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 #define TUNINGDBNAME "vactune.tdb"
  37
  38 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  39
  40 struct ctdb_vacuum_child_context {
  41         struct ctdb_vacuum_child_context *next, *prev;
  42         struct ctdb_vacuum_handle *vacuum_handle;
  43         /* fd child writes status to */
  44         int fd[2];
  45         pid_t child_pid;
  46         enum vacuum_child_status status;
  47         struct timeval start_time;
  48 };
  49
  50 struct ctdb_vacuum_handle {
  51         struct ctdb_db_context *ctdb_db;
  52         struct ctdb_vacuum_child_context *child_ctx;
  53         uint32_t fast_path_count;
  54 };
  55
  56
  57 /*  a list of records to possibly delete */
  58 struct vacuum_data {
  59         uint32_t vacuum_limit;
  60         uint32_t repack_limit;
  61         struct ctdb_context *ctdb;
  62         struct ctdb_db_context *ctdb_db;
  63         struct tdb_context *dest_db;
  64         trbt_tree_t *delete_tree;
  65         uint32_t delete_count;
  66         struct ctdb_marshall_buffer **list;
  67         struct timeval start;
  68         bool traverse_error;
  69         bool vacuum;
  70         uint32_t total;
  71         uint32_t vacuumed;
  72         uint32_t copied;
  73         uint32_t fast_added_to_vacuum_fetch_list;
  74         uint32_t fast_added_to_delete_tree;
  75         uint32_t fast_deleted;
  76         uint32_t fast_skipped;
  77         uint32_t fast_error;
  78         uint32_t fast_total;
  79         uint32_t full_added_to_vacuum_fetch_list;
  80         uint32_t full_added_to_delete_tree;
  81         uint32_t full_skipped;
  82         uint32_t full_error;
  83         uint32_t full_total;
  84 };
  85
  86 /* tuning information stored for every db */
  87 struct vacuum_tuning_data {
  88         uint32_t last_num_repack;
  89         uint32_t last_num_empty;
  90         uint32_t last_interval;
  91         uint32_t new_interval;
  92         struct timeval last_start;
  93         double   last_duration;
  94 };
  95
  96 /* this structure contains the information for one record to be deleted */
  97 struct delete_record_data {
  98         struct ctdb_context *ctdb;
  99         struct ctdb_db_context *ctdb_db;
 100         struct ctdb_ltdb_header hdr;
 101         TDB_DATA key;
 102 };
 103
 104 struct delete_records_list {
 105         struct ctdb_marshall_buffer *records;
 106 };
 107
 108 /**
 109  * Store key and header in a tree, indexed by the key hash.
 110  */
 111 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 112                                                struct ctdb_db_context *ctdb_db,
 113                                                trbt_tree_t *tree,
 114                                                const struct ctdb_ltdb_header *hdr,
 115                                                TDB_DATA key)
 116 {
 117         struct delete_record_data *dd;
 118         uint32_t hash;
 119
 120         dd = talloc_zero(tree, struct delete_record_data);
 121         if (dd == NULL) {
 122                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 123                 return -1;
 124         }
 125
 126         dd->ctdb      = ctdb;
 127         dd->ctdb_db   = ctdb_db;
 128         dd->key.dsize = key.dsize;
 129         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 130         if (dd->key.dptr == NULL) {
 131                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 132                 return -1;
 133         }
 134
 135         dd->hdr = *hdr;
 136
 137         hash = ctdb_hash(&key);
 138
 139         trbt_insert32(tree, hash, dd);
 140
 141         return 0;
 142 }
 143
 144 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
 145                                      struct ctdb_ltdb_header *hdr)
 146 {
 147         struct ctdb_context *ctdb = vdata->ctdb;
 148         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 149         uint32_t hash;
 150         int ret;
 151
 152         hash = ctdb_hash(&key);
 153
 154         if (trbt_lookup32(vdata->delete_tree, hash)) {
 155                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 156                 return 0;
 157         }
 158
 159         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 160                                                   vdata->delete_tree,
 161                                                   hdr, key);
 162         if (ret != 0) {
 163                 return -1;
 164         }
 165
 166         vdata->delete_count++;
 167
 168         return 0;
 169 }
 170
 171 /**
 172  * Add a record to the list of records to be sent
 173  * to their lmaster with VACUUM_FETCH.
 174  */
 175 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 176                                            TDB_DATA key)
 177 {
 178         struct ctdb_context *ctdb = vdata->ctdb;
 179         struct ctdb_rec_data *rec;
 180         uint32_t lmaster;
 181         size_t old_size;
 182
 183         lmaster = ctdb_lmaster(ctdb, &key);
 184
 185         rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
 186         if (rec == NULL) {
 187                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 188                 vdata->traverse_error = true;
 189                 return -1;
 190         }
 191
 192         old_size = talloc_get_size(vdata->list[lmaster]);
 193         vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
 194                                                    old_size + rec->length);
 195         if (vdata->list[lmaster] == NULL) {
 196                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 197                 vdata->traverse_error = true;
 198                 return -1;
 199         }
 200
 201         vdata->list[lmaster]->count++;
 202         memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
 203         talloc_free(rec);
 204
 205         vdata->total++;
 206
 207         return 0;
 208 }
 209
 210
 211 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 212                               struct timeval t, void *private_data);
 213
 214
 215 /*
 216  * traverse function for gathering the records that can be deleted
 217  */
 218 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 219 {
 220         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 221         struct ctdb_context *ctdb = vdata->ctdb;
 222         uint32_t lmaster;
 223         struct ctdb_ltdb_header *hdr;
 224         int res = 0;
 225
 226         vdata->full_total++;
 227
 228         lmaster = ctdb_lmaster(ctdb, &key);
 229         if (lmaster >= ctdb->num_nodes) {
 230                 vdata->full_error++;
 231                 DEBUG(DEBUG_CRIT, (__location__
 232                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 233                                    " with hash[%u]!\n",
 234                                    (unsigned)lmaster,
 235                                    (unsigned)ctdb->num_nodes,
 236                                    (unsigned)ctdb_hash(&key)));
 237                 return -1;
 238         }
 239
 240         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 241                 /* it is not a deleted record */
 242                 vdata->full_skipped++;
 243                 return 0;
 244         }
 245
 246         hdr = (struct ctdb_ltdb_header *)data.dptr;
 247
 248         if (hdr->dmaster != ctdb->pnn) {
 249                 vdata->full_skipped++;
 250                 return 0;
 251         }
 252
 253         if (lmaster == ctdb->pnn) {
 254                 /*
 255                  * We are both lmaster and dmaster, and the record is empty.
 256                  * So we should be able to delete it.
 257                  */
 258                 res = add_record_to_delete_tree(vdata, key, hdr);
 259                 if (res != 0) {
 260                         vdata->full_error++;
 261                 } else {
 262                         vdata->full_added_to_delete_tree++;
 263                 }
 264         } else {
 265                 /*
 266                  * We are not lmaster.
 267                  * Add the record to the blob ready to send to the nodes.
 268                  */
 269                 res = add_record_to_vacuum_fetch_list(vdata, key);
 270                 if (res != 0) {
 271                         vdata->full_error++;
 272                 } else {
 273                         vdata->full_added_to_vacuum_fetch_list++;
 274                 }
 275         }
 276
 277         return res;
 278 }
 279
 280 /*
 281  * traverse the tree of records to delete and marshall them into
 282  * a blob
 283  */
 284 static int delete_traverse(void *param, void *data)
 285 {
 286         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 287         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 288         struct ctdb_rec_data *rec;
 289         size_t old_size;
 290
 291         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 292         if (rec == NULL) {
 293                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 294                 return 0;
 295         }
 296
 297         old_size = talloc_get_size(recs->records);
 298         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 299         if (recs->records == NULL) {
 300                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 301                 return 0;
 302         }
 303         recs->records->count++;
 304         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 305         return 0;
 306 }
 307
 308 /**
 309  * traverse function for the traversal of the delete_queue,
 310  * the fast-path vacuuming list.
 311  *
 312  *  - If the record has been migrated off the node
 313  *    or has been revived (filled with data) on the node,
 314  *    then skip the record.
 315  *
 316  *  - If the current node is the record's lmaster and it is
 317  *    a record that has never been migrated with data, then
 318  *    delete the record from the local tdb.
 319  *
 320  *  - If the current node is the record's lmaster and it has
 321  *    been migrated with data, then schedule it for the normal
 322  *    vacuuming procedure (i.e. add it to the delete_list).
 323  *
 324  *  - If the current node is NOT the record's lmaster then
 325  *    add it to the list of records that are to be sent to
 326  *    the lmaster with the VACUUM_FETCH message.
 327  */
 328 static int delete_queue_traverse(void *param, void *data)
 329 {
 330         struct delete_record_data *dd =
 331                 talloc_get_type(data, struct delete_record_data);
 332         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 333         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 334         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 335         int res;
 336         struct ctdb_ltdb_header *header;
 337         TDB_DATA tdb_data;
 338         uint32_t lmaster;
 339
 340         vdata->fast_total++;
 341
 342         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 343         if (res != 0) {
 344                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 345                 vdata->fast_error++;
 346                 return 0;
 347         }
 348
 349         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 350         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 351                 /* Does not exist or not a ctdb record. Skip. */
 352                 goto skipped;
 353         }
 354
 355         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 356                 /* The record has been recycled (filled with data). Skip. */
 357                 goto skipped;
 358         }
 359
 360         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 361
 362         if (header->dmaster != ctdb->pnn) {
 363                 /* The record has been migrated off the node. Skip. */
 364                 goto skipped;
 365         }
 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_tree(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_tree++;
 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 from local "
 419                                "data base.\n"));
 420                         vdata->fast_error++;
 421                 } else {
 422                         vdata->fast_deleted++;
 423                 }
 424         }
 425
 426         goto done;
 427
 428 skipped:
 429         vdata->fast_skipped++;
 430
 431 done:
 432         if (tdb_data.dptr != NULL) {
 433                 free(tdb_data.dptr);
 434         }
 435         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 436
 437         return 0;
 438 }
 439
 440 /**
 441  * Vacuum a DB:
 442  *  - Always do the fast vacuuming run, which traverses
 443  *    the in-memory delete queue: these records have been
 444  *    scheduled for deletion.
 445  *  - Only if explicitly requested, the database is traversed
 446  *    in order to use the traditional heuristics on empty records
 447  *    to trigger deletion.
 448  *    This is done only every VacuumFastPathCount'th vacuuming run.
 449  *
 450  * The traverse runs fill two lists:
 451  *
 452  * - The delete_list:
 453  *   This is the list of empty records the current
 454  *   node is lmaster and dmaster for. These records are later
 455  *   deleted first on other nodes and then locally.
 456  *
 457  *   The fast vacuuming run has a short cut for those records
 458  *   that have never been migrated with data: these records
 459  *   are immediately deleted locally, since they have left
 460  *   no trace on other nodes.
 461  *
 462  * - The vacuum_fetch lists
 463  *   (one for each other lmaster node):
 464  *   The records in this list are sent for deletion to
 465  *   their lmaster in a bulk VACUUM_FETCH message.
 466  *
 467  *   The lmaster then migrates all these records to itelf
 468  *   so that they can be vacuumed there.
 469  *
 470  * This executes in the child context.
 471  */
 472 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 473                           struct vacuum_data *vdata,
 474                           bool full_vacuum_run)
 475 {
 476         struct ctdb_context *ctdb = ctdb_db->ctdb;
 477         const char *name = ctdb_db->db_name;
 478         int ret, i, pnn;
 479
 480         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 481                            "%s db_id[0x%08x]\n",
 482                            full_vacuum_run ? "full" : "fast",
 483                            ctdb_db->db_name, ctdb_db->db_id));
 484
 485         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 486         if (ret != 0) {
 487                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 488                 return ret;
 489         }
 490
 491         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 492         if (pnn == -1) {
 493                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 494                 return -1;
 495         }
 496
 497         ctdb->pnn = pnn;
 498
 499         vdata->fast_added_to_delete_tree = 0;
 500         vdata->fast_added_to_vacuum_fetch_list = 0;
 501         vdata->fast_deleted = 0;
 502         vdata->fast_skipped = 0;
 503         vdata->fast_error = 0;
 504         vdata->fast_total = 0;
 505         vdata->full_added_to_delete_tree = 0;
 506         vdata->full_added_to_vacuum_fetch_list = 0;
 507         vdata->full_skipped = 0;
 508         vdata->full_error = 0;
 509         vdata->full_total = 0;
 510
 511         /* the list needs to be of length num_nodes */
 512         vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
 513         if (vdata->list == NULL) {
 514                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 515                 return -1;
 516         }
 517         for (i = 0; i < ctdb->num_nodes; i++) {
 518                 vdata->list[i] = (struct ctdb_marshall_buffer *)
 519                         talloc_zero_size(vdata->list,
 520                                          offsetof(struct ctdb_marshall_buffer, data));
 521                 if (vdata->list[i] == NULL) {
 522                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 523                         return -1;
 524                 }
 525                 vdata->list[i]->db_id = ctdb_db->db_id;
 526         }
 527
 528         /*
 529          * Traverse the delete_queue.
 530          * This builds the same lists as the db traverse.
 531          */
 532         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 533
 534         if (vdata->fast_total > 0) {
 535                 DEBUG(DEBUG_INFO,
 536                       (__location__
 537                        " fast vacuuming delete_queue traverse statistics: "
 538                        "db[%s] "
 539                        "total[%u] "
 540                        "del[%u] "
 541                        "skp[%u] "
 542                        "err[%u] "
 543                        "adt[%u] "
 544                        "avf[%u]\n",
 545                        ctdb_db->db_name,
 546                        (unsigned)vdata->fast_total,
 547                        (unsigned)vdata->fast_deleted,
 548                        (unsigned)vdata->fast_skipped,
 549                        (unsigned)vdata->fast_error,
 550                        (unsigned)vdata->fast_added_to_delete_tree,
 551                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 552         }
 553
 554         /*
 555          * read-only traverse of the database, looking for records that
 556          * might be able to be vacuumed.
 557          *
 558          * This is not done each time but only every tunable
 559          * VacuumFastPathCount times.
 560          */
 561         if (full_vacuum_run) {
 562                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 563                 if (ret == -1 || vdata->traverse_error) {
 564                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 565                         return -1;
 566                 }
 567                 if (vdata->full_total > 0) {
 568                         DEBUG(DEBUG_INFO,
 569                               (__location__
 570                                " full vacuuming db traverse statistics: "
 571                                "db[%s] "
 572                                "total[%u] "
 573                                "skp[%u] "
 574                                "err[%u] "
 575                                "adt[%u] "
 576                                "avf[%u]\n",
 577                                ctdb_db->db_name,
 578                                (unsigned)vdata->full_total,
 579                                (unsigned)vdata->full_skipped,
 580                                (unsigned)vdata->full_error,
 581                                (unsigned)vdata->full_added_to_delete_tree,
 582                                (unsigned)vdata->full_added_to_vacuum_fetch_list));
 583                 }
 584         }
 585
 586         /*
 587          * For records where we are not the lmaster,
 588          * tell the lmaster to fetch the record.
 589          */
 590         for (i = 0; i < ctdb->num_nodes; i++) {
 591                 TDB_DATA data;
 592
 593                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 594                         continue;
 595                 }
 596
 597                 if (vdata->list[i]->count == 0) {
 598                         continue;
 599                 }
 600
 601                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 602                                    vdata->list[i]->count, ctdb->nodes[i]->pnn,
 603                                    name));
 604
 605                 data.dsize = talloc_get_size(vdata->list[i]);
 606                 data.dptr  = (void *)vdata->list[i];
 607                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 608                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 609                                           "fetch message to %u\n",
 610                                           ctdb->nodes[i]->pnn));
 611                         return -1;
 612                 }
 613         }
 614
 615         /* Process all records we can delete (if any) */
 616         if (vdata->delete_count > 0) {
 617                 struct delete_records_list *recs;
 618                 TDB_DATA indata, outdata;
 619                 int32_t res;
 620                 struct ctdb_node_map *nodemap;
 621                 uint32_t *active_nodes;
 622                 int num_active_nodes;
 623
 624                 recs = talloc_zero(vdata, struct delete_records_list);
 625                 if (recs == NULL) {
 626                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 627                         return -1;
 628                 }
 629                 recs->records = (struct ctdb_marshall_buffer *)
 630                         talloc_zero_size(vdata,
 631                                     offsetof(struct ctdb_marshall_buffer, data));
 632                 if (recs->records == NULL) {
 633                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 634                         return -1;
 635                 }
 636                 recs->records->db_id = ctdb_db->db_id;
 637
 638                 /*
 639                  * traverse the tree of all records we want to delete and
 640                  * create a blob we can send to the other nodes.
 641                  */
 642                 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
 643
 644                 indata.dsize = talloc_get_size(recs->records);
 645                 indata.dptr  = (void *)recs->records;
 646
 647                 /*
 648                  * now tell all the active nodes to delete all these records
 649                  * (if possible)
 650                  */
 651
 652                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 653                                            CTDB_CURRENT_NODE,
 654                                            recs, /* talloc context */
 655                                            &nodemap);
 656                 if (ret != 0) {
 657                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 658                         return -1;
 659                 }
 660
 661                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 662                                                     nodemap, /* talloc context */
 663                                                     false /* include self */);
 664                 /* yuck! ;-) */
 665                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 666
 667                 for (i = 0; i < num_active_nodes; i++) {
 668                         struct ctdb_marshall_buffer *records;
 669                         struct ctdb_rec_data *rec;
 670
 671                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 672                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 673                                         indata, recs, &outdata, &res,
 674                                         NULL, NULL);
 675                         if (ret != 0 || res != 0) {
 676                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 677                                                   "node %u: ret[%d] res[%d]\n",
 678                                                   active_nodes[i], ret, res));
 679                                 return -1;
 680                         }
 681
 682                         /*
 683                          * outdata countains the list of records coming back
 684                          * from the node which the node could not delete
 685                          */
 686                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 687                         rec = (struct ctdb_rec_data *)&records->data[0];
 688                         while (records->count-- > 1) {
 689                                 TDB_DATA reckey, recdata;
 690                                 struct ctdb_ltdb_header *rechdr;
 691
 692                                 reckey.dptr = &rec->data[0];
 693                                 reckey.dsize = rec->keylen;
 694                                 recdata.dptr = &rec->data[reckey.dsize];
 695                                 recdata.dsize = rec->datalen;
 696
 697                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 698                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 699                                         return -1;
 700                                 }
 701                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 702                                 recdata.dptr += sizeof(*rechdr);
 703                                 recdata.dsize -= sizeof(*rechdr);
 704
 705                                 /*
 706                                  * that other node couldnt delete the record
 707                                  * so we should delete it and thereby remove it from the tree
 708                                  */
 709                                 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
 710
 711                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 712                         }
 713                 }
 714
 715                 /* free nodemap and active_nodes */
 716                 talloc_free(nodemap);
 717
 718                 /*
 719                  * The only records remaining in the tree would be those
 720                  * records where all other nodes could successfully
 721                  * delete them, so we can safely delete them on the
 722                  * lmaster as well. Deletion implictely happens while
 723                  * we repack the database. The repack algorithm revisits
 724                  * the tree in order to find the records that don't need
 725                  * to be copied / repacked.
 726                  */
 727         }
 728
 729         /* this ensures we run our event queue */
 730         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 731
 732         return 0;
 733 }
 734
 735
 736 /*
 737  * traverse function for repacking
 738  */
 739 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 740 {
 741         struct vacuum_data *vdata = (struct vacuum_data *)private;
 742
 743         if (vdata->vacuum) {
 744                 uint32_t hash = ctdb_hash(&key);
 745                 struct delete_record_data *kd;
 746                 /*
 747                  * check if we can ignore this record because it's in the delete_tree
 748                  */
 749                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
 750                 /*
 751                  * there might be hash collisions so we have to compare the keys here to be sure
 752                  */
 753                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 754                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 755                         /*
 756                          * we have to check if the record hasn't changed in the meantime in order to
 757                          * savely remove it from the database
 758                          */
 759                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 760                                 hdr->dmaster == kd->ctdb->pnn &&
 761                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 762                                 kd->hdr.rsn == hdr->rsn) {
 763                                 vdata->vacuumed++;
 764                                 return 0;
 765                         }
 766                 }
 767         }
 768         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 769                 vdata->traverse_error = true;
 770                 return -1;
 771         }
 772         vdata->copied++;
 773         return 0;
 774 }
 775
 776 /*
 777  * repack a tdb
 778  */
 779 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 780 {
 781         struct tdb_context *tmp_db;
 782
 783         if (tdb_transaction_start(tdb) != 0) {
 784                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 785                 return -1;
 786         }
 787
 788         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 789                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 790                           O_RDWR|O_CREAT, 0);
 791         if (tmp_db == NULL) {
 792                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 793                 tdb_transaction_cancel(tdb);
 794                 return -1;
 795         }
 796
 797         vdata->traverse_error = false;
 798         vdata->dest_db = tmp_db;
 799         vdata->vacuum = true;
 800         vdata->vacuumed = 0;
 801         vdata->copied = 0;
 802
 803         /*
 804          * repack and vacuum on-the-fly by not writing the records that are
 805          * no longer needed
 806          */
 807         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 808                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 809                 tdb_transaction_cancel(tdb);
 810                 tdb_close(tmp_db);
 811                 return -1;
 812         }
 813
 814         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 815
 816         if (vdata->traverse_error) {
 817                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 818                 tdb_transaction_cancel(tdb);
 819                 tdb_close(tmp_db);
 820                 return -1;
 821         }
 822
 823         if (tdb_wipe_all(tdb) != 0) {
 824                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 825                 tdb_transaction_cancel(tdb);
 826                 tdb_close(tmp_db);
 827                 return -1;
 828         }
 829
 830         vdata->traverse_error = false;
 831         vdata->dest_db = tdb;
 832         vdata->vacuum = false;
 833         vdata->copied = 0;
 834
 835         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 836                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 837                 tdb_transaction_cancel(tdb);
 838                 tdb_close(tmp_db);
 839                 return -1;
 840         }
 841
 842         if (vdata->traverse_error) {
 843                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 844                 tdb_transaction_cancel(tdb);
 845                 tdb_close(tmp_db);
 846                 return -1;
 847         }
 848
 849         tdb_close(tmp_db);
 850
 851
 852         if (tdb_transaction_commit(tdb) != 0) {
 853                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 854                 return -1;
 855         }
 856         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 857
 858         return 0;
 859 }
 860
 861 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
 862 {
 863         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 864         TDB_CONTEXT *tune_tdb;
 865         TDB_DATA key, value;
 866         struct vacuum_tuning_data tdata;
 867         struct vacuum_tuning_data *tptr;
 868         char *vac_dbname;
 869         int flags;
 870
 871         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
 872                                      ctdb_db->ctdb->db_directory_state,
 873                                      TUNINGDBNAME, ctdb_db->ctdb->pnn);
 874         if (vac_dbname == NULL) {
 875                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
 876                 talloc_free(tmp_ctx);
 877                 return -1;
 878         }
 879
 880         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
 881         flags |= TDB_DISALLOW_NESTING;
 882         tune_tdb = tdb_open(vac_dbname, 0,
 883                             flags,
 884                             O_RDWR|O_CREAT, 0600);
 885         if (tune_tdb == NULL) {
 886                 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
 887                 talloc_free(tmp_ctx);
 888                 return -1;
 889         }
 890
 891         if (tdb_transaction_start(tune_tdb) != 0) {
 892                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 893                 tdb_close(tune_tdb);
 894                 return -1;
 895         }
 896         key.dptr = discard_const(ctdb_db->db_name);
 897         key.dsize = strlen(ctdb_db->db_name);
 898         value = tdb_fetch(tune_tdb, key);
 899
 900         if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
 901                 tptr = (struct vacuum_tuning_data *)value.dptr;
 902                 tdata = *tptr;
 903
 904                 /*
 905                  * re-calc new vacuum interval:
 906                  * in case no limit was reached we continuously increase the interval
 907                  * until vacuum_max_interval is reached
 908                  * in case a limit was reached we divide the current interval by 2
 909                  * unless vacuum_min_interval is reached
 910                  */
 911                 if (freelist < vdata->repack_limit &&
 912                     vdata->delete_count < vdata->vacuum_limit) {
 913                         if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
 914                                 tdata.new_interval = tdata.last_interval * 110 / 100;
 915                                 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
 916                                         tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 917                         }
 918                 } else {
 919                         tdata.new_interval = tdata.last_interval / 2;
 920                         if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
 921                                 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
 922                                 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
 923                         }
 924                         DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
 925                                          tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 926                 }
 927                 tdata.last_interval = tdata.new_interval;
 928         } else {
 929                 DEBUG(DEBUG_DEBUG,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
 930                 tdata.last_num_repack = freelist;
 931                 tdata.last_num_empty = vdata->delete_count;
 932                 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 933         }
 934
 935         if (value.dptr != NULL) {
 936                 free(value.dptr);
 937         }
 938
 939         tdata.last_start = vdata->start;
 940         tdata.last_duration = timeval_elapsed(&vdata->start);
 941
 942         value.dptr = (unsigned char *)&tdata;
 943         value.dsize = sizeof(tdata);
 944
 945         if (tdb_store(tune_tdb, key, value, 0) != 0) {
 946                 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
 947                 tdb_transaction_cancel(tune_tdb);
 948                 tdb_close(tune_tdb);
 949                 talloc_free(tmp_ctx);
 950                 return -1;
 951         }
 952         tdb_transaction_commit(tune_tdb);
 953         tdb_close(tune_tdb);
 954         talloc_free(tmp_ctx);
 955
 956         return 0;
 957 }
 958
 959 /*
 960  * repack and vaccum a db
 961  * called from the child context
 962  */
 963 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 964                                      TALLOC_CTX *mem_ctx,
 965                                      bool full_vacuum_run)
 966 {
 967         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
 968         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
 969         const char *name = ctdb_db->db_name;
 970         int freelist_size;
 971         struct vacuum_data *vdata;
 972
 973         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
 974         if (freelist_size == -1) {
 975                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
 976                 return -1;
 977         }
 978
 979         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 980         if (vdata == NULL) {
 981                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 982                 return -1;
 983         }
 984
 985         vdata->ctdb = ctdb_db->ctdb;
 986         vdata->vacuum_limit = vacuum_limit;
 987         vdata->repack_limit = repack_limit;
 988         vdata->delete_tree = trbt_create(vdata, 0);
 989         vdata->ctdb_db = ctdb_db;
 990         if (vdata->delete_tree == NULL) {
 991                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 992                 talloc_free(vdata);
 993                 return -1;
 994         }
 995
 996         vdata->start = timeval_current();
 997
 998         /*
 999          * gather all records that can be deleted in vdata
1000          */
1001         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1002                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1003         }
1004
1005         /*
1006          * decide if a repack is necessary
1007          */
1008         if (freelist_size < repack_limit && vdata->delete_count < vacuum_limit)
1009         {
1010                 update_tuning_db(ctdb_db, vdata, freelist_size);
1011                 talloc_free(vdata);
1012                 return 0;
1013         }
1014
1015         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1016                         name, freelist_size, vdata->delete_count));
1017
1018         /*
1019          * repack and implicitely get rid of the records we can delete
1020          */
1021         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1022                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1023                 update_tuning_db(ctdb_db, vdata, freelist_size);
1024                 talloc_free(vdata);
1025                 return -1;
1026         }
1027         update_tuning_db(ctdb_db, vdata, freelist_size);
1028         talloc_free(vdata);
1029
1030         return 0;
1031 }
1032
1033 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1034 {
1035         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1036         TDB_CONTEXT *tdb;
1037         TDB_DATA key, value;
1038         char *vac_dbname;
1039         uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
1040         struct ctdb_context *ctdb = ctdb_db->ctdb;
1041         int flags;
1042
1043         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
1044         if (vac_dbname == NULL) {
1045                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
1046                 talloc_free(tmp_ctx);
1047                 return interval;
1048         }
1049
1050         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
1051         flags |= TDB_DISALLOW_NESTING;
1052         tdb = tdb_open(vac_dbname, 0,
1053                        flags,
1054                        O_RDWR|O_CREAT, 0600);
1055         if (!tdb) {
1056                 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval. Errno : %s (%d)\n", vac_dbname, strerror(errno), errno));
1057                 talloc_free(tmp_ctx);
1058                 return interval;
1059         }
1060
1061         key.dptr = discard_const(ctdb_db->db_name);
1062         key.dsize = strlen(ctdb_db->db_name);
1063
1064         value = tdb_fetch(tdb, key);
1065
1066         if (value.dptr != NULL) {
1067                 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
1068                         struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
1069
1070                         interval = tptr->new_interval;
1071
1072                         if (interval < ctdb->tunable.vacuum_min_interval) {
1073                                 interval = ctdb->tunable.vacuum_min_interval;
1074                         }
1075                         if (interval > ctdb->tunable.vacuum_max_interval) {
1076                                 interval = ctdb->tunable.vacuum_max_interval;
1077                         }
1078                 }
1079                 free(value.dptr);
1080         }
1081         tdb_close(tdb);
1082
1083         talloc_free(tmp_ctx);
1084
1085         return interval;
1086 }
1087
1088 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1089 {
1090         double l = timeval_elapsed(&child_ctx->start_time);
1091         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1092         struct ctdb_context *ctdb = ctdb_db->ctdb;
1093
1094         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1095
1096         if (child_ctx->child_pid != -1) {
1097                 kill(child_ctx->child_pid, SIGKILL);
1098         } else {
1099                 /* Bump the number of successful fast-path runs. */
1100                 child_ctx->vacuum_handle->fast_path_count++;
1101         }
1102
1103         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1104
1105         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1106                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1107                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1108
1109         return 0;
1110 }
1111
1112 /*
1113  * this event is generated when a vacuum child process times out
1114  */
1115 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1116                                          struct timeval t, void *private_data)
1117 {
1118         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1119
1120         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1121
1122         child_ctx->status = VACUUM_TIMEOUT;
1123
1124         talloc_free(child_ctx);
1125 }
1126
1127
1128 /*
1129  * this event is generated when a vacuum child process has completed
1130  */
1131 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1132                              uint16_t flags, void *private_data)
1133 {
1134         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1135         char c = 0;
1136         int ret;
1137
1138         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1139         child_ctx->child_pid = -1;
1140
1141         ret = read(child_ctx->fd[0], &c, 1);
1142         if (ret != 1 || c != 0) {
1143                 child_ctx->status = VACUUM_ERROR;
1144                 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));
1145         } else {
1146                 child_ctx->status = VACUUM_OK;
1147         }
1148
1149         talloc_free(child_ctx);
1150 }
1151
1152 /*
1153  * this event is called every time we need to start a new vacuum process
1154  */
1155 static void
1156 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1157                                struct timeval t, void *private_data)
1158 {
1159         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1160         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1161         struct ctdb_context *ctdb = ctdb_db->ctdb;
1162         struct ctdb_vacuum_child_context *child_ctx;
1163         struct tevent_fd *fde;
1164         int ret;
1165
1166         /* we dont vacuum if we are in recovery mode, or db frozen */
1167         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1168             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1169                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1170                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1171                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1172                                    ? "freeze pending"
1173                                    : "frozen"));
1174                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1175                 return;
1176         }
1177
1178         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1179         if (child_ctx == NULL) {
1180                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1181                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1182         }
1183
1184
1185         ret = pipe(child_ctx->fd);
1186         if (ret != 0) {
1187                 talloc_free(child_ctx);
1188                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1189                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1190                 return;
1191         }
1192
1193         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1194                 vacuum_handle->fast_path_count = 0;
1195         }
1196
1197         child_ctx->child_pid = ctdb_fork(ctdb);
1198         if (child_ctx->child_pid == (pid_t)-1) {
1199                 close(child_ctx->fd[0]);
1200                 close(child_ctx->fd[1]);
1201                 talloc_free(child_ctx);
1202                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1203                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1204                 return;
1205         }
1206
1207
1208         if (child_ctx->child_pid == 0) {
1209                 char cc = 0;
1210                 bool full_vacuum_run = false;
1211                 close(child_ctx->fd[0]);
1212
1213                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1214
1215                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1216                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1217                         _exit(1);
1218                 }
1219
1220                 /*
1221                  * repack the db
1222                  */
1223                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1224                     (vacuum_handle->fast_path_count == 0))
1225                 {
1226                         full_vacuum_run = true;
1227                 }
1228                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1229                                                full_vacuum_run);
1230
1231                 write(child_ctx->fd[1], &cc, 1);
1232                 _exit(0);
1233         }
1234
1235         set_close_on_exec(child_ctx->fd[0]);
1236         close(child_ctx->fd[1]);
1237
1238         child_ctx->status = VACUUM_RUNNING;
1239         child_ctx->start_time = timeval_current();
1240
1241         DLIST_ADD(ctdb->vacuumers, child_ctx);
1242         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1243
1244         /*
1245          * Clear the fastpath vacuuming list in the parent.
1246          */
1247         talloc_free(ctdb_db->delete_queue);
1248         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1249         if (ctdb_db->delete_queue == NULL) {
1250                 /* fatal here? ... */
1251                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1252                                  "in parent context. Shutting down\n");
1253         }
1254
1255         event_add_timed(ctdb->ev, child_ctx,
1256                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1257                 vacuum_child_timeout, child_ctx);
1258
1259         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1260
1261         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1262                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1263         tevent_fd_set_auto_close(fde);
1264
1265         vacuum_handle->child_ctx = child_ctx;
1266         child_ctx->vacuum_handle = vacuum_handle;
1267 }
1268
1269 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1270 {
1271         /* Simply free them all. */
1272         while (ctdb->vacuumers) {
1273                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1274                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1275                            (int)ctdb->vacuumers->child_pid));
1276                 /* vacuum_child_destructor kills it, removes from list */
1277                 talloc_free(ctdb->vacuumers);
1278         }
1279 }
1280
1281 /* this function initializes the vacuuming context for a database
1282  * starts the vacuuming events
1283  */
1284 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1285 {
1286         if (ctdb_db->persistent != 0) {
1287                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1288                 return 0;
1289         }
1290
1291         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1292         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1293
1294         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1295         ctdb_db->vacuum_handle->fast_path_count = 0;
1296
1297         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1298                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1299                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1300
1301         return 0;
1302 }
1303
1304 /**
1305  * Insert a record into the ctdb_db context's delete queue,
1306  * handling hash collisions.
1307  */
1308 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1309                                            const struct ctdb_ltdb_header *hdr,
1310                                            TDB_DATA key)
1311 {
1312         struct delete_record_data *kd;
1313         uint32_t hash;
1314         int ret;
1315
1316         hash = (uint32_t)ctdb_hash(&key);
1317
1318         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1319                            "db_id[0x%08x] "
1320                            "key_hash[0x%08x] "
1321                            "lmaster[%u] "
1322                            "migrated_with_data[%s]\n",
1323                             ctdb_db->db_name, ctdb_db->db_id,
1324                             hash,
1325                             ctdb_lmaster(ctdb_db->ctdb, &key),
1326                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1327
1328         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1329         if (kd != NULL) {
1330                 if ((kd->key.dsize != key.dsize) ||
1331                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1332                 {
1333                         DEBUG(DEBUG_INFO,
1334                               ("schedule for deletion: Hash collision (0x%08x)."
1335                                " Skipping the record.\n", hash));
1336                         return 0;
1337                 } else {
1338                         DEBUG(DEBUG_DEBUG,
1339                               ("schedule for deletion: Overwriting entry for "
1340                                "key with hash 0x%08x.\n", hash));
1341                 }
1342         }
1343
1344         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1345                                                   ctdb_db->delete_queue,
1346                                                   hdr, key);
1347         if (ret != 0) {
1348                 return -1;
1349         }
1350
1351         return 0;
1352 }
1353
1354 /**
1355  * Schedule a record for deletetion.
1356  * Called from the parent context.
1357  */
1358 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1359                                            TDB_DATA indata)
1360 {
1361         struct ctdb_control_schedule_for_deletion *dd;
1362         struct ctdb_db_context *ctdb_db;
1363         int ret;
1364         TDB_DATA key;
1365
1366         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1367
1368         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1369         if (ctdb_db == NULL) {
1370                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1371                                   dd->db_id));
1372                 return -1;
1373         }
1374
1375         key.dsize = dd->keylen;
1376         key.dptr = dd->key;
1377
1378         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1379
1380         return ret;
1381 }
1382
1383 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1384                                          const struct ctdb_ltdb_header *hdr,
1385                                          TDB_DATA key)
1386 {
1387         int ret;
1388         struct ctdb_control_schedule_for_deletion *dd;
1389         TDB_DATA indata;
1390         int32_t status;
1391
1392         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1393                 /* main daemon - directly queue */
1394                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1395
1396                 return ret;
1397         }
1398
1399         /* child process: send the main daemon a control */
1400
1401         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1402         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1403         if (indata.dptr == NULL) {
1404                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1405                 return -1;
1406         }
1407         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1408         dd->db_id = ctdb_db->db_id;
1409         dd->hdr = *hdr;
1410         dd->keylen = key.dsize;
1411         memcpy(dd->key, key.dptr, key.dsize);
1412
1413         ret = ctdb_control(ctdb_db->ctdb,
1414                            CTDB_CURRENT_NODE,
1415                            ctdb_db->db_id,
1416                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1417                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1418                            indata,
1419                            NULL, /* mem_ctx */
1420                            NULL, /* outdata */
1421                            &status,
1422                            NULL, /* timeout : NULL == wait forever */
1423                            NULL); /* error message */
1424
1425         talloc_free(indata.dptr);
1426
1427         if (ret != 0 || status != 0) {
1428                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1429                                   "SCHEDULE_FOR_DELETION "
1430                                   "control.\n"));
1431                 if (status != 0) {
1432                         ret = -1;
1433                 }
1434         }
1435
1436         return ret;
1437 }