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  * Full vacuum run:
 573  * read-only traverse of the database, looking for records that
 574  * might be able to be vacuumed.
 575  *
 576  * This is not done each time but only every tunable
 577  * VacuumFastPathCount times.
 578  */
 579 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
 580                                struct vacuum_data *vdata,
 581                                bool full_vacuum_run)
 582 {
 583         int ret;
 584
 585         if (!full_vacuum_run) {
 586                 return 0;
 587         }
 588
 589         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 590         if (ret == -1 || vdata->traverse_error) {
 591                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 592                                   "'%s'\n", ctdb_db->db_name));
 593                 return -1;
 594         }
 595
 596         if (vdata->full_total > 0) {
 597                 DEBUG(DEBUG_INFO,
 598                       (__location__
 599                        " full vacuuming db traverse statistics: "
 600                        "db[%s] "
 601                        "total[%u] "
 602                        "skp[%u] "
 603                        "err[%u] "
 604                        "adl[%u] "
 605                        "avf[%u]\n",
 606                        ctdb_db->db_name,
 607                        (unsigned)vdata->full_total,
 608                        (unsigned)vdata->full_skipped,
 609                        (unsigned)vdata->full_error,
 610                        (unsigned)vdata->full_added_to_delete_list,
 611                        (unsigned)vdata->full_added_to_vacuum_fetch_list));
 612         }
 613
 614         return 0;
 615 }
 616
 617 /**
 618  * Vacuum a DB:
 619  *  - Always do the fast vacuuming run, which traverses
 620  *    the in-memory delete queue: these records have been
 621  *    scheduled for deletion.
 622  *  - Only if explicitly requested, the database is traversed
 623  *    in order to use the traditional heuristics on empty records
 624  *    to trigger deletion.
 625  *    This is done only every VacuumFastPathCount'th vacuuming run.
 626  *
 627  * The traverse runs fill two lists:
 628  *
 629  * - The delete_list:
 630  *   This is the list of empty records the current
 631  *   node is lmaster and dmaster for. These records are later
 632  *   deleted first on other nodes and then locally.
 633  *
 634  *   The fast vacuuming run has a short cut for those records
 635  *   that have never been migrated with data: these records
 636  *   are immediately deleted locally, since they have left
 637  *   no trace on other nodes.
 638  *
 639  * - The vacuum_fetch lists
 640  *   (one for each other lmaster node):
 641  *   The records in this list are sent for deletion to
 642  *   their lmaster in a bulk VACUUM_FETCH message.
 643  *
 644  *   The lmaster then migrates all these records to itelf
 645  *   so that they can be vacuumed there.
 646  *
 647  * This executes in the child context.
 648  */
 649 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 650                           struct vacuum_data *vdata,
 651                           bool full_vacuum_run)
 652 {
 653         struct ctdb_context *ctdb = ctdb_db->ctdb;
 654         const char *name = ctdb_db->db_name;
 655         int ret, i, pnn;
 656
 657         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 658                            "%s db_id[0x%08x]\n",
 659                            full_vacuum_run ? "full" : "fast",
 660                            ctdb_db->db_name, ctdb_db->db_id));
 661
 662         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 663         if (ret != 0) {
 664                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 665                 return ret;
 666         }
 667
 668         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 669         if (pnn == -1) {
 670                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 671                 return -1;
 672         }
 673
 674         ctdb->pnn = pnn;
 675
 676         vdata->fast_added_to_delete_list = 0;
 677         vdata->fast_added_to_vacuum_fetch_list = 0;
 678         vdata->fast_deleted = 0;
 679         vdata->fast_skipped = 0;
 680         vdata->fast_error = 0;
 681         vdata->fast_total = 0;
 682         vdata->full_added_to_delete_list = 0;
 683         vdata->full_added_to_vacuum_fetch_list = 0;
 684         vdata->full_skipped = 0;
 685         vdata->full_error = 0;
 686         vdata->full_total = 0;
 687         vdata->delete_count = 0;
 688         vdata->delete_left = 0;
 689         vdata->delete_remote_error = 0;
 690         vdata->delete_local_error = 0;
 691         vdata->delete_skipped = 0;
 692         vdata->delete_deleted = 0;
 693
 694         /* the list needs to be of length num_nodes */
 695         vdata->vacuum_fetch_list = talloc_array(vdata,
 696                                                 struct ctdb_marshall_buffer *,
 697                                                 ctdb->num_nodes);
 698         if (vdata->vacuum_fetch_list == NULL) {
 699                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 700                 return -1;
 701         }
 702         for (i = 0; i < ctdb->num_nodes; i++) {
 703                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 704                         talloc_zero_size(vdata->vacuum_fetch_list,
 705                                          offsetof(struct ctdb_marshall_buffer, data));
 706                 if (vdata->vacuum_fetch_list[i] == NULL) {
 707                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 708                         return -1;
 709                 }
 710                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 711         }
 712
 713         ctdb_vacuum_db_fast(ctdb_db, vdata);
 714
 715         ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
 716         if (ret != 0) {
 717                 return ret;
 718         }
 719
 720         /*
 721          * For records where we are not the lmaster,
 722          * tell the lmaster to fetch the record.
 723          */
 724         for (i = 0; i < ctdb->num_nodes; i++) {
 725                 TDB_DATA data;
 726                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 727
 728                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 729                         continue;
 730                 }
 731
 732                 if (vfl->count == 0) {
 733                         continue;
 734                 }
 735
 736                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 737                                    vfl->count, ctdb->nodes[i]->pnn,
 738                                    name));
 739
 740                 data.dsize = talloc_get_size(vfl);
 741                 data.dptr  = (void *)vfl;
 742                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 743                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 744                                           "fetch message to %u\n",
 745                                           ctdb->nodes[i]->pnn));
 746                         return -1;
 747                 }
 748         }
 749
 750         /* Process all records we can delete (if any) */
 751         if (vdata->delete_count > 0) {
 752                 struct delete_records_list *recs;
 753                 TDB_DATA indata, outdata;
 754                 int32_t res;
 755                 struct ctdb_node_map *nodemap;
 756                 uint32_t *active_nodes;
 757                 int num_active_nodes;
 758
 759                 vdata->delete_left = vdata->delete_count;
 760
 761                 recs = talloc_zero(vdata, struct delete_records_list);
 762                 if (recs == NULL) {
 763                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 764                         return -1;
 765                 }
 766                 recs->records = (struct ctdb_marshall_buffer *)
 767                         talloc_zero_size(vdata,
 768                                     offsetof(struct ctdb_marshall_buffer, data));
 769                 if (recs->records == NULL) {
 770                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 771                         return -1;
 772                 }
 773                 recs->records->db_id = ctdb_db->db_id;
 774
 775                 /*
 776                  * traverse the tree of all records we want to delete and
 777                  * create a blob we can send to the other nodes.
 778                  */
 779                 trbt_traversearray32(vdata->delete_list, 1,
 780                                      delete_marshall_traverse, recs);
 781
 782                 indata.dsize = talloc_get_size(recs->records);
 783                 indata.dptr  = (void *)recs->records;
 784
 785                 /*
 786                  * now tell all the active nodes to delete all these records
 787                  * (if possible)
 788                  */
 789
 790                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 791                                            CTDB_CURRENT_NODE,
 792                                            recs, /* talloc context */
 793                                            &nodemap);
 794                 if (ret != 0) {
 795                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 796                         return -1;
 797                 }
 798
 799                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 800                                                     nodemap, /* talloc context */
 801                                                     false /* include self */);
 802                 /* yuck! ;-) */
 803                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 804
 805                 for (i = 0; i < num_active_nodes; i++) {
 806                         struct ctdb_marshall_buffer *records;
 807                         struct ctdb_rec_data *rec;
 808
 809                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 810                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 811                                         indata, recs, &outdata, &res,
 812                                         NULL, NULL);
 813                         if (ret != 0 || res != 0) {
 814                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 815                                                   "node %u: ret[%d] res[%d]\n",
 816                                                   active_nodes[i], ret, res));
 817                                 return -1;
 818                         }
 819
 820                         /*
 821                          * outdata contains the list of records coming back
 822                          * from the node: These are the records that the
 823                          * remote node could not delete.
 824                          */
 825                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 826                         rec = (struct ctdb_rec_data *)&records->data[0];
 827                         while (records->count-- > 1) {
 828                                 TDB_DATA reckey, recdata;
 829                                 struct ctdb_ltdb_header *rechdr;
 830                                 struct delete_record_data *dd;
 831
 832                                 reckey.dptr = &rec->data[0];
 833                                 reckey.dsize = rec->keylen;
 834                                 recdata.dptr = &rec->data[reckey.dsize];
 835                                 recdata.dsize = rec->datalen;
 836
 837                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 838                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 839                                         return -1;
 840                                 }
 841                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 842                                 recdata.dptr += sizeof(*rechdr);
 843                                 recdata.dsize -= sizeof(*rechdr);
 844
 845                                 dd = (struct delete_record_data *)trbt_lookup32(
 846                                                 vdata->delete_list,
 847                                                 ctdb_hash(&reckey));
 848                                 if (dd != NULL) {
 849                                         /*
 850                                          * The other node could not delete the
 851                                          * record and it is the first node that
 852                                          * failed. So we should remove it from
 853                                          * the tree and update statistics.
 854                                          */
 855                                         talloc_free(dd);
 856                                         vdata->delete_remote_error++;
 857                                         vdata->delete_left--;
 858                                 }
 859
 860                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 861                         }
 862                 }
 863
 864                 /* free nodemap and active_nodes */
 865                 talloc_free(nodemap);
 866         }
 867
 868         if (vdata->delete_left > 0) {
 869                 /*
 870                  * The only records remaining in the tree are those
 871                  * records which all other nodes could successfully
 872                  * delete, so we can safely delete them on the
 873                  * lmaster as well.
 874                  */
 875                 trbt_traversearray32(vdata->delete_list, 1,
 876                                      delete_record_traverse, vdata);
 877         }
 878
 879         /* this ensures we run our event queue */
 880         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 881
 882         return 0;
 883 }
 884
 885
 886 /*
 887  * traverse function for repacking
 888  */
 889 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 890 {
 891         struct vacuum_data *vdata = (struct vacuum_data *)private;
 892
 893         if (vdata->vacuum) {
 894                 uint32_t hash = ctdb_hash(&key);
 895                 struct delete_record_data *kd;
 896                 /*
 897                  * check if we can ignore this record because it's in the delete_list
 898                  */
 899                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
 900                 /*
 901                  * there might be hash collisions so we have to compare the keys here to be sure
 902                  */
 903                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 904                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 905                         /*
 906                          * we have to check if the record hasn't changed in the meantime in order to
 907                          * savely remove it from the database
 908                          */
 909                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 910                                 hdr->dmaster == kd->ctdb->pnn &&
 911                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 912                                 kd->hdr.rsn == hdr->rsn) {
 913                                 vdata->vacuumed++;
 914                                 return 0;
 915                         }
 916                 }
 917         }
 918         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 919                 vdata->traverse_error = true;
 920                 return -1;
 921         }
 922         vdata->copied++;
 923         return 0;
 924 }
 925
 926 /*
 927  * repack a tdb
 928  */
 929 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 930 {
 931         struct tdb_context *tmp_db;
 932
 933         if (tdb_transaction_start(tdb) != 0) {
 934                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 935                 return -1;
 936         }
 937
 938         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 939                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 940                           O_RDWR|O_CREAT, 0);
 941         if (tmp_db == NULL) {
 942                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 943                 tdb_transaction_cancel(tdb);
 944                 return -1;
 945         }
 946
 947         vdata->traverse_error = false;
 948         vdata->dest_db = tmp_db;
 949         vdata->vacuum = true;
 950         vdata->vacuumed = 0;
 951         vdata->copied = 0;
 952
 953         /*
 954          * repack and vacuum on-the-fly by not writing the records that are
 955          * no longer needed
 956          */
 957         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 958                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 959                 tdb_transaction_cancel(tdb);
 960                 tdb_close(tmp_db);
 961                 return -1;
 962         }
 963
 964         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 965
 966         if (vdata->traverse_error) {
 967                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 968                 tdb_transaction_cancel(tdb);
 969                 tdb_close(tmp_db);
 970                 return -1;
 971         }
 972
 973         if (tdb_wipe_all(tdb) != 0) {
 974                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 975                 tdb_transaction_cancel(tdb);
 976                 tdb_close(tmp_db);
 977                 return -1;
 978         }
 979
 980         vdata->traverse_error = false;
 981         vdata->dest_db = tdb;
 982         vdata->vacuum = false;
 983         vdata->copied = 0;
 984
 985         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 986                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 987                 tdb_transaction_cancel(tdb);
 988                 tdb_close(tmp_db);
 989                 return -1;
 990         }
 991
 992         if (vdata->traverse_error) {
 993                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 994                 tdb_transaction_cancel(tdb);
 995                 tdb_close(tmp_db);
 996                 return -1;
 997         }
 998
 999         tdb_close(tmp_db);
1000
1001
1002         if (tdb_transaction_commit(tdb) != 0) {
1003                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1004                 return -1;
1005         }
1006         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1007
1008         return 0;
1009 }
1010
1011 /*
1012  * repack and vaccum a db
1013  * called from the child context
1014  */
1015 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1016                                      TALLOC_CTX *mem_ctx,
1017                                      bool full_vacuum_run)
1018 {
1019         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1020         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1021         const char *name = ctdb_db->db_name;
1022         int freelist_size;
1023         struct vacuum_data *vdata;
1024
1025         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1026         if (freelist_size == -1) {
1027                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1028                 return -1;
1029         }
1030
1031         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1032         if (vdata == NULL) {
1033                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1034                 return -1;
1035         }
1036
1037         vdata->ctdb = ctdb_db->ctdb;
1038         vdata->vacuum_limit = vacuum_limit;
1039         vdata->repack_limit = repack_limit;
1040         vdata->delete_list = trbt_create(vdata, 0);
1041         vdata->ctdb_db = ctdb_db;
1042         if (vdata->delete_list == NULL) {
1043                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1044                 talloc_free(vdata);
1045                 return -1;
1046         }
1047
1048         vdata->start = timeval_current();
1049
1050         /*
1051          * gather all records that can be deleted in vdata
1052          */
1053         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1054                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1055         }
1056
1057         /*
1058          * decide if a repack is necessary
1059          */
1060         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1061         {
1062                 talloc_free(vdata);
1063                 return 0;
1064         }
1065
1066         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1067                         name, freelist_size, vdata->delete_left));
1068
1069         /*
1070          * repack and implicitely get rid of the records we can delete
1071          */
1072         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1073                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1074                 talloc_free(vdata);
1075                 return -1;
1076         }
1077         talloc_free(vdata);
1078
1079         return 0;
1080 }
1081
1082 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1083 {
1084         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1085
1086         return interval;
1087 }
1088
1089 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1090 {
1091         double l = timeval_elapsed(&child_ctx->start_time);
1092         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1093         struct ctdb_context *ctdb = ctdb_db->ctdb;
1094
1095         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1096
1097         if (child_ctx->child_pid != -1) {
1098                 kill(child_ctx->child_pid, SIGKILL);
1099         } else {
1100                 /* Bump the number of successful fast-path runs. */
1101                 child_ctx->vacuum_handle->fast_path_count++;
1102         }
1103
1104         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1105
1106         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1107                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1108                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1109
1110         return 0;
1111 }
1112
1113 /*
1114  * this event is generated when a vacuum child process times out
1115  */
1116 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1117                                          struct timeval t, void *private_data)
1118 {
1119         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1120
1121         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1122
1123         child_ctx->status = VACUUM_TIMEOUT;
1124
1125         talloc_free(child_ctx);
1126 }
1127
1128
1129 /*
1130  * this event is generated when a vacuum child process has completed
1131  */
1132 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1133                              uint16_t flags, void *private_data)
1134 {
1135         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1136         char c = 0;
1137         int ret;
1138
1139         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1140         child_ctx->child_pid = -1;
1141
1142         ret = read(child_ctx->fd[0], &c, 1);
1143         if (ret != 1 || c != 0) {
1144                 child_ctx->status = VACUUM_ERROR;
1145                 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));
1146         } else {
1147                 child_ctx->status = VACUUM_OK;
1148         }
1149
1150         talloc_free(child_ctx);
1151 }
1152
1153 /*
1154  * this event is called every time we need to start a new vacuum process
1155  */
1156 static void
1157 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1158                                struct timeval t, void *private_data)
1159 {
1160         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1161         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1162         struct ctdb_context *ctdb = ctdb_db->ctdb;
1163         struct ctdb_vacuum_child_context *child_ctx;
1164         struct tevent_fd *fde;
1165         int ret;
1166
1167         /* we dont vacuum if we are in recovery mode, or db frozen */
1168         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1169             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1170                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1171                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1172                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1173                                    ? "freeze pending"
1174                                    : "frozen"));
1175                 event_add_timed(ctdb->ev, vacuum_handle,
1176                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1177                         ctdb_vacuum_event, vacuum_handle);
1178                 return;
1179         }
1180
1181         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1182         if (child_ctx == NULL) {
1183                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1184                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1185         }
1186
1187
1188         ret = pipe(child_ctx->fd);
1189         if (ret != 0) {
1190                 talloc_free(child_ctx);
1191                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1192                 event_add_timed(ctdb->ev, vacuum_handle,
1193                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1194                         ctdb_vacuum_event, vacuum_handle);
1195                 return;
1196         }
1197
1198         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1199                 vacuum_handle->fast_path_count = 0;
1200         }
1201
1202         child_ctx->child_pid = ctdb_fork(ctdb);
1203         if (child_ctx->child_pid == (pid_t)-1) {
1204                 close(child_ctx->fd[0]);
1205                 close(child_ctx->fd[1]);
1206                 talloc_free(child_ctx);
1207                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1208                 event_add_timed(ctdb->ev, vacuum_handle,
1209                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1210                         ctdb_vacuum_event, vacuum_handle);
1211                 return;
1212         }
1213
1214
1215         if (child_ctx->child_pid == 0) {
1216                 char cc = 0;
1217                 bool full_vacuum_run = false;
1218                 close(child_ctx->fd[0]);
1219
1220                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1221
1222                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1223                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1224                         _exit(1);
1225                 }
1226
1227                 /*
1228                  * repack the db
1229                  */
1230                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1231                     (vacuum_handle->fast_path_count == 0))
1232                 {
1233                         full_vacuum_run = true;
1234                 }
1235                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1236                                                full_vacuum_run);
1237
1238                 write(child_ctx->fd[1], &cc, 1);
1239                 _exit(0);
1240         }
1241
1242         set_close_on_exec(child_ctx->fd[0]);
1243         close(child_ctx->fd[1]);
1244
1245         child_ctx->status = VACUUM_RUNNING;
1246         child_ctx->start_time = timeval_current();
1247
1248         DLIST_ADD(ctdb->vacuumers, child_ctx);
1249         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1250
1251         /*
1252          * Clear the fastpath vacuuming list in the parent.
1253          */
1254         talloc_free(ctdb_db->delete_queue);
1255         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1256         if (ctdb_db->delete_queue == NULL) {
1257                 /* fatal here? ... */
1258                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1259                                  "in parent context. Shutting down\n");
1260         }
1261
1262         event_add_timed(ctdb->ev, child_ctx,
1263                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1264                 vacuum_child_timeout, child_ctx);
1265
1266         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1267
1268         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1269                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1270         tevent_fd_set_auto_close(fde);
1271
1272         vacuum_handle->child_ctx = child_ctx;
1273         child_ctx->vacuum_handle = vacuum_handle;
1274 }
1275
1276 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1277 {
1278         /* Simply free them all. */
1279         while (ctdb->vacuumers) {
1280                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1281                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1282                            (int)ctdb->vacuumers->child_pid));
1283                 /* vacuum_child_destructor kills it, removes from list */
1284                 talloc_free(ctdb->vacuumers);
1285         }
1286 }
1287
1288 /* this function initializes the vacuuming context for a database
1289  * starts the vacuuming events
1290  */
1291 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1292 {
1293         if (ctdb_db->persistent != 0) {
1294                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1295                 return 0;
1296         }
1297
1298         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1299         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1300
1301         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1302         ctdb_db->vacuum_handle->fast_path_count = 0;
1303
1304         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1305                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1306                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1307
1308         return 0;
1309 }
1310
1311 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1312                                             const struct ctdb_ltdb_header *hdr,
1313                                             const TDB_DATA key)
1314 {
1315         struct delete_record_data *kd;
1316         uint32_t hash;
1317
1318         hash = (uint32_t)ctdb_hash(&key);
1319
1320         DEBUG(DEBUG_DEBUG, (__location__
1321                             " remove_record_from_delete_queue: db[%s] "
1322                             "db_id[0x%08x] "
1323                             "key_hash[0x%08x] "
1324                             "lmaster[%u] "
1325                             "migrated_with_data[%s]\n",
1326                              ctdb_db->db_name, ctdb_db->db_id,
1327                              hash,
1328                              ctdb_lmaster(ctdb_db->ctdb, &key),
1329                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1330
1331         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1332         if (kd == NULL) {
1333                 return;
1334         }
1335         if (kd->key.dsize != key.dsize) {
1336                 return;
1337         }
1338         if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1339                 return;
1340         }
1341
1342         talloc_free(kd);
1343
1344         return;
1345 }
1346
1347 /**
1348  * Insert a record into the ctdb_db context's delete queue,
1349  * handling hash collisions.
1350  */
1351 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1352                                            const struct ctdb_ltdb_header *hdr,
1353                                            TDB_DATA key)
1354 {
1355         struct delete_record_data *kd;
1356         uint32_t hash;
1357         int ret;
1358
1359         hash = (uint32_t)ctdb_hash(&key);
1360
1361         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1362                            "db_id[0x%08x] "
1363                            "key_hash[0x%08x] "
1364                            "lmaster[%u] "
1365                            "migrated_with_data[%s]\n",
1366                             ctdb_db->db_name, ctdb_db->db_id,
1367                             hash,
1368                             ctdb_lmaster(ctdb_db->ctdb, &key),
1369                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1370
1371         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1372         if (kd != NULL) {
1373                 if ((kd->key.dsize != key.dsize) ||
1374                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1375                 {
1376                         DEBUG(DEBUG_INFO,
1377                               ("schedule for deletion: Hash collision (0x%08x)."
1378                                " Skipping the record.\n", hash));
1379                         return 0;
1380                 } else {
1381                         DEBUG(DEBUG_DEBUG,
1382                               ("schedule for deletion: Overwriting entry for "
1383                                "key with hash 0x%08x.\n", hash));
1384                 }
1385         }
1386
1387         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1388                                                   ctdb_db->delete_queue,
1389                                                   hdr, key);
1390         if (ret != 0) {
1391                 return -1;
1392         }
1393
1394         return 0;
1395 }
1396
1397 /**
1398  * Schedule a record for deletetion.
1399  * Called from the parent context.
1400  */
1401 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1402                                            TDB_DATA indata)
1403 {
1404         struct ctdb_control_schedule_for_deletion *dd;
1405         struct ctdb_db_context *ctdb_db;
1406         int ret;
1407         TDB_DATA key;
1408
1409         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1410
1411         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1412         if (ctdb_db == NULL) {
1413                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1414                                   dd->db_id));
1415                 return -1;
1416         }
1417
1418         key.dsize = dd->keylen;
1419         key.dptr = dd->key;
1420
1421         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1422
1423         return ret;
1424 }
1425
1426 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1427                                          const struct ctdb_ltdb_header *hdr,
1428                                          TDB_DATA key)
1429 {
1430         int ret;
1431         struct ctdb_control_schedule_for_deletion *dd;
1432         TDB_DATA indata;
1433         int32_t status;
1434
1435         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1436                 /* main daemon - directly queue */
1437                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1438
1439                 return ret;
1440         }
1441
1442         /* child process: send the main daemon a control */
1443
1444         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1445         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1446         if (indata.dptr == NULL) {
1447                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1448                 return -1;
1449         }
1450         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1451         dd->db_id = ctdb_db->db_id;
1452         dd->hdr = *hdr;
1453         dd->keylen = key.dsize;
1454         memcpy(dd->key, key.dptr, key.dsize);
1455
1456         ret = ctdb_control(ctdb_db->ctdb,
1457                            CTDB_CURRENT_NODE,
1458                            ctdb_db->db_id,
1459                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1460                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1461                            indata,
1462                            NULL, /* mem_ctx */
1463                            NULL, /* outdata */
1464                            &status,
1465                            NULL, /* timeout : NULL == wait forever */
1466                            NULL); /* error message */
1467
1468         talloc_free(indata.dptr);
1469
1470         if (ret != 0 || status != 0) {
1471                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1472                                   "SCHEDULE_FOR_DELETION "
1473                                   "control.\n"));
1474                 if (status != 0) {
1475                         ret = -1;
1476                 }
1477         }
1478
1479         return ret;
1480 }
1481
1482 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1483                                          const struct ctdb_ltdb_header *hdr,
1484                                          const TDB_DATA key)
1485 {
1486         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1487                 /*
1488                  * Only remove the record from the delete queue if called
1489                  * in the main daemon.
1490                  */
1491                 return;
1492         }
1493
1494         remove_record_from_delete_queue(ctdb_db, hdr, key);
1495
1496         return;
1497 }