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