ctdb/server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5    Copyright (C) Michael Adam 2010-2011
   6    Copyright (C) Stefan Metzmacher 2010-2011
   7
   8    This program is free software; you can redistribute it and/or modify
   9    it under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3 of the License, or
  11    (at your option) any later version.
  12
  13    This program is distributed in the hope that it will be useful,
  14    but WITHOUT ANY WARRANTY; without even the implied warranty of
  15    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16    GNU General Public License for more details.
  17
  18    You should have received a copy of the GNU General Public License
  19    along with this program; if not, see <http://www.gnu.org/licenses/>.
  20 */
  21
  22 #include "includes.h"
  23 #include "lib/tevent/tevent.h"
  24 #include "lib/tdb/include/tdb.h"
  25 #include "system/network.h"
  26 #include "system/filesys.h"
  27 #include "system/dir.h"
  28 #include "../include/ctdb_private.h"
  29 #include "db_wrap.h"
  30 #include "lib/util/dlinklist.h"
  31 #include "lib/tevent/tevent.h"
  32 #include "../include/ctdb_private.h"
  33 #include "../common/rb_tree.h"
  34
  35 #define TIMELIMIT() timeval_current_ofs(10, 0)
  36
  37 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  38
  39 struct ctdb_vacuum_child_context {
  40         struct ctdb_vacuum_child_context *next, *prev;
  41         struct ctdb_vacuum_handle *vacuum_handle;
  42         /* fd child writes status to */
  43         int fd[2];
  44         pid_t child_pid;
  45         enum vacuum_child_status status;
  46         struct timeval start_time;
  47 };
  48
  49 struct ctdb_vacuum_handle {
  50         struct ctdb_db_context *ctdb_db;
  51         struct ctdb_vacuum_child_context *child_ctx;
  52         uint32_t fast_path_count;
  53 };
  54
  55
  56 /*  a list of records to possibly delete */
  57 struct vacuum_data {
  58         uint32_t vacuum_limit;
  59         uint32_t repack_limit;
  60         struct ctdb_context *ctdb;
  61         struct ctdb_db_context *ctdb_db;
  62         struct tdb_context *dest_db;
  63         trbt_tree_t *delete_list;
  64         uint32_t delete_count;
  65         struct ctdb_marshall_buffer **vacuum_fetch_list;
  66         struct timeval start;
  67         bool traverse_error;
  68         bool vacuum;
  69         uint32_t total;
  70         uint32_t vacuumed;
  71         uint32_t copied;
  72         uint32_t fast_added_to_vacuum_fetch_list;
  73         uint32_t fast_added_to_delete_list;
  74         uint32_t fast_deleted;
  75         uint32_t fast_skipped;
  76         uint32_t fast_error;
  77         uint32_t fast_total;
  78         uint32_t full_added_to_vacuum_fetch_list;
  79         uint32_t full_added_to_delete_list;
  80         uint32_t full_skipped;
  81         uint32_t full_error;
  82         uint32_t full_total;
  83         uint32_t delete_left;
  84         uint32_t delete_remote_error;
  85         uint32_t delete_local_error;
  86         uint32_t delete_deleted;
  87         uint32_t delete_skipped;
  88 };
  89
  90 /* this structure contains the information for one record to be deleted */
  91 struct delete_record_data {
  92         struct ctdb_context *ctdb;
  93         struct ctdb_db_context *ctdb_db;
  94         struct ctdb_ltdb_header hdr;
  95         TDB_DATA key;
  96 };
  97
  98 struct delete_records_list {
  99         struct ctdb_marshall_buffer *records;
 100 };
 101
 102 /**
 103  * Store key and header in a tree, indexed by the key hash.
 104  */
 105 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 106                                                struct ctdb_db_context *ctdb_db,
 107                                                trbt_tree_t *tree,
 108                                                const struct ctdb_ltdb_header *hdr,
 109                                                TDB_DATA key)
 110 {
 111         struct delete_record_data *dd;
 112         uint32_t hash;
 113
 114         dd = talloc_zero(tree, struct delete_record_data);
 115         if (dd == NULL) {
 116                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 117                 return -1;
 118         }
 119
 120         dd->ctdb      = ctdb;
 121         dd->ctdb_db   = ctdb_db;
 122         dd->key.dsize = key.dsize;
 123         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 124         if (dd->key.dptr == NULL) {
 125                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 126                 return -1;
 127         }
 128
 129         dd->hdr = *hdr;
 130
 131         hash = ctdb_hash(&key);
 132
 133         trbt_insert32(tree, hash, dd);
 134
 135         return 0;
 136 }
 137
 138 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 139                                      struct ctdb_ltdb_header *hdr)
 140 {
 141         struct ctdb_context *ctdb = vdata->ctdb;
 142         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 143         uint32_t hash;
 144         int ret;
 145
 146         hash = ctdb_hash(&key);
 147
 148         if (trbt_lookup32(vdata->delete_list, hash)) {
 149                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 150                 return 0;
 151         }
 152
 153         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 154                                                   vdata->delete_list,
 155                                                   hdr, key);
 156         if (ret != 0) {
 157                 return -1;
 158         }
 159
 160         vdata->delete_count++;
 161
 162         return 0;
 163 }
 164
 165 /**
 166  * Add a record to the list of records to be sent
 167  * to their lmaster with VACUUM_FETCH.
 168  */
 169 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 170                                            TDB_DATA key)
 171 {
 172         struct ctdb_context *ctdb = vdata->ctdb;
 173         struct ctdb_rec_data *rec;
 174         uint32_t lmaster;
 175         size_t old_size;
 176         struct ctdb_marshall_buffer *vfl;
 177
 178         lmaster = ctdb_lmaster(ctdb, &key);
 179
 180         vfl = vdata->vacuum_fetch_list[lmaster];
 181
 182         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 183         if (rec == NULL) {
 184                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 185                 vdata->traverse_error = true;
 186                 return -1;
 187         }
 188
 189         old_size = talloc_get_size(vfl);
 190         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 191         if (vfl == NULL) {
 192                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 193                 vdata->traverse_error = true;
 194                 return -1;
 195         }
 196         vdata->vacuum_fetch_list[lmaster] = vfl;
 197
 198         vfl->count++;
 199         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 200         talloc_free(rec);
 201
 202         vdata->total++;
 203
 204         return 0;
 205 }
 206
 207
 208 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 209                               struct timeval t, void *private_data);
 210
 211
 212 /*
 213  * traverse function for gathering the records that can be deleted
 214  */
 215 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 216 {
 217         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 218         struct ctdb_context *ctdb = vdata->ctdb;
 219         uint32_t lmaster;
 220         struct ctdb_ltdb_header *hdr;
 221         int res = 0;
 222
 223         vdata->full_total++;
 224
 225         lmaster = ctdb_lmaster(ctdb, &key);
 226         if (lmaster >= ctdb->num_nodes) {
 227                 vdata->full_error++;
 228                 DEBUG(DEBUG_CRIT, (__location__
 229                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 230                                    " with hash[%u]!\n",
 231                                    (unsigned)lmaster,
 232                                    (unsigned)ctdb->num_nodes,
 233                                    (unsigned)ctdb_hash(&key)));
 234                 return -1;
 235         }
 236
 237         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 238                 /* it is not a deleted record */
 239                 vdata->full_skipped++;
 240                 return 0;
 241         }
 242
 243         hdr = (struct ctdb_ltdb_header *)data.dptr;
 244
 245         if (hdr->dmaster != ctdb->pnn) {
 246                 vdata->full_skipped++;
 247                 return 0;
 248         }
 249
 250         if (lmaster == ctdb->pnn) {
 251                 /*
 252                  * We are both lmaster and dmaster, and the record is empty.
 253                  * So we should be able to delete it.
 254                  */
 255                 res = add_record_to_delete_list(vdata, key, hdr);
 256                 if (res != 0) {
 257                         vdata->full_error++;
 258                 } else {
 259                         vdata->full_added_to_delete_list++;
 260                 }
 261         } else {
 262                 /*
 263                  * We are not lmaster.
 264                  * Add the record to the blob ready to send to the nodes.
 265                  */
 266                 res = add_record_to_vacuum_fetch_list(vdata, key);
 267                 if (res != 0) {
 268                         vdata->full_error++;
 269                 } else {
 270                         vdata->full_added_to_vacuum_fetch_list++;
 271                 }
 272         }
 273
 274         return res;
 275 }
 276
 277 /*
 278  * traverse the tree of records to delete and marshall them into
 279  * a blob
 280  */
 281 static int delete_marshall_traverse(void *param, void *data)
 282 {
 283         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 284         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 285         struct ctdb_rec_data *rec;
 286         size_t old_size;
 287
 288         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 289         if (rec == NULL) {
 290                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 291                 return 0;
 292         }
 293
 294         old_size = talloc_get_size(recs->records);
 295         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 296         if (recs->records == NULL) {
 297                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 298                 return 0;
 299         }
 300         recs->records->count++;
 301         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 302         return 0;
 303 }
 304
 305 /**
 306  * traverse function for the traversal of the delete_queue,
 307  * the fast-path vacuuming list.
 308  *
 309  *  - If the record has been migrated off the node
 310  *    or has been revived (filled with data) on the node,
 311  *    then skip the record.
 312  *
 313  *  - If the current node is the record's lmaster and it is
 314  *    a record that has never been migrated with data, then
 315  *    delete the record from the local tdb.
 316  *
 317  *  - If the current node is the record's lmaster and it has
 318  *    been migrated with data, then schedule it for the normal
 319  *    vacuuming procedure (i.e. add it to the delete_list).
 320  *
 321  *  - If the current node is NOT the record's lmaster then
 322  *    add it to the list of records that are to be sent to
 323  *    the lmaster with the VACUUM_FETCH message.
 324  */
 325 static int delete_queue_traverse(void *param, void *data)
 326 {
 327         struct delete_record_data *dd =
 328                 talloc_get_type(data, struct delete_record_data);
 329         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 330         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 331         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 332         int res;
 333         struct ctdb_ltdb_header *header;
 334         TDB_DATA tdb_data;
 335         uint32_t lmaster;
 336
 337         vdata->fast_total++;
 338
 339         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 340         if (res != 0) {
 341                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 342                 vdata->fast_error++;
 343                 return 0;
 344         }
 345
 346         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 347         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 348                 /* Does not exist or not a ctdb record. Skip. */
 349                 goto skipped;
 350         }
 351
 352         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 353                 /* The record has been recycled (filled with data). Skip. */
 354                 goto skipped;
 355         }
 356
 357         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 358
 359         if (header->dmaster != ctdb->pnn) {
 360                 /* The record has been migrated off the node. Skip. */
 361                 goto skipped;
 362         }
 363
 364
 365         if (header->rsn != dd->hdr.rsn) {
 366                 /*
 367                  * The record has been migrated off the node and back again.
 368                  * But not requeued for deletion. Skip it.
 369                  */
 370                 goto skipped;
 371         }
 372
 373         /*
 374          * We are dmaster, and the record has no data, and it has
 375          * not been migrated after it has been queued for deletion.
 376          *
 377          * At this stage, the record could still have been revived locally
 378          * and last been written with empty data. This can only be
 379          * fixed with the addition of an active or delete flag. (TODO)
 380          */
 381
 382         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 383
 384         if (lmaster != ctdb->pnn) {
 385                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 386
 387                 if (res != 0) {
 388                         DEBUG(DEBUG_ERR,
 389                               (__location__ " Error adding record to list "
 390                                "of records to send to lmaster.\n"));
 391                         vdata->fast_error++;
 392                 } else {
 393                         vdata->fast_added_to_vacuum_fetch_list++;
 394                 }
 395                 goto done;
 396         }
 397
 398         /* use header->flags or dd->hdr.flags ?? */
 399         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 400                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 401
 402                 if (res != 0) {
 403                         DEBUG(DEBUG_ERR,
 404                               (__location__ " Error adding record to list "
 405                                "of records for deletion on lmaster.\n"));
 406                         vdata->fast_error++;
 407                 } else {
 408                         vdata->fast_added_to_delete_list++;
 409                 }
 410         } else {
 411                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 412
 413                 if (res != 0) {
 414                         DEBUG(DEBUG_ERR,
 415                               (__location__ " Error deleting record from local "
 416                                "data base.\n"));
 417                         vdata->fast_error++;
 418                 } else {
 419                         vdata->fast_deleted++;
 420                 }
 421         }
 422
 423         goto done;
 424
 425 skipped:
 426         vdata->fast_skipped++;
 427
 428 done:
 429         if (tdb_data.dptr != NULL) {
 430                 free(tdb_data.dptr);
 431         }
 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         TDB_DATA tdb_data;
 452         uint32_t lmaster;
 453         bool deleted = false;
 454
 455         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 456         if (res != 0) {
 457                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 458                 vdata->delete_local_error++;
 459                 return 0;
 460         }
 461
 462         /*
 463          * Verify that the record is still empty, its RSN has not
 464          * changed and that we are still its lmaster and dmaster.
 465          */
 466
 467         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 468         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 469                 /* Does not exist or not a ctdb record. Skip. */
 470                 vdata->delete_skipped++;
 471                 goto done;
 472         }
 473
 474         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 475                 /* The record has been recycled (filled with data). Skip. */
 476                 vdata->delete_skipped++;
 477                 goto done;
 478         }
 479
 480         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 481
 482         if (header->dmaster != ctdb->pnn) {
 483                 /* The record has been migrated off the node. Skip. */
 484                 vdata->delete_skipped++;
 485                 goto done;
 486         }
 487
 488
 489         if (header->rsn != dd->hdr.rsn) {
 490                 /*
 491                  * The record has been migrated off the node and back again.
 492                  * But not requeued for deletion. Skip it.
 493                  */
 494                 vdata->delete_skipped++;
 495                 goto done;
 496         }
 497
 498         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 499
 500         if (lmaster != ctdb->pnn) {
 501                 /* we are not lmaster - strange */
 502                 vdata->delete_skipped++;
 503                 goto done;
 504         }
 505
 506         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 507
 508         if (res != 0) {
 509                 DEBUG(DEBUG_ERR,
 510                       (__location__ " Error deleting record from local "
 511                        "data base.\n"));
 512                 vdata->delete_local_error++;
 513                 goto done;
 514         }
 515
 516         deleted = true;
 517
 518 done:
 519         if (tdb_data.dptr != NULL) {
 520                 free(tdb_data.dptr);
 521         }
 522
 523         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 524
 525         if (deleted) {
 526                 /*
 527                  * successfully deleted the record locally.
 528                  * remove it from the list and update statistics.
 529                  */
 530                 talloc_free(dd);
 531                 vdata->delete_deleted++;
 532                 vdata->delete_left--;
 533         }
 534
 535         return 0;
 536 }
 537
 538 /**
 539  * Fast vacuuming run:
 540  * Traverse the delete_queue.
 541  * This fills the same lists as the database traverse.
 542  */
 543 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
 544                                 struct vacuum_data *vdata)
 545 {
 546         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 547
 548         if (vdata->fast_total > 0) {
 549                 DEBUG(DEBUG_INFO,
 550                       (__location__
 551                        " fast vacuuming delete_queue traverse statistics: "
 552                        "db[%s] "
 553                        "total[%u] "
 554                        "del[%u] "
 555                        "skp[%u] "
 556                        "err[%u] "
 557                        "adl[%u] "
 558                        "avf[%u]\n",
 559                        ctdb_db->db_name,
 560                        (unsigned)vdata->fast_total,
 561                        (unsigned)vdata->fast_deleted,
 562                        (unsigned)vdata->fast_skipped,
 563                        (unsigned)vdata->fast_error,
 564                        (unsigned)vdata->fast_added_to_delete_list,
 565                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 566         }
 567
 568         return;
 569 }
 570
 571 /**
 572  * Vacuum a DB:
 573  *  - Always do the fast vacuuming run, which traverses
 574  *    the in-memory delete queue: these records have been
 575  *    scheduled for deletion.
 576  *  - Only if explicitly requested, the database is traversed
 577  *    in order to use the traditional heuristics on empty records
 578  *    to trigger deletion.
 579  *    This is done only every VacuumFastPathCount'th vacuuming run.
 580  *
 581  * The traverse runs fill two lists:
 582  *
 583  * - The delete_list:
 584  *   This is the list of empty records the current
 585  *   node is lmaster and dmaster for. These records are later
 586  *   deleted first on other nodes and then locally.
 587  *
 588  *   The fast vacuuming run has a short cut for those records
 589  *   that have never been migrated with data: these records
 590  *   are immediately deleted locally, since they have left
 591  *   no trace on other nodes.
 592  *
 593  * - The vacuum_fetch lists
 594  *   (one for each other lmaster node):
 595  *   The records in this list are sent for deletion to
 596  *   their lmaster in a bulk VACUUM_FETCH message.
 597  *
 598  *   The lmaster then migrates all these records to itelf
 599  *   so that they can be vacuumed there.
 600  *
 601  * This executes in the child context.
 602  */
 603 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 604                           struct vacuum_data *vdata,
 605                           bool full_vacuum_run)
 606 {
 607         struct ctdb_context *ctdb = ctdb_db->ctdb;
 608         const char *name = ctdb_db->db_name;
 609         int ret, i, pnn;
 610
 611         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 612                            "%s db_id[0x%08x]\n",
 613                            full_vacuum_run ? "full" : "fast",
 614                            ctdb_db->db_name, ctdb_db->db_id));
 615
 616         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 617         if (ret != 0) {
 618                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 619                 return ret;
 620         }
 621
 622         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 623         if (pnn == -1) {
 624                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 625                 return -1;
 626         }
 627
 628         ctdb->pnn = pnn;
 629
 630         vdata->fast_added_to_delete_list = 0;
 631         vdata->fast_added_to_vacuum_fetch_list = 0;
 632         vdata->fast_deleted = 0;
 633         vdata->fast_skipped = 0;
 634         vdata->fast_error = 0;
 635         vdata->fast_total = 0;
 636         vdata->full_added_to_delete_list = 0;
 637         vdata->full_added_to_vacuum_fetch_list = 0;
 638         vdata->full_skipped = 0;
 639         vdata->full_error = 0;
 640         vdata->full_total = 0;
 641         vdata->delete_count = 0;
 642         vdata->delete_left = 0;
 643         vdata->delete_remote_error = 0;
 644         vdata->delete_local_error = 0;
 645         vdata->delete_skipped = 0;
 646         vdata->delete_deleted = 0;
 647
 648         /* the list needs to be of length num_nodes */
 649         vdata->vacuum_fetch_list = talloc_array(vdata,
 650                                                 struct ctdb_marshall_buffer *,
 651                                                 ctdb->num_nodes);
 652         if (vdata->vacuum_fetch_list == NULL) {
 653                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 654                 return -1;
 655         }
 656         for (i = 0; i < ctdb->num_nodes; i++) {
 657                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 658                         talloc_zero_size(vdata->vacuum_fetch_list,
 659                                          offsetof(struct ctdb_marshall_buffer, data));
 660                 if (vdata->vacuum_fetch_list[i] == NULL) {
 661                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 662                         return -1;
 663                 }
 664                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 665         }
 666
 667         ctdb_vacuum_db_fast(ctdb_db, vdata);
 668
 669         /*
 670          * read-only traverse of the database, looking for records that
 671          * might be able to be vacuumed.
 672          *
 673          * This is not done each time but only every tunable
 674          * VacuumFastPathCount times.
 675          */
 676         if (full_vacuum_run) {
 677                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 678                 if (ret == -1 || vdata->traverse_error) {
 679                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 680                         return -1;
 681                 }
 682                 if (vdata->full_total > 0) {
 683                         DEBUG(DEBUG_INFO,
 684                               (__location__
 685                                " full vacuuming db traverse statistics: "
 686                                "db[%s] "
 687                                "total[%u] "
 688                                "skp[%u] "
 689                                "err[%u] "
 690                                "adl[%u] "
 691                                "avf[%u]\n",
 692                                ctdb_db->db_name,
 693                                (unsigned)vdata->full_total,
 694                                (unsigned)vdata->full_skipped,
 695                                (unsigned)vdata->full_error,
 696                                (unsigned)vdata->full_added_to_delete_list,
 697                                (unsigned)vdata->full_added_to_vacuum_fetch_list));
 698                 }
 699         }
 700
 701         /*
 702          * For records where we are not the lmaster,
 703          * tell the lmaster to fetch the record.
 704          */
 705         for (i = 0; i < ctdb->num_nodes; i++) {
 706                 TDB_DATA data;
 707                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 708
 709                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 710                         continue;
 711                 }
 712
 713                 if (vfl->count == 0) {
 714                         continue;
 715                 }
 716
 717                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 718                                    vfl->count, ctdb->nodes[i]->pnn,
 719                                    name));
 720
 721                 data.dsize = talloc_get_size(vfl);
 722                 data.dptr  = (void *)vfl;
 723                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 724                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 725                                           "fetch message to %u\n",
 726                                           ctdb->nodes[i]->pnn));
 727                         return -1;
 728                 }
 729         }
 730
 731         /* Process all records we can delete (if any) */
 732         if (vdata->delete_count > 0) {
 733                 struct delete_records_list *recs;
 734                 TDB_DATA indata, outdata;
 735                 int32_t res;
 736                 struct ctdb_node_map *nodemap;
 737                 uint32_t *active_nodes;
 738                 int num_active_nodes;
 739
 740                 vdata->delete_left = vdata->delete_count;
 741
 742                 recs = talloc_zero(vdata, struct delete_records_list);
 743                 if (recs == NULL) {
 744                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 745                         return -1;
 746                 }
 747                 recs->records = (struct ctdb_marshall_buffer *)
 748                         talloc_zero_size(vdata,
 749                                     offsetof(struct ctdb_marshall_buffer, data));
 750                 if (recs->records == NULL) {
 751                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 752                         return -1;
 753                 }
 754                 recs->records->db_id = ctdb_db->db_id;
 755
 756                 /*
 757                  * traverse the tree of all records we want to delete and
 758                  * create a blob we can send to the other nodes.
 759                  */
 760                 trbt_traversearray32(vdata->delete_list, 1,
 761                                      delete_marshall_traverse, recs);
 762
 763                 indata.dsize = talloc_get_size(recs->records);
 764                 indata.dptr  = (void *)recs->records;
 765
 766                 /*
 767                  * now tell all the active nodes to delete all these records
 768                  * (if possible)
 769                  */
 770
 771                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 772                                            CTDB_CURRENT_NODE,
 773                                            recs, /* talloc context */
 774                                            &nodemap);
 775                 if (ret != 0) {
 776                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 777                         return -1;
 778                 }
 779
 780                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 781                                                     nodemap, /* talloc context */
 782                                                     false /* include self */);
 783                 /* yuck! ;-) */
 784                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 785
 786                 for (i = 0; i < num_active_nodes; i++) {
 787                         struct ctdb_marshall_buffer *records;
 788                         struct ctdb_rec_data *rec;
 789
 790                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 791                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 792                                         indata, recs, &outdata, &res,
 793                                         NULL, NULL);
 794                         if (ret != 0 || res != 0) {
 795                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 796                                                   "node %u: ret[%d] res[%d]\n",
 797                                                   active_nodes[i], ret, res));
 798                                 return -1;
 799                         }
 800
 801                         /*
 802                          * outdata contains the list of records coming back
 803                          * from the node: These are the records that the
 804                          * remote node could not delete.
 805                          */
 806                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 807                         rec = (struct ctdb_rec_data *)&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                                         return -1;
 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->delete_remote_error++;
 838                                         vdata->delete_left--;
 839                                 }
 840
 841                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 842                         }
 843                 }
 844
 845                 /* free nodemap and active_nodes */
 846                 talloc_free(nodemap);
 847         }
 848
 849         if (vdata->delete_left > 0) {
 850                 /*
 851                  * The only records remaining in the tree are those
 852                  * records which all other nodes could successfully
 853                  * delete, so we can safely delete them on the
 854                  * lmaster as well.
 855                  */
 856                 trbt_traversearray32(vdata->delete_list, 1,
 857                                      delete_record_traverse, vdata);
 858         }
 859
 860         /* this ensures we run our event queue */
 861         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 862
 863         return 0;
 864 }
 865
 866
 867 /*
 868  * traverse function for repacking
 869  */
 870 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 871 {
 872         struct vacuum_data *vdata = (struct vacuum_data *)private;
 873
 874         if (vdata->vacuum) {
 875                 uint32_t hash = ctdb_hash(&key);
 876                 struct delete_record_data *kd;
 877                 /*
 878                  * check if we can ignore this record because it's in the delete_list
 879                  */
 880                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
 881                 /*
 882                  * there might be hash collisions so we have to compare the keys here to be sure
 883                  */
 884                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 885                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 886                         /*
 887                          * we have to check if the record hasn't changed in the meantime in order to
 888                          * savely remove it from the database
 889                          */
 890                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 891                                 hdr->dmaster == kd->ctdb->pnn &&
 892                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 893                                 kd->hdr.rsn == hdr->rsn) {
 894                                 vdata->vacuumed++;
 895                                 return 0;
 896                         }
 897                 }
 898         }
 899         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 900                 vdata->traverse_error = true;
 901                 return -1;
 902         }
 903         vdata->copied++;
 904         return 0;
 905 }
 906
 907 /*
 908  * repack a tdb
 909  */
 910 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 911 {
 912         struct tdb_context *tmp_db;
 913
 914         if (tdb_transaction_start(tdb) != 0) {
 915                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 916                 return -1;
 917         }
 918
 919         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 920                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 921                           O_RDWR|O_CREAT, 0);
 922         if (tmp_db == NULL) {
 923                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 924                 tdb_transaction_cancel(tdb);
 925                 return -1;
 926         }
 927
 928         vdata->traverse_error = false;
 929         vdata->dest_db = tmp_db;
 930         vdata->vacuum = true;
 931         vdata->vacuumed = 0;
 932         vdata->copied = 0;
 933
 934         /*
 935          * repack and vacuum on-the-fly by not writing the records that are
 936          * no longer needed
 937          */
 938         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 939                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 940                 tdb_transaction_cancel(tdb);
 941                 tdb_close(tmp_db);
 942                 return -1;
 943         }
 944
 945         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 946
 947         if (vdata->traverse_error) {
 948                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 949                 tdb_transaction_cancel(tdb);
 950                 tdb_close(tmp_db);
 951                 return -1;
 952         }
 953
 954         if (tdb_wipe_all(tdb) != 0) {
 955                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 956                 tdb_transaction_cancel(tdb);
 957                 tdb_close(tmp_db);
 958                 return -1;
 959         }
 960
 961         vdata->traverse_error = false;
 962         vdata->dest_db = tdb;
 963         vdata->vacuum = false;
 964         vdata->copied = 0;
 965
 966         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 967                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 968                 tdb_transaction_cancel(tdb);
 969                 tdb_close(tmp_db);
 970                 return -1;
 971         }
 972
 973         if (vdata->traverse_error) {
 974                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 975                 tdb_transaction_cancel(tdb);
 976                 tdb_close(tmp_db);
 977                 return -1;
 978         }
 979
 980         tdb_close(tmp_db);
 981
 982
 983         if (tdb_transaction_commit(tdb) != 0) {
 984                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 985                 return -1;
 986         }
 987         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 988
 989         return 0;
 990 }
 991
 992 /*
 993  * repack and vaccum a db
 994  * called from the child context
 995  */
 996 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 997                                      TALLOC_CTX *mem_ctx,
 998                                      bool full_vacuum_run)
 999 {
1000         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1001         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1002         const char *name = ctdb_db->db_name;
1003         int freelist_size;
1004         struct vacuum_data *vdata;
1005
1006         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1007         if (freelist_size == -1) {
1008                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1009                 return -1;
1010         }
1011
1012         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1013         if (vdata == NULL) {
1014                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1015                 return -1;
1016         }
1017
1018         vdata->ctdb = ctdb_db->ctdb;
1019         vdata->vacuum_limit = vacuum_limit;
1020         vdata->repack_limit = repack_limit;
1021         vdata->delete_list = trbt_create(vdata, 0);
1022         vdata->ctdb_db = ctdb_db;
1023         if (vdata->delete_list == NULL) {
1024                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1025                 talloc_free(vdata);
1026                 return -1;
1027         }
1028
1029         vdata->start = timeval_current();
1030
1031         /*
1032          * gather all records that can be deleted in vdata
1033          */
1034         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1035                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1036         }
1037
1038         /*
1039          * decide if a repack is necessary
1040          */
1041         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1042         {
1043                 talloc_free(vdata);
1044                 return 0;
1045         }
1046
1047         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1048                         name, freelist_size, vdata->delete_left));
1049
1050         /*
1051          * repack and implicitely get rid of the records we can delete
1052          */
1053         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1054                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1055                 talloc_free(vdata);
1056                 return -1;
1057         }
1058         talloc_free(vdata);
1059
1060         return 0;
1061 }
1062
1063 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1064 {
1065         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1066
1067         return interval;
1068 }
1069
1070 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1071 {
1072         double l = timeval_elapsed(&child_ctx->start_time);
1073         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1074         struct ctdb_context *ctdb = ctdb_db->ctdb;
1075
1076         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1077
1078         if (child_ctx->child_pid != -1) {
1079                 kill(child_ctx->child_pid, SIGKILL);
1080         } else {
1081                 /* Bump the number of successful fast-path runs. */
1082                 child_ctx->vacuum_handle->fast_path_count++;
1083         }
1084
1085         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1086
1087         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1088                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1089                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1090
1091         return 0;
1092 }
1093
1094 /*
1095  * this event is generated when a vacuum child process times out
1096  */
1097 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1098                                          struct timeval t, void *private_data)
1099 {
1100         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1101
1102         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1103
1104         child_ctx->status = VACUUM_TIMEOUT;
1105
1106         talloc_free(child_ctx);
1107 }
1108
1109
1110 /*
1111  * this event is generated when a vacuum child process has completed
1112  */
1113 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1114                              uint16_t flags, void *private_data)
1115 {
1116         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1117         char c = 0;
1118         int ret;
1119
1120         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1121         child_ctx->child_pid = -1;
1122
1123         ret = read(child_ctx->fd[0], &c, 1);
1124         if (ret != 1 || c != 0) {
1125                 child_ctx->status = VACUUM_ERROR;
1126                 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));
1127         } else {
1128                 child_ctx->status = VACUUM_OK;
1129         }
1130
1131         talloc_free(child_ctx);
1132 }
1133
1134 /*
1135  * this event is called every time we need to start a new vacuum process
1136  */
1137 static void
1138 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1139                                struct timeval t, void *private_data)
1140 {
1141         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1142         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1143         struct ctdb_context *ctdb = ctdb_db->ctdb;
1144         struct ctdb_vacuum_child_context *child_ctx;
1145         struct tevent_fd *fde;
1146         int ret;
1147
1148         /* we dont vacuum if we are in recovery mode, or db frozen */
1149         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1150             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1151                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1152                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1153                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1154                                    ? "freeze pending"
1155                                    : "frozen"));
1156                 event_add_timed(ctdb->ev, vacuum_handle,
1157                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1158                         ctdb_vacuum_event, vacuum_handle);
1159                 return;
1160         }
1161
1162         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1163         if (child_ctx == NULL) {
1164                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1165                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1166         }
1167
1168
1169         ret = pipe(child_ctx->fd);
1170         if (ret != 0) {
1171                 talloc_free(child_ctx);
1172                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1173                 event_add_timed(ctdb->ev, vacuum_handle,
1174                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1175                         ctdb_vacuum_event, vacuum_handle);
1176                 return;
1177         }
1178
1179         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1180                 vacuum_handle->fast_path_count = 0;
1181         }
1182
1183         child_ctx->child_pid = ctdb_fork(ctdb);
1184         if (child_ctx->child_pid == (pid_t)-1) {
1185                 close(child_ctx->fd[0]);
1186                 close(child_ctx->fd[1]);
1187                 talloc_free(child_ctx);
1188                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1189                 event_add_timed(ctdb->ev, vacuum_handle,
1190                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1191                         ctdb_vacuum_event, vacuum_handle);
1192                 return;
1193         }
1194
1195
1196         if (child_ctx->child_pid == 0) {
1197                 char cc = 0;
1198                 bool full_vacuum_run = false;
1199                 close(child_ctx->fd[0]);
1200
1201                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1202
1203                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1204                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1205                         _exit(1);
1206                 }
1207
1208                 /*
1209                  * repack the db
1210                  */
1211                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1212                     (vacuum_handle->fast_path_count == 0))
1213                 {
1214                         full_vacuum_run = true;
1215                 }
1216                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1217                                                full_vacuum_run);
1218
1219                 write(child_ctx->fd[1], &cc, 1);
1220                 _exit(0);
1221         }
1222
1223         set_close_on_exec(child_ctx->fd[0]);
1224         close(child_ctx->fd[1]);
1225
1226         child_ctx->status = VACUUM_RUNNING;
1227         child_ctx->start_time = timeval_current();
1228
1229         DLIST_ADD(ctdb->vacuumers, child_ctx);
1230         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1231
1232         /*
1233          * Clear the fastpath vacuuming list in the parent.
1234          */
1235         talloc_free(ctdb_db->delete_queue);
1236         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1237         if (ctdb_db->delete_queue == NULL) {
1238                 /* fatal here? ... */
1239                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1240                                  "in parent context. Shutting down\n");
1241         }
1242
1243         event_add_timed(ctdb->ev, child_ctx,
1244                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1245                 vacuum_child_timeout, child_ctx);
1246
1247         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1248
1249         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1250                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1251         tevent_fd_set_auto_close(fde);
1252
1253         vacuum_handle->child_ctx = child_ctx;
1254         child_ctx->vacuum_handle = vacuum_handle;
1255 }
1256
1257 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1258 {
1259         /* Simply free them all. */
1260         while (ctdb->vacuumers) {
1261                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1262                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1263                            (int)ctdb->vacuumers->child_pid));
1264                 /* vacuum_child_destructor kills it, removes from list */
1265                 talloc_free(ctdb->vacuumers);
1266         }
1267 }
1268
1269 /* this function initializes the vacuuming context for a database
1270  * starts the vacuuming events
1271  */
1272 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1273 {
1274         if (ctdb_db->persistent != 0) {
1275                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1276                 return 0;
1277         }
1278
1279         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1280         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1281
1282         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1283         ctdb_db->vacuum_handle->fast_path_count = 0;
1284
1285         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1286                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1287                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1288
1289         return 0;
1290 }
1291
1292 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1293                                             const struct ctdb_ltdb_header *hdr,
1294                                             const TDB_DATA key)
1295 {
1296         struct delete_record_data *kd;
1297         uint32_t hash;
1298
1299         hash = (uint32_t)ctdb_hash(&key);
1300
1301         DEBUG(DEBUG_DEBUG, (__location__
1302                             " remove_record_from_delete_queue: db[%s] "
1303                             "db_id[0x%08x] "
1304                             "key_hash[0x%08x] "
1305                             "lmaster[%u] "
1306                             "migrated_with_data[%s]\n",
1307                              ctdb_db->db_name, ctdb_db->db_id,
1308                              hash,
1309                              ctdb_lmaster(ctdb_db->ctdb, &key),
1310                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1311
1312         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1313         if (kd == NULL) {
1314                 return;
1315         }
1316         if (kd->key.dsize != key.dsize) {
1317                 return;
1318         }
1319         if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1320                 return;
1321         }
1322
1323         talloc_free(kd);
1324
1325         return;
1326 }
1327
1328 /**
1329  * Insert a record into the ctdb_db context's delete queue,
1330  * handling hash collisions.
1331  */
1332 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1333                                            const struct ctdb_ltdb_header *hdr,
1334                                            TDB_DATA key)
1335 {
1336         struct delete_record_data *kd;
1337         uint32_t hash;
1338         int ret;
1339
1340         hash = (uint32_t)ctdb_hash(&key);
1341
1342         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1343                            "db_id[0x%08x] "
1344                            "key_hash[0x%08x] "
1345                            "lmaster[%u] "
1346                            "migrated_with_data[%s]\n",
1347                             ctdb_db->db_name, ctdb_db->db_id,
1348                             hash,
1349                             ctdb_lmaster(ctdb_db->ctdb, &key),
1350                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1351
1352         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1353         if (kd != NULL) {
1354                 if ((kd->key.dsize != key.dsize) ||
1355                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1356                 {
1357                         DEBUG(DEBUG_INFO,
1358                               ("schedule for deletion: Hash collision (0x%08x)."
1359                                " Skipping the record.\n", hash));
1360                         return 0;
1361                 } else {
1362                         DEBUG(DEBUG_DEBUG,
1363                               ("schedule for deletion: Overwriting entry for "
1364                                "key with hash 0x%08x.\n", hash));
1365                 }
1366         }
1367
1368         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1369                                                   ctdb_db->delete_queue,
1370                                                   hdr, key);
1371         if (ret != 0) {
1372                 return -1;
1373         }
1374
1375         return 0;
1376 }
1377
1378 /**
1379  * Schedule a record for deletetion.
1380  * Called from the parent context.
1381  */
1382 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1383                                            TDB_DATA indata)
1384 {
1385         struct ctdb_control_schedule_for_deletion *dd;
1386         struct ctdb_db_context *ctdb_db;
1387         int ret;
1388         TDB_DATA key;
1389
1390         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1391
1392         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1393         if (ctdb_db == NULL) {
1394                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1395                                   dd->db_id));
1396                 return -1;
1397         }
1398
1399         key.dsize = dd->keylen;
1400         key.dptr = dd->key;
1401
1402         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1403
1404         return ret;
1405 }
1406
1407 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1408                                          const struct ctdb_ltdb_header *hdr,
1409                                          TDB_DATA key)
1410 {
1411         int ret;
1412         struct ctdb_control_schedule_for_deletion *dd;
1413         TDB_DATA indata;
1414         int32_t status;
1415
1416         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1417                 /* main daemon - directly queue */
1418                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1419
1420                 return ret;
1421         }
1422
1423         /* child process: send the main daemon a control */
1424
1425         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1426         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1427         if (indata.dptr == NULL) {
1428                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1429                 return -1;
1430         }
1431         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1432         dd->db_id = ctdb_db->db_id;
1433         dd->hdr = *hdr;
1434         dd->keylen = key.dsize;
1435         memcpy(dd->key, key.dptr, key.dsize);
1436
1437         ret = ctdb_control(ctdb_db->ctdb,
1438                            CTDB_CURRENT_NODE,
1439                            ctdb_db->db_id,
1440                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1441                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1442                            indata,
1443                            NULL, /* mem_ctx */
1444                            NULL, /* outdata */
1445                            &status,
1446                            NULL, /* timeout : NULL == wait forever */
1447                            NULL); /* error message */
1448
1449         talloc_free(indata.dptr);
1450
1451         if (ret != 0 || status != 0) {
1452                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1453                                   "SCHEDULE_FOR_DELETION "
1454                                   "control.\n"));
1455                 if (status != 0) {
1456                         ret = -1;
1457                 }
1458         }
1459
1460         return ret;
1461 }
1462
1463 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1464                                          const struct ctdb_ltdb_header *hdr,
1465                                          const TDB_DATA key)
1466 {
1467         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1468                 /*
1469                  * Only remove the record from the delete queue if called
1470                  * in the main daemon.
1471                  */
1472                 return;
1473         }
1474
1475         remove_record_from_delete_queue(ctdb_db, hdr, key);
1476
1477         return;
1478 }