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 "replace.h"
  23 #include "system/network.h"
  24 #include "system/filesys.h"
  25 #include "system/time.h"
  26
  27 #include <talloc.h>
  28 #include <tevent.h>
  29
  30 #include "lib/tdb_wrap/tdb_wrap.h"
  31 #include "lib/util/dlinklist.h"
  32 #include "lib/util/debug.h"
  33 #include "lib/util/samba_util.h"
  34 #include "lib/util/sys_rw.h"
  35 #include "lib/util/util_process.h"
  36
  37 #include "ctdb_private.h"
  38 #include "ctdb_client.h"
  39
  40 #include "common/rb_tree.h"
  41 #include "common/common.h"
  42 #include "common/logging.h"
  43
  44 #define TIMELIMIT() timeval_current_ofs(10, 0)
  45
  46 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  47
  48 struct ctdb_vacuum_child_context {
  49         struct ctdb_vacuum_child_context *next, *prev;
  50         struct ctdb_vacuum_handle *vacuum_handle;
  51         /* fd child writes status to */
  52         int fd[2];
  53         pid_t child_pid;
  54         enum vacuum_child_status status;
  55         struct timeval start_time;
  56 };
  57
  58 struct ctdb_vacuum_handle {
  59         struct ctdb_db_context *ctdb_db;
  60         struct ctdb_vacuum_child_context *child_ctx;
  61         uint32_t fast_path_count;
  62 };
  63
  64
  65 /*  a list of records to possibly delete */
  66 struct vacuum_data {
  67         struct ctdb_context *ctdb;
  68         struct ctdb_db_context *ctdb_db;
  69         struct tdb_context *dest_db;
  70         trbt_tree_t *delete_list;
  71         struct ctdb_marshall_buffer **vacuum_fetch_list;
  72         struct timeval start;
  73         bool traverse_error;
  74         bool vacuum;
  75         struct {
  76                 struct {
  77                         uint32_t added_to_vacuum_fetch_list;
  78                         uint32_t added_to_delete_list;
  79                         uint32_t deleted;
  80                         uint32_t skipped;
  81                         uint32_t error;
  82                         uint32_t total;
  83                 } delete_queue;
  84                 struct {
  85                         uint32_t scheduled;
  86                         uint32_t skipped;
  87                         uint32_t error;
  88                         uint32_t total;
  89                 } db_traverse;
  90                 struct {
  91                         uint32_t total;
  92                         uint32_t remote_error;
  93                         uint32_t local_error;
  94                         uint32_t deleted;
  95                         uint32_t skipped;
  96                         uint32_t left;
  97                 } delete_list;
  98                 struct {
  99                         uint32_t vacuumed;
 100                         uint32_t copied;
 101                 } repack;
 102         } count;
 103 };
 104
 105 /* this structure contains the information for one record to be deleted */
 106 struct delete_record_data {
 107         struct ctdb_context *ctdb;
 108         struct ctdb_db_context *ctdb_db;
 109         struct ctdb_ltdb_header hdr;
 110         uint32_t remote_fail_count;
 111         TDB_DATA key;
 112         uint8_t keydata[1];
 113 };
 114
 115 struct delete_records_list {
 116         struct ctdb_marshall_buffer *records;
 117         struct vacuum_data *vdata;
 118 };
 119
 120 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 121                                            const struct ctdb_ltdb_header *hdr,
 122                                            TDB_DATA key);
 123
 124 /**
 125  * Store key and header in a tree, indexed by the key hash.
 126  */
 127 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 128                                                struct ctdb_db_context *ctdb_db,
 129                                                trbt_tree_t *tree,
 130                                                const struct ctdb_ltdb_header *hdr,
 131                                                TDB_DATA key)
 132 {
 133         struct delete_record_data *dd;
 134         uint32_t hash;
 135         size_t len;
 136
 137         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 138
 139         dd = (struct delete_record_data *)talloc_size(tree, len);
 140         if (dd == NULL) {
 141                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 142                 return -1;
 143         }
 144         talloc_set_name_const(dd, "struct delete_record_data");
 145
 146         dd->ctdb      = ctdb;
 147         dd->ctdb_db   = ctdb_db;
 148         dd->key.dsize = key.dsize;
 149         dd->key.dptr  = dd->keydata;
 150         memcpy(dd->keydata, key.dptr, key.dsize);
 151
 152         dd->hdr = *hdr;
 153         dd->remote_fail_count = 0;
 154
 155         hash = ctdb_hash(&key);
 156
 157         trbt_insert32(tree, hash, dd);
 158
 159         return 0;
 160 }
 161
 162 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 163                                      struct ctdb_ltdb_header *hdr)
 164 {
 165         struct ctdb_context *ctdb = vdata->ctdb;
 166         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 167         uint32_t hash;
 168         int ret;
 169
 170         hash = ctdb_hash(&key);
 171
 172         if (trbt_lookup32(vdata->delete_list, hash)) {
 173                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 174                 return 0;
 175         }
 176
 177         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 178                                                   vdata->delete_list,
 179                                                   hdr, key);
 180         if (ret != 0) {
 181                 return -1;
 182         }
 183
 184         vdata->count.delete_list.total++;
 185
 186         return 0;
 187 }
 188
 189 /**
 190  * Add a record to the list of records to be sent
 191  * to their lmaster with VACUUM_FETCH.
 192  */
 193 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 194                                            TDB_DATA key)
 195 {
 196         struct ctdb_context *ctdb = vdata->ctdb;
 197         uint32_t lmaster;
 198         struct ctdb_marshall_buffer *vfl;
 199
 200         lmaster = ctdb_lmaster(ctdb, &key);
 201
 202         vfl = vdata->vacuum_fetch_list[lmaster];
 203
 204         vfl = ctdb_marshall_add(ctdb, vfl, vfl->db_id, ctdb->pnn,
 205                                 key, NULL, tdb_null);
 206         if (vfl == NULL) {
 207                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 208                 vdata->traverse_error = true;
 209                 return -1;
 210         }
 211
 212         vdata->vacuum_fetch_list[lmaster] = vfl;
 213
 214         return 0;
 215 }
 216
 217
 218 static void ctdb_vacuum_event(struct tevent_context *ev,
 219                               struct tevent_timer *te,
 220                               struct timeval t, void *private_data);
 221
 222 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 223 {
 224         struct ctdb_ltdb_header *header =
 225                 (struct ctdb_ltdb_header *)private_data;
 226
 227         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 228                 return -1;
 229         }
 230
 231         *header = *(struct ctdb_ltdb_header *)data.dptr;
 232
 233         return 0;
 234 }
 235
 236 /*
 237  * traverse function for gathering the records that can be deleted
 238  */
 239 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 240                            void *private_data)
 241 {
 242         struct vacuum_data *vdata = talloc_get_type(private_data,
 243                                                     struct vacuum_data);
 244         struct ctdb_context *ctdb = vdata->ctdb;
 245         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 246         uint32_t lmaster;
 247         struct ctdb_ltdb_header *hdr;
 248         int res = 0;
 249
 250         vdata->count.db_traverse.total++;
 251
 252         lmaster = ctdb_lmaster(ctdb, &key);
 253         if (lmaster >= ctdb->num_nodes) {
 254                 vdata->count.db_traverse.error++;
 255                 DEBUG(DEBUG_CRIT, (__location__
 256                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 257                                    " with hash[%u]!\n",
 258                                    (unsigned)lmaster,
 259                                    (unsigned)ctdb->num_nodes,
 260                                    (unsigned)ctdb_hash(&key)));
 261                 return -1;
 262         }
 263
 264         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 265                 /* it is not a deleted record */
 266                 vdata->count.db_traverse.skipped++;
 267                 return 0;
 268         }
 269
 270         hdr = (struct ctdb_ltdb_header *)data.dptr;
 271
 272         if (hdr->dmaster != ctdb->pnn) {
 273                 vdata->count.db_traverse.skipped++;
 274                 return 0;
 275         }
 276
 277         /*
 278          * Add the record to this process's delete_queue for processing
 279          * in the subsequent traverse in the fast vacuum run.
 280          */
 281         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 282         if (res != 0) {
 283                 vdata->count.db_traverse.error++;
 284         } else {
 285                 vdata->count.db_traverse.scheduled++;
 286         }
 287
 288         return 0;
 289 }
 290
 291 /*
 292  * traverse the tree of records to delete and marshall them into
 293  * a blob
 294  */
 295 static int delete_marshall_traverse(void *param, void *data)
 296 {
 297         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 298         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 299         struct ctdb_marshall_buffer *m;
 300
 301         m = ctdb_marshall_add(recs, recs->records, recs->records->db_id,
 302                               recs->records->db_id,
 303                               dd->key, &dd->hdr, tdb_null);
 304         if (m == NULL) {
 305                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 306                 return -1;
 307         }
 308
 309         recs->records = m;
 310         return 0;
 311 }
 312
 313 /**
 314  * traverse function for the traversal of the delete_queue,
 315  * the fast-path vacuuming list.
 316  *
 317  *  - If the record has been migrated off the node
 318  *    or has been revived (filled with data) on the node,
 319  *    then skip the record.
 320  *
 321  *  - If the current node is the record's lmaster and it is
 322  *    a record that has never been migrated with data, then
 323  *    delete the record from the local tdb.
 324  *
 325  *  - If the current node is the record's lmaster and it has
 326  *    been migrated with data, then schedule it for the normal
 327  *    vacuuming procedure (i.e. add it to the delete_list).
 328  *
 329  *  - If the current node is NOT the record's lmaster then
 330  *    add it to the list of records that are to be sent to
 331  *    the lmaster with the VACUUM_FETCH message.
 332  */
 333 static int delete_queue_traverse(void *param, void *data)
 334 {
 335         struct delete_record_data *dd =
 336                 talloc_get_type(data, struct delete_record_data);
 337         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 338         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 339         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 340         int res;
 341         struct ctdb_ltdb_header header;
 342         uint32_t lmaster;
 343         uint32_t hash = ctdb_hash(&(dd->key));
 344
 345         vdata->count.delete_queue.total++;
 346
 347         res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
 348         if (res != 0) {
 349                 vdata->count.delete_queue.error++;
 350                 return 0;
 351         }
 352
 353         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 354                                vacuum_record_parser, &header);
 355         if (res != 0) {
 356                 goto skipped;
 357         }
 358
 359         if (header.dmaster != ctdb->pnn) {
 360                 /* The record has been migrated off the node. Skip. */
 361                 goto skipped;
 362         }
 363
 364         if (header.rsn != dd->hdr.rsn) {
 365                 /*
 366                  * The record has been migrated off the node and back again.
 367                  * But not requeued for deletion. Skip it.
 368                  */
 369                 goto skipped;
 370         }
 371
 372         /*
 373          * We are dmaster, and the record has no data, and it has
 374          * not been migrated after it has been queued for deletion.
 375          *
 376          * At this stage, the record could still have been revived locally
 377          * and last been written with empty data. This can only be
 378          * fixed with the addition of an active or delete flag. (TODO)
 379          */
 380
 381         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 382
 383         if (lmaster != ctdb->pnn) {
 384                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 385
 386                 if (res != 0) {
 387                         DEBUG(DEBUG_ERR,
 388                               (__location__ " Error adding record to list "
 389                                "of records to send to lmaster.\n"));
 390                         vdata->count.delete_queue.error++;
 391                 } else {
 392                         vdata->count.delete_queue.added_to_vacuum_fetch_list++;
 393                 }
 394                 goto done;
 395         }
 396
 397         /* use header->flags or dd->hdr.flags ?? */
 398         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 399                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 400
 401                 if (res != 0) {
 402                         DEBUG(DEBUG_ERR,
 403                               (__location__ " Error adding record to list "
 404                                "of records for deletion on lmaster.\n"));
 405                         vdata->count.delete_queue.error++;
 406                 } else {
 407                         vdata->count.delete_queue.added_to_delete_list++;
 408                 }
 409         } else {
 410                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 411
 412                 if (res != 0) {
 413                         DEBUG(DEBUG_ERR,
 414                               (__location__ " Error deleting record with key "
 415                                "hash [0x%08x] from local data base db[%s].\n",
 416                                hash, ctdb_db->db_name));
 417                         vdata->count.delete_queue.error++;
 418                         goto done;
 419                 }
 420
 421                 DEBUG(DEBUG_DEBUG,
 422                       (__location__ " Deleted record with key hash "
 423                        "[0x%08x] from local data base db[%s].\n",
 424                        hash, ctdb_db->db_name));
 425                 vdata->count.delete_queue.deleted++;
 426         }
 427
 428         goto done;
 429
 430 skipped:
 431         vdata->count.delete_queue.skipped++;
 432
 433 done:
 434         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 435
 436         return 0;
 437 }
 438
 439 /**
 440  * Delete the records that we are lmaster and dmaster for and
 441  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 442  * control.
 443  */
 444 static int delete_record_traverse(void *param, void *data)
 445 {
 446         struct delete_record_data *dd =
 447                 talloc_get_type(data, struct delete_record_data);
 448         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 449         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 450         struct ctdb_context *ctdb = ctdb_db->ctdb;
 451         int res;
 452         struct ctdb_ltdb_header header;
 453         uint32_t lmaster;
 454         uint32_t hash = ctdb_hash(&(dd->key));
 455
 456         if (dd->remote_fail_count > 0) {
 457                 vdata->count.delete_list.remote_error++;
 458                 vdata->count.delete_list.left--;
 459                 talloc_free(dd);
 460                 return 0;
 461         }
 462
 463         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 464         if (res != 0) {
 465                 DEBUG(DEBUG_ERR,
 466                       (__location__ " Error getting chainlock on record with "
 467                        "key hash [0x%08x] on database db[%s].\n",
 468                        hash, ctdb_db->db_name));
 469                 vdata->count.delete_list.local_error++;
 470                 vdata->count.delete_list.left--;
 471                 talloc_free(dd);
 472                 return 0;
 473         }
 474
 475         /*
 476          * Verify that the record is still empty, its RSN has not
 477          * changed and that we are still its lmaster and dmaster.
 478          */
 479
 480         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 481                                vacuum_record_parser, &header);
 482         if (res != 0) {
 483                 goto skip;
 484         }
 485
 486         if (header.flags & CTDB_REC_RO_FLAGS) {
 487                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 488                                    "on database db[%s] has read-only flags. "
 489                                    "skipping.\n",
 490                                    hash, ctdb_db->db_name));
 491                 goto skip;
 492         }
 493
 494         if (header.dmaster != ctdb->pnn) {
 495                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 496                                    "on database db[%s] has been migrated away. "
 497                                    "skipping.\n",
 498                                    hash, ctdb_db->db_name));
 499                 goto skip;
 500         }
 501
 502         if (header.rsn != dd->hdr.rsn) {
 503                 /*
 504                  * The record has been migrated off the node and back again.
 505                  * But not requeued for deletion. Skip it.
 506                  */
 507                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 508                                    "on database db[%s] seems to have been "
 509                                    "migrated away and back again (with empty "
 510                                    "data). skipping.\n",
 511                                    hash, ctdb_db->db_name));
 512                 goto skip;
 513         }
 514
 515         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 516
 517         if (lmaster != ctdb->pnn) {
 518                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 519                                    "delete list (key hash [0x%08x], db[%s]). "
 520                                    "Strange! skipping.\n",
 521                                    hash, ctdb_db->db_name));
 522                 goto skip;
 523         }
 524
 525         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 526
 527         if (res != 0) {
 528                 DEBUG(DEBUG_ERR,
 529                       (__location__ " Error deleting record with key hash "
 530                        "[0x%08x] from local data base db[%s].\n",
 531                        hash, ctdb_db->db_name));
 532                 vdata->count.delete_list.local_error++;
 533                 goto done;
 534         }
 535
 536         DEBUG(DEBUG_DEBUG,
 537               (__location__ " Deleted record with key hash [0x%08x] from "
 538                "local data base db[%s].\n", hash, ctdb_db->db_name));
 539
 540         vdata->count.delete_list.deleted++;
 541         goto done;
 542
 543 skip:
 544         vdata->count.delete_list.skipped++;
 545
 546 done:
 547         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 548
 549         talloc_free(dd);
 550         vdata->count.delete_list.left--;
 551
 552         return 0;
 553 }
 554
 555 /**
 556  * Traverse the delete_queue.
 557  * Records are either deleted directly or filled
 558  * into the delete list or the vacuum fetch lists
 559  * for further processing.
 560  */
 561 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 562                                       struct vacuum_data *vdata)
 563 {
 564         uint32_t sum;
 565         int ret;
 566
 567         ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
 568                                    delete_queue_traverse, vdata);
 569
 570         if (ret != 0) {
 571                 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
 572                       "the delete queue.\n"));
 573         }
 574
 575         sum = vdata->count.delete_queue.deleted
 576             + vdata->count.delete_queue.skipped
 577             + vdata->count.delete_queue.error
 578             + vdata->count.delete_queue.added_to_delete_list
 579             + vdata->count.delete_queue.added_to_vacuum_fetch_list;
 580
 581         if (vdata->count.delete_queue.total != sum) {
 582                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 583                       "counts for db[%s]: total[%u] != sum[%u]\n",
 584                       ctdb_db->db_name,
 585                       (unsigned)vdata->count.delete_queue.total,
 586                       (unsigned)sum));
 587         }
 588
 589         if (vdata->count.delete_queue.total > 0) {
 590                 DEBUG(DEBUG_INFO,
 591                       (__location__
 592                        " fast vacuuming delete_queue traverse statistics: "
 593                        "db[%s] "
 594                        "total[%u] "
 595                        "del[%u] "
 596                        "skp[%u] "
 597                        "err[%u] "
 598                        "adl[%u] "
 599                        "avf[%u]\n",
 600                        ctdb_db->db_name,
 601                        (unsigned)vdata->count.delete_queue.total,
 602                        (unsigned)vdata->count.delete_queue.deleted,
 603                        (unsigned)vdata->count.delete_queue.skipped,
 604                        (unsigned)vdata->count.delete_queue.error,
 605                        (unsigned)vdata->count.delete_queue.added_to_delete_list,
 606                        (unsigned)vdata->count.delete_queue.added_to_vacuum_fetch_list));
 607         }
 608
 609         return;
 610 }
 611
 612 /**
 613  * read-only traverse of the database, looking for records that
 614  * might be able to be vacuumed.
 615  *
 616  * This is not done each time but only every tunable
 617  * VacuumFastPathCount times.
 618  */
 619 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 620                                     struct vacuum_data *vdata)
 621 {
 622         int ret;
 623
 624         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 625         if (ret == -1 || vdata->traverse_error) {
 626                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 627                                   "'%s'\n", ctdb_db->db_name));
 628                 return;
 629         }
 630
 631         if (vdata->count.db_traverse.total > 0) {
 632                 DEBUG(DEBUG_INFO,
 633                       (__location__
 634                        " full vacuuming db traverse statistics: "
 635                        "db[%s] "
 636                        "total[%u] "
 637                        "skp[%u] "
 638                        "err[%u] "
 639                        "sched[%u]\n",
 640                        ctdb_db->db_name,
 641                        (unsigned)vdata->count.db_traverse.total,
 642                        (unsigned)vdata->count.db_traverse.skipped,
 643                        (unsigned)vdata->count.db_traverse.error,
 644                        (unsigned)vdata->count.db_traverse.scheduled));
 645         }
 646
 647         return;
 648 }
 649
 650 /**
 651  * Process the vacuum fetch lists:
 652  * For records for which we are not the lmaster, tell the lmaster to
 653  * fetch the record.
 654  */
 655 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 656                                             struct vacuum_data *vdata)
 657 {
 658         unsigned int i;
 659         struct ctdb_context *ctdb = ctdb_db->ctdb;
 660
 661         for (i = 0; i < ctdb->num_nodes; i++) {
 662                 TDB_DATA data;
 663                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 664
 665                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 666                         continue;
 667                 }
 668
 669                 if (vfl->count == 0) {
 670                         continue;
 671                 }
 672
 673                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 674                                    vfl->count, ctdb->nodes[i]->pnn,
 675                                    ctdb_db->db_name));
 676
 677                 data = ctdb_marshall_finish(vfl);
 678                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 679                                              CTDB_SRVID_VACUUM_FETCH,
 680                                              data) != 0)
 681                 {
 682                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 683                                           "fetch message to %u\n",
 684                                           ctdb->nodes[i]->pnn));
 685                 }
 686         }
 687
 688         return;
 689 }
 690
 691 /**
 692  * Process the delete list:
 693  *
 694  * This is the last step of vacuuming that consistently deletes
 695  * those records that have been migrated with data and can hence
 696  * not be deleted when leaving a node.
 697  *
 698  * In this step, the lmaster does the final deletion of those empty
 699  * records that it is also dmaster for. It has ususally received
 700  * at least some of these records previously from the former dmasters
 701  * with the vacuum fetch message.
 702  *
 703  *  1) Send the records to all active nodes with the TRY_DELETE_RECORDS
 704  *     control. The remote notes delete their local copy.
 705  *  2) The lmaster locally deletes its copies of all records that
 706  *     could successfully be deleted remotely in step #2.
 707  */
 708 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 709                                      struct vacuum_data *vdata)
 710 {
 711         int ret, i;
 712         struct ctdb_context *ctdb = ctdb_db->ctdb;
 713         struct delete_records_list *recs;
 714         TDB_DATA indata;
 715         struct ctdb_node_map_old *nodemap;
 716         uint32_t *active_nodes;
 717         int num_active_nodes;
 718         TALLOC_CTX *tmp_ctx;
 719         uint32_t sum;
 720
 721         if (vdata->count.delete_list.total == 0) {
 722                 return;
 723         }
 724
 725         tmp_ctx = talloc_new(vdata);
 726         if (tmp_ctx == NULL) {
 727                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 728                 return;
 729         }
 730
 731         vdata->count.delete_list.left = vdata->count.delete_list.total;
 732
 733         /*
 734          * get the list of currently active nodes
 735          */
 736
 737         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 738                                    CTDB_CURRENT_NODE,
 739                                    tmp_ctx,
 740                                    &nodemap);
 741         if (ret != 0) {
 742                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 743                 goto done;
 744         }
 745
 746         active_nodes = list_of_active_nodes(ctdb, nodemap,
 747                                             nodemap, /* talloc context */
 748                                             false /* include self */);
 749         /* yuck! ;-) */
 750         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 751
 752         /*
 753          * Now delete the records all active nodes in a two-phase process:
 754          * 1) tell all active remote nodes to delete all their copy
 755          * 2) if all remote nodes deleted their record copy, delete it locally
 756          */
 757
 758         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 759         if (recs == NULL) {
 760                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 761                 goto done;
 762         }
 763
 764         /*
 765          * Step 1:
 766          * Send all records to all active nodes for deletion.
 767          */
 768
 769         /*
 770          * Create a marshall blob from the remaining list of records to delete.
 771          */
 772
 773         recs->records = (struct ctdb_marshall_buffer *)
 774                 talloc_zero_size(recs,
 775                                  offsetof(struct ctdb_marshall_buffer, data));
 776         if (recs->records == NULL) {
 777                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 778                 goto done;
 779         }
 780         recs->records->db_id = ctdb_db->db_id;
 781
 782         ret = trbt_traversearray32(vdata->delete_list, 1,
 783                                    delete_marshall_traverse, recs);
 784         if (ret != 0) {
 785                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
 786                       "delete list for second marshalling.\n"));
 787                 goto done;
 788         }
 789
 790         indata = ctdb_marshall_finish(recs->records);
 791
 792         for (i = 0; i < num_active_nodes; i++) {
 793                 struct ctdb_marshall_buffer *records;
 794                 struct ctdb_rec_data_old *rec;
 795                 int32_t res;
 796                 TDB_DATA outdata;
 797
 798                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 799                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 800                                 indata, recs, &outdata, &res,
 801                                 NULL, NULL);
 802                 if (ret != 0 || res != 0) {
 803                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
 804                                           "node %u: ret[%d] res[%d]\n",
 805                                           active_nodes[i], ret, res));
 806                         goto done;
 807                 }
 808
 809                 /*
 810                  * outdata contains the list of records coming back
 811                  * from the node: These are the records that the
 812                  * remote node could not delete. We remove these from
 813                  * the list to delete locally.
 814                  */
 815                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 816                 rec = (struct ctdb_rec_data_old *)&records->data[0];
 817                 while (records->count-- > 1) {
 818                         TDB_DATA reckey, recdata;
 819                         struct ctdb_ltdb_header *rechdr;
 820                         struct delete_record_data *dd;
 821
 822                         reckey.dptr = &rec->data[0];
 823                         reckey.dsize = rec->keylen;
 824                         recdata.dptr = &rec->data[reckey.dsize];
 825                         recdata.dsize = rec->datalen;
 826
 827                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 828                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 829                                 goto done;
 830                         }
 831                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 832                         recdata.dptr += sizeof(*rechdr);
 833                         recdata.dsize -= sizeof(*rechdr);
 834
 835                         dd = (struct delete_record_data *)trbt_lookup32(
 836                                         vdata->delete_list,
 837                                         ctdb_hash(&reckey));
 838                         if (dd != NULL) {
 839                                 /*
 840                                  * The remote node could not delete the
 841                                  * record.  Since other remote nodes can
 842                                  * also fail, we just mark the record.
 843                                  */
 844                                 dd->remote_fail_count++;
 845                         } else {
 846                                 DEBUG(DEBUG_ERR, (__location__ " Failed to "
 847                                       "find record with hash 0x%08x coming "
 848                                       "back from TRY_DELETE_RECORDS "
 849                                       "control in delete list.\n",
 850                                       ctdb_hash(&reckey)));
 851                         }
 852
 853                         rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
 854                 }
 855         }
 856
 857         /*
 858          * Step 2:
 859          * Delete the remaining records locally.
 860          *
 861          * These records have successfully been deleted on all
 862          * active remote nodes.
 863          */
 864
 865         ret = trbt_traversearray32(vdata->delete_list, 1,
 866                                    delete_record_traverse, vdata);
 867         if (ret != 0) {
 868                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
 869                       "delete list for deletion.\n"));
 870         }
 871
 872         if (vdata->count.delete_list.left != 0) {
 873                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
 874                       "there are %u records left for deletion after "
 875                       "processing delete list\n",
 876                       ctdb_db->db_name,
 877                       (unsigned)vdata->count.delete_list.left));
 878         }
 879
 880         sum = vdata->count.delete_list.deleted
 881             + vdata->count.delete_list.skipped
 882             + vdata->count.delete_list.remote_error
 883             + vdata->count.delete_list.local_error
 884             + vdata->count.delete_list.left;
 885
 886         if (vdata->count.delete_list.total != sum) {
 887                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
 888                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
 889                       ctdb_db->db_name,
 890                       (unsigned)vdata->count.delete_list.total,
 891                       (unsigned)sum));
 892         }
 893
 894         if (vdata->count.delete_list.total > 0) {
 895                 DEBUG(DEBUG_INFO,
 896                       (__location__
 897                        " vacuum delete list statistics: "
 898                        "db[%s] "
 899                        "total[%u] "
 900                        "del[%u] "
 901                        "skip[%u] "
 902                        "rem.err[%u] "
 903                        "loc.err[%u] "
 904                        "left[%u]\n",
 905                        ctdb_db->db_name,
 906                        (unsigned)vdata->count.delete_list.total,
 907                        (unsigned)vdata->count.delete_list.deleted,
 908                        (unsigned)vdata->count.delete_list.skipped,
 909                        (unsigned)vdata->count.delete_list.remote_error,
 910                        (unsigned)vdata->count.delete_list.local_error,
 911                        (unsigned)vdata->count.delete_list.left));
 912         }
 913
 914 done:
 915         talloc_free(tmp_ctx);
 916
 917         return;
 918 }
 919
 920 /**
 921  * initialize the vacuum_data
 922  */
 923 static struct vacuum_data *ctdb_vacuum_init_vacuum_data(
 924                                         struct ctdb_db_context *ctdb_db,
 925                                         TALLOC_CTX *mem_ctx)
 926 {
 927         unsigned int i;
 928         struct ctdb_context *ctdb = ctdb_db->ctdb;
 929         struct vacuum_data *vdata;
 930
 931         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 932         if (vdata == NULL) {
 933                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 934                 return NULL;
 935         }
 936
 937         vdata->ctdb = ctdb_db->ctdb;
 938         vdata->ctdb_db = ctdb_db;
 939         vdata->delete_list = trbt_create(vdata, 0);
 940         if (vdata->delete_list == NULL) {
 941                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 942                 goto fail;
 943         }
 944
 945         vdata->start = timeval_current();
 946
 947         vdata->count.delete_queue.added_to_delete_list = 0;
 948         vdata->count.delete_queue.added_to_vacuum_fetch_list = 0;
 949         vdata->count.delete_queue.deleted = 0;
 950         vdata->count.delete_queue.skipped = 0;
 951         vdata->count.delete_queue.error = 0;
 952         vdata->count.delete_queue.total = 0;
 953         vdata->count.db_traverse.scheduled = 0;
 954         vdata->count.db_traverse.skipped = 0;
 955         vdata->count.db_traverse.error = 0;
 956         vdata->count.db_traverse.total = 0;
 957         vdata->count.delete_list.total = 0;
 958         vdata->count.delete_list.left = 0;
 959         vdata->count.delete_list.remote_error = 0;
 960         vdata->count.delete_list.local_error = 0;
 961         vdata->count.delete_list.skipped = 0;
 962         vdata->count.delete_list.deleted = 0;
 963
 964         /* the list needs to be of length num_nodes */
 965         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
 966                                                 struct ctdb_marshall_buffer *,
 967                                                 ctdb->num_nodes);
 968         if (vdata->vacuum_fetch_list == NULL) {
 969                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 970                 goto fail;
 971         }
 972         for (i = 0; i < ctdb->num_nodes; i++) {
 973                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 974                         talloc_zero_size(vdata->vacuum_fetch_list,
 975                                          offsetof(struct ctdb_marshall_buffer, data));
 976                 if (vdata->vacuum_fetch_list[i] == NULL) {
 977                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 978                         talloc_free(vdata);
 979                         return NULL;
 980                 }
 981                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 982         }
 983
 984         return vdata;
 985
 986 fail:
 987         talloc_free(vdata);
 988         return NULL;
 989 }
 990
 991 /**
 992  * Vacuum a DB:
 993  *  - Always do the fast vacuuming run, which traverses
 994  *    the in-memory delete queue: these records have been
 995  *    scheduled for deletion.
 996  *  - Only if explicitly requested, the database is traversed
 997  *    in order to use the traditional heuristics on empty records
 998  *    to trigger deletion.
 999  *    This is done only every VacuumFastPathCount'th vacuuming run.
1000  *
1001  * The traverse runs fill two lists:
1002  *
1003  * - The delete_list:
1004  *   This is the list of empty records the current
1005  *   node is lmaster and dmaster for. These records are later
1006  *   deleted first on other nodes and then locally.
1007  *
1008  *   The fast vacuuming run has a short cut for those records
1009  *   that have never been migrated with data: these records
1010  *   are immediately deleted locally, since they have left
1011  *   no trace on other nodes.
1012  *
1013  * - The vacuum_fetch lists
1014  *   (one for each other lmaster node):
1015  *   The records in this list are sent for deletion to
1016  *   their lmaster in a bulk VACUUM_FETCH message.
1017  *
1018  *   The lmaster then migrates all these records to itelf
1019  *   so that they can be vacuumed there.
1020  *
1021  * This executes in the child context.
1022  */
1023 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1024                           bool full_vacuum_run)
1025 {
1026         struct ctdb_context *ctdb = ctdb_db->ctdb;
1027         int ret, pnn;
1028         struct vacuum_data *vdata;
1029         TALLOC_CTX *tmp_ctx;
1030
1031         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1032                            "%s db_id[0x%08x]\n",
1033                            full_vacuum_run ? "full" : "fast",
1034                            ctdb_db->db_name, ctdb_db->db_id));
1035
1036         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1037         if (ret != 0) {
1038                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1039                 return ret;
1040         }
1041
1042         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1043         if (pnn == -1) {
1044                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1045                 return -1;
1046         }
1047
1048         ctdb->pnn = pnn;
1049
1050         tmp_ctx = talloc_new(ctdb_db);
1051         if (tmp_ctx == NULL) {
1052                 DEBUG(DEBUG_ERR, ("Out of memory!\n"));
1053                 return -1;
1054         }
1055
1056         vdata = ctdb_vacuum_init_vacuum_data(ctdb_db, tmp_ctx);
1057         if (vdata == NULL) {
1058                 talloc_free(tmp_ctx);
1059                 return -1;
1060         }
1061
1062         if (full_vacuum_run) {
1063                 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1064         }
1065
1066         ctdb_process_delete_queue(ctdb_db, vdata);
1067
1068         ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1069
1070         ctdb_process_delete_list(ctdb_db, vdata);
1071
1072         talloc_free(tmp_ctx);
1073
1074         /* this ensures we run our event queue */
1075         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1076
1077         return 0;
1078 }
1079
1080 /*
1081  * repack and vaccum a db
1082  * called from the child context
1083  */
1084 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1085                                      bool full_vacuum_run)
1086 {
1087         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1088         const char *name = ctdb_db->db_name;
1089         int freelist_size = 0;
1090         int ret;
1091
1092         if (ctdb_vacuum_db(ctdb_db, full_vacuum_run) != 0) {
1093                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1094         }
1095
1096         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1097         if (freelist_size == -1) {
1098                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1099                 return -1;
1100         }
1101
1102         /*
1103          * decide if a repack is necessary
1104          */
1105         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1106         {
1107                 return 0;
1108         }
1109
1110         DEBUG(DEBUG_INFO, ("Repacking %s with %u freelist entries\n",
1111                            name, freelist_size));
1112
1113         ret = tdb_repack(ctdb_db->ltdb->tdb);
1114         if (ret != 0) {
1115                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1116                 return -1;
1117         }
1118
1119         return 0;
1120 }
1121
1122 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1123 {
1124         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1125
1126         return interval;
1127 }
1128
1129 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1130 {
1131         double l = timeval_elapsed(&child_ctx->start_time);
1132         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1133         struct ctdb_context *ctdb = ctdb_db->ctdb;
1134
1135         CTDB_UPDATE_DB_LATENCY(ctdb_db, "vacuum", vacuum.latency, l);
1136         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1137
1138         if (child_ctx->child_pid != -1) {
1139                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1140         } else {
1141                 /* Bump the number of successful fast-path runs. */
1142                 child_ctx->vacuum_handle->fast_path_count++;
1143         }
1144
1145         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1146
1147         tevent_add_timer(ctdb->ev, child_ctx->vacuum_handle,
1148                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1149                          ctdb_vacuum_event, child_ctx->vacuum_handle);
1150
1151         return 0;
1152 }
1153
1154 /*
1155  * this event is generated when a vacuum child process times out
1156  */
1157 static void vacuum_child_timeout(struct tevent_context *ev,
1158                                  struct tevent_timer *te,
1159                                  struct timeval t, void *private_data)
1160 {
1161         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1162
1163         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1164
1165         child_ctx->status = VACUUM_TIMEOUT;
1166
1167         talloc_free(child_ctx);
1168 }
1169
1170
1171 /*
1172  * this event is generated when a vacuum child process has completed
1173  */
1174 static void vacuum_child_handler(struct tevent_context *ev,
1175                                  struct tevent_fd *fde,
1176                                  uint16_t flags, void *private_data)
1177 {
1178         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1179         char c = 0;
1180         int ret;
1181
1182         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1183         child_ctx->child_pid = -1;
1184
1185         ret = sys_read(child_ctx->fd[0], &c, 1);
1186         if (ret != 1 || c != 0) {
1187                 child_ctx->status = VACUUM_ERROR;
1188                 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));
1189         } else {
1190                 child_ctx->status = VACUUM_OK;
1191         }
1192
1193         talloc_free(child_ctx);
1194 }
1195
1196 /*
1197  * this event is called every time we need to start a new vacuum process
1198  */
1199 static void ctdb_vacuum_event(struct tevent_context *ev,
1200                               struct tevent_timer *te,
1201                               struct timeval t, void *private_data)
1202 {
1203         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1204         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1205         struct ctdb_context *ctdb = ctdb_db->ctdb;
1206         struct ctdb_vacuum_child_context *child_ctx;
1207         struct tevent_fd *fde;
1208         bool full_vacuum_run = false;
1209         int ret;
1210
1211         /* we don't vacuum if we are in recovery mode, or db frozen */
1212         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1213             ctdb_db_frozen(ctdb_db)) {
1214                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1215                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ?
1216                                         "in recovery" : "frozen"));
1217                 tevent_add_timer(ctdb->ev, vacuum_handle,
1218                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1219                                  ctdb_vacuum_event, vacuum_handle);
1220                 return;
1221         }
1222
1223         /* Do not allow multiple vacuuming child processes to be active at the
1224          * same time.  If there is vacuuming child process active, delay
1225          * new vacuuming event to stagger vacuuming events.
1226          */
1227         if (ctdb->vacuumers != NULL) {
1228                 tevent_add_timer(ctdb->ev, vacuum_handle,
1229                                  timeval_current_ofs(0, 500*1000),
1230                                  ctdb_vacuum_event, vacuum_handle);
1231                 return;
1232         }
1233
1234         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1235         if (child_ctx == NULL) {
1236                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1237                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1238         }
1239
1240
1241         ret = pipe(child_ctx->fd);
1242         if (ret != 0) {
1243                 talloc_free(child_ctx);
1244                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1245                 tevent_add_timer(ctdb->ev, vacuum_handle,
1246                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1247                                  ctdb_vacuum_event, vacuum_handle);
1248                 return;
1249         }
1250
1251         if (vacuum_handle->fast_path_count >=
1252             ctdb->tunable.vacuum_fast_path_count) {
1253                 if (ctdb->tunable.vacuum_fast_path_count > 0) {
1254                         full_vacuum_run = true;
1255                 }
1256                 vacuum_handle->fast_path_count = 0;
1257         }
1258
1259         child_ctx->child_pid = ctdb_fork(ctdb);
1260         if (child_ctx->child_pid == (pid_t)-1) {
1261                 close(child_ctx->fd[0]);
1262                 close(child_ctx->fd[1]);
1263                 talloc_free(child_ctx);
1264                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1265                 tevent_add_timer(ctdb->ev, vacuum_handle,
1266                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1267                                  ctdb_vacuum_event, vacuum_handle);
1268                 return;
1269         }
1270
1271
1272         if (child_ctx->child_pid == 0) {
1273                 char cc = 0;
1274                 close(child_ctx->fd[0]);
1275
1276                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1277                 prctl_set_comment("ctdb_vacuum");
1278                 if (switch_from_server_to_client(ctdb) != 0) {
1279                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1280                         _exit(1);
1281                 }
1282
1283                 cc = ctdb_vacuum_and_repack_db(ctdb_db, full_vacuum_run);
1284
1285                 sys_write(child_ctx->fd[1], &cc, 1);
1286                 _exit(0);
1287         }
1288
1289         set_close_on_exec(child_ctx->fd[0]);
1290         close(child_ctx->fd[1]);
1291
1292         child_ctx->status = VACUUM_RUNNING;
1293         child_ctx->start_time = timeval_current();
1294
1295         DLIST_ADD(ctdb->vacuumers, child_ctx);
1296         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1297
1298         /*
1299          * Clear the fastpath vacuuming list in the parent.
1300          */
1301         talloc_free(ctdb_db->delete_queue);
1302         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1303         if (ctdb_db->delete_queue == NULL) {
1304                 /* fatal here? ... */
1305                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1306                                  "in parent context. Shutting down\n");
1307         }
1308
1309         tevent_add_timer(ctdb->ev, child_ctx,
1310                          timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1311                          vacuum_child_timeout, child_ctx);
1312
1313         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1314
1315         fde = tevent_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1316                             TEVENT_FD_READ, vacuum_child_handler, child_ctx);
1317         tevent_fd_set_auto_close(fde);
1318
1319         vacuum_handle->child_ctx = child_ctx;
1320         child_ctx->vacuum_handle = vacuum_handle;
1321 }
1322
1323 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1324 {
1325         /* Simply free them all. */
1326         while (ctdb->vacuumers) {
1327                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1328                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1329                            (int)ctdb->vacuumers->child_pid));
1330                 /* vacuum_child_destructor kills it, removes from list */
1331                 talloc_free(ctdb->vacuumers);
1332         }
1333 }
1334
1335 /* this function initializes the vacuuming context for a database
1336  * starts the vacuuming events
1337  */
1338 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1339 {
1340         if (! ctdb_db_volatile(ctdb_db)) {
1341                 DEBUG(DEBUG_ERR,
1342                       ("Vacuuming is disabled for non-volatile database %s\n",
1343                        ctdb_db->db_name));
1344                 return 0;
1345         }
1346
1347         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1348         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1349
1350         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1351         ctdb_db->vacuum_handle->fast_path_count = 0;
1352
1353         tevent_add_timer(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1354                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1355                          ctdb_vacuum_event, ctdb_db->vacuum_handle);
1356
1357         return 0;
1358 }
1359
1360 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1361                                             const struct ctdb_ltdb_header *hdr,
1362                                             const TDB_DATA key)
1363 {
1364         struct delete_record_data *kd;
1365         uint32_t hash;
1366
1367         hash = (uint32_t)ctdb_hash(&key);
1368
1369         DEBUG(DEBUG_DEBUG, (__location__
1370                             " remove_record_from_delete_queue: "
1371                             "db[%s] "
1372                             "db_id[0x%08x] "
1373                             "key_hash[0x%08x] "
1374                             "lmaster[%u] "
1375                             "migrated_with_data[%s]\n",
1376                              ctdb_db->db_name, ctdb_db->db_id,
1377                              hash,
1378                              ctdb_lmaster(ctdb_db->ctdb, &key),
1379                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1380
1381         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1382         if (kd == NULL) {
1383                 DEBUG(DEBUG_DEBUG, (__location__
1384                                     " remove_record_from_delete_queue: "
1385                                     "record not in queue (hash[0x%08x])\n.",
1386                                     hash));
1387                 return;
1388         }
1389
1390         if ((kd->key.dsize != key.dsize) ||
1391             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1392         {
1393                 DEBUG(DEBUG_DEBUG, (__location__
1394                                     " remove_record_from_delete_queue: "
1395                                     "hash collision for key with hash[0x%08x] "
1396                                     "in db[%s] - skipping\n",
1397                                     hash, ctdb_db->db_name));
1398                 return;
1399         }
1400
1401         DEBUG(DEBUG_DEBUG, (__location__
1402                             " remove_record_from_delete_queue: "
1403                             "removing key with hash[0x%08x]\n",
1404                              hash));
1405
1406         talloc_free(kd);
1407
1408         return;
1409 }
1410
1411 /**
1412  * Insert a record into the ctdb_db context's delete queue,
1413  * handling hash collisions.
1414  */
1415 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1416                                            const struct ctdb_ltdb_header *hdr,
1417                                            TDB_DATA key)
1418 {
1419         struct delete_record_data *kd;
1420         uint32_t hash;
1421         int ret;
1422
1423         hash = (uint32_t)ctdb_hash(&key);
1424
1425         DEBUG(DEBUG_DEBUG, (__location__ " schedule for deletion: db[%s] "
1426                             "db_id[0x%08x] "
1427                             "key_hash[0x%08x] "
1428                             "lmaster[%u] "
1429                             "migrated_with_data[%s]\n",
1430                             ctdb_db->db_name, ctdb_db->db_id,
1431                             hash,
1432                             ctdb_lmaster(ctdb_db->ctdb, &key),
1433                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1434
1435         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1436         if (kd != NULL) {
1437                 if ((kd->key.dsize != key.dsize) ||
1438                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1439                 {
1440                         DEBUG(DEBUG_INFO,
1441                               (__location__ " schedule for deletion: "
1442                                "hash collision for key hash [0x%08x]. "
1443                                "Skipping the record.\n", hash));
1444                         return 0;
1445                 } else {
1446                         DEBUG(DEBUG_DEBUG,
1447                               (__location__ " schedule for deletion: "
1448                                "updating entry for key with hash [0x%08x].\n",
1449                                hash));
1450                 }
1451         }
1452
1453         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1454                                                   ctdb_db->delete_queue,
1455                                                   hdr, key);
1456         if (ret != 0) {
1457                 DEBUG(DEBUG_INFO,
1458                       (__location__ " schedule for deletion: error "
1459                        "inserting key with hash [0x%08x] into delete queue\n",
1460                        hash));
1461                 return -1;
1462         }
1463
1464         return 0;
1465 }
1466
1467 /**
1468  * Schedule a record for deletetion.
1469  * Called from the parent context.
1470  */
1471 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1472                                            TDB_DATA indata)
1473 {
1474         struct ctdb_control_schedule_for_deletion *dd;
1475         struct ctdb_db_context *ctdb_db;
1476         int ret;
1477         TDB_DATA key;
1478
1479         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1480
1481         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1482         if (ctdb_db == NULL) {
1483                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1484                                   dd->db_id));
1485                 return -1;
1486         }
1487
1488         key.dsize = dd->keylen;
1489         key.dptr = dd->key;
1490
1491         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1492
1493         return ret;
1494 }
1495
1496 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1497                                          const struct ctdb_ltdb_header *hdr,
1498                                          TDB_DATA key)
1499 {
1500         int ret;
1501         struct ctdb_control_schedule_for_deletion *dd;
1502         TDB_DATA indata;
1503         int32_t status;
1504
1505         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1506                 /* main daemon - directly queue */
1507                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1508
1509                 return ret;
1510         }
1511
1512         /* if we don't have a connection to the daemon we can not send
1513            a control. For example sometimes from update_record control child
1514            process.
1515         */
1516         if (!ctdb_db->ctdb->can_send_controls) {
1517                 return -1;
1518         }
1519
1520
1521         /* child process: send the main daemon a control */
1522         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1523         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1524         if (indata.dptr == NULL) {
1525                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1526                 return -1;
1527         }
1528         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1529         dd->db_id = ctdb_db->db_id;
1530         dd->hdr = *hdr;
1531         dd->keylen = key.dsize;
1532         memcpy(dd->key, key.dptr, key.dsize);
1533
1534         ret = ctdb_control(ctdb_db->ctdb,
1535                            CTDB_CURRENT_NODE,
1536                            ctdb_db->db_id,
1537                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1538                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1539                            indata,
1540                            NULL, /* mem_ctx */
1541                            NULL, /* outdata */
1542                            &status,
1543                            NULL, /* timeout : NULL == wait forever */
1544                            NULL); /* error message */
1545
1546         talloc_free(indata.dptr);
1547
1548         if (ret != 0 || status != 0) {
1549                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1550                                   "SCHEDULE_FOR_DELETION "
1551                                   "control.\n"));
1552                 if (status != 0) {
1553                         ret = -1;
1554                 }
1555         }
1556
1557         return ret;
1558 }
1559
1560 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1561                                          const struct ctdb_ltdb_header *hdr,
1562                                          const TDB_DATA key)
1563 {
1564         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1565                 /*
1566                  * Only remove the record from the delete queue if called
1567                  * in the main daemon.
1568                  */
1569                 return;
1570         }
1571
1572         remove_record_from_delete_queue(ctdb_db, hdr, key);
1573
1574         return;
1575 }