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         uint32_t hash = ctdb_hash(&(dd->key));
 337
 338         vdata->fast_total++;
 339
 340         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 341         if (res != 0) {
 342                 DEBUG(DEBUG_ERR,
 343                       (__location__ " Error getting chainlock on record with "
 344                        "key hash [0x%08x] on database db[%s].\n",
 345                        hash, ctdb_db->db_name));
 346                 vdata->fast_error++;
 347                 return 0;
 348         }
 349
 350         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 351         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 352                 /* Does not exist or not a ctdb record. Skip. */
 353                 goto skipped;
 354         }
 355
 356         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 357                 /* The record has been recycled (filled with data). Skip. */
 358                 goto skipped;
 359         }
 360
 361         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 362
 363         if (header->dmaster != ctdb->pnn) {
 364                 /* The record has been migrated off the node. Skip. */
 365                 goto skipped;
 366         }
 367
 368         if (header->rsn != dd->hdr.rsn) {
 369                 /*
 370                  * The record has been migrated off the node and back again.
 371                  * But not requeued for deletion. Skip it.
 372                  */
 373                 goto skipped;
 374         }
 375
 376         /*
 377          * We are dmaster, and the record has no data, and it has
 378          * not been migrated after it has been queued for deletion.
 379          *
 380          * At this stage, the record could still have been revived locally
 381          * and last been written with empty data. This can only be
 382          * fixed with the addition of an active or delete flag. (TODO)
 383          */
 384
 385         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 386
 387         if (lmaster != ctdb->pnn) {
 388                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 389
 390                 if (res != 0) {
 391                         DEBUG(DEBUG_ERR,
 392                               (__location__ " Error adding record to list "
 393                                "of records to send to lmaster.\n"));
 394                         vdata->fast_error++;
 395                 } else {
 396                         vdata->fast_added_to_vacuum_fetch_list++;
 397                 }
 398                 goto done;
 399         }
 400
 401         /* use header->flags or dd->hdr.flags ?? */
 402         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 403                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 404
 405                 if (res != 0) {
 406                         DEBUG(DEBUG_ERR,
 407                               (__location__ " Error adding record to list "
 408                                "of records for deletion on lmaster.\n"));
 409                         vdata->fast_error++;
 410                 } else {
 411                         vdata->fast_added_to_delete_list++;
 412                 }
 413         } else {
 414                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 415
 416                 if (res != 0) {
 417                         DEBUG(DEBUG_ERR,
 418                               (__location__ " Error deleting record with key "
 419                                "hash [0x%08x] from local data base db[%s].\n",
 420                                hash, ctdb_db->db_name));
 421                         vdata->fast_error++;
 422                 } else {
 423                         DEBUG(DEBUG_DEBUG,
 424                               (__location__ " Deleted record with key hash "
 425                                "[0x%08x] from local data base db[%s].\n",
 426                                hash, ctdb_db->db_name));
 427                         vdata->fast_deleted++;
 428                 }
 429         }
 430
 431         goto done;
 432
 433 skipped:
 434         vdata->fast_skipped++;
 435
 436 done:
 437         if (tdb_data.dptr != NULL) {
 438                 free(tdb_data.dptr);
 439         }
 440         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 441
 442         return 0;
 443 }
 444
 445 /**
 446  * Delete the records that we are lmaster and dmaster for and
 447  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 448  * control.
 449  */
 450 static int delete_record_traverse(void *param, void *data)
 451 {
 452         struct delete_record_data *dd =
 453                 talloc_get_type(data, struct delete_record_data);
 454         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 455         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 456         struct ctdb_context *ctdb = ctdb_db->ctdb;
 457         int res;
 458         struct ctdb_ltdb_header *header;
 459         TDB_DATA tdb_data;
 460         uint32_t lmaster;
 461         bool deleted = false;
 462         uint32_t hash = ctdb_hash(&(dd->key));
 463
 464         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 465         if (res != 0) {
 466                 DEBUG(DEBUG_ERR,
 467                       (__location__ " Error getting chainlock on record with "
 468                        "key hash [0x%08x] on database db[%s].\n",
 469                        hash, ctdb_db->db_name));
 470                 vdata->delete_local_error++;
 471                 return 0;
 472         }
 473
 474         /*
 475          * Verify that the record is still empty, its RSN has not
 476          * changed and that we are still its lmaster and dmaster.
 477          */
 478
 479         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 480         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 481                 /* Does not exist or not a ctdb record. Skip. */
 482                 vdata->delete_skipped++;
 483                 goto done;
 484         }
 485
 486         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 487                 /* The record has been recycled (filled with data). Skip. */
 488                 vdata->delete_skipped++;
 489                 goto done;
 490         }
 491
 492         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 493
 494         if (header->dmaster != ctdb->pnn) {
 495                 /* The record has been migrated off the node. Skip. */
 496                 vdata->delete_skipped++;
 497                 goto done;
 498         }
 499
 500
 501         if (header->rsn != dd->hdr.rsn) {
 502                 /*
 503                  * The record has been migrated off the node and back again.
 504                  * But not requeued for deletion. Skip it.
 505                  */
 506                 vdata->delete_skipped++;
 507                 goto done;
 508         }
 509
 510         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 511
 512         if (lmaster != ctdb->pnn) {
 513                 /* we are not lmaster - strange */
 514                 vdata->delete_skipped++;
 515                 goto done;
 516         }
 517
 518         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 519
 520         if (res != 0) {
 521                 DEBUG(DEBUG_ERR,
 522                       (__location__ " Error deleting record with key hash "
 523                        "[0x%08x] from local data base db[%s].\n",
 524                        hash, ctdb_db->db_name));
 525                 vdata->delete_local_error++;
 526                 goto done;
 527         }
 528
 529         deleted = true;
 530
 531         DEBUG(DEBUG_DEBUG,
 532               (__location__ " Deleted record with key hash [0x%08x] from "
 533                "local data base db[%s].\n", hash, ctdb_db->db_name));
 534
 535 done:
 536         if (tdb_data.dptr != NULL) {
 537                 free(tdb_data.dptr);
 538         }
 539
 540         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 541
 542         if (deleted) {
 543                 /*
 544                  * successfully deleted the record locally.
 545                  * remove it from the list and update statistics.
 546                  */
 547                 talloc_free(dd);
 548                 vdata->delete_deleted++;
 549                 vdata->delete_left--;
 550         }
 551
 552         return 0;
 553 }
 554
 555 /**
 556  * Fast vacuuming run:
 557  * Traverse the delete_queue.
 558  * This fills the same lists as the database traverse.
 559  */
 560 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
 561                                 struct vacuum_data *vdata)
 562 {
 563         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 564
 565         if (vdata->fast_total > 0) {
 566                 DEBUG(DEBUG_INFO,
 567                       (__location__
 568                        " fast vacuuming delete_queue traverse statistics: "
 569                        "db[%s] "
 570                        "total[%u] "
 571                        "del[%u] "
 572                        "skp[%u] "
 573                        "err[%u] "
 574                        "adl[%u] "
 575                        "avf[%u]\n",
 576                        ctdb_db->db_name,
 577                        (unsigned)vdata->fast_total,
 578                        (unsigned)vdata->fast_deleted,
 579                        (unsigned)vdata->fast_skipped,
 580                        (unsigned)vdata->fast_error,
 581                        (unsigned)vdata->fast_added_to_delete_list,
 582                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 583         }
 584
 585         return;
 586 }
 587
 588 /**
 589  * Full vacuum run:
 590  * read-only traverse of the database, looking for records that
 591  * might be able to be vacuumed.
 592  *
 593  * This is not done each time but only every tunable
 594  * VacuumFastPathCount times.
 595  */
 596 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
 597                                struct vacuum_data *vdata,
 598                                bool full_vacuum_run)
 599 {
 600         int ret;
 601
 602         if (!full_vacuum_run) {
 603                 return 0;
 604         }
 605
 606         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 607         if (ret == -1 || vdata->traverse_error) {
 608                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 609                                   "'%s'\n", ctdb_db->db_name));
 610                 return -1;
 611         }
 612
 613         if (vdata->full_total > 0) {
 614                 DEBUG(DEBUG_INFO,
 615                       (__location__
 616                        " full vacuuming db traverse statistics: "
 617                        "db[%s] "
 618                        "total[%u] "
 619                        "skp[%u] "
 620                        "err[%u] "
 621                        "adl[%u] "
 622                        "avf[%u]\n",
 623                        ctdb_db->db_name,
 624                        (unsigned)vdata->full_total,
 625                        (unsigned)vdata->full_skipped,
 626                        (unsigned)vdata->full_error,
 627                        (unsigned)vdata->full_added_to_delete_list,
 628                        (unsigned)vdata->full_added_to_vacuum_fetch_list));
 629         }
 630
 631         return 0;
 632 }
 633
 634 /**
 635  * Process the vacuum fetch lists:
 636  * For records for which we are not the lmaster, tell the lmaster to
 637  * fetch the record.
 638  */
 639 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 640                                            struct vacuum_data *vdata)
 641 {
 642         int i;
 643         struct ctdb_context *ctdb = ctdb_db->ctdb;
 644
 645         for (i = 0; i < ctdb->num_nodes; i++) {
 646                 TDB_DATA data;
 647                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 648
 649                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 650                         continue;
 651                 }
 652
 653                 if (vfl->count == 0) {
 654                         continue;
 655                 }
 656
 657                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 658                                    vfl->count, ctdb->nodes[i]->pnn,
 659                                    ctdb_db->db_name));
 660
 661                 data.dsize = talloc_get_size(vfl);
 662                 data.dptr  = (void *)vfl;
 663                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 664                                              CTDB_SRVID_VACUUM_FETCH,
 665                                              data) != 0)
 666                 {
 667                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 668                                           "fetch message to %u\n",
 669                                           ctdb->nodes[i]->pnn));
 670                         return -1;
 671                 }
 672         }
 673
 674         return 0;
 675 }
 676
 677 /**
 678  * Proces the delete list:
 679  * Send the records to delete to all other nodes with the
 680  * try_delete_records control.
 681  */
 682 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 683                                     struct vacuum_data *vdata)
 684 {
 685         int ret, i;
 686         struct ctdb_context *ctdb = ctdb_db->ctdb;
 687
 688         vdata->delete_left = vdata->delete_count;
 689
 690         if (vdata->delete_count > 0) {
 691                 struct delete_records_list *recs;
 692                 TDB_DATA indata, outdata;
 693                 int32_t res;
 694                 struct ctdb_node_map *nodemap;
 695                 uint32_t *active_nodes;
 696                 int num_active_nodes;
 697
 698                 recs = talloc_zero(vdata, struct delete_records_list);
 699                 if (recs == NULL) {
 700                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 701                         return -1;
 702                 }
 703                 recs->records = (struct ctdb_marshall_buffer *)
 704                         talloc_zero_size(vdata,
 705                                     offsetof(struct ctdb_marshall_buffer, data));
 706                 if (recs->records == NULL) {
 707                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 708                         return -1;
 709                 }
 710                 recs->records->db_id = ctdb_db->db_id;
 711
 712                 /*
 713                  * traverse the tree of all records we want to delete and
 714                  * create a blob we can send to the other nodes.
 715                  */
 716                 trbt_traversearray32(vdata->delete_list, 1,
 717                                      delete_marshall_traverse, recs);
 718
 719                 indata.dsize = talloc_get_size(recs->records);
 720                 indata.dptr  = (void *)recs->records;
 721
 722                 /*
 723                  * now tell all the active nodes to delete all these records
 724                  * (if possible)
 725                  */
 726
 727                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 728                                            CTDB_CURRENT_NODE,
 729                                            recs, /* talloc context */
 730                                            &nodemap);
 731                 if (ret != 0) {
 732                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 733                         return -1;
 734                 }
 735
 736                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 737                                                     nodemap, /* talloc context */
 738                                                     false /* include self */);
 739                 /* yuck! ;-) */
 740                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 741
 742                 for (i = 0; i < num_active_nodes; i++) {
 743                         struct ctdb_marshall_buffer *records;
 744                         struct ctdb_rec_data *rec;
 745
 746                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 747                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 748                                         indata, recs, &outdata, &res,
 749                                         NULL, NULL);
 750                         if (ret != 0 || res != 0) {
 751                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 752                                                   "node %u: ret[%d] res[%d]\n",
 753                                                   active_nodes[i], ret, res));
 754                                 return -1;
 755                         }
 756
 757                         /*
 758                          * outdata contains the list of records coming back
 759                          * from the node: These are the records that the
 760                          * remote node could not delete.
 761                          *
 762                          * NOTE: There is a problem here:
 763                          *
 764                          * When a node failed to delete the record, but
 765                          * others succeeded, we may have created gaps in the
 766                          * history of the record. Hence when a node dies, an
 767                          * closed file handle might be resurrected or an open
 768                          * file handle might be lost, leading to blocked access
 769                          * or data corruption.
 770                          *
 771                          * TODO: This needs to be fixed!
 772                          */
 773                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 774                         rec = (struct ctdb_rec_data *)&records->data[0];
 775                         while (records->count-- > 1) {
 776                                 TDB_DATA reckey, recdata;
 777                                 struct ctdb_ltdb_header *rechdr;
 778                                 struct delete_record_data *dd;
 779
 780                                 reckey.dptr = &rec->data[0];
 781                                 reckey.dsize = rec->keylen;
 782                                 recdata.dptr = &rec->data[reckey.dsize];
 783                                 recdata.dsize = rec->datalen;
 784
 785                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 786                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 787                                         return -1;
 788                                 }
 789                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 790                                 recdata.dptr += sizeof(*rechdr);
 791                                 recdata.dsize -= sizeof(*rechdr);
 792
 793                                 dd = (struct delete_record_data *)trbt_lookup32(
 794                                                 vdata->delete_list,
 795                                                 ctdb_hash(&reckey));
 796                                 if (dd != NULL) {
 797                                         /*
 798                                          * The other node could not delete the
 799                                          * record and it is the first node that
 800                                          * failed. So we should remove it from
 801                                          * the tree and update statistics.
 802                                          */
 803                                         talloc_free(dd);
 804                                         vdata->delete_remote_error++;
 805                                         vdata->delete_left--;
 806                                 }
 807
 808                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 809                         }
 810                 }
 811
 812                 /* free nodemap and active_nodes */
 813                 talloc_free(nodemap);
 814         }
 815
 816         if (vdata->delete_left > 0) {
 817                 /*
 818                  * The only records remaining in the tree are those
 819                  * records which all other nodes could successfully
 820                  * delete, so we can safely delete them on the
 821                  * lmaster as well.
 822                  */
 823                 trbt_traversearray32(vdata->delete_list, 1,
 824                                      delete_record_traverse, vdata);
 825         }
 826
 827         if (vdata->delete_count > 0) {
 828                 DEBUG(DEBUG_INFO,
 829                       (__location__
 830                        " vacuum delete list statistics: "
 831                        "db[%s] "
 832                        "coll[%u] "
 833                        "rem.err[%u] "
 834                        "loc.err[%u] "
 835                        "skip[%u] "
 836                        "del[%u] "
 837                        "left[%u]\n",
 838                        ctdb_db->db_name,
 839                        (unsigned)vdata->delete_count,
 840                        (unsigned)vdata->delete_remote_error,
 841                        (unsigned)vdata->delete_local_error,
 842                        (unsigned)vdata->delete_skipped,
 843                        (unsigned)vdata->delete_deleted,
 844                        (unsigned)vdata->delete_left));
 845         }
 846
 847         return 0;
 848 }
 849
 850 /**
 851  * initialize the vacuum_data
 852  */
 853 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
 854                                         struct vacuum_data *vdata)
 855 {
 856         int i;
 857         struct ctdb_context *ctdb = ctdb_db->ctdb;
 858
 859         vdata->fast_added_to_delete_list = 0;
 860         vdata->fast_added_to_vacuum_fetch_list = 0;
 861         vdata->fast_deleted = 0;
 862         vdata->fast_skipped = 0;
 863         vdata->fast_error = 0;
 864         vdata->fast_total = 0;
 865         vdata->full_added_to_delete_list = 0;
 866         vdata->full_added_to_vacuum_fetch_list = 0;
 867         vdata->full_skipped = 0;
 868         vdata->full_error = 0;
 869         vdata->full_total = 0;
 870         vdata->delete_count = 0;
 871         vdata->delete_left = 0;
 872         vdata->delete_remote_error = 0;
 873         vdata->delete_local_error = 0;
 874         vdata->delete_skipped = 0;
 875         vdata->delete_deleted = 0;
 876
 877         /* the list needs to be of length num_nodes */
 878         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
 879                                                 struct ctdb_marshall_buffer *,
 880                                                 ctdb->num_nodes);
 881         if (vdata->vacuum_fetch_list == NULL) {
 882                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 883                 return -1;
 884         }
 885         for (i = 0; i < ctdb->num_nodes; i++) {
 886                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 887                         talloc_zero_size(vdata->vacuum_fetch_list,
 888                                          offsetof(struct ctdb_marshall_buffer, data));
 889                 if (vdata->vacuum_fetch_list[i] == NULL) {
 890                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 891                         return -1;
 892                 }
 893                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 894         }
 895
 896         return 0;
 897 }
 898
 899 /**
 900  * Vacuum a DB:
 901  *  - Always do the fast vacuuming run, which traverses
 902  *    the in-memory delete queue: these records have been
 903  *    scheduled for deletion.
 904  *  - Only if explicitly requested, the database is traversed
 905  *    in order to use the traditional heuristics on empty records
 906  *    to trigger deletion.
 907  *    This is done only every VacuumFastPathCount'th vacuuming run.
 908  *
 909  * The traverse runs fill two lists:
 910  *
 911  * - The delete_list:
 912  *   This is the list of empty records the current
 913  *   node is lmaster and dmaster for. These records are later
 914  *   deleted first on other nodes and then locally.
 915  *
 916  *   The fast vacuuming run has a short cut for those records
 917  *   that have never been migrated with data: these records
 918  *   are immediately deleted locally, since they have left
 919  *   no trace on other nodes.
 920  *
 921  * - The vacuum_fetch lists
 922  *   (one for each other lmaster node):
 923  *   The records in this list are sent for deletion to
 924  *   their lmaster in a bulk VACUUM_FETCH message.
 925  *
 926  *   The lmaster then migrates all these records to itelf
 927  *   so that they can be vacuumed there.
 928  *
 929  * This executes in the child context.
 930  */
 931 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 932                           struct vacuum_data *vdata,
 933                           bool full_vacuum_run)
 934 {
 935         struct ctdb_context *ctdb = ctdb_db->ctdb;
 936         int ret, pnn;
 937
 938         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 939                            "%s db_id[0x%08x]\n",
 940                            full_vacuum_run ? "full" : "fast",
 941                            ctdb_db->db_name, ctdb_db->db_id));
 942
 943         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 944         if (ret != 0) {
 945                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 946                 return ret;
 947         }
 948
 949         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 950         if (pnn == -1) {
 951                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 952                 return -1;
 953         }
 954
 955         ctdb->pnn = pnn;
 956
 957         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
 958         if (ret != 0) {
 959                 return ret;
 960         }
 961
 962         ctdb_vacuum_db_fast(ctdb_db, vdata);
 963
 964         ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
 965         if (ret != 0) {
 966                 return ret;
 967         }
 968
 969         ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
 970         if (ret != 0) {
 971                 return ret;
 972         }
 973
 974         ret = ctdb_process_delete_list(ctdb_db, vdata);
 975         if (ret != 0) {
 976                 return ret;
 977         }
 978
 979         /* this ensures we run our event queue */
 980         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 981
 982         return 0;
 983 }
 984
 985
 986 /*
 987  * traverse function for repacking
 988  */
 989 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 990 {
 991         struct vacuum_data *vdata = (struct vacuum_data *)private;
 992
 993         if (vdata->vacuum) {
 994                 uint32_t hash = ctdb_hash(&key);
 995                 struct delete_record_data *kd;
 996                 /*
 997                  * check if we can ignore this record because it's in the delete_list
 998                  */
 999                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1000                 /*
1001                  * there might be hash collisions so we have to compare the keys here to be sure
1002                  */
1003                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1004                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1005                         /*
1006                          * we have to check if the record hasn't changed in the meantime in order to
1007                          * savely remove it from the database
1008                          */
1009                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1010                                 hdr->dmaster == kd->ctdb->pnn &&
1011                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1012                                 kd->hdr.rsn == hdr->rsn) {
1013                                 vdata->vacuumed++;
1014                                 return 0;
1015                         }
1016                 }
1017         }
1018         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1019                 vdata->traverse_error = true;
1020                 return -1;
1021         }
1022         vdata->copied++;
1023         return 0;
1024 }
1025
1026 /*
1027  * repack a tdb
1028  */
1029 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1030 {
1031         struct tdb_context *tmp_db;
1032
1033         if (tdb_transaction_start(tdb) != 0) {
1034                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1035                 return -1;
1036         }
1037
1038         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1039                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1040                           O_RDWR|O_CREAT, 0);
1041         if (tmp_db == NULL) {
1042                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1043                 tdb_transaction_cancel(tdb);
1044                 return -1;
1045         }
1046
1047         vdata->traverse_error = false;
1048         vdata->dest_db = tmp_db;
1049         vdata->vacuum = true;
1050         vdata->vacuumed = 0;
1051         vdata->copied = 0;
1052
1053         /*
1054          * repack and vacuum on-the-fly by not writing the records that are
1055          * no longer needed
1056          */
1057         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1058                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1059                 tdb_transaction_cancel(tdb);
1060                 tdb_close(tmp_db);
1061                 return -1;
1062         }
1063
1064         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1065
1066         if (vdata->traverse_error) {
1067                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1068                 tdb_transaction_cancel(tdb);
1069                 tdb_close(tmp_db);
1070                 return -1;
1071         }
1072
1073         if (tdb_wipe_all(tdb) != 0) {
1074                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1075                 tdb_transaction_cancel(tdb);
1076                 tdb_close(tmp_db);
1077                 return -1;
1078         }
1079
1080         vdata->traverse_error = false;
1081         vdata->dest_db = tdb;
1082         vdata->vacuum = false;
1083         vdata->copied = 0;
1084
1085         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1086                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1087                 tdb_transaction_cancel(tdb);
1088                 tdb_close(tmp_db);
1089                 return -1;
1090         }
1091
1092         if (vdata->traverse_error) {
1093                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1094                 tdb_transaction_cancel(tdb);
1095                 tdb_close(tmp_db);
1096                 return -1;
1097         }
1098
1099         tdb_close(tmp_db);
1100
1101
1102         if (tdb_transaction_commit(tdb) != 0) {
1103                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1104                 return -1;
1105         }
1106         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1107
1108         return 0;
1109 }
1110
1111 /*
1112  * repack and vaccum a db
1113  * called from the child context
1114  */
1115 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1116                                      TALLOC_CTX *mem_ctx,
1117                                      bool full_vacuum_run)
1118 {
1119         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1120         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1121         const char *name = ctdb_db->db_name;
1122         int freelist_size;
1123         struct vacuum_data *vdata;
1124
1125         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1126         if (freelist_size == -1) {
1127                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1128                 return -1;
1129         }
1130
1131         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1132         if (vdata == NULL) {
1133                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1134                 return -1;
1135         }
1136
1137         vdata->ctdb = ctdb_db->ctdb;
1138         vdata->vacuum_limit = vacuum_limit;
1139         vdata->repack_limit = repack_limit;
1140         vdata->delete_list = trbt_create(vdata, 0);
1141         vdata->ctdb_db = ctdb_db;
1142         if (vdata->delete_list == NULL) {
1143                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1144                 talloc_free(vdata);
1145                 return -1;
1146         }
1147
1148         vdata->start = timeval_current();
1149
1150         /*
1151          * gather all records that can be deleted in vdata
1152          */
1153         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1154                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1155         }
1156
1157         /*
1158          * decide if a repack is necessary
1159          */
1160         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1161         {
1162                 talloc_free(vdata);
1163                 return 0;
1164         }
1165
1166         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1167                         name, freelist_size, vdata->delete_left));
1168
1169         /*
1170          * repack and implicitely get rid of the records we can delete
1171          */
1172         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1173                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1174                 talloc_free(vdata);
1175                 return -1;
1176         }
1177         talloc_free(vdata);
1178
1179         return 0;
1180 }
1181
1182 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1183 {
1184         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1185
1186         return interval;
1187 }
1188
1189 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1190 {
1191         double l = timeval_elapsed(&child_ctx->start_time);
1192         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1193         struct ctdb_context *ctdb = ctdb_db->ctdb;
1194
1195         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1196
1197         if (child_ctx->child_pid != -1) {
1198                 kill(child_ctx->child_pid, SIGKILL);
1199         } else {
1200                 /* Bump the number of successful fast-path runs. */
1201                 child_ctx->vacuum_handle->fast_path_count++;
1202         }
1203
1204         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1205
1206         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1207                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1208                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1209
1210         return 0;
1211 }
1212
1213 /*
1214  * this event is generated when a vacuum child process times out
1215  */
1216 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1217                                          struct timeval t, void *private_data)
1218 {
1219         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1220
1221         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1222
1223         child_ctx->status = VACUUM_TIMEOUT;
1224
1225         talloc_free(child_ctx);
1226 }
1227
1228
1229 /*
1230  * this event is generated when a vacuum child process has completed
1231  */
1232 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1233                              uint16_t flags, void *private_data)
1234 {
1235         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1236         char c = 0;
1237         int ret;
1238
1239         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1240         child_ctx->child_pid = -1;
1241
1242         ret = read(child_ctx->fd[0], &c, 1);
1243         if (ret != 1 || c != 0) {
1244                 child_ctx->status = VACUUM_ERROR;
1245                 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));
1246         } else {
1247                 child_ctx->status = VACUUM_OK;
1248         }
1249
1250         talloc_free(child_ctx);
1251 }
1252
1253 /*
1254  * this event is called every time we need to start a new vacuum process
1255  */
1256 static void
1257 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1258                                struct timeval t, void *private_data)
1259 {
1260         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1261         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1262         struct ctdb_context *ctdb = ctdb_db->ctdb;
1263         struct ctdb_vacuum_child_context *child_ctx;
1264         struct tevent_fd *fde;
1265         int ret;
1266
1267         /* we dont vacuum if we are in recovery mode, or db frozen */
1268         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1269             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1270                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1271                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1272                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1273                                    ? "freeze pending"
1274                                    : "frozen"));
1275                 event_add_timed(ctdb->ev, vacuum_handle,
1276                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1277                         ctdb_vacuum_event, vacuum_handle);
1278                 return;
1279         }
1280
1281         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1282         if (child_ctx == NULL) {
1283                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1284                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1285         }
1286
1287
1288         ret = pipe(child_ctx->fd);
1289         if (ret != 0) {
1290                 talloc_free(child_ctx);
1291                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1292                 event_add_timed(ctdb->ev, vacuum_handle,
1293                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1294                         ctdb_vacuum_event, vacuum_handle);
1295                 return;
1296         }
1297
1298         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1299                 vacuum_handle->fast_path_count = 0;
1300         }
1301
1302         child_ctx->child_pid = ctdb_fork(ctdb);
1303         if (child_ctx->child_pid == (pid_t)-1) {
1304                 close(child_ctx->fd[0]);
1305                 close(child_ctx->fd[1]);
1306                 talloc_free(child_ctx);
1307                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1308                 event_add_timed(ctdb->ev, vacuum_handle,
1309                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1310                         ctdb_vacuum_event, vacuum_handle);
1311                 return;
1312         }
1313
1314
1315         if (child_ctx->child_pid == 0) {
1316                 char cc = 0;
1317                 bool full_vacuum_run = false;
1318                 close(child_ctx->fd[0]);
1319
1320                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1321
1322                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1323                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1324                         _exit(1);
1325                 }
1326
1327                 /*
1328                  * repack the db
1329                  */
1330                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1331                     (vacuum_handle->fast_path_count == 0))
1332                 {
1333                         full_vacuum_run = true;
1334                 }
1335                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1336                                                full_vacuum_run);
1337
1338                 write(child_ctx->fd[1], &cc, 1);
1339                 _exit(0);
1340         }
1341
1342         set_close_on_exec(child_ctx->fd[0]);
1343         close(child_ctx->fd[1]);
1344
1345         child_ctx->status = VACUUM_RUNNING;
1346         child_ctx->start_time = timeval_current();
1347
1348         DLIST_ADD(ctdb->vacuumers, child_ctx);
1349         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1350
1351         /*
1352          * Clear the fastpath vacuuming list in the parent.
1353          */
1354         talloc_free(ctdb_db->delete_queue);
1355         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1356         if (ctdb_db->delete_queue == NULL) {
1357                 /* fatal here? ... */
1358                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1359                                  "in parent context. Shutting down\n");
1360         }
1361
1362         event_add_timed(ctdb->ev, child_ctx,
1363                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1364                 vacuum_child_timeout, child_ctx);
1365
1366         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1367
1368         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1369                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1370         tevent_fd_set_auto_close(fde);
1371
1372         vacuum_handle->child_ctx = child_ctx;
1373         child_ctx->vacuum_handle = vacuum_handle;
1374 }
1375
1376 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1377 {
1378         /* Simply free them all. */
1379         while (ctdb->vacuumers) {
1380                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1381                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1382                            (int)ctdb->vacuumers->child_pid));
1383                 /* vacuum_child_destructor kills it, removes from list */
1384                 talloc_free(ctdb->vacuumers);
1385         }
1386 }
1387
1388 /* this function initializes the vacuuming context for a database
1389  * starts the vacuuming events
1390  */
1391 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1392 {
1393         if (ctdb_db->persistent != 0) {
1394                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1395                 return 0;
1396         }
1397
1398         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1399         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1400
1401         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1402         ctdb_db->vacuum_handle->fast_path_count = 0;
1403
1404         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1405                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1406                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1407
1408         return 0;
1409 }
1410
1411 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1412                                             const struct ctdb_ltdb_header *hdr,
1413                                             const TDB_DATA key)
1414 {
1415         struct delete_record_data *kd;
1416         uint32_t hash;
1417
1418         hash = (uint32_t)ctdb_hash(&key);
1419
1420         DEBUG(DEBUG_DEBUG, (__location__
1421                             " remove_record_from_delete_queue: "
1422                             "db[%s] "
1423                             "db_id[0x%08x] "
1424                             "key_hash[0x%08x] "
1425                             "lmaster[%u] "
1426                             "migrated_with_data[%s]\n",
1427                              ctdb_db->db_name, ctdb_db->db_id,
1428                              hash,
1429                              ctdb_lmaster(ctdb_db->ctdb, &key),
1430                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1431
1432         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1433         if (kd == NULL) {
1434                 DEBUG(DEBUG_DEBUG, (__location__
1435                                     " remove_record_from_delete_queue: "
1436                                     "record not in queue (hash[0x%08x])\n.",
1437                                     hash));
1438                 return;
1439         }
1440
1441         if ((kd->key.dsize != key.dsize) ||
1442             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1443         {
1444                 DEBUG(DEBUG_DEBUG, (__location__
1445                                     " remove_record_from_delete_queue: "
1446                                     "hash collision for key with hash[0x%08x] "
1447                                     "in db[%s] - skipping\n",
1448                                     hash, ctdb_db->db_name));
1449                 return;
1450         }
1451
1452         DEBUG(DEBUG_DEBUG, (__location__
1453                             " remove_record_from_delete_queue: "
1454                             "removing key with hash[0x%08x]\n",
1455                              hash));
1456
1457         talloc_free(kd);
1458
1459         return;
1460 }
1461
1462 /**
1463  * Insert a record into the ctdb_db context's delete queue,
1464  * handling hash collisions.
1465  */
1466 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1467                                            const struct ctdb_ltdb_header *hdr,
1468                                            TDB_DATA key)
1469 {
1470         struct delete_record_data *kd;
1471         uint32_t hash;
1472         int ret;
1473
1474         hash = (uint32_t)ctdb_hash(&key);
1475
1476         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1477                            "db_id[0x%08x] "
1478                            "key_hash[0x%08x] "
1479                            "lmaster[%u] "
1480                            "migrated_with_data[%s]\n",
1481                             ctdb_db->db_name, ctdb_db->db_id,
1482                             hash,
1483                             ctdb_lmaster(ctdb_db->ctdb, &key),
1484                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1485
1486         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1487         if (kd != NULL) {
1488                 if ((kd->key.dsize != key.dsize) ||
1489                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1490                 {
1491                         DEBUG(DEBUG_INFO,
1492                               (__location__ " schedule for deletion: "
1493                                "hash collision for key hash [0x%08x]. "
1494                                "Skipping the record.\n", hash));
1495                         return 0;
1496                 } else {
1497                         DEBUG(DEBUG_DEBUG,
1498                               (__location__ " schedule for deletion: "
1499                                "updating entry for key with hash [0x%08x].\n",
1500                                hash));
1501                 }
1502         }
1503
1504         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1505                                                   ctdb_db->delete_queue,
1506                                                   hdr, key);
1507         if (ret != 0) {
1508                 DEBUG(DEBUG_INFO,
1509                       (__location__ " schedule for deletion: error "
1510                        "inserting key with hash [0x%08x] into delete queue\n",
1511                        hash));
1512                 return -1;
1513         }
1514
1515         return 0;
1516 }
1517
1518 /**
1519  * Schedule a record for deletetion.
1520  * Called from the parent context.
1521  */
1522 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1523                                            TDB_DATA indata)
1524 {
1525         struct ctdb_control_schedule_for_deletion *dd;
1526         struct ctdb_db_context *ctdb_db;
1527         int ret;
1528         TDB_DATA key;
1529
1530         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1531
1532         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1533         if (ctdb_db == NULL) {
1534                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1535                                   dd->db_id));
1536                 return -1;
1537         }
1538
1539         key.dsize = dd->keylen;
1540         key.dptr = dd->key;
1541
1542         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1543
1544         return ret;
1545 }
1546
1547 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1548                                          const struct ctdb_ltdb_header *hdr,
1549                                          TDB_DATA key)
1550 {
1551         int ret;
1552         struct ctdb_control_schedule_for_deletion *dd;
1553         TDB_DATA indata;
1554         int32_t status;
1555
1556         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1557                 /* main daemon - directly queue */
1558                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1559
1560                 return ret;
1561         }
1562
1563         /* child process: send the main daemon a control */
1564
1565         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1566         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1567         if (indata.dptr == NULL) {
1568                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1569                 return -1;
1570         }
1571         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1572         dd->db_id = ctdb_db->db_id;
1573         dd->hdr = *hdr;
1574         dd->keylen = key.dsize;
1575         memcpy(dd->key, key.dptr, key.dsize);
1576
1577         ret = ctdb_control(ctdb_db->ctdb,
1578                            CTDB_CURRENT_NODE,
1579                            ctdb_db->db_id,
1580                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1581                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1582                            indata,
1583                            NULL, /* mem_ctx */
1584                            NULL, /* outdata */
1585                            &status,
1586                            NULL, /* timeout : NULL == wait forever */
1587                            NULL); /* error message */
1588
1589         talloc_free(indata.dptr);
1590
1591         if (ret != 0 || status != 0) {
1592                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1593                                   "SCHEDULE_FOR_DELETION "
1594                                   "control.\n"));
1595                 if (status != 0) {
1596                         ret = -1;
1597                 }
1598         }
1599
1600         return ret;
1601 }
1602
1603 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1604                                          const struct ctdb_ltdb_header *hdr,
1605                                          const TDB_DATA key)
1606 {
1607         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1608                 /*
1609                  * Only remove the record from the delete queue if called
1610                  * in the main daemon.
1611                  */
1612                 return;
1613         }
1614
1615         remove_record_from_delete_queue(ctdb_db, hdr, key);
1616
1617         return;
1618 }