server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5    Copyright (C) Michael Adam 2010-2011
   6    Copyright (C) Stefan Metzmacher 2010-2011
   7
   8    This program is free software; you can redistribute it and/or modify
   9    it under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3 of the License, or
  11    (at your option) any later version.
  12
  13    This program is distributed in the hope that it will be useful,
  14    but WITHOUT ANY WARRANTY; without even the implied warranty of
  15    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16    GNU General Public License for more details.
  17
  18    You should have received a copy of the GNU General Public License
  19    along with this program; if not, see <http://www.gnu.org/licenses/>.
  20 */
  21
  22 #include "includes.h"
  23 #include "lib/tevent/tevent.h"
  24 #include "lib/tdb/include/tdb.h"
  25 #include "system/network.h"
  26 #include "system/filesys.h"
  27 #include "system/dir.h"
  28 #include "../include/ctdb_private.h"
  29 #include "db_wrap.h"
  30 #include "lib/util/dlinklist.h"
  31 #include "lib/tevent/tevent.h"
  32 #include "../include/ctdb_private.h"
  33 #include "../common/rb_tree.h"
  34
  35 #define TIMELIMIT() timeval_current_ofs(10, 0)
  36
  37 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  38
  39 struct ctdb_vacuum_child_context {
  40         struct ctdb_vacuum_child_context *next, *prev;
  41         struct ctdb_vacuum_handle *vacuum_handle;
  42         /* fd child writes status to */
  43         int fd[2];
  44         pid_t child_pid;
  45         enum vacuum_child_status status;
  46         struct timeval start_time;
  47 };
  48
  49 struct ctdb_vacuum_handle {
  50         struct ctdb_db_context *ctdb_db;
  51         struct ctdb_vacuum_child_context *child_ctx;
  52         uint32_t fast_path_count;
  53 };
  54
  55
  56 /*  a list of records to possibly delete */
  57 struct vacuum_data {
  58         uint32_t vacuum_limit;
  59         uint32_t repack_limit;
  60         struct ctdb_context *ctdb;
  61         struct ctdb_db_context *ctdb_db;
  62         struct tdb_context *dest_db;
  63         trbt_tree_t *delete_tree;
  64         uint32_t delete_count;
  65         struct ctdb_marshall_buffer **list;
  66         struct timeval start;
  67         bool traverse_error;
  68         bool vacuum;
  69         uint32_t total;
  70         uint32_t vacuumed;
  71         uint32_t copied;
  72         uint32_t fast_added_to_vacuum_fetch_list;
  73         uint32_t fast_added_to_delete_tree;
  74         uint32_t fast_deleted;
  75         uint32_t fast_skipped;
  76         uint32_t fast_error;
  77         uint32_t fast_total;
  78         uint32_t full_added_to_vacuum_fetch_list;
  79         uint32_t full_added_to_delete_tree;
  80         uint32_t full_skipped;
  81         uint32_t full_error;
  82         uint32_t full_total;
  83 };
  84
  85 /* tuning information stored for every db */
  86 struct vacuum_tuning_data {
  87         uint32_t last_num_repack;
  88         uint32_t last_num_empty;
  89         uint32_t last_interval;
  90         uint32_t new_interval;
  91         struct timeval last_start;
  92         double   last_duration;
  93 };
  94
  95 /* this structure contains the information for one record to be deleted */
  96 struct delete_record_data {
  97         struct ctdb_context *ctdb;
  98         struct ctdb_db_context *ctdb_db;
  99         struct ctdb_ltdb_header hdr;
 100         TDB_DATA key;
 101 };
 102
 103 struct delete_records_list {
 104         struct ctdb_marshall_buffer *records;
 105 };
 106
 107 /**
 108  * Store key and header in a tree, indexed by the key hash.
 109  */
 110 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 111                                                struct ctdb_db_context *ctdb_db,
 112                                                trbt_tree_t *tree,
 113                                                const struct ctdb_ltdb_header *hdr,
 114                                                TDB_DATA key)
 115 {
 116         struct delete_record_data *dd;
 117         uint32_t hash;
 118
 119         dd = talloc_zero(tree, struct delete_record_data);
 120         if (dd == NULL) {
 121                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 122                 return -1;
 123         }
 124
 125         dd->ctdb      = ctdb;
 126         dd->ctdb_db   = ctdb_db;
 127         dd->key.dsize = key.dsize;
 128         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 129         if (dd->key.dptr == NULL) {
 130                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 131                 return -1;
 132         }
 133
 134         dd->hdr = *hdr;
 135
 136         hash = ctdb_hash(&key);
 137
 138         trbt_insert32(tree, hash, dd);
 139
 140         return 0;
 141 }
 142
 143 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
 144                                      struct ctdb_ltdb_header *hdr)
 145 {
 146         struct ctdb_context *ctdb = vdata->ctdb;
 147         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 148         uint32_t hash;
 149         int ret;
 150
 151         hash = ctdb_hash(&key);
 152
 153         if (trbt_lookup32(vdata->delete_tree, hash)) {
 154                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 155                 return 0;
 156         }
 157
 158         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 159                                                   vdata->delete_tree,
 160                                                   hdr, key);
 161         if (ret != 0) {
 162                 return -1;
 163         }
 164
 165         vdata->delete_count++;
 166
 167         return 0;
 168 }
 169
 170 /**
 171  * Add a record to the list of records to be sent
 172  * to their lmaster with VACUUM_FETCH.
 173  */
 174 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 175                                            TDB_DATA key)
 176 {
 177         struct ctdb_context *ctdb = vdata->ctdb;
 178         struct ctdb_rec_data *rec;
 179         uint32_t lmaster;
 180         size_t old_size;
 181
 182         lmaster = ctdb_lmaster(ctdb, &key);
 183
 184         rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
 185         if (rec == NULL) {
 186                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 187                 vdata->traverse_error = true;
 188                 return -1;
 189         }
 190
 191         old_size = talloc_get_size(vdata->list[lmaster]);
 192         vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
 193                                                    old_size + rec->length);
 194         if (vdata->list[lmaster] == NULL) {
 195                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 196                 vdata->traverse_error = true;
 197                 return -1;
 198         }
 199
 200         vdata->list[lmaster]->count++;
 201         memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
 202         talloc_free(rec);
 203
 204         vdata->total++;
 205
 206         return 0;
 207 }
 208
 209
 210 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 211                               struct timeval t, void *private_data);
 212
 213
 214 /*
 215  * traverse function for gathering the records that can be deleted
 216  */
 217 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 218 {
 219         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 220         struct ctdb_context *ctdb = vdata->ctdb;
 221         uint32_t lmaster;
 222         struct ctdb_ltdb_header *hdr;
 223         int res = 0;
 224
 225         vdata->full_total++;
 226
 227         lmaster = ctdb_lmaster(ctdb, &key);
 228         if (lmaster >= ctdb->num_nodes) {
 229                 vdata->full_error++;
 230                 DEBUG(DEBUG_CRIT, (__location__
 231                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 232                                    " with hash[%u]!\n",
 233                                    (unsigned)lmaster,
 234                                    (unsigned)ctdb->num_nodes,
 235                                    (unsigned)ctdb_hash(&key)));
 236                 return -1;
 237         }
 238
 239         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 240                 /* it is not a deleted record */
 241                 vdata->full_skipped++;
 242                 return 0;
 243         }
 244
 245         hdr = (struct ctdb_ltdb_header *)data.dptr;
 246
 247         if (hdr->dmaster != ctdb->pnn) {
 248                 vdata->full_skipped++;
 249                 return 0;
 250         }
 251
 252         if (lmaster == ctdb->pnn) {
 253                 /*
 254                  * We are both lmaster and dmaster, and the record is empty.
 255                  * So we should be able to delete it.
 256                  */
 257                 res = add_record_to_delete_tree(vdata, key, hdr);
 258                 if (res != 0) {
 259                         vdata->full_error++;
 260                 } else {
 261                         vdata->full_added_to_delete_tree++;
 262                 }
 263         } else {
 264                 /*
 265                  * We are not lmaster.
 266                  * Add the record to the blob ready to send to the nodes.
 267                  */
 268                 res = add_record_to_vacuum_fetch_list(vdata, key);
 269                 if (res != 0) {
 270                         vdata->full_error++;
 271                 } else {
 272                         vdata->full_added_to_vacuum_fetch_list++;
 273                 }
 274         }
 275
 276         return res;
 277 }
 278
 279 /*
 280  * traverse the tree of records to delete and marshall them into
 281  * a blob
 282  */
 283 static int delete_traverse(void *param, void *data)
 284 {
 285         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 286         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 287         struct ctdb_rec_data *rec;
 288         size_t old_size;
 289
 290         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 291         if (rec == NULL) {
 292                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 293                 return 0;
 294         }
 295
 296         old_size = talloc_get_size(recs->records);
 297         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 298         if (recs->records == NULL) {
 299                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 300                 return 0;
 301         }
 302         recs->records->count++;
 303         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 304         return 0;
 305 }
 306
 307 /**
 308  * traverse function for the traversal of the delete_queue,
 309  * the fast-path vacuuming list.
 310  *
 311  *  - If the record has been migrated off the node
 312  *    or has been revived (filled with data) on the node,
 313  *    then skip the record.
 314  *
 315  *  - If the current node is the record's lmaster and it is
 316  *    a record that has never been migrated with data, then
 317  *    delete the record from the local tdb.
 318  *
 319  *  - If the current node is the record's lmaster and it has
 320  *    been migrated with data, then schedule it for the normal
 321  *    vacuuming procedure (i.e. add it to the delete_list).
 322  *
 323  *  - If the current node is NOT the record's lmaster then
 324  *    add it to the list of records that are to be sent to
 325  *    the lmaster with the VACUUM_FETCH message.
 326  */
 327 static int delete_queue_traverse(void *param, void *data)
 328 {
 329         struct delete_record_data *dd =
 330                 talloc_get_type(data, struct delete_record_data);
 331         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 332         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 333         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 334         int res;
 335         struct ctdb_ltdb_header *header;
 336         TDB_DATA tdb_data;
 337         uint32_t lmaster;
 338
 339         vdata->fast_total++;
 340
 341         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 342         if (res != 0) {
 343                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 344                 vdata->fast_error++;
 345                 return 0;
 346         }
 347
 348         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 349         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 350                 /* Does not exist or not a ctdb record. Skip. */
 351                 goto skipped;
 352         }
 353
 354         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 355                 /* The record has been recycled (filled with data). Skip. */
 356                 goto skipped;
 357         }
 358
 359         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 360
 361         if (header->dmaster != ctdb->pnn) {
 362                 /* The record has been migrated off the node. Skip. */
 363                 goto skipped;
 364         }
 365
 366
 367         if (header->rsn != dd->hdr.rsn) {
 368                 /*
 369                  * The record has been migrated off the node and back again.
 370                  * But not requeued for deletion. Skip it.
 371                  */
 372                 goto skipped;
 373         }
 374
 375         /*
 376          * We are dmaster, and the record has no data, and it has
 377          * not been migrated after it has been queued for deletion.
 378          *
 379          * At this stage, the record could still have been revived locally
 380          * and last been written with empty data. This can only be
 381          * fixed with the addition of an active or delete flag. (TODO)
 382          */
 383
 384         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 385
 386         if (lmaster != ctdb->pnn) {
 387                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 388
 389                 if (res != 0) {
 390                         DEBUG(DEBUG_ERR,
 391                               (__location__ " Error adding record to list "
 392                                "of records to send to lmaster.\n"));
 393                         vdata->fast_error++;
 394                 } else {
 395                         vdata->fast_added_to_vacuum_fetch_list++;
 396                 }
 397                 goto done;
 398         }
 399
 400         /* use header->flags or dd->hdr.flags ?? */
 401         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 402                 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
 403
 404                 if (res != 0) {
 405                         DEBUG(DEBUG_ERR,
 406                               (__location__ " Error adding record to list "
 407                                "of records for deletion on lmaster.\n"));
 408                         vdata->fast_error++;
 409                 } else {
 410                         vdata->fast_added_to_delete_tree++;
 411                 }
 412         } else {
 413                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 414
 415                 if (res != 0) {
 416                         DEBUG(DEBUG_ERR,
 417                               (__location__ " Error deleting record from local "
 418                                "data base.\n"));
 419                         vdata->fast_error++;
 420                 } else {
 421                         vdata->fast_deleted++;
 422                 }
 423         }
 424
 425         goto done;
 426
 427 skipped:
 428         vdata->fast_skipped++;
 429
 430 done:
 431         if (tdb_data.dptr != NULL) {
 432                 free(tdb_data.dptr);
 433         }
 434         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 435
 436         return 0;
 437 }
 438
 439 /**
 440  * Vacuum a DB:
 441  *  - Always do the fast vacuuming run, which traverses
 442  *    the in-memory delete queue: these records have been
 443  *    scheduled for deletion.
 444  *  - Only if explicitly requested, the database is traversed
 445  *    in order to use the traditional heuristics on empty records
 446  *    to trigger deletion.
 447  *    This is done only every VacuumFastPathCount'th vacuuming run.
 448  *
 449  * The traverse runs fill two lists:
 450  *
 451  * - The delete_list:
 452  *   This is the list of empty records the current
 453  *   node is lmaster and dmaster for. These records are later
 454  *   deleted first on other nodes and then locally.
 455  *
 456  *   The fast vacuuming run has a short cut for those records
 457  *   that have never been migrated with data: these records
 458  *   are immediately deleted locally, since they have left
 459  *   no trace on other nodes.
 460  *
 461  * - The vacuum_fetch lists
 462  *   (one for each other lmaster node):
 463  *   The records in this list are sent for deletion to
 464  *   their lmaster in a bulk VACUUM_FETCH message.
 465  *
 466  *   The lmaster then migrates all these records to itelf
 467  *   so that they can be vacuumed there.
 468  *
 469  * This executes in the child context.
 470  */
 471 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 472                           struct vacuum_data *vdata,
 473                           bool full_vacuum_run)
 474 {
 475         struct ctdb_context *ctdb = ctdb_db->ctdb;
 476         const char *name = ctdb_db->db_name;
 477         int ret, i, pnn;
 478
 479         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 480                            "%s db_id[0x%08x]\n",
 481                            full_vacuum_run ? "full" : "fast",
 482                            ctdb_db->db_name, ctdb_db->db_id));
 483
 484         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 485         if (ret != 0) {
 486                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 487                 return ret;
 488         }
 489
 490         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 491         if (pnn == -1) {
 492                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 493                 return -1;
 494         }
 495
 496         ctdb->pnn = pnn;
 497
 498         vdata->fast_added_to_delete_tree = 0;
 499         vdata->fast_added_to_vacuum_fetch_list = 0;
 500         vdata->fast_deleted = 0;
 501         vdata->fast_skipped = 0;
 502         vdata->fast_error = 0;
 503         vdata->fast_total = 0;
 504         vdata->full_added_to_delete_tree = 0;
 505         vdata->full_added_to_vacuum_fetch_list = 0;
 506         vdata->full_skipped = 0;
 507         vdata->full_error = 0;
 508         vdata->full_total = 0;
 509
 510         /* the list needs to be of length num_nodes */
 511         vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
 512         if (vdata->list == NULL) {
 513                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 514                 return -1;
 515         }
 516         for (i = 0; i < ctdb->num_nodes; i++) {
 517                 vdata->list[i] = (struct ctdb_marshall_buffer *)
 518                         talloc_zero_size(vdata->list,
 519                                          offsetof(struct ctdb_marshall_buffer, data));
 520                 if (vdata->list[i] == NULL) {
 521                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 522                         return -1;
 523                 }
 524                 vdata->list[i]->db_id = ctdb_db->db_id;
 525         }
 526
 527         /*
 528          * Traverse the delete_queue.
 529          * This builds the same lists as the db traverse.
 530          */
 531         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 532
 533         if (vdata->fast_total > 0) {
 534                 DEBUG(DEBUG_INFO,
 535                       (__location__
 536                        " fast vacuuming delete_queue traverse statistics: "
 537                        "db[%s] "
 538                        "total[%u] "
 539                        "del[%u] "
 540                        "skp[%u] "
 541                        "err[%u] "
 542                        "adt[%u] "
 543                        "avf[%u]\n",
 544                        ctdb_db->db_name,
 545                        (unsigned)vdata->fast_total,
 546                        (unsigned)vdata->fast_deleted,
 547                        (unsigned)vdata->fast_skipped,
 548                        (unsigned)vdata->fast_error,
 549                        (unsigned)vdata->fast_added_to_delete_tree,
 550                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 551         }
 552
 553         /*
 554          * read-only traverse of the database, looking for records that
 555          * might be able to be vacuumed.
 556          *
 557          * This is not done each time but only every tunable
 558          * VacuumFastPathCount times.
 559          */
 560         if (full_vacuum_run) {
 561                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 562                 if (ret == -1 || vdata->traverse_error) {
 563                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 564                         return -1;
 565                 }
 566                 if (vdata->full_total > 0) {
 567                         DEBUG(DEBUG_INFO,
 568                               (__location__
 569                                " full vacuuming db traverse statistics: "
 570                                "db[%s] "
 571                                "total[%u] "
 572                                "skp[%u] "
 573                                "err[%u] "
 574                                "adt[%u] "
 575                                "avf[%u]\n",
 576                                ctdb_db->db_name,
 577                                (unsigned)vdata->full_total,
 578                                (unsigned)vdata->full_skipped,
 579                                (unsigned)vdata->full_error,
 580                                (unsigned)vdata->full_added_to_delete_tree,
 581                                (unsigned)vdata->full_added_to_vacuum_fetch_list));
 582                 }
 583         }
 584
 585         /*
 586          * For records where we are not the lmaster,
 587          * tell the lmaster to fetch the record.
 588          */
 589         for (i = 0; i < ctdb->num_nodes; i++) {
 590                 TDB_DATA data;
 591
 592                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 593                         continue;
 594                 }
 595
 596                 if (vdata->list[i]->count == 0) {
 597                         continue;
 598                 }
 599
 600                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 601                                    vdata->list[i]->count, ctdb->nodes[i]->pnn,
 602                                    name));
 603
 604                 data.dsize = talloc_get_size(vdata->list[i]);
 605                 data.dptr  = (void *)vdata->list[i];
 606                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 607                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 608                                           "fetch message to %u\n",
 609                                           ctdb->nodes[i]->pnn));
 610                         return -1;
 611                 }
 612         }
 613
 614         /* Process all records we can delete (if any) */
 615         if (vdata->delete_count > 0) {
 616                 struct delete_records_list *recs;
 617                 TDB_DATA indata, outdata;
 618                 int32_t res;
 619                 struct ctdb_node_map *nodemap;
 620                 uint32_t *active_nodes;
 621                 int num_active_nodes;
 622
 623                 recs = talloc_zero(vdata, struct delete_records_list);
 624                 if (recs == NULL) {
 625                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 626                         return -1;
 627                 }
 628                 recs->records = (struct ctdb_marshall_buffer *)
 629                         talloc_zero_size(vdata,
 630                                     offsetof(struct ctdb_marshall_buffer, data));
 631                 if (recs->records == NULL) {
 632                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 633                         return -1;
 634                 }
 635                 recs->records->db_id = ctdb_db->db_id;
 636
 637                 /*
 638                  * traverse the tree of all records we want to delete and
 639                  * create a blob we can send to the other nodes.
 640                  */
 641                 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
 642
 643                 indata.dsize = talloc_get_size(recs->records);
 644                 indata.dptr  = (void *)recs->records;
 645
 646                 /*
 647                  * now tell all the active nodes to delete all these records
 648                  * (if possible)
 649                  */
 650
 651                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 652                                            CTDB_CURRENT_NODE,
 653                                            recs, /* talloc context */
 654                                            &nodemap);
 655                 if (ret != 0) {
 656                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 657                         return -1;
 658                 }
 659
 660                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 661                                                     nodemap, /* talloc context */
 662                                                     false /* include self */);
 663                 /* yuck! ;-) */
 664                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 665
 666                 for (i = 0; i < num_active_nodes; i++) {
 667                         struct ctdb_marshall_buffer *records;
 668                         struct ctdb_rec_data *rec;
 669
 670                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 671                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 672                                         indata, recs, &outdata, &res,
 673                                         NULL, NULL);
 674                         if (ret != 0 || res != 0) {
 675                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 676                                                   "node %u: ret[%d] res[%d]\n",
 677                                                   active_nodes[i], ret, res));
 678                                 return -1;
 679                         }
 680
 681                         /*
 682                          * outdata countains the list of records coming back
 683                          * from the node which the node could not delete
 684                          */
 685                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 686                         rec = (struct ctdb_rec_data *)&records->data[0];
 687                         while (records->count-- > 1) {
 688                                 TDB_DATA reckey, recdata;
 689                                 struct ctdb_ltdb_header *rechdr;
 690
 691                                 reckey.dptr = &rec->data[0];
 692                                 reckey.dsize = rec->keylen;
 693                                 recdata.dptr = &rec->data[reckey.dsize];
 694                                 recdata.dsize = rec->datalen;
 695
 696                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 697                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 698                                         return -1;
 699                                 }
 700                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 701                                 recdata.dptr += sizeof(*rechdr);
 702                                 recdata.dsize -= sizeof(*rechdr);
 703
 704                                 /*
 705                                  * that other node couldnt delete the record
 706                                  * so we should delete it and thereby remove it from the tree
 707                                  */
 708                                 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
 709
 710                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 711                         }
 712                 }
 713
 714                 /* free nodemap and active_nodes */
 715                 talloc_free(nodemap);
 716
 717                 /*
 718                  * The only records remaining in the tree would be those
 719                  * records where all other nodes could successfully
 720                  * delete them, so we can safely delete them on the
 721                  * lmaster as well. Deletion implictely happens while
 722                  * we repack the database. The repack algorithm revisits
 723                  * the tree in order to find the records that don't need
 724                  * to be copied / repacked.
 725                  */
 726         }
 727
 728         /* this ensures we run our event queue */
 729         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 730
 731         return 0;
 732 }
 733
 734
 735 /*
 736  * traverse function for repacking
 737  */
 738 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 739 {
 740         struct vacuum_data *vdata = (struct vacuum_data *)private;
 741
 742         if (vdata->vacuum) {
 743                 uint32_t hash = ctdb_hash(&key);
 744                 struct delete_record_data *kd;
 745                 /*
 746                  * check if we can ignore this record because it's in the delete_tree
 747                  */
 748                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
 749                 /*
 750                  * there might be hash collisions so we have to compare the keys here to be sure
 751                  */
 752                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 753                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 754                         /*
 755                          * we have to check if the record hasn't changed in the meantime in order to
 756                          * savely remove it from the database
 757                          */
 758                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 759                                 hdr->dmaster == kd->ctdb->pnn &&
 760                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 761                                 kd->hdr.rsn == hdr->rsn) {
 762                                 vdata->vacuumed++;
 763                                 return 0;
 764                         }
 765                 }
 766         }
 767         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 768                 vdata->traverse_error = true;
 769                 return -1;
 770         }
 771         vdata->copied++;
 772         return 0;
 773 }
 774
 775 /*
 776  * repack a tdb
 777  */
 778 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 779 {
 780         struct tdb_context *tmp_db;
 781
 782         if (tdb_transaction_start(tdb) != 0) {
 783                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 784                 return -1;
 785         }
 786
 787         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 788                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 789                           O_RDWR|O_CREAT, 0);
 790         if (tmp_db == NULL) {
 791                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 792                 tdb_transaction_cancel(tdb);
 793                 return -1;
 794         }
 795
 796         vdata->traverse_error = false;
 797         vdata->dest_db = tmp_db;
 798         vdata->vacuum = true;
 799         vdata->vacuumed = 0;
 800         vdata->copied = 0;
 801
 802         /*
 803          * repack and vacuum on-the-fly by not writing the records that are
 804          * no longer needed
 805          */
 806         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 807                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 808                 tdb_transaction_cancel(tdb);
 809                 tdb_close(tmp_db);
 810                 return -1;
 811         }
 812
 813         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 814
 815         if (vdata->traverse_error) {
 816                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 817                 tdb_transaction_cancel(tdb);
 818                 tdb_close(tmp_db);
 819                 return -1;
 820         }
 821
 822         if (tdb_wipe_all(tdb) != 0) {
 823                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 824                 tdb_transaction_cancel(tdb);
 825                 tdb_close(tmp_db);
 826                 return -1;
 827         }
 828
 829         vdata->traverse_error = false;
 830         vdata->dest_db = tdb;
 831         vdata->vacuum = false;
 832         vdata->copied = 0;
 833
 834         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 835                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 836                 tdb_transaction_cancel(tdb);
 837                 tdb_close(tmp_db);
 838                 return -1;
 839         }
 840
 841         if (vdata->traverse_error) {
 842                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 843                 tdb_transaction_cancel(tdb);
 844                 tdb_close(tmp_db);
 845                 return -1;
 846         }
 847
 848         tdb_close(tmp_db);
 849
 850
 851         if (tdb_transaction_commit(tdb) != 0) {
 852                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 853                 return -1;
 854         }
 855         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 856
 857         return 0;
 858 }
 859
 860 /*
 861  * repack and vaccum a db
 862  * called from the child context
 863  */
 864 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 865                                      TALLOC_CTX *mem_ctx,
 866                                      bool full_vacuum_run)
 867 {
 868         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
 869         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
 870         const char *name = ctdb_db->db_name;
 871         int freelist_size;
 872         struct vacuum_data *vdata;
 873
 874         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
 875         if (freelist_size == -1) {
 876                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
 877                 return -1;
 878         }
 879
 880         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 881         if (vdata == NULL) {
 882                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 883                 return -1;
 884         }
 885
 886         vdata->ctdb = ctdb_db->ctdb;
 887         vdata->vacuum_limit = vacuum_limit;
 888         vdata->repack_limit = repack_limit;
 889         vdata->delete_tree = trbt_create(vdata, 0);
 890         vdata->ctdb_db = ctdb_db;
 891         if (vdata->delete_tree == NULL) {
 892                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 893                 talloc_free(vdata);
 894                 return -1;
 895         }
 896
 897         vdata->start = timeval_current();
 898
 899         /*
 900          * gather all records that can be deleted in vdata
 901          */
 902         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
 903                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
 904         }
 905
 906         /*
 907          * decide if a repack is necessary
 908          */
 909         if (freelist_size < repack_limit && vdata->delete_count < vacuum_limit)
 910         {
 911                 talloc_free(vdata);
 912                 return 0;
 913         }
 914
 915         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
 916                         name, freelist_size, vdata->delete_count));
 917
 918         /*
 919          * repack and implicitely get rid of the records we can delete
 920          */
 921         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
 922                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
 923                 talloc_free(vdata);
 924                 return -1;
 925         }
 926         talloc_free(vdata);
 927
 928         return 0;
 929 }
 930
 931 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
 932 {
 933         uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 934
 935         return interval;
 936 }
 937
 938 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
 939 {
 940         double l = timeval_elapsed(&child_ctx->start_time);
 941         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
 942         struct ctdb_context *ctdb = ctdb_db->ctdb;
 943
 944         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
 945
 946         if (child_ctx->child_pid != -1) {
 947                 kill(child_ctx->child_pid, SIGKILL);
 948         } else {
 949                 /* Bump the number of successful fast-path runs. */
 950                 child_ctx->vacuum_handle->fast_path_count++;
 951         }
 952
 953         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
 954
 955         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
 956                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
 957                         ctdb_vacuum_event, child_ctx->vacuum_handle);
 958
 959         return 0;
 960 }
 961
 962 /*
 963  * this event is generated when a vacuum child process times out
 964  */
 965 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
 966                                          struct timeval t, void *private_data)
 967 {
 968         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
 969
 970         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
 971
 972         child_ctx->status = VACUUM_TIMEOUT;
 973
 974         talloc_free(child_ctx);
 975 }
 976
 977
 978 /*
 979  * this event is generated when a vacuum child process has completed
 980  */
 981 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
 982                              uint16_t flags, void *private_data)
 983 {
 984         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
 985         char c = 0;
 986         int ret;
 987
 988         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
 989         child_ctx->child_pid = -1;
 990
 991         ret = read(child_ctx->fd[0], &c, 1);
 992         if (ret != 1 || c != 0) {
 993                 child_ctx->status = VACUUM_ERROR;
 994                 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));
 995         } else {
 996                 child_ctx->status = VACUUM_OK;
 997         }
 998
 999         talloc_free(child_ctx);
1000 }
1001
1002 /*
1003  * this event is called every time we need to start a new vacuum process
1004  */
1005 static void
1006 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1007                                struct timeval t, void *private_data)
1008 {
1009         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1010         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1011         struct ctdb_context *ctdb = ctdb_db->ctdb;
1012         struct ctdb_vacuum_child_context *child_ctx;
1013         struct tevent_fd *fde;
1014         int ret;
1015
1016         /* we dont vacuum if we are in recovery mode, or db frozen */
1017         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1018             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1019                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1020                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1021                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1022                                    ? "freeze pending"
1023                                    : "frozen"));
1024                 event_add_timed(ctdb->ev, vacuum_handle,
1025                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1026                         ctdb_vacuum_event, vacuum_handle);
1027                 return;
1028         }
1029
1030         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1031         if (child_ctx == NULL) {
1032                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1033                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1034         }
1035
1036
1037         ret = pipe(child_ctx->fd);
1038         if (ret != 0) {
1039                 talloc_free(child_ctx);
1040                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1041                 event_add_timed(ctdb->ev, vacuum_handle,
1042                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1043                         ctdb_vacuum_event, vacuum_handle);
1044                 return;
1045         }
1046
1047         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1048                 vacuum_handle->fast_path_count = 0;
1049         }
1050
1051         child_ctx->child_pid = ctdb_fork(ctdb);
1052         if (child_ctx->child_pid == (pid_t)-1) {
1053                 close(child_ctx->fd[0]);
1054                 close(child_ctx->fd[1]);
1055                 talloc_free(child_ctx);
1056                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1057                 event_add_timed(ctdb->ev, vacuum_handle,
1058                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1059                         ctdb_vacuum_event, vacuum_handle);
1060                 return;
1061         }
1062
1063
1064         if (child_ctx->child_pid == 0) {
1065                 char cc = 0;
1066                 bool full_vacuum_run = false;
1067                 close(child_ctx->fd[0]);
1068
1069                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1070
1071                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1072                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1073                         _exit(1);
1074                 }
1075
1076                 /*
1077                  * repack the db
1078                  */
1079                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1080                     (vacuum_handle->fast_path_count == 0))
1081                 {
1082                         full_vacuum_run = true;
1083                 }
1084                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1085                                                full_vacuum_run);
1086
1087                 write(child_ctx->fd[1], &cc, 1);
1088                 _exit(0);
1089         }
1090
1091         set_close_on_exec(child_ctx->fd[0]);
1092         close(child_ctx->fd[1]);
1093
1094         child_ctx->status = VACUUM_RUNNING;
1095         child_ctx->start_time = timeval_current();
1096
1097         DLIST_ADD(ctdb->vacuumers, child_ctx);
1098         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1099
1100         /*
1101          * Clear the fastpath vacuuming list in the parent.
1102          */
1103         talloc_free(ctdb_db->delete_queue);
1104         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1105         if (ctdb_db->delete_queue == NULL) {
1106                 /* fatal here? ... */
1107                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1108                                  "in parent context. Shutting down\n");
1109         }
1110
1111         event_add_timed(ctdb->ev, child_ctx,
1112                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1113                 vacuum_child_timeout, child_ctx);
1114
1115         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1116
1117         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1118                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1119         tevent_fd_set_auto_close(fde);
1120
1121         vacuum_handle->child_ctx = child_ctx;
1122         child_ctx->vacuum_handle = vacuum_handle;
1123 }
1124
1125 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1126 {
1127         /* Simply free them all. */
1128         while (ctdb->vacuumers) {
1129                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1130                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1131                            (int)ctdb->vacuumers->child_pid));
1132                 /* vacuum_child_destructor kills it, removes from list */
1133                 talloc_free(ctdb->vacuumers);
1134         }
1135 }
1136
1137 /* this function initializes the vacuuming context for a database
1138  * starts the vacuuming events
1139  */
1140 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1141 {
1142         if (ctdb_db->persistent != 0) {
1143                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1144                 return 0;
1145         }
1146
1147         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1148         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1149
1150         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1151         ctdb_db->vacuum_handle->fast_path_count = 0;
1152
1153         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1154                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1155                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1156
1157         return 0;
1158 }
1159
1160 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1161                                             const struct ctdb_ltdb_header *hdr,
1162                                             const TDB_DATA key)
1163 {
1164         struct delete_record_data *kd;
1165         uint32_t hash;
1166
1167         hash = (uint32_t)ctdb_hash(&key);
1168
1169         DEBUG(DEBUG_DEBUG, (__location__
1170                             " remove_record_from_delete_queue: db[%s] "
1171                             "db_id[0x%08x] "
1172                             "key_hash[0x%08x] "
1173                             "lmaster[%u] "
1174                             "migrated_with_data[%s]\n",
1175                              ctdb_db->db_name, ctdb_db->db_id,
1176                              hash,
1177                              ctdb_lmaster(ctdb_db->ctdb, &key),
1178                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1179
1180         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1181         if (kd == NULL) {
1182                 return;
1183         }
1184         if (kd->key.dsize != key.dsize) {
1185                 return;
1186         }
1187         if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1188                 return;
1189         }
1190
1191         talloc_free(kd);
1192
1193         return;
1194 }
1195
1196 /**
1197  * Insert a record into the ctdb_db context's delete queue,
1198  * handling hash collisions.
1199  */
1200 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1201                                            const struct ctdb_ltdb_header *hdr,
1202                                            TDB_DATA key)
1203 {
1204         struct delete_record_data *kd;
1205         uint32_t hash;
1206         int ret;
1207
1208         hash = (uint32_t)ctdb_hash(&key);
1209
1210         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1211                            "db_id[0x%08x] "
1212                            "key_hash[0x%08x] "
1213                            "lmaster[%u] "
1214                            "migrated_with_data[%s]\n",
1215                             ctdb_db->db_name, ctdb_db->db_id,
1216                             hash,
1217                             ctdb_lmaster(ctdb_db->ctdb, &key),
1218                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1219
1220         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1221         if (kd != NULL) {
1222                 if ((kd->key.dsize != key.dsize) ||
1223                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1224                 {
1225                         DEBUG(DEBUG_INFO,
1226                               ("schedule for deletion: Hash collision (0x%08x)."
1227                                " Skipping the record.\n", hash));
1228                         return 0;
1229                 } else {
1230                         DEBUG(DEBUG_DEBUG,
1231                               ("schedule for deletion: Overwriting entry for "
1232                                "key with hash 0x%08x.\n", hash));
1233                 }
1234         }
1235
1236         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1237                                                   ctdb_db->delete_queue,
1238                                                   hdr, key);
1239         if (ret != 0) {
1240                 return -1;
1241         }
1242
1243         return 0;
1244 }
1245
1246 /**
1247  * Schedule a record for deletetion.
1248  * Called from the parent context.
1249  */
1250 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1251                                            TDB_DATA indata)
1252 {
1253         struct ctdb_control_schedule_for_deletion *dd;
1254         struct ctdb_db_context *ctdb_db;
1255         int ret;
1256         TDB_DATA key;
1257
1258         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1259
1260         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1261         if (ctdb_db == NULL) {
1262                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1263                                   dd->db_id));
1264                 return -1;
1265         }
1266
1267         key.dsize = dd->keylen;
1268         key.dptr = dd->key;
1269
1270         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1271
1272         return ret;
1273 }
1274
1275 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1276                                          const struct ctdb_ltdb_header *hdr,
1277                                          TDB_DATA key)
1278 {
1279         int ret;
1280         struct ctdb_control_schedule_for_deletion *dd;
1281         TDB_DATA indata;
1282         int32_t status;
1283
1284         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1285                 /* main daemon - directly queue */
1286                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1287
1288                 return ret;
1289         }
1290
1291         /* child process: send the main daemon a control */
1292
1293         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1294         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1295         if (indata.dptr == NULL) {
1296                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1297                 return -1;
1298         }
1299         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1300         dd->db_id = ctdb_db->db_id;
1301         dd->hdr = *hdr;
1302         dd->keylen = key.dsize;
1303         memcpy(dd->key, key.dptr, key.dsize);
1304
1305         ret = ctdb_control(ctdb_db->ctdb,
1306                            CTDB_CURRENT_NODE,
1307                            ctdb_db->db_id,
1308                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1309                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1310                            indata,
1311                            NULL, /* mem_ctx */
1312                            NULL, /* outdata */
1313                            &status,
1314                            NULL, /* timeout : NULL == wait forever */
1315                            NULL); /* error message */
1316
1317         talloc_free(indata.dptr);
1318
1319         if (ret != 0 || status != 0) {
1320                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1321                                   "SCHEDULE_FOR_DELETION "
1322                                   "control.\n"));
1323                 if (status != 0) {
1324                         ret = -1;
1325                 }
1326         }
1327
1328         return ret;
1329 }
1330
1331 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1332                                          const struct ctdb_ltdb_header *hdr,
1333                                          const TDB_DATA key)
1334 {
1335         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1336                 /*
1337                  * Only remove the record from the delete queue if called
1338                  * in the main daemon.
1339                  */
1340                 return;
1341         }
1342
1343         remove_record_from_delete_queue(ctdb_db, hdr, key);
1344
1345         return;
1346 }