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  * Process the vacuum fetch lists:
 619  * For records for which we are not the lmaster, tell the lmaster to
 620  * fetch the record.
 621  */
 622 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 623                                            struct vacuum_data *vdata)
 624 {
 625         int i;
 626         struct ctdb_context *ctdb = ctdb_db->ctdb;
 627
 628         for (i = 0; i < ctdb->num_nodes; i++) {
 629                 TDB_DATA data;
 630                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 631
 632                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 633                         continue;
 634                 }
 635
 636                 if (vfl->count == 0) {
 637                         continue;
 638                 }
 639
 640                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 641                                    vfl->count, ctdb->nodes[i]->pnn,
 642                                    ctdb_db->db_name));
 643
 644                 data.dsize = talloc_get_size(vfl);
 645                 data.dptr  = (void *)vfl;
 646                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 647                                              CTDB_SRVID_VACUUM_FETCH,
 648                                              data) != 0)
 649                 {
 650                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 651                                           "fetch message to %u\n",
 652                                           ctdb->nodes[i]->pnn));
 653                         return -1;
 654                 }
 655         }
 656
 657         return 0;
 658 }
 659
 660 /**
 661  * Vacuum a DB:
 662  *  - Always do the fast vacuuming run, which traverses
 663  *    the in-memory delete queue: these records have been
 664  *    scheduled for deletion.
 665  *  - Only if explicitly requested, the database is traversed
 666  *    in order to use the traditional heuristics on empty records
 667  *    to trigger deletion.
 668  *    This is done only every VacuumFastPathCount'th vacuuming run.
 669  *
 670  * The traverse runs fill two lists:
 671  *
 672  * - The delete_list:
 673  *   This is the list of empty records the current
 674  *   node is lmaster and dmaster for. These records are later
 675  *   deleted first on other nodes and then locally.
 676  *
 677  *   The fast vacuuming run has a short cut for those records
 678  *   that have never been migrated with data: these records
 679  *   are immediately deleted locally, since they have left
 680  *   no trace on other nodes.
 681  *
 682  * - The vacuum_fetch lists
 683  *   (one for each other lmaster node):
 684  *   The records in this list are sent for deletion to
 685  *   their lmaster in a bulk VACUUM_FETCH message.
 686  *
 687  *   The lmaster then migrates all these records to itelf
 688  *   so that they can be vacuumed there.
 689  *
 690  * This executes in the child context.
 691  */
 692 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 693                           struct vacuum_data *vdata,
 694                           bool full_vacuum_run)
 695 {
 696         struct ctdb_context *ctdb = ctdb_db->ctdb;
 697         int ret, i, pnn;
 698
 699         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 700                            "%s db_id[0x%08x]\n",
 701                            full_vacuum_run ? "full" : "fast",
 702                            ctdb_db->db_name, ctdb_db->db_id));
 703
 704         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 705         if (ret != 0) {
 706                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 707                 return ret;
 708         }
 709
 710         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 711         if (pnn == -1) {
 712                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 713                 return -1;
 714         }
 715
 716         ctdb->pnn = pnn;
 717
 718         vdata->fast_added_to_delete_list = 0;
 719         vdata->fast_added_to_vacuum_fetch_list = 0;
 720         vdata->fast_deleted = 0;
 721         vdata->fast_skipped = 0;
 722         vdata->fast_error = 0;
 723         vdata->fast_total = 0;
 724         vdata->full_added_to_delete_list = 0;
 725         vdata->full_added_to_vacuum_fetch_list = 0;
 726         vdata->full_skipped = 0;
 727         vdata->full_error = 0;
 728         vdata->full_total = 0;
 729         vdata->delete_count = 0;
 730         vdata->delete_left = 0;
 731         vdata->delete_remote_error = 0;
 732         vdata->delete_local_error = 0;
 733         vdata->delete_skipped = 0;
 734         vdata->delete_deleted = 0;
 735
 736         /* the list needs to be of length num_nodes */
 737         vdata->vacuum_fetch_list = talloc_array(vdata,
 738                                                 struct ctdb_marshall_buffer *,
 739                                                 ctdb->num_nodes);
 740         if (vdata->vacuum_fetch_list == NULL) {
 741                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 742                 return -1;
 743         }
 744         for (i = 0; i < ctdb->num_nodes; i++) {
 745                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 746                         talloc_zero_size(vdata->vacuum_fetch_list,
 747                                          offsetof(struct ctdb_marshall_buffer, data));
 748                 if (vdata->vacuum_fetch_list[i] == NULL) {
 749                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 750                         return -1;
 751                 }
 752                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 753         }
 754
 755         ctdb_vacuum_db_fast(ctdb_db, vdata);
 756
 757         ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
 758         if (ret != 0) {
 759                 return ret;
 760         }
 761
 762         ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
 763         if (ret != 0) {
 764                 return ret;
 765         }
 766
 767         /* Process all records we can delete (if any) */
 768         if (vdata->delete_count > 0) {
 769                 struct delete_records_list *recs;
 770                 TDB_DATA indata, outdata;
 771                 int32_t res;
 772                 struct ctdb_node_map *nodemap;
 773                 uint32_t *active_nodes;
 774                 int num_active_nodes;
 775
 776                 vdata->delete_left = vdata->delete_count;
 777
 778                 recs = talloc_zero(vdata, struct delete_records_list);
 779                 if (recs == NULL) {
 780                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 781                         return -1;
 782                 }
 783                 recs->records = (struct ctdb_marshall_buffer *)
 784                         talloc_zero_size(vdata,
 785                                     offsetof(struct ctdb_marshall_buffer, data));
 786                 if (recs->records == NULL) {
 787                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 788                         return -1;
 789                 }
 790                 recs->records->db_id = ctdb_db->db_id;
 791
 792                 /*
 793                  * traverse the tree of all records we want to delete and
 794                  * create a blob we can send to the other nodes.
 795                  */
 796                 trbt_traversearray32(vdata->delete_list, 1,
 797                                      delete_marshall_traverse, recs);
 798
 799                 indata.dsize = talloc_get_size(recs->records);
 800                 indata.dptr  = (void *)recs->records;
 801
 802                 /*
 803                  * now tell all the active nodes to delete all these records
 804                  * (if possible)
 805                  */
 806
 807                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 808                                            CTDB_CURRENT_NODE,
 809                                            recs, /* talloc context */
 810                                            &nodemap);
 811                 if (ret != 0) {
 812                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 813                         return -1;
 814                 }
 815
 816                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 817                                                     nodemap, /* talloc context */
 818                                                     false /* include self */);
 819                 /* yuck! ;-) */
 820                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 821
 822                 for (i = 0; i < num_active_nodes; i++) {
 823                         struct ctdb_marshall_buffer *records;
 824                         struct ctdb_rec_data *rec;
 825
 826                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 827                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 828                                         indata, recs, &outdata, &res,
 829                                         NULL, NULL);
 830                         if (ret != 0 || res != 0) {
 831                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 832                                                   "node %u: ret[%d] res[%d]\n",
 833                                                   active_nodes[i], ret, res));
 834                                 return -1;
 835                         }
 836
 837                         /*
 838                          * outdata contains the list of records coming back
 839                          * from the node: These are the records that the
 840                          * remote node could not delete.
 841                          */
 842                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 843                         rec = (struct ctdb_rec_data *)&records->data[0];
 844                         while (records->count-- > 1) {
 845                                 TDB_DATA reckey, recdata;
 846                                 struct ctdb_ltdb_header *rechdr;
 847                                 struct delete_record_data *dd;
 848
 849                                 reckey.dptr = &rec->data[0];
 850                                 reckey.dsize = rec->keylen;
 851                                 recdata.dptr = &rec->data[reckey.dsize];
 852                                 recdata.dsize = rec->datalen;
 853
 854                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 855                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 856                                         return -1;
 857                                 }
 858                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 859                                 recdata.dptr += sizeof(*rechdr);
 860                                 recdata.dsize -= sizeof(*rechdr);
 861
 862                                 dd = (struct delete_record_data *)trbt_lookup32(
 863                                                 vdata->delete_list,
 864                                                 ctdb_hash(&reckey));
 865                                 if (dd != NULL) {
 866                                         /*
 867                                          * The other node could not delete the
 868                                          * record and it is the first node that
 869                                          * failed. So we should remove it from
 870                                          * the tree and update statistics.
 871                                          */
 872                                         talloc_free(dd);
 873                                         vdata->delete_remote_error++;
 874                                         vdata->delete_left--;
 875                                 }
 876
 877                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 878                         }
 879                 }
 880
 881                 /* free nodemap and active_nodes */
 882                 talloc_free(nodemap);
 883         }
 884
 885         if (vdata->delete_left > 0) {
 886                 /*
 887                  * The only records remaining in the tree are those
 888                  * records which all other nodes could successfully
 889                  * delete, so we can safely delete them on the
 890                  * lmaster as well.
 891                  */
 892                 trbt_traversearray32(vdata->delete_list, 1,
 893                                      delete_record_traverse, vdata);
 894         }
 895
 896         /* this ensures we run our event queue */
 897         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 898
 899         return 0;
 900 }
 901
 902
 903 /*
 904  * traverse function for repacking
 905  */
 906 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 907 {
 908         struct vacuum_data *vdata = (struct vacuum_data *)private;
 909
 910         if (vdata->vacuum) {
 911                 uint32_t hash = ctdb_hash(&key);
 912                 struct delete_record_data *kd;
 913                 /*
 914                  * check if we can ignore this record because it's in the delete_list
 915                  */
 916                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
 917                 /*
 918                  * there might be hash collisions so we have to compare the keys here to be sure
 919                  */
 920                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 921                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 922                         /*
 923                          * we have to check if the record hasn't changed in the meantime in order to
 924                          * savely remove it from the database
 925                          */
 926                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 927                                 hdr->dmaster == kd->ctdb->pnn &&
 928                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 929                                 kd->hdr.rsn == hdr->rsn) {
 930                                 vdata->vacuumed++;
 931                                 return 0;
 932                         }
 933                 }
 934         }
 935         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 936                 vdata->traverse_error = true;
 937                 return -1;
 938         }
 939         vdata->copied++;
 940         return 0;
 941 }
 942
 943 /*
 944  * repack a tdb
 945  */
 946 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 947 {
 948         struct tdb_context *tmp_db;
 949
 950         if (tdb_transaction_start(tdb) != 0) {
 951                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 952                 return -1;
 953         }
 954
 955         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 956                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 957                           O_RDWR|O_CREAT, 0);
 958         if (tmp_db == NULL) {
 959                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 960                 tdb_transaction_cancel(tdb);
 961                 return -1;
 962         }
 963
 964         vdata->traverse_error = false;
 965         vdata->dest_db = tmp_db;
 966         vdata->vacuum = true;
 967         vdata->vacuumed = 0;
 968         vdata->copied = 0;
 969
 970         /*
 971          * repack and vacuum on-the-fly by not writing the records that are
 972          * no longer needed
 973          */
 974         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 975                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 976                 tdb_transaction_cancel(tdb);
 977                 tdb_close(tmp_db);
 978                 return -1;
 979         }
 980
 981         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 982
 983         if (vdata->traverse_error) {
 984                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 985                 tdb_transaction_cancel(tdb);
 986                 tdb_close(tmp_db);
 987                 return -1;
 988         }
 989
 990         if (tdb_wipe_all(tdb) != 0) {
 991                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 992                 tdb_transaction_cancel(tdb);
 993                 tdb_close(tmp_db);
 994                 return -1;
 995         }
 996
 997         vdata->traverse_error = false;
 998         vdata->dest_db = tdb;
 999         vdata->vacuum = false;
1000         vdata->copied = 0;
1001
1002         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1003                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1004                 tdb_transaction_cancel(tdb);
1005                 tdb_close(tmp_db);
1006                 return -1;
1007         }
1008
1009         if (vdata->traverse_error) {
1010                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1011                 tdb_transaction_cancel(tdb);
1012                 tdb_close(tmp_db);
1013                 return -1;
1014         }
1015
1016         tdb_close(tmp_db);
1017
1018
1019         if (tdb_transaction_commit(tdb) != 0) {
1020                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1021                 return -1;
1022         }
1023         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1024
1025         return 0;
1026 }
1027
1028 /*
1029  * repack and vaccum a db
1030  * called from the child context
1031  */
1032 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1033                                      TALLOC_CTX *mem_ctx,
1034                                      bool full_vacuum_run)
1035 {
1036         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1037         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1038         const char *name = ctdb_db->db_name;
1039         int freelist_size;
1040         struct vacuum_data *vdata;
1041
1042         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1043         if (freelist_size == -1) {
1044                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1045                 return -1;
1046         }
1047
1048         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1049         if (vdata == NULL) {
1050                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1051                 return -1;
1052         }
1053
1054         vdata->ctdb = ctdb_db->ctdb;
1055         vdata->vacuum_limit = vacuum_limit;
1056         vdata->repack_limit = repack_limit;
1057         vdata->delete_list = trbt_create(vdata, 0);
1058         vdata->ctdb_db = ctdb_db;
1059         if (vdata->delete_list == NULL) {
1060                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1061                 talloc_free(vdata);
1062                 return -1;
1063         }
1064
1065         vdata->start = timeval_current();
1066
1067         /*
1068          * gather all records that can be deleted in vdata
1069          */
1070         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1071                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1072         }
1073
1074         /*
1075          * decide if a repack is necessary
1076          */
1077         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1078         {
1079                 talloc_free(vdata);
1080                 return 0;
1081         }
1082
1083         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1084                         name, freelist_size, vdata->delete_left));
1085
1086         /*
1087          * repack and implicitely get rid of the records we can delete
1088          */
1089         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1090                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1091                 talloc_free(vdata);
1092                 return -1;
1093         }
1094         talloc_free(vdata);
1095
1096         return 0;
1097 }
1098
1099 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1100 {
1101         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1102
1103         return interval;
1104 }
1105
1106 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1107 {
1108         double l = timeval_elapsed(&child_ctx->start_time);
1109         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1110         struct ctdb_context *ctdb = ctdb_db->ctdb;
1111
1112         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1113
1114         if (child_ctx->child_pid != -1) {
1115                 kill(child_ctx->child_pid, SIGKILL);
1116         } else {
1117                 /* Bump the number of successful fast-path runs. */
1118                 child_ctx->vacuum_handle->fast_path_count++;
1119         }
1120
1121         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1122
1123         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1124                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1125                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1126
1127         return 0;
1128 }
1129
1130 /*
1131  * this event is generated when a vacuum child process times out
1132  */
1133 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1134                                          struct timeval t, void *private_data)
1135 {
1136         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1137
1138         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1139
1140         child_ctx->status = VACUUM_TIMEOUT;
1141
1142         talloc_free(child_ctx);
1143 }
1144
1145
1146 /*
1147  * this event is generated when a vacuum child process has completed
1148  */
1149 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1150                              uint16_t flags, void *private_data)
1151 {
1152         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1153         char c = 0;
1154         int ret;
1155
1156         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1157         child_ctx->child_pid = -1;
1158
1159         ret = read(child_ctx->fd[0], &c, 1);
1160         if (ret != 1 || c != 0) {
1161                 child_ctx->status = VACUUM_ERROR;
1162                 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));
1163         } else {
1164                 child_ctx->status = VACUUM_OK;
1165         }
1166
1167         talloc_free(child_ctx);
1168 }
1169
1170 /*
1171  * this event is called every time we need to start a new vacuum process
1172  */
1173 static void
1174 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1175                                struct timeval t, void *private_data)
1176 {
1177         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1178         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1179         struct ctdb_context *ctdb = ctdb_db->ctdb;
1180         struct ctdb_vacuum_child_context *child_ctx;
1181         struct tevent_fd *fde;
1182         int ret;
1183
1184         /* we dont vacuum if we are in recovery mode, or db frozen */
1185         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1186             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1187                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1188                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1189                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1190                                    ? "freeze pending"
1191                                    : "frozen"));
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         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1199         if (child_ctx == NULL) {
1200                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1201                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1202         }
1203
1204
1205         ret = pipe(child_ctx->fd);
1206         if (ret != 0) {
1207                 talloc_free(child_ctx);
1208                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1209                 event_add_timed(ctdb->ev, vacuum_handle,
1210                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1211                         ctdb_vacuum_event, vacuum_handle);
1212                 return;
1213         }
1214
1215         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1216                 vacuum_handle->fast_path_count = 0;
1217         }
1218
1219         child_ctx->child_pid = ctdb_fork(ctdb);
1220         if (child_ctx->child_pid == (pid_t)-1) {
1221                 close(child_ctx->fd[0]);
1222                 close(child_ctx->fd[1]);
1223                 talloc_free(child_ctx);
1224                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1225                 event_add_timed(ctdb->ev, vacuum_handle,
1226                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1227                         ctdb_vacuum_event, vacuum_handle);
1228                 return;
1229         }
1230
1231
1232         if (child_ctx->child_pid == 0) {
1233                 char cc = 0;
1234                 bool full_vacuum_run = false;
1235                 close(child_ctx->fd[0]);
1236
1237                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1238
1239                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1240                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1241                         _exit(1);
1242                 }
1243
1244                 /*
1245                  * repack the db
1246                  */
1247                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1248                     (vacuum_handle->fast_path_count == 0))
1249                 {
1250                         full_vacuum_run = true;
1251                 }
1252                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1253                                                full_vacuum_run);
1254
1255                 write(child_ctx->fd[1], &cc, 1);
1256                 _exit(0);
1257         }
1258
1259         set_close_on_exec(child_ctx->fd[0]);
1260         close(child_ctx->fd[1]);
1261
1262         child_ctx->status = VACUUM_RUNNING;
1263         child_ctx->start_time = timeval_current();
1264
1265         DLIST_ADD(ctdb->vacuumers, child_ctx);
1266         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1267
1268         /*
1269          * Clear the fastpath vacuuming list in the parent.
1270          */
1271         talloc_free(ctdb_db->delete_queue);
1272         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1273         if (ctdb_db->delete_queue == NULL) {
1274                 /* fatal here? ... */
1275                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1276                                  "in parent context. Shutting down\n");
1277         }
1278
1279         event_add_timed(ctdb->ev, child_ctx,
1280                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1281                 vacuum_child_timeout, child_ctx);
1282
1283         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1284
1285         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1286                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1287         tevent_fd_set_auto_close(fde);
1288
1289         vacuum_handle->child_ctx = child_ctx;
1290         child_ctx->vacuum_handle = vacuum_handle;
1291 }
1292
1293 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1294 {
1295         /* Simply free them all. */
1296         while (ctdb->vacuumers) {
1297                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1298                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1299                            (int)ctdb->vacuumers->child_pid));
1300                 /* vacuum_child_destructor kills it, removes from list */
1301                 talloc_free(ctdb->vacuumers);
1302         }
1303 }
1304
1305 /* this function initializes the vacuuming context for a database
1306  * starts the vacuuming events
1307  */
1308 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1309 {
1310         if (ctdb_db->persistent != 0) {
1311                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1312                 return 0;
1313         }
1314
1315         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1316         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1317
1318         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1319         ctdb_db->vacuum_handle->fast_path_count = 0;
1320
1321         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1322                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1323                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1324
1325         return 0;
1326 }
1327
1328 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1329                                             const struct ctdb_ltdb_header *hdr,
1330                                             const TDB_DATA key)
1331 {
1332         struct delete_record_data *kd;
1333         uint32_t hash;
1334
1335         hash = (uint32_t)ctdb_hash(&key);
1336
1337         DEBUG(DEBUG_DEBUG, (__location__
1338                             " remove_record_from_delete_queue: db[%s] "
1339                             "db_id[0x%08x] "
1340                             "key_hash[0x%08x] "
1341                             "lmaster[%u] "
1342                             "migrated_with_data[%s]\n",
1343                              ctdb_db->db_name, ctdb_db->db_id,
1344                              hash,
1345                              ctdb_lmaster(ctdb_db->ctdb, &key),
1346                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1347
1348         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1349         if (kd == NULL) {
1350                 return;
1351         }
1352         if (kd->key.dsize != key.dsize) {
1353                 return;
1354         }
1355         if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1356                 return;
1357         }
1358
1359         talloc_free(kd);
1360
1361         return;
1362 }
1363
1364 /**
1365  * Insert a record into the ctdb_db context's delete queue,
1366  * handling hash collisions.
1367  */
1368 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1369                                            const struct ctdb_ltdb_header *hdr,
1370                                            TDB_DATA key)
1371 {
1372         struct delete_record_data *kd;
1373         uint32_t hash;
1374         int ret;
1375
1376         hash = (uint32_t)ctdb_hash(&key);
1377
1378         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1379                            "db_id[0x%08x] "
1380                            "key_hash[0x%08x] "
1381                            "lmaster[%u] "
1382                            "migrated_with_data[%s]\n",
1383                             ctdb_db->db_name, ctdb_db->db_id,
1384                             hash,
1385                             ctdb_lmaster(ctdb_db->ctdb, &key),
1386                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1387
1388         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1389         if (kd != NULL) {
1390                 if ((kd->key.dsize != key.dsize) ||
1391                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1392                 {
1393                         DEBUG(DEBUG_INFO,
1394                               ("schedule for deletion: Hash collision (0x%08x)."
1395                                " Skipping the record.\n", hash));
1396                         return 0;
1397                 } else {
1398                         DEBUG(DEBUG_DEBUG,
1399                               ("schedule for deletion: Overwriting entry for "
1400                                "key with hash 0x%08x.\n", hash));
1401                 }
1402         }
1403
1404         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1405                                                   ctdb_db->delete_queue,
1406                                                   hdr, key);
1407         if (ret != 0) {
1408                 return -1;
1409         }
1410
1411         return 0;
1412 }
1413
1414 /**
1415  * Schedule a record for deletetion.
1416  * Called from the parent context.
1417  */
1418 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1419                                            TDB_DATA indata)
1420 {
1421         struct ctdb_control_schedule_for_deletion *dd;
1422         struct ctdb_db_context *ctdb_db;
1423         int ret;
1424         TDB_DATA key;
1425
1426         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1427
1428         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1429         if (ctdb_db == NULL) {
1430                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1431                                   dd->db_id));
1432                 return -1;
1433         }
1434
1435         key.dsize = dd->keylen;
1436         key.dptr = dd->key;
1437
1438         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1439
1440         return ret;
1441 }
1442
1443 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1444                                          const struct ctdb_ltdb_header *hdr,
1445                                          TDB_DATA key)
1446 {
1447         int ret;
1448         struct ctdb_control_schedule_for_deletion *dd;
1449         TDB_DATA indata;
1450         int32_t status;
1451
1452         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1453                 /* main daemon - directly queue */
1454                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1455
1456                 return ret;
1457         }
1458
1459         /* child process: send the main daemon a control */
1460
1461         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1462         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1463         if (indata.dptr == NULL) {
1464                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1465                 return -1;
1466         }
1467         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1468         dd->db_id = ctdb_db->db_id;
1469         dd->hdr = *hdr;
1470         dd->keylen = key.dsize;
1471         memcpy(dd->key, key.dptr, key.dsize);
1472
1473         ret = ctdb_control(ctdb_db->ctdb,
1474                            CTDB_CURRENT_NODE,
1475                            ctdb_db->db_id,
1476                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1477                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1478                            indata,
1479                            NULL, /* mem_ctx */
1480                            NULL, /* outdata */
1481                            &status,
1482                            NULL, /* timeout : NULL == wait forever */
1483                            NULL); /* error message */
1484
1485         talloc_free(indata.dptr);
1486
1487         if (ret != 0 || status != 0) {
1488                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1489                                   "SCHEDULE_FOR_DELETION "
1490                                   "control.\n"));
1491                 if (status != 0) {
1492                         ret = -1;
1493                 }
1494         }
1495
1496         return ret;
1497 }
1498
1499 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1500                                          const struct ctdb_ltdb_header *hdr,
1501                                          const TDB_DATA key)
1502 {
1503         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1504                 /*
1505                  * Only remove the record from the delete queue if called
1506                  * in the main daemon.
1507                  */
1508                 return;
1509         }
1510
1511         remove_record_from_delete_queue(ctdb_db, hdr, key);
1512
1513         return;
1514 }