server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program; if not, see <http://www.gnu.org/licenses/>.
  18 */
  19
  20 #include "includes.h"
  21 #include "lib/events/events.h"
  22 #include "lib/tdb/include/tdb.h"
  23 #include "system/network.h"
  24 #include "system/filesys.h"
  25 #include "system/dir.h"
  26 #include "../include/ctdb_private.h"
  27 #include "db_wrap.h"
  28 #include "lib/util/dlinklist.h"
  29 #include "lib/events/events.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34 #define TUNINGDBNAME "vactune.tdb"
  35
  36 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  37
  38 struct ctdb_vacuum_child_context {
  39         struct ctdb_vacuum_handle *vacuum_handle;
  40         int fd[2];
  41         pid_t child_pid;
  42         enum vacuum_child_status status;
  43         struct timeval start_time;
  44 };
  45
  46 struct ctdb_vacuum_handle {
  47         struct ctdb_db_context *ctdb_db;
  48         struct ctdb_vacuum_child_context *child_ctx;
  49         uint32_t fast_path_count;
  50 };
  51
  52
  53 /*  a list of records to possibly delete */
  54 struct vacuum_data {
  55         uint32_t vacuum_limit;
  56         uint32_t repack_limit;
  57         struct ctdb_context *ctdb;
  58         struct ctdb_db_context *ctdb_db;
  59         struct tdb_context *dest_db;
  60         trbt_tree_t *delete_tree;
  61         uint32_t delete_count;
  62         struct ctdb_marshall_buffer **list;
  63         struct timeval start;
  64         bool traverse_error;
  65         bool vacuum;
  66         uint32_t total;
  67         uint32_t vacuumed;
  68         uint32_t copied;
  69         uint32_t fast_added_to_vacuum_fetch_list;
  70         uint32_t fast_added_to_delete_tree;
  71         uint32_t fast_deleted;
  72         uint32_t fast_skipped;
  73         uint32_t fast_error;
  74         uint32_t fast_total;
  75         uint32_t full_added_to_vacuum_fetch_list;
  76         uint32_t full_added_to_delete_tree;
  77         uint32_t full_skipped;
  78         uint32_t full_error;
  79         uint32_t full_total;
  80 };
  81
  82 /* tuning information stored for every db */
  83 struct vacuum_tuning_data {
  84         uint32_t last_num_repack;
  85         uint32_t last_num_empty;
  86         uint32_t last_interval;
  87         uint32_t new_interval;
  88         struct timeval last_start;
  89         double   last_duration;
  90 };
  91
  92 /* this structure contains the information for one record to be deleted */
  93 struct delete_record_data {
  94         struct ctdb_context *ctdb;
  95         struct ctdb_db_context *ctdb_db;
  96         struct ctdb_ltdb_header hdr;
  97         TDB_DATA key;
  98 };
  99
 100 struct delete_records_list {
 101         struct ctdb_marshall_buffer *records;
 102 };
 103
 104 /**
 105  * Store key and header in a tree, indexed by the key hash.
 106  */
 107 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 108                                                struct ctdb_db_context *ctdb_db,
 109                                                trbt_tree_t *tree,
 110                                                const struct ctdb_ltdb_header *hdr,
 111                                                TDB_DATA key)
 112 {
 113         struct delete_record_data *dd;
 114         uint32_t hash;
 115
 116         dd = talloc_zero(tree, struct delete_record_data);
 117         if (dd == NULL) {
 118                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 119                 return -1;
 120         }
 121
 122         dd->ctdb      = ctdb;
 123         dd->ctdb_db   = ctdb_db;
 124         dd->key.dsize = key.dsize;
 125         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 126         if (dd->key.dptr == NULL) {
 127                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 128                 return -1;
 129         }
 130
 131         dd->hdr = *hdr;
 132
 133         hash = (uint32_t)tdb_jenkins_hash(&key);
 134
 135         trbt_insert32(tree, hash, dd);
 136
 137         return 0;
 138 }
 139
 140 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
 141                                      struct ctdb_ltdb_header *hdr)
 142 {
 143         struct ctdb_context *ctdb = vdata->ctdb;
 144         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 145         uint32_t hash;
 146         int ret;
 147
 148         hash = (uint32_t)tdb_jenkins_hash(&key);
 149
 150         if (trbt_lookup32(vdata->delete_tree, hash)) {
 151                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 152                 return 0;
 153         }
 154
 155         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 156                                                   vdata->delete_tree,
 157                                                   hdr, key);
 158         if (ret != 0) {
 159                 return -1;
 160         }
 161
 162         vdata->delete_count++;
 163
 164         return 0;
 165 }
 166
 167 /**
 168  * Add a record to the list of records to be sent
 169  * to their lmaster with VACUUM_FETCH.
 170  */
 171 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 172                                            TDB_DATA key)
 173 {
 174         struct ctdb_context *ctdb = vdata->ctdb;
 175         struct ctdb_rec_data *rec;
 176         uint32_t lmaster;
 177         size_t old_size;
 178
 179         lmaster = ctdb_lmaster(ctdb, &key);
 180
 181         rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
 182         if (rec == NULL) {
 183                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 184                 vdata->traverse_error = true;
 185                 return -1;
 186         }
 187
 188         old_size = talloc_get_size(vdata->list[lmaster]);
 189         vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
 190                                                    old_size + rec->length);
 191         if (vdata->list[lmaster] == NULL) {
 192                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 193                 vdata->traverse_error = true;
 194                 return -1;
 195         }
 196
 197         vdata->list[lmaster]->count++;
 198         memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
 199         talloc_free(rec);
 200
 201         vdata->total++;
 202
 203         return 0;
 204 }
 205
 206
 207 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 208                                                           struct timeval t, void *private_data);
 209
 210
 211 /*
 212  * traverse function for gathering the records that can be deleted
 213  */
 214 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 215 {
 216         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 217         struct ctdb_context *ctdb = vdata->ctdb;
 218         uint32_t lmaster;
 219         struct ctdb_ltdb_header *hdr;
 220         int res = 0;
 221
 222         vdata->full_total++;
 223
 224         lmaster = ctdb_lmaster(ctdb, &key);
 225         if (lmaster >= ctdb->num_nodes) {
 226                 vdata->full_error++;
 227                 DEBUG(DEBUG_CRIT, (__location__
 228                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 229                                    " with hash[%u]!\n",
 230                                    (unsigned)lmaster,
 231                                    (unsigned)ctdb->num_nodes,
 232                                    (unsigned)ctdb_hash(&key)));
 233                 return -1;
 234         }
 235
 236         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 237                 /* its not a deleted record */
 238                 vdata->full_skipped++;
 239                 return 0;
 240         }
 241
 242         hdr = (struct ctdb_ltdb_header *)data.dptr;
 243
 244         if (hdr->dmaster != ctdb->pnn) {
 245                 vdata->full_skipped++;
 246                 return 0;
 247         }
 248
 249         if (lmaster == ctdb->pnn) {
 250                 /*
 251                  * We are both lmaster and dmaster, and the record * is empty.
 252                  * So we should be able to delete it.
 253                  */
 254                 res = add_record_to_delete_tree(vdata, key, hdr);
 255                 if (res != 0) {
 256                         vdata->full_error++;
 257                 } else {
 258                         vdata->full_added_to_delete_tree++;
 259                 }
 260         } else {
 261                 /*
 262                  * We are not lmaster.
 263                  * Add the record to the blob ready to send to the nodes.
 264                  */
 265                 res = add_record_to_vacuum_fetch_list(vdata, key);
 266                 if (res != 0) {
 267                         vdata->full_error++;
 268                 } else {
 269                         vdata->full_added_to_vacuum_fetch_list++;
 270                 }
 271         }
 272
 273         return res;
 274 }
 275
 276 /*
 277  * traverse the tree of records to delete and marshall them into
 278  * a blob
 279  */
 280 static void delete_traverse(void *param, void *data)
 281 {
 282         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 283         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 284         struct ctdb_rec_data *rec;
 285         size_t old_size;
 286
 287         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 288         if (rec == NULL) {
 289                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 290                 return;
 291         }
 292
 293         old_size = talloc_get_size(recs->records);
 294         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 295         if (recs->records == NULL) {
 296                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 297                 return;
 298         }
 299         recs->records->count++;
 300         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 301 }
 302
 303 /**
 304  * traverse function for the traversal of the delete_queue,
 305  * the fast-path vacuuming list.
 306  *
 307  *  - If the record has been migrated off the node
 308  *    or has been revived (filled with data) on the node,
 309  *    then skip the record.
 310  *
 311  *  - If the current node is the record's lmaster and it is
 312  *    a record that has never been migrated with data, then
 313  *    delete the record from the local tdb.
 314  *
 315  *  - If the current node is the record's lmaster and it has
 316  *    been migrated with data, then schedule it for the normal
 317  *    vacuuming procedure (i.e. add it to the delete_list).
 318  *
 319  *  - If the current node is NOT the record's lmaster then
 320  *    add it to the list of records that are to be sent to
 321  *    the lmaster with the VACUUM_FETCH message.
 322  */
 323 static void delete_queue_traverse(void *param, void *data)
 324 {
 325         struct delete_record_data *dd =
 326                 talloc_get_type(data, struct delete_record_data);
 327         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 328         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 329         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 330         int res;
 331         struct ctdb_ltdb_header *header;
 332         TDB_DATA tdb_data;
 333         uint32_t lmaster;
 334
 335         vdata->fast_total++;
 336
 337         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 338         if (res != 0) {
 339                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 340                 vdata->fast_error++;
 341                 return;
 342         }
 343
 344         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 345         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 346                 /* Does not exist or not a ctdb record. Skip. */
 347                 goto skipped;
 348         }
 349
 350         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 351                 /* The record has been recycled (filled with data). Skip. */
 352                 goto skipped;
 353         }
 354
 355         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 356
 357         if (header->dmaster != ctdb->pnn) {
 358                 /* The record has been migrated off the node. Skip. */
 359                 goto skipped;
 360         }
 361
 362
 363         if (header->rsn != dd->hdr.rsn) {
 364                 /*
 365                  * The record has been migrated off the node and back again.
 366                  * But not requeued for deletion. Skip it.
 367                  */
 368                 goto skipped;
 369         }
 370
 371         /*
 372          * We are dmaster, and the record has no data, and it has
 373          * not been migrated after it has been queued for deletion.
 374          *
 375          * At this stage, the record could still have been revived locally
 376          * and last been written with empty data. This can only be
 377          * fixed with the addition of an active or delete flag. (TODO)
 378          */
 379
 380         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 381
 382         if (lmaster != ctdb->pnn) {
 383                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 384
 385                 if (res != 0) {
 386                         DEBUG(DEBUG_ERR,
 387                               (__location__ " Error adding record to list "
 388                                "of records to send to lmaster.\n"));
 389                         vdata->fast_error++;
 390                 } else {
 391                         vdata->fast_added_to_vacuum_fetch_list++;
 392                 }
 393                 goto done;
 394         }
 395
 396         /* use header->flags or dd->hdr.flags ?? */
 397         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 398                 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
 399
 400                 if (res != 0) {
 401                         DEBUG(DEBUG_ERR,
 402                               (__location__ " Error adding record to list "
 403                                "of records for deletion on lmaster.\n"));
 404                         vdata->fast_error++;
 405                 } else {
 406                         vdata->fast_added_to_delete_tree++;
 407                 }
 408         } else {
 409                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 410
 411                 if (res != 0) {
 412                         DEBUG(DEBUG_ERR,
 413                               (__location__ " Error deleting record from local "
 414                                "data base.\n"));
 415                         vdata->fast_error++;
 416                 } else {
 417                         vdata->fast_deleted++;
 418                 }
 419         }
 420
 421         goto done;
 422
 423 skipped:
 424         vdata->fast_skipped++;
 425
 426 done:
 427         if (tdb_data.dptr != NULL) {
 428                 free(tdb_data.dptr);
 429         }
 430         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 431
 432         return;
 433 }
 434
 435 /*
 436  * read-only traverse the database in order to find
 437  * records that can be deleted and try to delete these
 438  * records on the other nodes
 439  * this executes in the child context
 440  */
 441 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 442                           struct vacuum_data *vdata,
 443                           bool full_vacuum_run)
 444 {
 445         struct ctdb_context *ctdb = ctdb_db->ctdb;
 446         const char *name = ctdb_db->db_name;
 447         int ret, i, pnn;
 448
 449         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 450                            "%s db_id[0x%08x]\n",
 451                            full_vacuum_run ? "full" : "fast",
 452                            ctdb_db->db_name, ctdb_db->db_id));
 453
 454         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 455         if (ret != 0) {
 456                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 457                 return ret;
 458         }
 459
 460         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 461         if (pnn == -1) {
 462                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 463                 return -1;
 464         }
 465
 466         ctdb->pnn = pnn;
 467
 468         vdata->fast_added_to_delete_tree = 0;
 469         vdata->fast_added_to_vacuum_fetch_list = 0;
 470         vdata->fast_deleted = 0;
 471         vdata->fast_skipped = 0;
 472         vdata->fast_error = 0;
 473         vdata->fast_total = 0;
 474         vdata->full_added_to_delete_tree = 0;
 475         vdata->full_added_to_vacuum_fetch_list = 0;
 476         vdata->full_skipped = 0;
 477         vdata->full_error = 0;
 478         vdata->full_total = 0;
 479
 480         /* the list needs to be of length num_nodes */
 481         vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
 482         if (vdata->list == NULL) {
 483                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 484                 return -1;
 485         }
 486         for (i = 0; i < ctdb->num_nodes; i++) {
 487                 vdata->list[i] = (struct ctdb_marshall_buffer *)
 488                         talloc_zero_size(vdata->list,
 489                                                          offsetof(struct ctdb_marshall_buffer, data));
 490                 if (vdata->list[i] == NULL) {
 491                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 492                         return -1;
 493                 }
 494                 vdata->list[i]->db_id = ctdb_db->db_id;
 495         }
 496
 497         /*
 498          * Traverse the delete_queue.
 499          * This builds the same lists as the db traverse.
 500          */
 501         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 502
 503         if (vdata->fast_total > 0) {
 504                 DEBUG(DEBUG_INFO,
 505                       (__location__
 506                        " fast vacuuming delete_queue traverse statistics: "
 507                        "db[%s] "
 508                        "total[%u] "
 509                        "del[%u] "
 510                        "skp[%u] "
 511                        "err[%u] "
 512                        "adt[%u] "
 513                        "avf[%u]\n",
 514                        ctdb_db->db_name,
 515                        (unsigned)vdata->fast_total,
 516                        (unsigned)vdata->fast_deleted,
 517                        (unsigned)vdata->fast_skipped,
 518                        (unsigned)vdata->fast_error,
 519                        (unsigned)vdata->fast_added_to_delete_tree,
 520                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 521         }
 522
 523         /*
 524          * read-only traverse of the database, looking for records that
 525          * might be able to be vacuumed.
 526          *
 527          * This is not done each time but only every tunable
 528          * VacuumFastPathCount times.
 529          */
 530         if (full_vacuum_run) {
 531                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 532                 if (ret == -1 || vdata->traverse_error) {
 533                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 534                         return -1;
 535                 }
 536                 if (vdata->full_total > 0) {
 537                         DEBUG(DEBUG_INFO,
 538                               (__location__
 539                                " full vacuuming db traverse statistics: "
 540                                "db[%s] "
 541                                "total[%u] "
 542                                "skp[%u] "
 543                                "err[%u] "
 544                                "adt[%u] "
 545                                "avf[%u]\n",
 546                                ctdb_db->db_name,
 547                                (unsigned)vdata->full_total,
 548                                (unsigned)vdata->full_skipped,
 549                                (unsigned)vdata->full_error,
 550                                (unsigned)vdata->full_added_to_delete_tree,
 551                                (unsigned)vdata->full_added_to_vacuum_fetch_list));
 552                 }
 553         }
 554
 555         /*
 556          * For records where we are not the lmaster,
 557          * tell the lmaster to fetch the record.
 558          */
 559         for (i = 0; i < ctdb->num_nodes; i++) {
 560                 TDB_DATA data;
 561
 562                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 563                         continue;
 564                 }
 565
 566                 if (vdata->list[i]->count == 0) {
 567                         continue;
 568                 }
 569
 570                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 571                                    vdata->list[i]->count, ctdb->nodes[i]->pnn,
 572                                    name));
 573
 574                 data.dsize = talloc_get_size(vdata->list[i]);
 575                 data.dptr  = (void *)vdata->list[i];
 576                 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 577                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 578                                           "fetch message to %u\n",
 579                                           ctdb->nodes[i]->pnn));
 580                         return -1;
 581                 }
 582         }
 583
 584         /* Process all records we can delete (if any) */
 585         if (vdata->delete_count > 0) {
 586                 struct delete_records_list *recs;
 587                 TDB_DATA indata, outdata;
 588                 int32_t res;
 589                 struct ctdb_node_map *nodemap;
 590                 uint32_t *active_nodes;
 591                 int num_active_nodes;
 592
 593                 recs = talloc_zero(vdata, struct delete_records_list);
 594                 if (recs == NULL) {
 595                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 596                         return -1;
 597                 }
 598                 recs->records = (struct ctdb_marshall_buffer *)
 599                         talloc_zero_size(vdata,
 600                                     offsetof(struct ctdb_marshall_buffer, data));
 601                 if (recs->records == NULL) {
 602                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 603                         return -1;
 604                 }
 605                 recs->records->db_id = ctdb_db->db_id;
 606
 607                 /*
 608                  * traverse the tree of all records we want to delete and
 609                  * create a blob we can send to the other nodes.
 610                  */
 611                 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
 612
 613                 indata.dsize = talloc_get_size(recs->records);
 614                 indata.dptr  = (void *)recs->records;
 615
 616                 /*
 617                  * now tell all the active nodes to delete all these records
 618                  * (if possible)
 619                  */
 620
 621                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 622                                            CTDB_CURRENT_NODE,
 623                                            recs, /* talloc context */
 624                                            &nodemap);
 625                 if (ret != 0) {
 626                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 627                         return -1;
 628                 }
 629
 630                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 631                                                     nodemap, /* talloc context */
 632                                                     false /* include self */);
 633                 /* yuck! ;-) */
 634                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 635
 636                 for (i = 0; i < num_active_nodes; i++) {
 637                         struct ctdb_marshall_buffer *records;
 638                         struct ctdb_rec_data *rec;
 639
 640                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 641                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 642                                         indata, recs, &outdata, &res,
 643                                         NULL, NULL);
 644                         if (ret != 0 || res != 0) {
 645                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 646                                                   "node %u: ret[%d] res[%d]\n",
 647                                                   active_nodes[i], ret, res));
 648                                 return -1;
 649                         }
 650
 651                         /*
 652                          * outdata countains the list of records coming back
 653                          * from the node which the node could not delete
 654                          */
 655                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 656                         rec = (struct ctdb_rec_data *)&records->data[0];
 657                         while (records->count-- > 1) {
 658                                 TDB_DATA reckey, recdata;
 659                                 struct ctdb_ltdb_header *rechdr;
 660
 661                                 reckey.dptr = &rec->data[0];
 662                                 reckey.dsize = rec->keylen;
 663                                 recdata.dptr = &rec->data[reckey.dsize];
 664                                 recdata.dsize = rec->datalen;
 665
 666                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 667                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 668                                         return -1;
 669                                 }
 670                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 671                                 recdata.dptr += sizeof(*rechdr);
 672                                 recdata.dsize -= sizeof(*rechdr);
 673
 674                                 /*
 675                                  * that other node couldnt delete the record
 676                                  * so we should delete it and thereby remove it from the tree
 677                                  */
 678                                 talloc_free(trbt_lookup32(vdata->delete_tree,
 679                                                           (uint32_t)tdb_jenkins_hash(&reckey)));
 680
 681                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 682                         }
 683                 }
 684
 685                 /* free nodemap and active_nodes */
 686                 talloc_free(nodemap);
 687
 688                 /*
 689                  * The only records remaining in the tree would be those
 690                  * records where all other nodes could successfully
 691                  * delete them, so we can safely delete them on the
 692                  * lmaster as well. Deletion implictely happens while
 693                  * we repack the database. The repack algorithm revisits
 694                  * the tree in order to find the records that don't need
 695                  * to be copied / repacked.
 696                  */
 697         }
 698
 699         /* this ensures we run our event queue */
 700         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 701
 702         return 0;
 703 }
 704
 705
 706 /*
 707  * traverse function for repacking
 708  */
 709 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 710 {
 711         struct vacuum_data *vdata = (struct vacuum_data *)private;
 712
 713         if (vdata->vacuum) {
 714                 uint32_t hash = (uint32_t)tdb_jenkins_hash(&key);
 715                 struct delete_record_data *kd;
 716                 /*
 717                  * check if we can ignore this record because it's in the delete_tree
 718                  */
 719                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
 720                 /*
 721                  * there might be hash collisions so we have to compare the keys here to be sure
 722                  */
 723                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 724                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 725                         /*
 726                          * we have to check if the record hasn't changed in the meantime in order to
 727                          * savely remove it from the database
 728                          */
 729                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 730                                 hdr->dmaster == kd->ctdb->pnn &&
 731                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 732                                 kd->hdr.rsn == hdr->rsn) {
 733                                 vdata->vacuumed++;
 734                                 return 0;
 735                         }
 736                 }
 737         }
 738         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 739                 vdata->traverse_error = true;
 740                 return -1;
 741         }
 742         vdata->copied++;
 743         return 0;
 744 }
 745
 746 /*
 747  * repack a tdb
 748  */
 749 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 750 {
 751         struct tdb_context *tmp_db;
 752
 753         if (tdb_transaction_start(tdb) != 0) {
 754                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 755                 return -1;
 756         }
 757
 758         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 759                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 760                           O_RDWR|O_CREAT, 0);
 761         if (tmp_db == NULL) {
 762                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 763                 tdb_transaction_cancel(tdb);
 764                 return -1;
 765         }
 766
 767         vdata->traverse_error = false;
 768         vdata->dest_db = tmp_db;
 769         vdata->vacuum = true;
 770         vdata->vacuumed = 0;
 771         vdata->copied = 0;
 772
 773         /*
 774          * repack and vacuum on-the-fly by not writing the records that are
 775          * no longer needed
 776          */
 777         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 778                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 779                 tdb_transaction_cancel(tdb);
 780                 tdb_close(tmp_db);
 781                 return -1;
 782         }
 783
 784         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 785
 786         if (vdata->traverse_error) {
 787                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 788                 tdb_transaction_cancel(tdb);
 789                 tdb_close(tmp_db);
 790                 return -1;
 791         }
 792
 793         if (tdb_wipe_all(tdb) != 0) {
 794                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 795                 tdb_transaction_cancel(tdb);
 796                 tdb_close(tmp_db);
 797                 return -1;
 798         }
 799
 800         vdata->traverse_error = false;
 801         vdata->dest_db = tdb;
 802         vdata->vacuum = false;
 803         vdata->copied = 0;
 804
 805         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 806                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 807                 tdb_transaction_cancel(tdb);
 808                 tdb_close(tmp_db);
 809                 return -1;
 810         }
 811
 812         if (vdata->traverse_error) {
 813                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 814                 tdb_transaction_cancel(tdb);
 815                 tdb_close(tmp_db);
 816                 return -1;
 817         }
 818
 819         tdb_close(tmp_db);
 820
 821
 822         if (tdb_transaction_commit(tdb) != 0) {
 823                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 824                 return -1;
 825         }
 826         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 827
 828         return 0;
 829 }
 830
 831 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
 832 {
 833         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 834         TDB_CONTEXT *tune_tdb;
 835         TDB_DATA key, value;
 836         struct vacuum_tuning_data tdata;
 837         struct vacuum_tuning_data *tptr;
 838         char *vac_dbname;
 839         int flags;
 840
 841         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
 842                                      ctdb_db->ctdb->db_directory_state,
 843                                      TUNINGDBNAME, ctdb_db->ctdb->pnn);
 844         if (vac_dbname == NULL) {
 845                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
 846                 talloc_free(tmp_ctx);
 847                 return -1;
 848         }
 849
 850         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
 851         flags |= TDB_DISALLOW_NESTING;
 852         tune_tdb = tdb_open(vac_dbname, 0,
 853                             flags,
 854                             O_RDWR|O_CREAT, 0600);
 855         if (tune_tdb == NULL) {
 856                 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
 857                 talloc_free(tmp_ctx);
 858                 return -1;
 859         }
 860
 861         if (tdb_transaction_start(tune_tdb) != 0) {
 862                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 863                 tdb_close(tune_tdb);
 864                 return -1;
 865         }
 866         key.dptr = discard_const(ctdb_db->db_name);
 867         key.dsize = strlen(ctdb_db->db_name);
 868         value = tdb_fetch(tune_tdb, key);
 869
 870         if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
 871                 tptr = (struct vacuum_tuning_data *)value.dptr;
 872                 tdata = *tptr;
 873
 874                 /*
 875                  * re-calc new vacuum interval:
 876                  * in case no limit was reached we continously increase the interval
 877                  * until vacuum_max_interval is reached
 878                  * in case a limit was reached we divide the current interval by 2
 879                  * unless vacuum_min_interval is reached
 880                  */
 881                 if (freelist < vdata->repack_limit &&
 882                     vdata->delete_count < vdata->vacuum_limit) {
 883                         if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
 884                                 tdata.new_interval = tdata.last_interval * 110 / 100;
 885                                 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
 886                                         tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 887                         }
 888                 } else {
 889                         tdata.new_interval = tdata.last_interval / 2;
 890                         if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
 891                                 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
 892                                 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
 893                         }
 894                         DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
 895                                          tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 896                 }
 897                 tdata.last_interval = tdata.new_interval;
 898         } else {
 899                 DEBUG(DEBUG_ERR,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
 900                 tdata.last_num_repack = freelist;
 901                 tdata.last_num_empty = vdata->delete_count;
 902                 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 903         }
 904
 905         if (value.dptr != NULL) {
 906                 free(value.dptr);
 907         }
 908
 909         tdata.last_start = vdata->start;
 910         tdata.last_duration = timeval_elapsed(&vdata->start);
 911
 912         value.dptr = (unsigned char *)&tdata;
 913         value.dsize = sizeof(tdata);
 914
 915         if (tdb_store(tune_tdb, key, value, 0) != 0) {
 916                 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
 917                 tdb_transaction_cancel(tune_tdb);
 918                 tdb_close(tune_tdb);
 919                 talloc_free(tmp_ctx);
 920                 return -1;
 921         }
 922         tdb_transaction_commit(tune_tdb);
 923         tdb_close(tune_tdb);
 924         talloc_free(tmp_ctx);
 925
 926         return 0;
 927 }
 928
 929 /*
 930  * repack and vaccum a db
 931  * called from the child context
 932  */
 933 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 934                                      TALLOC_CTX *mem_ctx,
 935                                      bool full_vacuum_run)
 936 {
 937         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
 938         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
 939         const char *name = ctdb_db->db_name;
 940         int size;
 941         struct vacuum_data *vdata;
 942
 943         size = tdb_freelist_size(ctdb_db->ltdb->tdb);
 944         if (size == -1) {
 945                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
 946                 return -1;
 947         }
 948
 949         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 950         if (vdata == NULL) {
 951                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 952                 return -1;
 953         }
 954
 955         vdata->ctdb = ctdb_db->ctdb;
 956         vdata->vacuum_limit = vacuum_limit;
 957         vdata->repack_limit = repack_limit;
 958         vdata->delete_tree = trbt_create(vdata, 0);
 959         if (vdata->delete_tree == NULL) {
 960                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 961                 talloc_free(vdata);
 962                 return -1;
 963         }
 964
 965         vdata->start = timeval_current();
 966
 967         /*
 968          * gather all records that can be deleted in vdata
 969          */
 970         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
 971                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
 972         }
 973
 974         /*
 975          * decide if a repack is necessary
 976          */
 977         if (size < repack_limit && vdata->delete_count < vacuum_limit) {
 978                 update_tuning_db(ctdb_db, vdata, size);
 979                 talloc_free(vdata);
 980                 return 0;
 981         }
 982
 983         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
 984                         name, size, vdata->delete_count));
 985
 986         /*
 987          * repack and implicitely get rid of the records we can delete
 988          */
 989         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
 990                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
 991                 update_tuning_db(ctdb_db, vdata, size);
 992                 talloc_free(vdata);
 993                 return -1;
 994         }
 995         update_tuning_db(ctdb_db, vdata, size);
 996         talloc_free(vdata);
 997
 998         return 0;
 999 }
1000
1001 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1002 {
1003         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1004         TDB_CONTEXT *tdb;
1005         TDB_DATA key, value;
1006         char *vac_dbname;
1007         uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
1008         struct ctdb_context *ctdb = ctdb_db->ctdb;
1009         int flags;
1010
1011         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
1012         if (vac_dbname == NULL) {
1013                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
1014                 talloc_free(tmp_ctx);
1015                 return interval;
1016         }
1017
1018         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
1019         flags |= TDB_DISALLOW_NESTING;
1020         tdb = tdb_open(vac_dbname, 0,
1021                        flags,
1022                        O_RDWR|O_CREAT, 0600);
1023         if (!tdb) {
1024                 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval\n", vac_dbname));
1025                 talloc_free(tmp_ctx);
1026                 return interval;
1027         }
1028
1029         key.dptr = discard_const(ctdb_db->db_name);
1030         key.dsize = strlen(ctdb_db->db_name);
1031
1032         value = tdb_fetch(tdb, key);
1033
1034         if (value.dptr != NULL) {
1035                 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
1036                         struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
1037
1038                         interval = tptr->new_interval;
1039
1040                         if (interval < ctdb->tunable.vacuum_min_interval) {
1041                                 interval = ctdb->tunable.vacuum_min_interval;
1042                         }
1043                         if (interval > ctdb->tunable.vacuum_max_interval) {
1044                                 interval = ctdb->tunable.vacuum_max_interval;
1045                         }
1046                 }
1047                 free(value.dptr);
1048         }
1049         tdb_close(tdb);
1050
1051         talloc_free(tmp_ctx);
1052
1053         return interval;
1054 }
1055
1056 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1057 {
1058         double l = timeval_elapsed(&child_ctx->start_time);
1059         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1060         struct ctdb_context *ctdb = ctdb_db->ctdb;
1061
1062         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1063
1064         if (child_ctx->child_pid != -1) {
1065                 kill(child_ctx->child_pid, SIGKILL);
1066         } else {
1067                 /* Bump the number of successful fast-path runs. */
1068                 child_ctx->vacuum_handle->fast_path_count++;
1069         }
1070
1071         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1072                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1073                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1074
1075         return 0;
1076 }
1077
1078 /*
1079  * this event is generated when a vacuum child process times out
1080  */
1081 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1082                                          struct timeval t, void *private_data)
1083 {
1084         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1085
1086         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1087
1088         child_ctx->status = VACUUM_TIMEOUT;
1089
1090         talloc_free(child_ctx);
1091 }
1092
1093
1094 /*
1095  * this event is generated when a vacuum child process has completed
1096  */
1097 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1098                              uint16_t flags, void *private_data)
1099 {
1100         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1101         char c = 0;
1102         int ret;
1103
1104         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1105         child_ctx->child_pid = -1;
1106
1107         ret = read(child_ctx->fd[0], &c, 1);
1108         if (ret != 1 || c != 0) {
1109                 child_ctx->status = VACUUM_ERROR;
1110                 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));
1111         } else {
1112                 child_ctx->status = VACUUM_OK;
1113         }
1114
1115         talloc_free(child_ctx);
1116 }
1117
1118 /*
1119  * this event is called every time we need to start a new vacuum process
1120  */
1121 static void
1122 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1123                                struct timeval t, void *private_data)
1124 {
1125         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1126         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1127         struct ctdb_context *ctdb = ctdb_db->ctdb;
1128         struct ctdb_vacuum_child_context *child_ctx;
1129         int ret;
1130
1131         /* we dont vacuum if we are in recovery mode */
1132         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE) {
1133                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1134                 return;
1135         }
1136
1137         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1138         if (child_ctx == NULL) {
1139                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1140                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1141         }
1142
1143
1144         ret = pipe(child_ctx->fd);
1145         if (ret != 0) {
1146                 talloc_free(child_ctx);
1147                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1148                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1149                 return;
1150         }
1151
1152         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1153                 vacuum_handle->fast_path_count = 0;
1154         }
1155
1156         child_ctx->child_pid = fork();
1157         if (child_ctx->child_pid == (pid_t)-1) {
1158                 close(child_ctx->fd[0]);
1159                 close(child_ctx->fd[1]);
1160                 talloc_free(child_ctx);
1161                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1162                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1163                 return;
1164         }
1165
1166
1167         if (child_ctx->child_pid == 0) {
1168                 char cc = 0;
1169                 bool full_vacuum_run = false;
1170                 close(child_ctx->fd[0]);
1171
1172                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1173
1174                 if (switch_from_server_to_client(ctdb) != 0) {
1175                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1176                         _exit(1);
1177                 }
1178
1179                 /*
1180                  * repack the db
1181                  */
1182                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1183                     (vacuum_handle->fast_path_count == 0))
1184                 {
1185                         full_vacuum_run = true;
1186                 }
1187                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1188                                                full_vacuum_run);
1189
1190                 write(child_ctx->fd[1], &cc, 1);
1191                 _exit(0);
1192         }
1193
1194         set_close_on_exec(child_ctx->fd[0]);
1195         close(child_ctx->fd[1]);
1196
1197         child_ctx->status = VACUUM_RUNNING;
1198         child_ctx->start_time = timeval_current();
1199
1200         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1201
1202         /*
1203          * Clear the fastpath vacuuming list in the parent.
1204          */
1205         talloc_free(ctdb_db->delete_queue);
1206         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1207         if (ctdb_db->delete_queue == NULL) {
1208                 /* fatal here? ... */
1209                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1210                                  "in parent context. Shutting down\n");
1211         }
1212
1213         event_add_timed(ctdb->ev, child_ctx,
1214                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1215                 vacuum_child_timeout, child_ctx);
1216
1217         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1218
1219         event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1220                 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1221                 vacuum_child_handler,
1222                 child_ctx);
1223
1224         vacuum_handle->child_ctx = child_ctx;
1225         child_ctx->vacuum_handle = vacuum_handle;
1226 }
1227
1228
1229 /* this function initializes the vacuuming context for a database
1230  * starts the vacuuming events
1231  */
1232 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1233 {
1234         if (ctdb_db->persistent != 0) {
1235                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1236                 return 0;
1237         }
1238
1239         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1240         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1241
1242         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1243         ctdb_db->vacuum_handle->fast_path_count = 0;
1244
1245         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1246                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1247                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1248
1249         return 0;
1250 }
1251
1252 /**
1253  * Insert a record into the ctdb_db context's delete queue,
1254  * handling hash collisions.
1255  */
1256 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1257                                            const struct ctdb_ltdb_header *hdr,
1258                                            TDB_DATA key)
1259 {
1260         struct delete_record_data *kd;
1261         uint32_t hash;
1262         int ret;
1263
1264         hash = (uint32_t)tdb_jenkins_hash(&key);
1265
1266         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1267                            "db_id[0x%08x] "
1268                            "key_hash[0x%08x] "
1269                            "lmaster[%u] "
1270                            "migrated_with_data[%s]\n",
1271                             ctdb_db->db_name, ctdb_db->db_id,
1272                             hash,
1273                             ctdb_lmaster(ctdb_db->ctdb, &key),
1274                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1275
1276         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1277         if (kd != NULL) {
1278                 if ((kd->key.dsize != key.dsize) ||
1279                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1280                 {
1281                         DEBUG(DEBUG_INFO,
1282                               ("schedule for deletion: Hash collision (0x%08x)."
1283                                " Skipping the record.\n", hash));
1284                         return 0;
1285                 } else {
1286                         DEBUG(DEBUG_DEBUG,
1287                               ("schedule for deletion: Overwriting entry for "
1288                                "key with hash 0x%08x.\n", hash));
1289                 }
1290         }
1291
1292         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1293                                                   ctdb_db->delete_queue,
1294                                                   hdr, key);
1295         if (ret != 0) {
1296                 return -1;
1297         }
1298
1299         return 0;
1300 }
1301
1302 /**
1303  * Schedule a record for deletetion.
1304  * Called from the parent context.
1305  */
1306 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1307                                            TDB_DATA indata)
1308 {
1309         struct ctdb_control_schedule_for_deletion *dd;
1310         struct ctdb_db_context *ctdb_db;
1311         int ret;
1312         TDB_DATA key;
1313
1314         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1315
1316         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1317         if (ctdb_db == NULL) {
1318                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1319                                   dd->db_id));
1320                 return -1;
1321         }
1322
1323         key.dsize = dd->keylen;
1324         key.dptr = dd->key;
1325
1326         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1327
1328         return ret;
1329 }
1330
1331 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1332                                          const struct ctdb_ltdb_header *hdr,
1333                                          TDB_DATA key)
1334 {
1335         int ret;
1336         struct ctdb_control_schedule_for_deletion *dd;
1337         TDB_DATA indata;
1338         int32_t status;
1339
1340         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1341                 /* main daemon - directly queue */
1342                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1343
1344                 return ret;
1345         }
1346
1347         /* child process: send the main daemon a control */
1348
1349         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1350         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1351         if (indata.dptr == NULL) {
1352                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1353                 return -1;
1354         }
1355         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1356         dd->db_id = ctdb_db->db_id;
1357         dd->hdr = *hdr;
1358         dd->keylen = key.dsize;
1359         memcpy(dd->key, key.dptr, key.dsize);
1360
1361         ret = ctdb_control(ctdb_db->ctdb,
1362                            CTDB_CURRENT_NODE,
1363                            ctdb_db->db_id,
1364                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1365                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1366                            indata,
1367                            NULL, /* mem_ctx */
1368                            NULL, /* outdata */
1369                            &status,
1370                            NULL, /* timeout : NULL == wait forever */
1371                            NULL); /* error message */
1372
1373         talloc_free(indata.dptr);
1374
1375         if (ret != 0 || status != 0) {
1376                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1377                                   "SCHEDULE_FOR_DELETION "
1378                                   "control.\n"));
1379                 if (status != 0) {
1380                         ret = -1;
1381                 }
1382         }
1383
1384         return ret;
1385 }