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         TDB_DATA key;
 111         uint8_t keydata[1];
 112 };
 113
 114 struct delete_records_list {
 115         struct ctdb_marshall_buffer *records;
 116         struct vacuum_data *vdata;
 117 };
 118
 119 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 120                                            const struct ctdb_ltdb_header *hdr,
 121                                            TDB_DATA key);
 122
 123 /**
 124  * Store key and header in a tree, indexed by the key hash.
 125  */
 126 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 127                                                struct ctdb_db_context *ctdb_db,
 128                                                trbt_tree_t *tree,
 129                                                const struct ctdb_ltdb_header *hdr,
 130                                                TDB_DATA key)
 131 {
 132         struct delete_record_data *dd;
 133         uint32_t hash;
 134         size_t len;
 135
 136         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 137
 138         dd = (struct delete_record_data *)talloc_size(tree, len);
 139         if (dd == NULL) {
 140                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 141                 return -1;
 142         }
 143         talloc_set_name_const(dd, "struct delete_record_data");
 144
 145         dd->ctdb      = ctdb;
 146         dd->ctdb_db   = ctdb_db;
 147         dd->key.dsize = key.dsize;
 148         dd->key.dptr  = dd->keydata;
 149         memcpy(dd->keydata, key.dptr, key.dsize);
 150
 151         dd->hdr = *hdr;
 152
 153         hash = ctdb_hash(&key);
 154
 155         trbt_insert32(tree, hash, dd);
 156
 157         return 0;
 158 }
 159
 160 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 161                                      struct ctdb_ltdb_header *hdr)
 162 {
 163         struct ctdb_context *ctdb = vdata->ctdb;
 164         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 165         uint32_t hash;
 166         int ret;
 167
 168         hash = ctdb_hash(&key);
 169
 170         if (trbt_lookup32(vdata->delete_list, hash)) {
 171                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 172                 return 0;
 173         }
 174
 175         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 176                                                   vdata->delete_list,
 177                                                   hdr, key);
 178         if (ret != 0) {
 179                 return -1;
 180         }
 181
 182         vdata->count.delete_list.total++;
 183
 184         return 0;
 185 }
 186
 187 /**
 188  * Add a record to the list of records to be sent
 189  * to their lmaster with VACUUM_FETCH.
 190  */
 191 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 192                                            TDB_DATA key)
 193 {
 194         struct ctdb_context *ctdb = vdata->ctdb;
 195         uint32_t lmaster;
 196         struct ctdb_marshall_buffer *vfl;
 197
 198         lmaster = ctdb_lmaster(ctdb, &key);
 199
 200         vfl = vdata->vacuum_fetch_list[lmaster];
 201
 202         vfl = ctdb_marshall_add(ctdb, vfl, vfl->db_id, ctdb->pnn,
 203                                 key, NULL, tdb_null);
 204         if (vfl == NULL) {
 205                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 206                 vdata->traverse_error = true;
 207                 return -1;
 208         }
 209
 210         vdata->vacuum_fetch_list[lmaster] = vfl;
 211
 212         return 0;
 213 }
 214
 215
 216 static void ctdb_vacuum_event(struct tevent_context *ev,
 217                               struct tevent_timer *te,
 218                               struct timeval t, void *private_data);
 219
 220 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 221 {
 222         struct ctdb_ltdb_header *header =
 223                 (struct ctdb_ltdb_header *)private_data;
 224
 225         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 226                 return -1;
 227         }
 228
 229         *header = *(struct ctdb_ltdb_header *)data.dptr;
 230
 231         return 0;
 232 }
 233
 234 /*
 235  * traverse function for gathering the records that can be deleted
 236  */
 237 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 238                            void *private_data)
 239 {
 240         struct vacuum_data *vdata = talloc_get_type(private_data,
 241                                                     struct vacuum_data);
 242         struct ctdb_context *ctdb = vdata->ctdb;
 243         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 244         uint32_t lmaster;
 245         struct ctdb_ltdb_header *hdr;
 246         int res = 0;
 247
 248         vdata->count.db_traverse.total++;
 249
 250         lmaster = ctdb_lmaster(ctdb, &key);
 251         if (lmaster >= ctdb->num_nodes) {
 252                 vdata->count.db_traverse.error++;
 253                 DEBUG(DEBUG_CRIT, (__location__
 254                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 255                                    " with hash[%u]!\n",
 256                                    (unsigned)lmaster,
 257                                    (unsigned)ctdb->num_nodes,
 258                                    (unsigned)ctdb_hash(&key)));
 259                 return -1;
 260         }
 261
 262         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 263                 /* it is not a deleted record */
 264                 vdata->count.db_traverse.skipped++;
 265                 return 0;
 266         }
 267
 268         hdr = (struct ctdb_ltdb_header *)data.dptr;
 269
 270         if (hdr->dmaster != ctdb->pnn) {
 271                 vdata->count.db_traverse.skipped++;
 272                 return 0;
 273         }
 274
 275         /*
 276          * Add the record to this process's delete_queue for processing
 277          * in the subsequent traverse in the fast vacuum run.
 278          */
 279         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 280         if (res != 0) {
 281                 vdata->count.db_traverse.error++;
 282         } else {
 283                 vdata->count.db_traverse.scheduled++;
 284         }
 285
 286         return 0;
 287 }
 288
 289 /*
 290  * traverse the tree of records to delete and marshall them into
 291  * a blob
 292  */
 293 static int delete_marshall_traverse(void *param, void *data)
 294 {
 295         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 296         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 297         struct ctdb_marshall_buffer *m;
 298
 299         m = ctdb_marshall_add(recs, recs->records, recs->records->db_id,
 300                               recs->records->db_id,
 301                               dd->key, &dd->hdr, tdb_null);
 302         if (m == NULL) {
 303                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 304                 return -1;
 305         }
 306
 307         recs->records = m;
 308         return 0;
 309 }
 310
 311 /**
 312  * traverse function for the traversal of the delete_queue,
 313  * the fast-path vacuuming list.
 314  *
 315  *  - If the record has been migrated off the node
 316  *    or has been revived (filled with data) on the node,
 317  *    then skip the record.
 318  *
 319  *  - If the current node is the record's lmaster and it is
 320  *    a record that has never been migrated with data, then
 321  *    delete the record from the local tdb.
 322  *
 323  *  - If the current node is the record's lmaster and it has
 324  *    been migrated with data, then schedule it for the normal
 325  *    vacuuming procedure (i.e. add it to the delete_list).
 326  *
 327  *  - If the current node is NOT the record's lmaster then
 328  *    add it to the list of records that are to be sent to
 329  *    the lmaster with the VACUUM_FETCH message.
 330  */
 331 static int delete_queue_traverse(void *param, void *data)
 332 {
 333         struct delete_record_data *dd =
 334                 talloc_get_type(data, struct delete_record_data);
 335         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 336         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 337         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 338         int res;
 339         struct ctdb_ltdb_header header;
 340         uint32_t lmaster;
 341         uint32_t hash = ctdb_hash(&(dd->key));
 342
 343         vdata->count.delete_queue.total++;
 344
 345         res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
 346         if (res != 0) {
 347                 vdata->count.delete_queue.error++;
 348                 return 0;
 349         }
 350
 351         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 352                                vacuum_record_parser, &header);
 353         if (res != 0) {
 354                 goto skipped;
 355         }
 356
 357         if (header.dmaster != ctdb->pnn) {
 358                 /* The record has been migrated off the node. Skip. */
 359                 goto skipped;
 360         }
 361
 362         if (header.rsn != dd->hdr.rsn) {
 363                 /*
 364                  * The record has been migrated off the node and back again.
 365                  * But not requeued for deletion. Skip it.
 366                  */
 367                 goto skipped;
 368         }
 369
 370         /*
 371          * We are dmaster, and the record has no data, and it has
 372          * not been migrated after it has been queued for deletion.
 373          *
 374          * At this stage, the record could still have been revived locally
 375          * and last been written with empty data. This can only be
 376          * fixed with the addition of an active or delete flag. (TODO)
 377          */
 378
 379         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 380
 381         if (lmaster != ctdb->pnn) {
 382                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 383
 384                 if (res != 0) {
 385                         DEBUG(DEBUG_ERR,
 386                               (__location__ " Error adding record to list "
 387                                "of records to send to lmaster.\n"));
 388                         vdata->count.delete_queue.error++;
 389                 } else {
 390                         vdata->count.delete_queue.added_to_vacuum_fetch_list++;
 391                 }
 392                 goto done;
 393         }
 394
 395         /* use header->flags or dd->hdr.flags ?? */
 396         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 397                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 398
 399                 if (res != 0) {
 400                         DEBUG(DEBUG_ERR,
 401                               (__location__ " Error adding record to list "
 402                                "of records for deletion on lmaster.\n"));
 403                         vdata->count.delete_queue.error++;
 404                 } else {
 405                         vdata->count.delete_queue.added_to_delete_list++;
 406                 }
 407         } else {
 408                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 409
 410                 if (res != 0) {
 411                         DEBUG(DEBUG_ERR,
 412                               (__location__ " Error deleting record with key "
 413                                "hash [0x%08x] from local data base db[%s].\n",
 414                                hash, ctdb_db->db_name));
 415                         vdata->count.delete_queue.error++;
 416                         goto done;
 417                 }
 418
 419                 DEBUG(DEBUG_DEBUG,
 420                       (__location__ " Deleted record with key hash "
 421                        "[0x%08x] from local data base db[%s].\n",
 422                        hash, ctdb_db->db_name));
 423                 vdata->count.delete_queue.deleted++;
 424         }
 425
 426         goto done;
 427
 428 skipped:
 429         vdata->count.delete_queue.skipped++;
 430
 431 done:
 432         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 433
 434         return 0;
 435 }
 436
 437 /**
 438  * Delete the records that we are lmaster and dmaster for and
 439  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 440  * control.
 441  */
 442 static int delete_record_traverse(void *param, void *data)
 443 {
 444         struct delete_record_data *dd =
 445                 talloc_get_type(data, struct delete_record_data);
 446         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 447         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 448         struct ctdb_context *ctdb = ctdb_db->ctdb;
 449         int res;
 450         struct ctdb_ltdb_header header;
 451         uint32_t lmaster;
 452         uint32_t hash = ctdb_hash(&(dd->key));
 453
 454         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 455         if (res != 0) {
 456                 DEBUG(DEBUG_ERR,
 457                       (__location__ " Error getting chainlock on record with "
 458                        "key hash [0x%08x] on database db[%s].\n",
 459                        hash, ctdb_db->db_name));
 460                 vdata->count.delete_list.local_error++;
 461                 vdata->count.delete_list.left--;
 462                 talloc_free(dd);
 463                 return 0;
 464         }
 465
 466         /*
 467          * Verify that the record is still empty, its RSN has not
 468          * changed and that we are still its lmaster and dmaster.
 469          */
 470
 471         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 472                                vacuum_record_parser, &header);
 473         if (res != 0) {
 474                 goto skip;
 475         }
 476
 477         if (header.flags & CTDB_REC_RO_FLAGS) {
 478                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 479                                    "on database db[%s] has read-only flags. "
 480                                    "skipping.\n",
 481                                    hash, ctdb_db->db_name));
 482                 goto skip;
 483         }
 484
 485         if (header.dmaster != ctdb->pnn) {
 486                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 487                                    "on database db[%s] has been migrated away. "
 488                                    "skipping.\n",
 489                                    hash, ctdb_db->db_name));
 490                 goto skip;
 491         }
 492
 493         if (header.rsn != dd->hdr.rsn) {
 494                 /*
 495                  * The record has been migrated off the node and back again.
 496                  * But not requeued for deletion. Skip it.
 497                  */
 498                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 499                                    "on database db[%s] seems to have been "
 500                                    "migrated away and back again (with empty "
 501                                    "data). skipping.\n",
 502                                    hash, ctdb_db->db_name));
 503                 goto skip;
 504         }
 505
 506         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 507
 508         if (lmaster != ctdb->pnn) {
 509                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 510                                    "delete list (key hash [0x%08x], db[%s]). "
 511                                    "Strange! skipping.\n",
 512                                    hash, ctdb_db->db_name));
 513                 goto skip;
 514         }
 515
 516         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 517
 518         if (res != 0) {
 519                 DEBUG(DEBUG_ERR,
 520                       (__location__ " Error deleting record with key hash "
 521                        "[0x%08x] from local data base db[%s].\n",
 522                        hash, ctdb_db->db_name));
 523                 vdata->count.delete_list.local_error++;
 524                 goto done;
 525         }
 526
 527         DEBUG(DEBUG_DEBUG,
 528               (__location__ " Deleted record with key hash [0x%08x] from "
 529                "local data base db[%s].\n", hash, ctdb_db->db_name));
 530
 531         vdata->count.delete_list.deleted++;
 532         goto done;
 533
 534 skip:
 535         vdata->count.delete_list.skipped++;
 536
 537 done:
 538         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 539
 540         talloc_free(dd);
 541         vdata->count.delete_list.left--;
 542
 543         return 0;
 544 }
 545
 546 /**
 547  * Traverse the delete_queue.
 548  * Records are either deleted directly or filled
 549  * into the delete list or the vacuum fetch lists
 550  * for further processing.
 551  */
 552 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 553                                       struct vacuum_data *vdata)
 554 {
 555         uint32_t sum;
 556         int ret;
 557
 558         ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
 559                                    delete_queue_traverse, vdata);
 560
 561         if (ret != 0) {
 562                 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
 563                       "the delete queue.\n"));
 564         }
 565
 566         sum = vdata->count.delete_queue.deleted
 567             + vdata->count.delete_queue.skipped
 568             + vdata->count.delete_queue.error
 569             + vdata->count.delete_queue.added_to_delete_list
 570             + vdata->count.delete_queue.added_to_vacuum_fetch_list;
 571
 572         if (vdata->count.delete_queue.total != sum) {
 573                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 574                       "counts for db[%s]: total[%u] != sum[%u]\n",
 575                       ctdb_db->db_name,
 576                       (unsigned)vdata->count.delete_queue.total,
 577                       (unsigned)sum));
 578         }
 579
 580         if (vdata->count.delete_queue.total > 0) {
 581                 DEBUG(DEBUG_INFO,
 582                       (__location__
 583                        " fast vacuuming delete_queue traverse statistics: "
 584                        "db[%s] "
 585                        "total[%u] "
 586                        "del[%u] "
 587                        "skp[%u] "
 588                        "err[%u] "
 589                        "adl[%u] "
 590                        "avf[%u]\n",
 591                        ctdb_db->db_name,
 592                        (unsigned)vdata->count.delete_queue.total,
 593                        (unsigned)vdata->count.delete_queue.deleted,
 594                        (unsigned)vdata->count.delete_queue.skipped,
 595                        (unsigned)vdata->count.delete_queue.error,
 596                        (unsigned)vdata->count.delete_queue.added_to_delete_list,
 597                        (unsigned)vdata->count.delete_queue.added_to_vacuum_fetch_list));
 598         }
 599
 600         return;
 601 }
 602
 603 /**
 604  * read-only traverse of the database, looking for records that
 605  * might be able to be vacuumed.
 606  *
 607  * This is not done each time but only every tunable
 608  * VacuumFastPathCount times.
 609  */
 610 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 611                                     struct vacuum_data *vdata)
 612 {
 613         int ret;
 614
 615         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 616         if (ret == -1 || vdata->traverse_error) {
 617                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 618                                   "'%s'\n", ctdb_db->db_name));
 619                 return;
 620         }
 621
 622         if (vdata->count.db_traverse.total > 0) {
 623                 DEBUG(DEBUG_INFO,
 624                       (__location__
 625                        " full vacuuming db traverse statistics: "
 626                        "db[%s] "
 627                        "total[%u] "
 628                        "skp[%u] "
 629                        "err[%u] "
 630                        "sched[%u]\n",
 631                        ctdb_db->db_name,
 632                        (unsigned)vdata->count.db_traverse.total,
 633                        (unsigned)vdata->count.db_traverse.skipped,
 634                        (unsigned)vdata->count.db_traverse.error,
 635                        (unsigned)vdata->count.db_traverse.scheduled));
 636         }
 637
 638         return;
 639 }
 640
 641 /**
 642  * Process the vacuum fetch lists:
 643  * For records for which we are not the lmaster, tell the lmaster to
 644  * fetch the record.
 645  */
 646 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 647                                             struct vacuum_data *vdata)
 648 {
 649         int i;
 650         struct ctdb_context *ctdb = ctdb_db->ctdb;
 651
 652         for (i = 0; i < ctdb->num_nodes; i++) {
 653                 TDB_DATA data;
 654                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 655
 656                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 657                         continue;
 658                 }
 659
 660                 if (vfl->count == 0) {
 661                         continue;
 662                 }
 663
 664                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 665                                    vfl->count, ctdb->nodes[i]->pnn,
 666                                    ctdb_db->db_name));
 667
 668                 data = ctdb_marshall_finish(vfl);
 669                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 670                                              CTDB_SRVID_VACUUM_FETCH,
 671                                              data) != 0)
 672                 {
 673                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 674                                           "fetch message to %u\n",
 675                                           ctdb->nodes[i]->pnn));
 676                 }
 677         }
 678
 679         return;
 680 }
 681
 682 /**
 683  * Process the delete list:
 684  *
 685  * This is the last step of vacuuming that consistently deletes
 686  * those records that have been migrated with data and can hence
 687  * not be deleted when leaving a node.
 688  *
 689  * In this step, the lmaster does the final deletion of those empty
 690  * records that it is also dmaster for. It has ususally received
 691  * at least some of these records previously from the former dmasters
 692  * with the vacuum fetch message.
 693  *
 694  *  1) Send the records to all active nodes with the TRY_DELETE_RECORDS
 695  *     control. The remote notes delete their local copy.
 696  *  2) The lmaster locally deletes its copies of all records that
 697  *     could successfully be deleted remotely in step #2.
 698  */
 699 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 700                                      struct vacuum_data *vdata)
 701 {
 702         int ret, i;
 703         struct ctdb_context *ctdb = ctdb_db->ctdb;
 704         struct delete_records_list *recs;
 705         TDB_DATA indata;
 706         struct ctdb_node_map_old *nodemap;
 707         uint32_t *active_nodes;
 708         int num_active_nodes;
 709         TALLOC_CTX *tmp_ctx;
 710         uint32_t sum;
 711
 712         if (vdata->count.delete_list.total == 0) {
 713                 return;
 714         }
 715
 716         tmp_ctx = talloc_new(vdata);
 717         if (tmp_ctx == NULL) {
 718                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 719                 return;
 720         }
 721
 722         vdata->count.delete_list.left = vdata->count.delete_list.total;
 723
 724         /*
 725          * get the list of currently active nodes
 726          */
 727
 728         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 729                                    CTDB_CURRENT_NODE,
 730                                    tmp_ctx,
 731                                    &nodemap);
 732         if (ret != 0) {
 733                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 734                 goto done;
 735         }
 736
 737         active_nodes = list_of_active_nodes(ctdb, nodemap,
 738                                             nodemap, /* talloc context */
 739                                             false /* include self */);
 740         /* yuck! ;-) */
 741         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 742
 743         /*
 744          * Now delete the records all active nodes in a two-phase process:
 745          * 1) tell all active remote nodes to delete all their copy
 746          * 2) if all remote nodes deleted their record copy, delete it locally
 747          */
 748
 749         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 750         if (recs == NULL) {
 751                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 752                 goto done;
 753         }
 754
 755         /*
 756          * Step 1:
 757          * Send all records to all active nodes for deletion.
 758          */
 759
 760         /*
 761          * Create a marshall blob from the remaining list of records to delete.
 762          */
 763
 764         recs->records = (struct ctdb_marshall_buffer *)
 765                 talloc_zero_size(recs,
 766                                  offsetof(struct ctdb_marshall_buffer, data));
 767         if (recs->records == NULL) {
 768                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 769                 goto done;
 770         }
 771         recs->records->db_id = ctdb_db->db_id;
 772
 773         ret = trbt_traversearray32(vdata->delete_list, 1,
 774                                    delete_marshall_traverse, recs);
 775         if (ret != 0) {
 776                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
 777                       "delete list for second marshalling.\n"));
 778                 goto done;
 779         }
 780
 781         indata = ctdb_marshall_finish(recs->records);
 782
 783         for (i = 0; i < num_active_nodes; i++) {
 784                 struct ctdb_marshall_buffer *records;
 785                 struct ctdb_rec_data_old *rec;
 786                 int32_t res;
 787                 TDB_DATA outdata;
 788
 789                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 790                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 791                                 indata, recs, &outdata, &res,
 792                                 NULL, NULL);
 793                 if (ret != 0 || res != 0) {
 794                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
 795                                           "node %u: ret[%d] res[%d]\n",
 796                                           active_nodes[i], ret, res));
 797                         goto done;
 798                 }
 799
 800                 /*
 801                  * outdata contains the list of records coming back
 802                  * from the node: These are the records that the
 803                  * remote node could not delete. We remove these from
 804                  * the list to delete locally.
 805                  */
 806                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 807                 rec = (struct ctdb_rec_data_old *)&records->data[0];
 808                 while (records->count-- > 1) {
 809                         TDB_DATA reckey, recdata;
 810                         struct ctdb_ltdb_header *rechdr;
 811                         struct delete_record_data *dd;
 812
 813                         reckey.dptr = &rec->data[0];
 814                         reckey.dsize = rec->keylen;
 815                         recdata.dptr = &rec->data[reckey.dsize];
 816                         recdata.dsize = rec->datalen;
 817
 818                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 819                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 820                                 goto done;
 821                         }
 822                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 823                         recdata.dptr += sizeof(*rechdr);
 824                         recdata.dsize -= sizeof(*rechdr);
 825
 826                         dd = (struct delete_record_data *)trbt_lookup32(
 827                                         vdata->delete_list,
 828                                         ctdb_hash(&reckey));
 829                         if (dd != NULL) {
 830                                 /*
 831                                  * The other node could not delete the
 832                                  * record and it is the first node that
 833                                  * failed. So we should remove it from
 834                                  * the tree and update statistics.
 835                                  */
 836                                 talloc_free(dd);
 837                                 vdata->count.delete_list.remote_error++;
 838                                 vdata->count.delete_list.left--;
 839                         } else {
 840                                 DEBUG(DEBUG_ERR, (__location__ " Failed to "
 841                                       "find record with hash 0x%08x coming "
 842                                       "back from TRY_DELETE_RECORDS "
 843                                       "control in delete list.\n",
 844                                       ctdb_hash(&reckey)));
 845                                 vdata->count.delete_list.local_error++;
 846                                 vdata->count.delete_list.left--;
 847                         }
 848
 849                         rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
 850                 }
 851         }
 852
 853         if (vdata->count.delete_list.left == 0) {
 854                 goto success;
 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 success:
 873
 874         if (vdata->count.delete_list.left != 0) {
 875                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
 876                       "there are %u records left for deletion after "
 877                       "processing delete list\n",
 878                       ctdb_db->db_name,
 879                       (unsigned)vdata->count.delete_list.left));
 880         }
 881
 882         sum = vdata->count.delete_list.deleted
 883             + vdata->count.delete_list.skipped
 884             + vdata->count.delete_list.remote_error
 885             + vdata->count.delete_list.local_error
 886             + vdata->count.delete_list.left;
 887
 888         if (vdata->count.delete_list.total != sum) {
 889                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
 890                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
 891                       ctdb_db->db_name,
 892                       (unsigned)vdata->count.delete_list.total,
 893                       (unsigned)sum));
 894         }
 895
 896         if (vdata->count.delete_list.total > 0) {
 897                 DEBUG(DEBUG_INFO,
 898                       (__location__
 899                        " vacuum delete list statistics: "
 900                        "db[%s] "
 901                        "total[%u] "
 902                        "del[%u] "
 903                        "skip[%u] "
 904                        "rem.err[%u] "
 905                        "loc.err[%u] "
 906                        "left[%u]\n",
 907                        ctdb_db->db_name,
 908                        (unsigned)vdata->count.delete_list.total,
 909                        (unsigned)vdata->count.delete_list.deleted,
 910                        (unsigned)vdata->count.delete_list.skipped,
 911                        (unsigned)vdata->count.delete_list.remote_error,
 912                        (unsigned)vdata->count.delete_list.local_error,
 913                        (unsigned)vdata->count.delete_list.left));
 914         }
 915
 916 done:
 917         talloc_free(tmp_ctx);
 918
 919         return;
 920 }
 921
 922 /**
 923  * initialize the vacuum_data
 924  */
 925 static struct vacuum_data *ctdb_vacuum_init_vacuum_data(
 926                                         struct ctdb_db_context *ctdb_db,
 927                                         TALLOC_CTX *mem_ctx)
 928 {
 929         int i;
 930         struct ctdb_context *ctdb = ctdb_db->ctdb;
 931         struct vacuum_data *vdata;
 932
 933         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 934         if (vdata == NULL) {
 935                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 936                 return NULL;
 937         }
 938
 939         vdata->ctdb = ctdb_db->ctdb;
 940         vdata->ctdb_db = ctdb_db;
 941         vdata->delete_list = trbt_create(vdata, 0);
 942         if (vdata->delete_list == NULL) {
 943                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 944                 goto fail;
 945         }
 946
 947         vdata->start = timeval_current();
 948
 949         vdata->count.delete_queue.added_to_delete_list = 0;
 950         vdata->count.delete_queue.added_to_vacuum_fetch_list = 0;
 951         vdata->count.delete_queue.deleted = 0;
 952         vdata->count.delete_queue.skipped = 0;
 953         vdata->count.delete_queue.error = 0;
 954         vdata->count.delete_queue.total = 0;
 955         vdata->count.db_traverse.scheduled = 0;
 956         vdata->count.db_traverse.skipped = 0;
 957         vdata->count.db_traverse.error = 0;
 958         vdata->count.db_traverse.total = 0;
 959         vdata->count.delete_list.total = 0;
 960         vdata->count.delete_list.left = 0;
 961         vdata->count.delete_list.remote_error = 0;
 962         vdata->count.delete_list.local_error = 0;
 963         vdata->count.delete_list.skipped = 0;
 964         vdata->count.delete_list.deleted = 0;
 965
 966         /* the list needs to be of length num_nodes */
 967         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
 968                                                 struct ctdb_marshall_buffer *,
 969                                                 ctdb->num_nodes);
 970         if (vdata->vacuum_fetch_list == NULL) {
 971                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 972                 goto fail;
 973         }
 974         for (i = 0; i < ctdb->num_nodes; i++) {
 975                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 976                         talloc_zero_size(vdata->vacuum_fetch_list,
 977                                          offsetof(struct ctdb_marshall_buffer, data));
 978                 if (vdata->vacuum_fetch_list[i] == NULL) {
 979                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 980                         talloc_free(vdata);
 981                         return NULL;
 982                 }
 983                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 984         }
 985
 986         return vdata;
 987
 988 fail:
 989         talloc_free(vdata);
 990         return NULL;
 991 }
 992
 993 /**
 994  * Vacuum a DB:
 995  *  - Always do the fast vacuuming run, which traverses
 996  *    the in-memory delete queue: these records have been
 997  *    scheduled for deletion.
 998  *  - Only if explicitly requested, the database is traversed
 999  *    in order to use the traditional heuristics on empty records
1000  *    to trigger deletion.
1001  *    This is done only every VacuumFastPathCount'th vacuuming run.
1002  *
1003  * The traverse runs fill two lists:
1004  *
1005  * - The delete_list:
1006  *   This is the list of empty records the current
1007  *   node is lmaster and dmaster for. These records are later
1008  *   deleted first on other nodes and then locally.
1009  *
1010  *   The fast vacuuming run has a short cut for those records
1011  *   that have never been migrated with data: these records
1012  *   are immediately deleted locally, since they have left
1013  *   no trace on other nodes.
1014  *
1015  * - The vacuum_fetch lists
1016  *   (one for each other lmaster node):
1017  *   The records in this list are sent for deletion to
1018  *   their lmaster in a bulk VACUUM_FETCH message.
1019  *
1020  *   The lmaster then migrates all these records to itelf
1021  *   so that they can be vacuumed there.
1022  *
1023  * This executes in the child context.
1024  */
1025 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1026                           bool full_vacuum_run)
1027 {
1028         struct ctdb_context *ctdb = ctdb_db->ctdb;
1029         int ret, pnn;
1030         struct vacuum_data *vdata;
1031         TALLOC_CTX *tmp_ctx;
1032
1033         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1034                            "%s db_id[0x%08x]\n",
1035                            full_vacuum_run ? "full" : "fast",
1036                            ctdb_db->db_name, ctdb_db->db_id));
1037
1038         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1039         if (ret != 0) {
1040                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1041                 return ret;
1042         }
1043
1044         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1045         if (pnn == -1) {
1046                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1047                 return -1;
1048         }
1049
1050         ctdb->pnn = pnn;
1051
1052         tmp_ctx = talloc_new(ctdb_db);
1053         if (tmp_ctx == NULL) {
1054                 DEBUG(DEBUG_ERR, ("Out of memory!\n"));
1055                 return -1;
1056         }
1057
1058         vdata = ctdb_vacuum_init_vacuum_data(ctdb_db, tmp_ctx);
1059         if (vdata == NULL) {
1060                 talloc_free(tmp_ctx);
1061                 return -1;
1062         }
1063
1064         if (full_vacuum_run) {
1065                 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1066         }
1067
1068         ctdb_process_delete_queue(ctdb_db, vdata);
1069
1070         ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1071
1072         ctdb_process_delete_list(ctdb_db, vdata);
1073
1074         talloc_free(tmp_ctx);
1075
1076         /* this ensures we run our event queue */
1077         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1078
1079         return 0;
1080 }
1081
1082 /*
1083  * repack and vaccum a db
1084  * called from the child context
1085  */
1086 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1087                                      bool full_vacuum_run)
1088 {
1089         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1090         const char *name = ctdb_db->db_name;
1091         int freelist_size = 0;
1092         int ret;
1093
1094         if (ctdb_vacuum_db(ctdb_db, full_vacuum_run) != 0) {
1095                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1096         }
1097
1098         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1099         if (freelist_size == -1) {
1100                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1101                 return -1;
1102         }
1103
1104         /*
1105          * decide if a repack is necessary
1106          */
1107         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1108         {
1109                 return 0;
1110         }
1111
1112         DEBUG(DEBUG_INFO, ("Repacking %s with %u freelist entries\n",
1113                            name, freelist_size));
1114
1115         ret = tdb_repack(ctdb_db->ltdb->tdb);
1116         if (ret != 0) {
1117                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1118                 return -1;
1119         }
1120
1121         return 0;
1122 }
1123
1124 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1125 {
1126         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1127
1128         return interval;
1129 }
1130
1131 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1132 {
1133         double l = timeval_elapsed(&child_ctx->start_time);
1134         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1135         struct ctdb_context *ctdb = ctdb_db->ctdb;
1136
1137         CTDB_UPDATE_DB_LATENCY(ctdb_db, "vacuum", vacuum.latency, l);
1138         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1139
1140         if (child_ctx->child_pid != -1) {
1141                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1142         } else {
1143                 /* Bump the number of successful fast-path runs. */
1144                 child_ctx->vacuum_handle->fast_path_count++;
1145         }
1146
1147         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1148
1149         tevent_add_timer(ctdb->ev, child_ctx->vacuum_handle,
1150                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1151                          ctdb_vacuum_event, child_ctx->vacuum_handle);
1152
1153         return 0;
1154 }
1155
1156 /*
1157  * this event is generated when a vacuum child process times out
1158  */
1159 static void vacuum_child_timeout(struct tevent_context *ev,
1160                                  struct tevent_timer *te,
1161                                  struct timeval t, void *private_data)
1162 {
1163         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1164
1165         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1166
1167         child_ctx->status = VACUUM_TIMEOUT;
1168
1169         talloc_free(child_ctx);
1170 }
1171
1172
1173 /*
1174  * this event is generated when a vacuum child process has completed
1175  */
1176 static void vacuum_child_handler(struct tevent_context *ev,
1177                                  struct tevent_fd *fde,
1178                                  uint16_t flags, void *private_data)
1179 {
1180         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1181         char c = 0;
1182         int ret;
1183
1184         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1185         child_ctx->child_pid = -1;
1186
1187         ret = sys_read(child_ctx->fd[0], &c, 1);
1188         if (ret != 1 || c != 0) {
1189                 child_ctx->status = VACUUM_ERROR;
1190                 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));
1191         } else {
1192                 child_ctx->status = VACUUM_OK;
1193         }
1194
1195         talloc_free(child_ctx);
1196 }
1197
1198 /*
1199  * this event is called every time we need to start a new vacuum process
1200  */
1201 static void ctdb_vacuum_event(struct tevent_context *ev,
1202                               struct tevent_timer *te,
1203                               struct timeval t, void *private_data)
1204 {
1205         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1206         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1207         struct ctdb_context *ctdb = ctdb_db->ctdb;
1208         struct ctdb_vacuum_child_context *child_ctx;
1209         struct tevent_fd *fde;
1210         int ret;
1211
1212         /* we don't vacuum if we are in recovery mode, or db frozen */
1213         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1214             ctdb_db_frozen(ctdb_db)) {
1215                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1216                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ?
1217                                         "in recovery" : "frozen"));
1218                 tevent_add_timer(ctdb->ev, vacuum_handle,
1219                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1220                                  ctdb_vacuum_event, vacuum_handle);
1221                 return;
1222         }
1223
1224         /* Do not allow multiple vacuuming child processes to be active at the
1225          * same time.  If there is vacuuming child process active, delay
1226          * new vacuuming event to stagger vacuuming events.
1227          */
1228         if (ctdb->vacuumers != NULL) {
1229                 tevent_add_timer(ctdb->ev, vacuum_handle,
1230                                  timeval_current_ofs(0, 500*1000),
1231                                  ctdb_vacuum_event, vacuum_handle);
1232                 return;
1233         }
1234
1235         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1236         if (child_ctx == NULL) {
1237                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1238                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1239         }
1240
1241
1242         ret = pipe(child_ctx->fd);
1243         if (ret != 0) {
1244                 talloc_free(child_ctx);
1245                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1246                 tevent_add_timer(ctdb->ev, vacuum_handle,
1247                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1248                                  ctdb_vacuum_event, vacuum_handle);
1249                 return;
1250         }
1251
1252         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1253                 vacuum_handle->fast_path_count = 0;
1254         }
1255
1256         child_ctx->child_pid = ctdb_fork(ctdb);
1257         if (child_ctx->child_pid == (pid_t)-1) {
1258                 close(child_ctx->fd[0]);
1259                 close(child_ctx->fd[1]);
1260                 talloc_free(child_ctx);
1261                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1262                 tevent_add_timer(ctdb->ev, vacuum_handle,
1263                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1264                                  ctdb_vacuum_event, vacuum_handle);
1265                 return;
1266         }
1267
1268
1269         if (child_ctx->child_pid == 0) {
1270                 char cc = 0;
1271                 bool full_vacuum_run = false;
1272                 close(child_ctx->fd[0]);
1273
1274                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1275                 prctl_set_comment("ctdb_vacuum");
1276                 if (switch_from_server_to_client(ctdb) != 0) {
1277                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1278                         _exit(1);
1279                 }
1280
1281                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1282                     (vacuum_handle->fast_path_count == 0))
1283                 {
1284                         full_vacuum_run = true;
1285                 }
1286                 cc = ctdb_vacuum_and_repack_db(ctdb_db, full_vacuum_run);
1287
1288                 sys_write(child_ctx->fd[1], &cc, 1);
1289                 _exit(0);
1290         }
1291
1292         set_close_on_exec(child_ctx->fd[0]);
1293         close(child_ctx->fd[1]);
1294
1295         child_ctx->status = VACUUM_RUNNING;
1296         child_ctx->start_time = timeval_current();
1297
1298         DLIST_ADD(ctdb->vacuumers, child_ctx);
1299         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1300
1301         /*
1302          * Clear the fastpath vacuuming list in the parent.
1303          */
1304         talloc_free(ctdb_db->delete_queue);
1305         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1306         if (ctdb_db->delete_queue == NULL) {
1307                 /* fatal here? ... */
1308                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1309                                  "in parent context. Shutting down\n");
1310         }
1311
1312         tevent_add_timer(ctdb->ev, child_ctx,
1313                          timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1314                          vacuum_child_timeout, child_ctx);
1315
1316         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1317
1318         fde = tevent_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1319                             TEVENT_FD_READ, vacuum_child_handler, child_ctx);
1320         tevent_fd_set_auto_close(fde);
1321
1322         vacuum_handle->child_ctx = child_ctx;
1323         child_ctx->vacuum_handle = vacuum_handle;
1324 }
1325
1326 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1327 {
1328         /* Simply free them all. */
1329         while (ctdb->vacuumers) {
1330                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1331                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1332                            (int)ctdb->vacuumers->child_pid));
1333                 /* vacuum_child_destructor kills it, removes from list */
1334                 talloc_free(ctdb->vacuumers);
1335         }
1336 }
1337
1338 /* this function initializes the vacuuming context for a database
1339  * starts the vacuuming events
1340  */
1341 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1342 {
1343         if (! ctdb_db_volatile(ctdb_db)) {
1344                 DEBUG(DEBUG_ERR,
1345                       ("Vacuuming is disabled for non-volatile database %s\n",
1346                        ctdb_db->db_name));
1347                 return 0;
1348         }
1349
1350         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1351         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1352
1353         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1354         ctdb_db->vacuum_handle->fast_path_count = 0;
1355
1356         tevent_add_timer(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1357                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1358                          ctdb_vacuum_event, ctdb_db->vacuum_handle);
1359
1360         return 0;
1361 }
1362
1363 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1364                                             const struct ctdb_ltdb_header *hdr,
1365                                             const TDB_DATA key)
1366 {
1367         struct delete_record_data *kd;
1368         uint32_t hash;
1369
1370         hash = (uint32_t)ctdb_hash(&key);
1371
1372         DEBUG(DEBUG_DEBUG, (__location__
1373                             " remove_record_from_delete_queue: "
1374                             "db[%s] "
1375                             "db_id[0x%08x] "
1376                             "key_hash[0x%08x] "
1377                             "lmaster[%u] "
1378                             "migrated_with_data[%s]\n",
1379                              ctdb_db->db_name, ctdb_db->db_id,
1380                              hash,
1381                              ctdb_lmaster(ctdb_db->ctdb, &key),
1382                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1383
1384         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1385         if (kd == NULL) {
1386                 DEBUG(DEBUG_DEBUG, (__location__
1387                                     " remove_record_from_delete_queue: "
1388                                     "record not in queue (hash[0x%08x])\n.",
1389                                     hash));
1390                 return;
1391         }
1392
1393         if ((kd->key.dsize != key.dsize) ||
1394             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1395         {
1396                 DEBUG(DEBUG_DEBUG, (__location__
1397                                     " remove_record_from_delete_queue: "
1398                                     "hash collision for key with hash[0x%08x] "
1399                                     "in db[%s] - skipping\n",
1400                                     hash, ctdb_db->db_name));
1401                 return;
1402         }
1403
1404         DEBUG(DEBUG_DEBUG, (__location__
1405                             " remove_record_from_delete_queue: "
1406                             "removing key with hash[0x%08x]\n",
1407                              hash));
1408
1409         talloc_free(kd);
1410
1411         return;
1412 }
1413
1414 /**
1415  * Insert a record into the ctdb_db context's delete queue,
1416  * handling hash collisions.
1417  */
1418 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1419                                            const struct ctdb_ltdb_header *hdr,
1420                                            TDB_DATA key)
1421 {
1422         struct delete_record_data *kd;
1423         uint32_t hash;
1424         int ret;
1425
1426         hash = (uint32_t)ctdb_hash(&key);
1427
1428         DEBUG(DEBUG_DEBUG, (__location__ " schedule for deletion: db[%s] "
1429                             "db_id[0x%08x] "
1430                             "key_hash[0x%08x] "
1431                             "lmaster[%u] "
1432                             "migrated_with_data[%s]\n",
1433                             ctdb_db->db_name, ctdb_db->db_id,
1434                             hash,
1435                             ctdb_lmaster(ctdb_db->ctdb, &key),
1436                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1437
1438         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1439         if (kd != NULL) {
1440                 if ((kd->key.dsize != key.dsize) ||
1441                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1442                 {
1443                         DEBUG(DEBUG_INFO,
1444                               (__location__ " schedule for deletion: "
1445                                "hash collision for key hash [0x%08x]. "
1446                                "Skipping the record.\n", hash));
1447                         return 0;
1448                 } else {
1449                         DEBUG(DEBUG_DEBUG,
1450                               (__location__ " schedule for deletion: "
1451                                "updating entry for key with hash [0x%08x].\n",
1452                                hash));
1453                 }
1454         }
1455
1456         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1457                                                   ctdb_db->delete_queue,
1458                                                   hdr, key);
1459         if (ret != 0) {
1460                 DEBUG(DEBUG_INFO,
1461                       (__location__ " schedule for deletion: error "
1462                        "inserting key with hash [0x%08x] into delete queue\n",
1463                        hash));
1464                 return -1;
1465         }
1466
1467         return 0;
1468 }
1469
1470 /**
1471  * Schedule a record for deletetion.
1472  * Called from the parent context.
1473  */
1474 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1475                                            TDB_DATA indata)
1476 {
1477         struct ctdb_control_schedule_for_deletion *dd;
1478         struct ctdb_db_context *ctdb_db;
1479         int ret;
1480         TDB_DATA key;
1481
1482         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1483
1484         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1485         if (ctdb_db == NULL) {
1486                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1487                                   dd->db_id));
1488                 return -1;
1489         }
1490
1491         key.dsize = dd->keylen;
1492         key.dptr = dd->key;
1493
1494         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1495
1496         return ret;
1497 }
1498
1499 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1500                                          const struct ctdb_ltdb_header *hdr,
1501                                          TDB_DATA key)
1502 {
1503         int ret;
1504         struct ctdb_control_schedule_for_deletion *dd;
1505         TDB_DATA indata;
1506         int32_t status;
1507
1508         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1509                 /* main daemon - directly queue */
1510                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1511
1512                 return ret;
1513         }
1514
1515         /* if we don't have a connection to the daemon we can not send
1516            a control. For example sometimes from update_record control child
1517            process.
1518         */
1519         if (!ctdb_db->ctdb->can_send_controls) {
1520                 return -1;
1521         }
1522
1523
1524         /* child process: send the main daemon a control */
1525         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1526         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1527         if (indata.dptr == NULL) {
1528                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1529                 return -1;
1530         }
1531         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1532         dd->db_id = ctdb_db->db_id;
1533         dd->hdr = *hdr;
1534         dd->keylen = key.dsize;
1535         memcpy(dd->key, key.dptr, key.dsize);
1536
1537         ret = ctdb_control(ctdb_db->ctdb,
1538                            CTDB_CURRENT_NODE,
1539                            ctdb_db->db_id,
1540                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1541                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1542                            indata,
1543                            NULL, /* mem_ctx */
1544                            NULL, /* outdata */
1545                            &status,
1546                            NULL, /* timeout : NULL == wait forever */
1547                            NULL); /* error message */
1548
1549         talloc_free(indata.dptr);
1550
1551         if (ret != 0 || status != 0) {
1552                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1553                                   "SCHEDULE_FOR_DELETION "
1554                                   "control.\n"));
1555                 if (status != 0) {
1556                         ret = -1;
1557                 }
1558         }
1559
1560         return ret;
1561 }
1562
1563 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1564                                          const struct ctdb_ltdb_header *hdr,
1565                                          const TDB_DATA key)
1566 {
1567         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1568                 /*
1569                  * Only remove the record from the delete queue if called
1570                  * in the main daemon.
1571                  */
1572                 return;
1573         }
1574
1575         remove_record_from_delete_queue(ctdb_db, hdr, key);
1576
1577         return;
1578 }