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/events/events.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/events/events.h"
  32 #include "../include/ctdb_private.h"
  33 #include "../common/rb_tree.h"
  34
  35 #define TIMELIMIT() timeval_current_ofs(10, 0)
  36
  37 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  38
  39 struct ctdb_vacuum_child_context {
  40         struct ctdb_vacuum_child_context *next, *prev;
  41         struct ctdb_vacuum_handle *vacuum_handle;
  42         /* fd child writes status to */
  43         int fd[2];
  44         pid_t child_pid;
  45         enum vacuum_child_status status;
  46         struct timeval start_time;
  47 };
  48
  49 struct ctdb_vacuum_handle {
  50         struct ctdb_db_context *ctdb_db;
  51         struct ctdb_vacuum_child_context *child_ctx;
  52         uint32_t fast_path_count;
  53 };
  54
  55
  56 /*  a list of records to possibly delete */
  57 struct vacuum_data {
  58         uint32_t vacuum_limit;
  59         uint32_t repack_limit;
  60         struct ctdb_context *ctdb;
  61         struct ctdb_db_context *ctdb_db;
  62         struct tdb_context *dest_db;
  63         trbt_tree_t *delete_list;
  64         uint32_t delete_count;
  65         struct ctdb_marshall_buffer **vacuum_fetch_list;
  66         struct timeval start;
  67         bool traverse_error;
  68         bool vacuum;
  69         uint32_t total;
  70         uint32_t vacuumed;
  71         uint32_t copied;
  72         uint32_t fast_added_to_vacuum_fetch_list;
  73         uint32_t fast_added_to_delete_list;
  74         uint32_t fast_deleted;
  75         uint32_t fast_skipped;
  76         uint32_t fast_error;
  77         uint32_t fast_total;
  78         uint32_t full_added_to_vacuum_fetch_list;
  79         uint32_t full_added_to_delete_list;
  80         uint32_t full_skipped;
  81         uint32_t full_error;
  82         uint32_t full_total;
  83         uint32_t delete_left;
  84         uint32_t delete_remote_error;
  85         uint32_t delete_local_error;
  86         uint32_t delete_deleted;
  87         uint32_t delete_skipped;
  88 };
  89
  90 /* this structure contains the information for one record to be deleted */
  91 struct delete_record_data {
  92         struct ctdb_context *ctdb;
  93         struct ctdb_db_context *ctdb_db;
  94         struct ctdb_ltdb_header hdr;
  95         TDB_DATA key;
  96 };
  97
  98 struct delete_records_list {
  99         struct ctdb_marshall_buffer *records;
 100 };
 101
 102 /**
 103  * Store key and header in a tree, indexed by the key hash.
 104  */
 105 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 106                                                struct ctdb_db_context *ctdb_db,
 107                                                trbt_tree_t *tree,
 108                                                const struct ctdb_ltdb_header *hdr,
 109                                                TDB_DATA key)
 110 {
 111         struct delete_record_data *dd;
 112         uint32_t hash;
 113
 114         dd = talloc_zero(tree, struct delete_record_data);
 115         if (dd == NULL) {
 116                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 117                 return -1;
 118         }
 119
 120         dd->ctdb      = ctdb;
 121         dd->ctdb_db   = ctdb_db;
 122         dd->key.dsize = key.dsize;
 123         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 124         if (dd->key.dptr == NULL) {
 125                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 126                 return -1;
 127         }
 128
 129         dd->hdr = *hdr;
 130
 131         hash = (uint32_t)tdb_jenkins_hash(&key);
 132
 133         trbt_insert32(tree, hash, dd);
 134
 135         return 0;
 136 }
 137
 138 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 139                                      struct ctdb_ltdb_header *hdr)
 140 {
 141         struct ctdb_context *ctdb = vdata->ctdb;
 142         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 143         uint32_t hash;
 144         int ret;
 145
 146         hash = (uint32_t)tdb_jenkins_hash(&key);
 147
 148         if (trbt_lookup32(vdata->delete_list, hash)) {
 149                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 150                 return 0;
 151         }
 152
 153         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 154                                                   vdata->delete_list,
 155                                                   hdr, key);
 156         if (ret != 0) {
 157                 return -1;
 158         }
 159
 160         vdata->delete_count++;
 161
 162         return 0;
 163 }
 164
 165 /**
 166  * Add a record to the list of records to be sent
 167  * to their lmaster with VACUUM_FETCH.
 168  */
 169 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 170                                            TDB_DATA key)
 171 {
 172         struct ctdb_context *ctdb = vdata->ctdb;
 173         struct ctdb_rec_data *rec;
 174         uint32_t lmaster;
 175         size_t old_size;
 176         struct ctdb_marshall_buffer *vfl;
 177
 178         lmaster = ctdb_lmaster(ctdb, &key);
 179
 180         vfl = vdata->vacuum_fetch_list[lmaster];
 181
 182         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 183         if (rec == NULL) {
 184                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 185                 vdata->traverse_error = true;
 186                 return -1;
 187         }
 188
 189         old_size = talloc_get_size(vfl);
 190         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 191         if (vfl == NULL) {
 192                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 193                 vdata->traverse_error = true;
 194                 return -1;
 195         }
 196         vdata->vacuum_fetch_list[lmaster] = vfl;
 197
 198         vfl->count++;
 199         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 200         talloc_free(rec);
 201
 202         vdata->total++;
 203
 204         return 0;
 205 }
 206
 207
 208 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 209                               struct timeval t, void *private_data);
 210
 211
 212 /*
 213  * traverse function for gathering the records that can be deleted
 214  */
 215 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 216 {
 217         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 218         struct ctdb_context *ctdb = vdata->ctdb;
 219         uint32_t lmaster;
 220         struct ctdb_ltdb_header *hdr;
 221         int res = 0;
 222
 223         vdata->full_total++;
 224
 225         lmaster = ctdb_lmaster(ctdb, &key);
 226         if (lmaster >= ctdb->num_nodes) {
 227                 vdata->full_error++;
 228                 DEBUG(DEBUG_CRIT, (__location__
 229                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 230                                    " with hash[%u]!\n",
 231                                    (unsigned)lmaster,
 232                                    (unsigned)ctdb->num_nodes,
 233                                    (unsigned)ctdb_hash(&key)));
 234                 return -1;
 235         }
 236
 237         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 238                 /* it is not a deleted record */
 239                 vdata->full_skipped++;
 240                 return 0;
 241         }
 242
 243         hdr = (struct ctdb_ltdb_header *)data.dptr;
 244
 245         if (hdr->dmaster != ctdb->pnn) {
 246                 vdata->full_skipped++;
 247                 return 0;
 248         }
 249
 250         if (lmaster == ctdb->pnn) {
 251                 /*
 252                  * We are both lmaster and dmaster, and the record is empty.
 253                  * So we should be able to delete it.
 254                  */
 255                 res = add_record_to_delete_list(vdata, key, hdr);
 256                 if (res != 0) {
 257                         vdata->full_error++;
 258                 } else {
 259                         vdata->full_added_to_delete_list++;
 260                 }
 261         } else {
 262                 /*
 263                  * We are not lmaster.
 264                  * Add the record to the blob ready to send to the nodes.
 265                  */
 266                 res = add_record_to_vacuum_fetch_list(vdata, key);
 267                 if (res != 0) {
 268                         vdata->full_error++;
 269                 } else {
 270                         vdata->full_added_to_vacuum_fetch_list++;
 271                 }
 272         }
 273
 274         return res;
 275 }
 276
 277 /*
 278  * traverse the tree of records to delete and marshall them into
 279  * a blob
 280  */
 281 static int delete_marshall_traverse(void *param, void *data)
 282 {
 283         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 284         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 285         struct ctdb_rec_data *rec;
 286         size_t old_size;
 287
 288         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 289         if (rec == NULL) {
 290                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 291                 return 0;
 292         }
 293
 294         old_size = talloc_get_size(recs->records);
 295         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 296         if (recs->records == NULL) {
 297                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 298                 return 0;
 299         }
 300         recs->records->count++;
 301         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 302         return 0;
 303 }
 304
 305 /**
 306  * traverse function for the traversal of the delete_queue,
 307  * the fast-path vacuuming list.
 308  *
 309  *  - If the record has been migrated off the node
 310  *    or has been revived (filled with data) on the node,
 311  *    then skip the record.
 312  *
 313  *  - If the current node is the record's lmaster and it is
 314  *    a record that has never been migrated with data, then
 315  *    delete the record from the local tdb.
 316  *
 317  *  - If the current node is the record's lmaster and it has
 318  *    been migrated with data, then schedule it for the normal
 319  *    vacuuming procedure (i.e. add it to the delete_list).
 320  *
 321  *  - If the current node is NOT the record's lmaster then
 322  *    add it to the list of records that are to be sent to
 323  *    the lmaster with the VACUUM_FETCH message.
 324  */
 325 static int delete_queue_traverse(void *param, void *data)
 326 {
 327         struct delete_record_data *dd =
 328                 talloc_get_type(data, struct delete_record_data);
 329         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 330         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 331         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 332         int res;
 333         struct ctdb_ltdb_header *header;
 334         TDB_DATA tdb_data;
 335         uint32_t lmaster;
 336         uint32_t hash = ctdb_hash(&(dd->key));
 337
 338         vdata->fast_total++;
 339
 340         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 341         if (res != 0) {
 342                 DEBUG(DEBUG_ERR,
 343                       (__location__ " Error getting chainlock on record with "
 344                        "key hash [0x%08x] on database db[%s].\n",
 345                        hash, ctdb_db->db_name));
 346                 vdata->fast_error++;
 347                 return 0;
 348         }
 349
 350         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 351         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 352                 /* Does not exist or not a ctdb record. Skip. */
 353                 goto skipped;
 354         }
 355
 356         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 357                 /* The record has been recycled (filled with data). Skip. */
 358                 goto skipped;
 359         }
 360
 361         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 362
 363         if (header->dmaster != ctdb->pnn) {
 364                 /* The record has been migrated off the node. Skip. */
 365                 goto skipped;
 366         }
 367
 368         if (header->rsn != dd->hdr.rsn) {
 369                 /*
 370                  * The record has been migrated off the node and back again.
 371                  * But not requeued for deletion. Skip it.
 372                  */
 373                 goto skipped;
 374         }
 375
 376         /*
 377          * We are dmaster, and the record has no data, and it has
 378          * not been migrated after it has been queued for deletion.
 379          *
 380          * At this stage, the record could still have been revived locally
 381          * and last been written with empty data. This can only be
 382          * fixed with the addition of an active or delete flag. (TODO)
 383          */
 384
 385         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 386
 387         if (lmaster != ctdb->pnn) {
 388                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 389
 390                 if (res != 0) {
 391                         DEBUG(DEBUG_ERR,
 392                               (__location__ " Error adding record to list "
 393                                "of records to send to lmaster.\n"));
 394                         vdata->fast_error++;
 395                 } else {
 396                         vdata->fast_added_to_vacuum_fetch_list++;
 397                 }
 398                 goto done;
 399         }
 400
 401         /* use header->flags or dd->hdr.flags ?? */
 402         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 403                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 404
 405                 if (res != 0) {
 406                         DEBUG(DEBUG_ERR,
 407                               (__location__ " Error adding record to list "
 408                                "of records for deletion on lmaster.\n"));
 409                         vdata->fast_error++;
 410                 } else {
 411                         vdata->fast_added_to_delete_list++;
 412                 }
 413         } else {
 414                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 415
 416                 if (res != 0) {
 417                         DEBUG(DEBUG_ERR,
 418                               (__location__ " Error deleting record with key "
 419                                "hash [0x%08x] from local data base db[%s].\n",
 420                                hash, ctdb_db->db_name));
 421                         vdata->fast_error++;
 422                 } else {
 423                         DEBUG(DEBUG_DEBUG,
 424                               (__location__ " Deleted record with key hash "
 425                                "[0x%08x] from local data base db[%s].\n",
 426                                hash, ctdb_db->db_name));
 427                         vdata->fast_deleted++;
 428                 }
 429         }
 430
 431         goto done;
 432
 433 skipped:
 434         vdata->fast_skipped++;
 435
 436 done:
 437         if (tdb_data.dptr != NULL) {
 438                 free(tdb_data.dptr);
 439         }
 440         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 441
 442         return 0;
 443 }
 444
 445 /**
 446  * Delete the records that we are lmaster and dmaster for and
 447  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 448  * control.
 449  */
 450 static int delete_record_traverse(void *param, void *data)
 451 {
 452         struct delete_record_data *dd =
 453                 talloc_get_type(data, struct delete_record_data);
 454         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 455         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 456         struct ctdb_context *ctdb = ctdb_db->ctdb;
 457         int res;
 458         struct ctdb_ltdb_header *header;
 459         TDB_DATA tdb_data;
 460         uint32_t lmaster;
 461         bool deleted = false;
 462
 463         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 464         if (res != 0) {
 465                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 466                 vdata->delete_local_error++;
 467                 return 0;
 468         }
 469
 470         /*
 471          * Verify that the record is still empty, its RSN has not
 472          * changed and that we are still its lmaster and dmaster.
 473          */
 474
 475         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 476         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 477                 /* Does not exist or not a ctdb record. Skip. */
 478                 vdata->delete_skipped++;
 479                 goto done;
 480         }
 481
 482         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 483                 /* The record has been recycled (filled with data). Skip. */
 484                 vdata->delete_skipped++;
 485                 goto done;
 486         }
 487
 488         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 489
 490         if (header->dmaster != ctdb->pnn) {
 491                 /* The record has been migrated off the node. Skip. */
 492                 vdata->delete_skipped++;
 493                 goto done;
 494         }
 495
 496
 497         if (header->rsn != dd->hdr.rsn) {
 498                 /*
 499                  * The record has been migrated off the node and back again.
 500                  * But not requeued for deletion. Skip it.
 501                  */
 502                 vdata->delete_skipped++;
 503                 goto done;
 504         }
 505
 506         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 507
 508         if (lmaster != ctdb->pnn) {
 509                 /* we are not lmaster - strange */
 510                 vdata->delete_skipped++;
 511                 goto done;
 512         }
 513
 514         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 515
 516         if (res != 0) {
 517                 DEBUG(DEBUG_ERR,
 518                       (__location__ " Error deleting record from local "
 519                        "data base.\n"));
 520                 vdata->delete_local_error++;
 521                 goto done;
 522         }
 523
 524         deleted = true;
 525
 526 done:
 527         if (tdb_data.dptr != NULL) {
 528                 free(tdb_data.dptr);
 529         }
 530
 531         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 532
 533         if (deleted) {
 534                 /*
 535                  * successfully deleted the record locally.
 536                  * remove it from the list and update statistics.
 537                  */
 538                 talloc_free(dd);
 539                 vdata->delete_deleted++;
 540                 vdata->delete_left--;
 541         }
 542
 543         return 0;
 544 }
 545
 546 /**
 547  * Fast vacuuming run:
 548  * Traverse the delete_queue.
 549  * This fills the same lists as the database traverse.
 550  */
 551 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
 552                                 struct vacuum_data *vdata)
 553 {
 554         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 555
 556         if (vdata->fast_total > 0) {
 557                 DEBUG(DEBUG_INFO,
 558                       (__location__
 559                        " fast vacuuming delete_queue traverse statistics: "
 560                        "db[%s] "
 561                        "total[%u] "
 562                        "del[%u] "
 563                        "skp[%u] "
 564                        "err[%u] "
 565                        "adl[%u] "
 566                        "avf[%u]\n",
 567                        ctdb_db->db_name,
 568                        (unsigned)vdata->fast_total,
 569                        (unsigned)vdata->fast_deleted,
 570                        (unsigned)vdata->fast_skipped,
 571                        (unsigned)vdata->fast_error,
 572                        (unsigned)vdata->fast_added_to_delete_list,
 573                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 574         }
 575
 576         return;
 577 }
 578
 579 /**
 580  * Full vacuum run:
 581  * read-only traverse of the database, looking for records that
 582  * might be able to be vacuumed.
 583  *
 584  * This is not done each time but only every tunable
 585  * VacuumFastPathCount times.
 586  */
 587 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
 588                                struct vacuum_data *vdata,
 589                                bool full_vacuum_run)
 590 {
 591         int ret;
 592
 593         if (!full_vacuum_run) {
 594                 return 0;
 595         }
 596
 597         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 598         if (ret == -1 || vdata->traverse_error) {
 599                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 600                                   "'%s'\n", ctdb_db->db_name));
 601                 return -1;
 602         }
 603
 604         if (vdata->full_total > 0) {
 605                 DEBUG(DEBUG_INFO,
 606                       (__location__
 607                        " full vacuuming db traverse statistics: "
 608                        "db[%s] "
 609                        "total[%u] "
 610                        "skp[%u] "
 611                        "err[%u] "
 612                        "adl[%u] "
 613                        "avf[%u]\n",
 614                        ctdb_db->db_name,
 615                        (unsigned)vdata->full_total,
 616                        (unsigned)vdata->full_skipped,
 617                        (unsigned)vdata->full_error,
 618                        (unsigned)vdata->full_added_to_delete_list,
 619                        (unsigned)vdata->full_added_to_vacuum_fetch_list));
 620         }
 621
 622         return 0;
 623 }
 624
 625 /**
 626  * Process the vacuum fetch lists:
 627  * For records for which we are not the lmaster, tell the lmaster to
 628  * fetch the record.
 629  */
 630 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 631                                            struct vacuum_data *vdata)
 632 {
 633         int i;
 634         struct ctdb_context *ctdb = ctdb_db->ctdb;
 635
 636         for (i = 0; i < ctdb->num_nodes; i++) {
 637                 TDB_DATA data;
 638                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 639
 640                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 641                         continue;
 642                 }
 643
 644                 if (vfl->count == 0) {
 645                         continue;
 646                 }
 647
 648                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 649                                    vfl->count, ctdb->nodes[i]->pnn,
 650                                    ctdb_db->db_name));
 651
 652                 data.dsize = talloc_get_size(vfl);
 653                 data.dptr  = (void *)vfl;
 654                 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn,
 655                                       CTDB_SRVID_VACUUM_FETCH,
 656                                       data) != 0)
 657                 {
 658                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 659                                           "fetch message to %u\n",
 660                                           ctdb->nodes[i]->pnn));
 661                         return -1;
 662                 }
 663         }
 664
 665         return 0;
 666 }
 667
 668 /**
 669  * Proces the delete list:
 670  * Send the records to delete to all other nodes with the
 671  * try_delete_records control.
 672  */
 673 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 674                                     struct vacuum_data *vdata)
 675 {
 676         int ret, i;
 677         struct ctdb_context *ctdb = ctdb_db->ctdb;
 678
 679         vdata->delete_left = vdata->delete_count;
 680
 681         if (vdata->delete_count > 0) {
 682                 struct delete_records_list *recs;
 683                 TDB_DATA indata, outdata;
 684                 int32_t res;
 685                 struct ctdb_node_map *nodemap;
 686                 uint32_t *active_nodes;
 687                 int num_active_nodes;
 688
 689                 recs = talloc_zero(vdata, struct delete_records_list);
 690                 if (recs == NULL) {
 691                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 692                         return -1;
 693                 }
 694                 recs->records = (struct ctdb_marshall_buffer *)
 695                         talloc_zero_size(vdata,
 696                                     offsetof(struct ctdb_marshall_buffer, data));
 697                 if (recs->records == NULL) {
 698                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 699                         return -1;
 700                 }
 701                 recs->records->db_id = ctdb_db->db_id;
 702
 703                 /*
 704                  * traverse the tree of all records we want to delete and
 705                  * create a blob we can send to the other nodes.
 706                  */
 707                 trbt_traversearray32(vdata->delete_list, 1,
 708                                      delete_marshall_traverse, recs);
 709
 710                 indata.dsize = talloc_get_size(recs->records);
 711                 indata.dptr  = (void *)recs->records;
 712
 713                 /*
 714                  * now tell all the active nodes to delete all these records
 715                  * (if possible)
 716                  */
 717
 718                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 719                                            CTDB_CURRENT_NODE,
 720                                            recs, /* talloc context */
 721                                            &nodemap);
 722                 if (ret != 0) {
 723                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 724                         return -1;
 725                 }
 726
 727                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 728                                                     nodemap, /* talloc context */
 729                                                     false /* include self */);
 730                 /* yuck! ;-) */
 731                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 732
 733                 for (i = 0; i < num_active_nodes; i++) {
 734                         struct ctdb_marshall_buffer *records;
 735                         struct ctdb_rec_data *rec;
 736
 737                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 738                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 739                                         indata, recs, &outdata, &res,
 740                                         NULL, NULL);
 741                         if (ret != 0 || res != 0) {
 742                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 743                                                   "node %u: ret[%d] res[%d]\n",
 744                                                   active_nodes[i], ret, res));
 745                                 return -1;
 746                         }
 747
 748                         /*
 749                          * outdata contains the list of records coming back
 750                          * from the node: These are the records that the
 751                          * remote node could not delete.
 752                          *
 753                          * NOTE: There is a problem here:
 754                          *
 755                          * When a node failed to delete the record, but
 756                          * others succeeded, we may have created gaps in the
 757                          * history of the record. Hence when a node dies, an
 758                          * closed file handle might be resurrected or an open
 759                          * file handle might be lost, leading to blocked access
 760                          * or data corruption.
 761                          *
 762                          * TODO: This needs to be fixed!
 763                          */
 764                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 765                         rec = (struct ctdb_rec_data *)&records->data[0];
 766                         while (records->count-- > 1) {
 767                                 TDB_DATA reckey, recdata;
 768                                 struct ctdb_ltdb_header *rechdr;
 769                                 struct delete_record_data *dd;
 770
 771                                 reckey.dptr = &rec->data[0];
 772                                 reckey.dsize = rec->keylen;
 773                                 recdata.dptr = &rec->data[reckey.dsize];
 774                                 recdata.dsize = rec->datalen;
 775
 776                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 777                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 778                                         return -1;
 779                                 }
 780                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 781                                 recdata.dptr += sizeof(*rechdr);
 782                                 recdata.dsize -= sizeof(*rechdr);
 783
 784                                 dd = (struct delete_record_data *)trbt_lookup32(
 785                                                 vdata->delete_list,
 786                                                 (uint32_t)tdb_jenkins_hash(&reckey));
 787                                 if (dd != NULL) {
 788                                         /*
 789                                          * The other node could not delete the
 790                                          * record and it is the first node that
 791                                          * failed. So we should remove it from
 792                                          * the tree and update statistics.
 793                                          */
 794                                         talloc_free(dd);
 795                                         vdata->delete_remote_error++;
 796                                         vdata->delete_left--;
 797                                 }
 798
 799                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 800                         }
 801                 }
 802
 803                 /* free nodemap and active_nodes */
 804                 talloc_free(nodemap);
 805         }
 806
 807         if (vdata->delete_left > 0) {
 808                 /*
 809                  * The only records remaining in the tree are those
 810                  * records which all other nodes could successfully
 811                  * delete, so we can safely delete them on the
 812                  * lmaster as well.
 813                  */
 814                 trbt_traversearray32(vdata->delete_list, 1,
 815                                      delete_record_traverse, vdata);
 816         }
 817
 818         if (vdata->delete_count > 0) {
 819                 DEBUG(DEBUG_INFO,
 820                       (__location__
 821                        " vacuum delete list statistics: "
 822                        "db[%s] "
 823                        "coll[%u] "
 824                        "rem.err[%u] "
 825                        "loc.err[%u] "
 826                        "skip[%u] "
 827                        "del[%u] "
 828                        "left[%u]\n",
 829                        ctdb_db->db_name,
 830                        (unsigned)vdata->delete_count,
 831                        (unsigned)vdata->delete_remote_error,
 832                        (unsigned)vdata->delete_local_error,
 833                        (unsigned)vdata->delete_skipped,
 834                        (unsigned)vdata->delete_deleted,
 835                        (unsigned)vdata->delete_left));
 836         }
 837
 838         return 0;
 839 }
 840
 841 /**
 842  * initialize the vacuum_data
 843  */
 844 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
 845                                         struct vacuum_data *vdata)
 846 {
 847         int i;
 848         struct ctdb_context *ctdb = ctdb_db->ctdb;
 849
 850         vdata->fast_added_to_delete_list = 0;
 851         vdata->fast_added_to_vacuum_fetch_list = 0;
 852         vdata->fast_deleted = 0;
 853         vdata->fast_skipped = 0;
 854         vdata->fast_error = 0;
 855         vdata->fast_total = 0;
 856         vdata->full_added_to_delete_list = 0;
 857         vdata->full_added_to_vacuum_fetch_list = 0;
 858         vdata->full_skipped = 0;
 859         vdata->full_error = 0;
 860         vdata->full_total = 0;
 861         vdata->delete_count = 0;
 862         vdata->delete_left = 0;
 863         vdata->delete_remote_error = 0;
 864         vdata->delete_local_error = 0;
 865         vdata->delete_skipped = 0;
 866         vdata->delete_deleted = 0;
 867
 868         /* the list needs to be of length num_nodes */
 869         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
 870                                                 struct ctdb_marshall_buffer *,
 871                                                 ctdb->num_nodes);
 872         if (vdata->vacuum_fetch_list == NULL) {
 873                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 874                 return -1;
 875         }
 876         for (i = 0; i < ctdb->num_nodes; i++) {
 877                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 878                         talloc_zero_size(vdata->vacuum_fetch_list,
 879                                          offsetof(struct ctdb_marshall_buffer, data));
 880                 if (vdata->vacuum_fetch_list[i] == NULL) {
 881                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 882                         return -1;
 883                 }
 884                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 885         }
 886
 887         return 0;
 888 }
 889
 890 /**
 891  * Vacuum a DB:
 892  *  - Always do the fast vacuuming run, which traverses
 893  *    the in-memory delete queue: these records have been
 894  *    scheduled for deletion.
 895  *  - Only if explicitly requested, the database is traversed
 896  *    in order to use the traditional heuristics on empty records
 897  *    to trigger deletion.
 898  *    This is done only every VacuumFastPathCount'th vacuuming run.
 899  *
 900  * The traverse runs fill two lists:
 901  *
 902  * - The delete_list:
 903  *   This is the list of empty records the current
 904  *   node is lmaster and dmaster for. These records are later
 905  *   deleted first on other nodes and then locally.
 906  *
 907  *   The fast vacuuming run has a short cut for those records
 908  *   that have never been migrated with data: these records
 909  *   are immediately deleted locally, since they have left
 910  *   no trace on other nodes.
 911  *
 912  * - The vacuum_fetch lists
 913  *   (one for each other lmaster node):
 914  *   The records in this list are sent for deletion to
 915  *   their lmaster in a bulk VACUUM_FETCH message.
 916  *
 917  *   The lmaster then migrates all these records to itelf
 918  *   so that they can be vacuumed there.
 919  *
 920  * This executes in the child context.
 921  */
 922 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 923                           struct vacuum_data *vdata,
 924                           bool full_vacuum_run)
 925 {
 926         struct ctdb_context *ctdb = ctdb_db->ctdb;
 927         int ret, pnn;
 928
 929         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 930                            "%s db_id[0x%08x]\n",
 931                            full_vacuum_run ? "full" : "fast",
 932                            ctdb_db->db_name, ctdb_db->db_id));
 933
 934         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 935         if (ret != 0) {
 936                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 937                 return ret;
 938         }
 939
 940         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 941         if (pnn == -1) {
 942                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 943                 return -1;
 944         }
 945
 946         ctdb->pnn = pnn;
 947
 948         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
 949         if (ret != 0) {
 950                 return ret;
 951         }
 952
 953         ctdb_vacuum_db_fast(ctdb_db, vdata);
 954
 955         ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
 956         if (ret != 0) {
 957                 return ret;
 958         }
 959
 960         ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
 961         if (ret != 0) {
 962                 return ret;
 963         }
 964
 965         ret = ctdb_process_delete_list(ctdb_db, vdata);
 966         if (ret != 0) {
 967                 return ret;
 968         }
 969
 970         /* this ensures we run our event queue */
 971         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 972
 973         return 0;
 974 }
 975
 976
 977 /*
 978  * traverse function for repacking
 979  */
 980 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 981 {
 982         struct vacuum_data *vdata = (struct vacuum_data *)private;
 983
 984         if (vdata->vacuum) {
 985                 uint32_t hash = (uint32_t)tdb_jenkins_hash(&key);
 986                 struct delete_record_data *kd;
 987                 /*
 988                  * check if we can ignore this record because it's in the delete_list
 989                  */
 990                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
 991                 /*
 992                  * there might be hash collisions so we have to compare the keys here to be sure
 993                  */
 994                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 995                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 996                         /*
 997                          * we have to check if the record hasn't changed in the meantime in order to
 998                          * savely remove it from the database
 999                          */
1000                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1001                                 hdr->dmaster == kd->ctdb->pnn &&
1002                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1003                                 kd->hdr.rsn == hdr->rsn) {
1004                                 vdata->vacuumed++;
1005                                 return 0;
1006                         }
1007                 }
1008         }
1009         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1010                 vdata->traverse_error = true;
1011                 return -1;
1012         }
1013         vdata->copied++;
1014         return 0;
1015 }
1016
1017 /*
1018  * repack a tdb
1019  */
1020 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1021 {
1022         struct tdb_context *tmp_db;
1023
1024         if (tdb_transaction_start(tdb) != 0) {
1025                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1026                 return -1;
1027         }
1028
1029         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1030                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1031                           O_RDWR|O_CREAT, 0);
1032         if (tmp_db == NULL) {
1033                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1034                 tdb_transaction_cancel(tdb);
1035                 return -1;
1036         }
1037
1038         vdata->traverse_error = false;
1039         vdata->dest_db = tmp_db;
1040         vdata->vacuum = true;
1041         vdata->vacuumed = 0;
1042         vdata->copied = 0;
1043
1044         /*
1045          * repack and vacuum on-the-fly by not writing the records that are
1046          * no longer needed
1047          */
1048         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1049                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1050                 tdb_transaction_cancel(tdb);
1051                 tdb_close(tmp_db);
1052                 return -1;
1053         }
1054
1055         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1056
1057         if (vdata->traverse_error) {
1058                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1059                 tdb_transaction_cancel(tdb);
1060                 tdb_close(tmp_db);
1061                 return -1;
1062         }
1063
1064         if (tdb_wipe_all(tdb) != 0) {
1065                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1066                 tdb_transaction_cancel(tdb);
1067                 tdb_close(tmp_db);
1068                 return -1;
1069         }
1070
1071         vdata->traverse_error = false;
1072         vdata->dest_db = tdb;
1073         vdata->vacuum = false;
1074         vdata->copied = 0;
1075
1076         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1077                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1078                 tdb_transaction_cancel(tdb);
1079                 tdb_close(tmp_db);
1080                 return -1;
1081         }
1082
1083         if (vdata->traverse_error) {
1084                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1085                 tdb_transaction_cancel(tdb);
1086                 tdb_close(tmp_db);
1087                 return -1;
1088         }
1089
1090         tdb_close(tmp_db);
1091
1092
1093         if (tdb_transaction_commit(tdb) != 0) {
1094                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1095                 return -1;
1096         }
1097         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1098
1099         return 0;
1100 }
1101
1102 /*
1103  * repack and vaccum a db
1104  * called from the child context
1105  */
1106 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1107                                      TALLOC_CTX *mem_ctx,
1108                                      bool full_vacuum_run)
1109 {
1110         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1111         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1112         const char *name = ctdb_db->db_name;
1113         int freelist_size;
1114         struct vacuum_data *vdata;
1115
1116         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1117         if (freelist_size == -1) {
1118                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1119                 return -1;
1120         }
1121
1122         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1123         if (vdata == NULL) {
1124                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1125                 return -1;
1126         }
1127
1128         vdata->ctdb = ctdb_db->ctdb;
1129         vdata->vacuum_limit = vacuum_limit;
1130         vdata->repack_limit = repack_limit;
1131         vdata->delete_list = trbt_create(vdata, 0);
1132         vdata->ctdb_db = ctdb_db;
1133         if (vdata->delete_list == NULL) {
1134                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1135                 talloc_free(vdata);
1136                 return -1;
1137         }
1138
1139         vdata->start = timeval_current();
1140
1141         /*
1142          * gather all records that can be deleted in vdata
1143          */
1144         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1145                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1146         }
1147
1148         /*
1149          * decide if a repack is necessary
1150          */
1151         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1152         {
1153                 talloc_free(vdata);
1154                 return 0;
1155         }
1156
1157         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1158                         name, freelist_size, vdata->delete_left));
1159
1160         /*
1161          * repack and implicitely get rid of the records we can delete
1162          */
1163         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1164                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1165                 talloc_free(vdata);
1166                 return -1;
1167         }
1168         talloc_free(vdata);
1169
1170         return 0;
1171 }
1172
1173 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1174 {
1175         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1176
1177         return interval;
1178 }
1179
1180 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1181 {
1182         double l = timeval_elapsed(&child_ctx->start_time);
1183         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1184         struct ctdb_context *ctdb = ctdb_db->ctdb;
1185
1186         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1187
1188         if (child_ctx->child_pid != -1) {
1189                 kill(child_ctx->child_pid, SIGKILL);
1190         } else {
1191                 /* Bump the number of successful fast-path runs. */
1192                 child_ctx->vacuum_handle->fast_path_count++;
1193         }
1194
1195         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1196
1197         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1198                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1199                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1200
1201         return 0;
1202 }
1203
1204 /*
1205  * this event is generated when a vacuum child process times out
1206  */
1207 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1208                                          struct timeval t, void *private_data)
1209 {
1210         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1211
1212         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1213
1214         child_ctx->status = VACUUM_TIMEOUT;
1215
1216         talloc_free(child_ctx);
1217 }
1218
1219
1220 /*
1221  * this event is generated when a vacuum child process has completed
1222  */
1223 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1224                              uint16_t flags, void *private_data)
1225 {
1226         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1227         char c = 0;
1228         int ret;
1229
1230         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1231         child_ctx->child_pid = -1;
1232
1233         ret = read(child_ctx->fd[0], &c, 1);
1234         if (ret != 1 || c != 0) {
1235                 child_ctx->status = VACUUM_ERROR;
1236                 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));
1237         } else {
1238                 child_ctx->status = VACUUM_OK;
1239         }
1240
1241         talloc_free(child_ctx);
1242 }
1243
1244 /*
1245  * this event is called every time we need to start a new vacuum process
1246  */
1247 static void
1248 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1249                                struct timeval t, void *private_data)
1250 {
1251         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1252         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1253         struct ctdb_context *ctdb = ctdb_db->ctdb;
1254         struct ctdb_vacuum_child_context *child_ctx;
1255         int ret;
1256
1257         /* we dont vacuum if we are in recovery mode, or db frozen */
1258         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1259             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1260                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1261                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1262                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1263                                    ? "freeze pending"
1264                                    : "frozen"));
1265                 event_add_timed(ctdb->ev, vacuum_handle,
1266                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1267                         ctdb_vacuum_event, vacuum_handle);
1268                 return;
1269         }
1270
1271         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1272         if (child_ctx == NULL) {
1273                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1274                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1275         }
1276
1277
1278         ret = pipe(child_ctx->fd);
1279         if (ret != 0) {
1280                 talloc_free(child_ctx);
1281                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1282                 event_add_timed(ctdb->ev, vacuum_handle,
1283                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1284                         ctdb_vacuum_event, vacuum_handle);
1285                 return;
1286         }
1287
1288         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1289                 vacuum_handle->fast_path_count = 0;
1290         }
1291
1292         child_ctx->child_pid = fork();
1293         if (child_ctx->child_pid == (pid_t)-1) {
1294                 close(child_ctx->fd[0]);
1295                 close(child_ctx->fd[1]);
1296                 talloc_free(child_ctx);
1297                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1298                 event_add_timed(ctdb->ev, vacuum_handle,
1299                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1300                         ctdb_vacuum_event, vacuum_handle);
1301                 return;
1302         }
1303
1304
1305         if (child_ctx->child_pid == 0) {
1306                 char cc = 0;
1307                 bool full_vacuum_run = false;
1308                 close(child_ctx->fd[0]);
1309
1310                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1311
1312                 if (switch_from_server_to_client(ctdb) != 0) {
1313                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1314                         _exit(1);
1315                 }
1316
1317                 /*
1318                  * repack the db
1319                  */
1320                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1321                     (vacuum_handle->fast_path_count == 0))
1322                 {
1323                         full_vacuum_run = true;
1324                 }
1325                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1326                                                full_vacuum_run);
1327
1328                 write(child_ctx->fd[1], &cc, 1);
1329                 _exit(0);
1330         }
1331
1332         set_close_on_exec(child_ctx->fd[0]);
1333         close(child_ctx->fd[1]);
1334
1335         child_ctx->status = VACUUM_RUNNING;
1336         child_ctx->start_time = timeval_current();
1337
1338         DLIST_ADD(ctdb->vacuumers, child_ctx);
1339         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1340
1341         /*
1342          * Clear the fastpath vacuuming list in the parent.
1343          */
1344         talloc_free(ctdb_db->delete_queue);
1345         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1346         if (ctdb_db->delete_queue == NULL) {
1347                 /* fatal here? ... */
1348                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1349                                  "in parent context. Shutting down\n");
1350         }
1351
1352         event_add_timed(ctdb->ev, child_ctx,
1353                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1354                 vacuum_child_timeout, child_ctx);
1355
1356         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1357
1358         event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1359                 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1360                 vacuum_child_handler,
1361                 child_ctx);
1362
1363         vacuum_handle->child_ctx = child_ctx;
1364         child_ctx->vacuum_handle = vacuum_handle;
1365 }
1366
1367 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1368 {
1369         /* Simply free them all. */
1370         while (ctdb->vacuumers) {
1371                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1372                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1373                            (int)ctdb->vacuumers->child_pid));
1374                 /* vacuum_child_destructor kills it, removes from list */
1375                 talloc_free(ctdb->vacuumers);
1376         }
1377 }
1378
1379 /* this function initializes the vacuuming context for a database
1380  * starts the vacuuming events
1381  */
1382 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1383 {
1384         if (ctdb_db->persistent != 0) {
1385                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1386                 return 0;
1387         }
1388
1389         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1390         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1391
1392         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1393         ctdb_db->vacuum_handle->fast_path_count = 0;
1394
1395         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1396                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1397                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1398
1399         return 0;
1400 }
1401
1402 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1403                                             const struct ctdb_ltdb_header *hdr,
1404                                             const TDB_DATA key)
1405 {
1406         struct delete_record_data *kd;
1407         uint32_t hash;
1408
1409         hash = (uint32_t)tdb_jenkins_hash(discard_const(&key));
1410
1411         DEBUG(DEBUG_DEBUG, (__location__
1412                             " remove_record_from_delete_queue: db[%s] "
1413                             "db_id[0x%08x] "
1414                             "key_hash[0x%08x] "
1415                             "lmaster[%u] "
1416                             "migrated_with_data[%s]\n",
1417                              ctdb_db->db_name, ctdb_db->db_id,
1418                              hash,
1419                              ctdb_lmaster(ctdb_db->ctdb, &key),
1420                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1421
1422         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1423         if (kd == NULL) {
1424                 return;
1425         }
1426         if (kd->key.dsize != key.dsize) {
1427                 return;
1428         }
1429         if (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0) {
1430                 return;
1431         }
1432
1433         talloc_free(kd);
1434
1435         return;
1436 }
1437
1438 /**
1439  * Insert a record into the ctdb_db context's delete queue,
1440  * handling hash collisions.
1441  */
1442 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1443                                            const struct ctdb_ltdb_header *hdr,
1444                                            TDB_DATA key)
1445 {
1446         struct delete_record_data *kd;
1447         uint32_t hash;
1448         int ret;
1449
1450         hash = (uint32_t)tdb_jenkins_hash(&key);
1451
1452         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1453                            "db_id[0x%08x] "
1454                            "key_hash[0x%08x] "
1455                            "lmaster[%u] "
1456                            "migrated_with_data[%s]\n",
1457                             ctdb_db->db_name, ctdb_db->db_id,
1458                             hash,
1459                             ctdb_lmaster(ctdb_db->ctdb, &key),
1460                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1461
1462         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1463         if (kd != NULL) {
1464                 if ((kd->key.dsize != key.dsize) ||
1465                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1466                 {
1467                         DEBUG(DEBUG_INFO,
1468                               (__location__ " schedule for deletion: "
1469                                "hash collision (0x%08x)."
1470                                " Skipping the record.\n", hash));
1471                         return 0;
1472                 } else {
1473                         DEBUG(DEBUG_DEBUG,
1474                               (__location__ " schedule for deletion: "
1475                                "updating entry for key with hash 0x%08x.\n",
1476                                hash));
1477                 }
1478         }
1479
1480         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1481                                                   ctdb_db->delete_queue,
1482                                                   hdr, key);
1483         if (ret != 0) {
1484                 DEBUG(DEBUG_INFO,
1485                       (__location__ " schedule for deletion: error "
1486                        "inserting key with hash 0x%08x into delete queue\n",
1487                        hash));
1488                 return -1;
1489         }
1490
1491         return 0;
1492 }
1493
1494 /**
1495  * Schedule a record for deletetion.
1496  * Called from the parent context.
1497  */
1498 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1499                                            TDB_DATA indata)
1500 {
1501         struct ctdb_control_schedule_for_deletion *dd;
1502         struct ctdb_db_context *ctdb_db;
1503         int ret;
1504         TDB_DATA key;
1505
1506         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1507
1508         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1509         if (ctdb_db == NULL) {
1510                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1511                                   dd->db_id));
1512                 return -1;
1513         }
1514
1515         key.dsize = dd->keylen;
1516         key.dptr = dd->key;
1517
1518         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1519
1520         return ret;
1521 }
1522
1523 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1524                                          const struct ctdb_ltdb_header *hdr,
1525                                          TDB_DATA key)
1526 {
1527         int ret;
1528         struct ctdb_control_schedule_for_deletion *dd;
1529         TDB_DATA indata;
1530         int32_t status;
1531
1532         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1533                 /* main daemon - directly queue */
1534                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1535
1536                 return ret;
1537         }
1538
1539         /* child process: send the main daemon a control */
1540
1541         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1542         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1543         if (indata.dptr == NULL) {
1544                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1545                 return -1;
1546         }
1547         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1548         dd->db_id = ctdb_db->db_id;
1549         dd->hdr = *hdr;
1550         dd->keylen = key.dsize;
1551         memcpy(dd->key, key.dptr, key.dsize);
1552
1553         ret = ctdb_control(ctdb_db->ctdb,
1554                            CTDB_CURRENT_NODE,
1555                            ctdb_db->db_id,
1556                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1557                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1558                            indata,
1559                            NULL, /* mem_ctx */
1560                            NULL, /* outdata */
1561                            &status,
1562                            NULL, /* timeout : NULL == wait forever */
1563                            NULL); /* error message */
1564
1565         talloc_free(indata.dptr);
1566
1567         if (ret != 0 || status != 0) {
1568                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1569                                   "SCHEDULE_FOR_DELETION "
1570                                   "control.\n"));
1571                 if (status != 0) {
1572                         ret = -1;
1573                 }
1574         }
1575
1576         return ret;
1577 }
1578
1579 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1580                                          const struct ctdb_ltdb_header *hdr,
1581                                          const TDB_DATA key)
1582 {
1583         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1584                 /*
1585                  * Only remove the record from the delete queue if called
1586                  * in the main daemon.
1587                  */
1588                 return;
1589         }
1590
1591         remove_record_from_delete_queue(ctdb_db, hdr, key);
1592
1593         return;
1594 }