1765                             "removing key with hash[0x%08x]\n",
1766                              hash));
1767
1768         talloc_free(kd);
1769
1770         return;
1771 }
1772
1773 /**
1774  * Insert a record into the ctdb_db context's delete queue,
1775  * handling hash collisions.
1776  */
1777 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1778                                            const struct ctdb_ltdb_header *hdr,
1779                                            TDB_DATA key)
1780 {
1781         struct delete_record_data *kd;
1782         uint32_t hash;
1783         int ret;
1784
1785         hash = (uint32_t)ctdb_hash(&key);
1786
1787         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1788                            "db_id[0x%08x] "
1789                            "key_hash[0x%08x] "
1790                            "lmaster[%u] "
1791                            "migrated_with_data[%s]\n",
1792                             ctdb_db->db_name, ctdb_db->db_id,
1793                             hash,
1794                             ctdb_lmaster(ctdb_db->ctdb, &key),
1795                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1796
1797         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1798         if (kd != NULL) {
1799                 if ((kd->key.dsize != key.dsize) ||
1800                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1801                 {
1802                         DEBUG(DEBUG_INFO,
1803                               (__location__ " schedule for deletion: "
1804                                "hash collision for key hash [0x%08x]. "
1805                                "Skipping the record.\n", hash));
1806                         return 0;
1807                 } else {
1808                         DEBUG(DEBUG_DEBUG,
1809                               (__location__ " schedule for deletion: "
1810                                "updating entry for key with hash [0x%08x].\n",
1811                                hash));
1812                 }
1813         }
1814
1815         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1816                                                   ctdb_db->delete_queue,
1817                                                   hdr, key);
1818         if (ret != 0) {
1819                 DEBUG(DEBUG_INFO,
1820                       (__location__ " schedule for deletion: error "
1821                        "inserting key with hash [0x%08x] into delete queue\n",
1822                        hash));
1823                 return -1;
1824         }
1825
1826         return 0;
1827 }
1828
1829 /**
1830  * Schedule a record for deletetion.
1831  * Called from the parent context.
1832  */
1833 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1834                                            TDB_DATA indata)
1835 {
1836         struct ctdb_control_schedule_for_deletion *dd;
1837         struct ctdb_db_context *ctdb_db;
1838         int ret;
1839         TDB_DATA key;
1840
1841         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1842
1843         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1844         if (ctdb_db == NULL) {
1845                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1846                                   dd->db_id));
1847                 return -1;
1848         }
1849
1850         key.dsize = dd->keylen;
1851         key.dptr = dd->key;
1852
1853         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1854
1855         return ret;
1856 }
1857
1858 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1859                                          const struct ctdb_ltdb_header *hdr,
1860                                          TDB_DATA key)
1861 {
1862         int ret;
1863         struct ctdb_control_schedule_for_deletion *dd;
1864         TDB_DATA indata;
1865         int32_t status;
1866
1867         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1868                 /* main daemon - directly queue */
1869                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1870
1871                 return ret;
1872         }
1873
1874         /* if we dont have a connection to the daemon we can not send
1875            a control. For example sometimes from update_record control child
1876            process.
1877         */
1878         if (!ctdb_db->ctdb->can_send_controls) {
1879                 return -1;
1880         }
1881
1882
1883         /* child process: send the main daemon a control */
1884         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1885         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1886         if (indata.dptr == NULL) {
1887                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1888                 return -1;
1889         }
1890         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1891         dd->db_id = ctdb_db->db_id;
1892         dd->hdr = *hdr;
1893         dd->keylen = key.dsize;
1894         memcpy(dd->key, key.dptr, key.dsize);
1895
1896         ret = ctdb_control(ctdb_db->ctdb,
1897                            CTDB_CURRENT_NODE,
1898                            ctdb_db->db_id,
1899                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1900                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1901                            indata,
1902                            NULL, /* mem_ctx */
1903                            NULL, /* outdata */
1904                            &status,
1905                            NULL, /* timeout : NULL == wait forever */
1906                            NULL); /* error message */
1907
1908         talloc_free(indata.dptr);
1909
1910         if (ret != 0 || status != 0) {
1911                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1912                                   "SCHEDULE_FOR_DELETION "
1913                                   "control.\n"));
1914                 if (status != 0) {
1915                         ret = -1;
1916                 }
1917         }
1918
1919         return ret;
1920 }
1921
1922 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1923                                          const struct ctdb_ltdb_header *hdr,
1924                                          const TDB_DATA key)
1925 {
1926         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1927                 /*
1928                  * Only remove the record from the delete queue if called
1929                  * in the main daemon.
1930                  */
1931                 return;
1932         }
1933
1934         remove_record_from_delete_queue(ctdb_db, hdr, key);
1935
1936         return;
1937 }