ctdb/server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program; if not, see <http://www.gnu.org/licenses/>.
  18 */
  19
  20 #include "includes.h"
  21 #include "lib/tevent/tevent.h"
  22 #include "lib/tdb/include/tdb.h"
  23 #include "system/network.h"
  24 #include "system/filesys.h"
  25 #include "system/dir.h"
  26 #include "../include/ctdb_private.h"
  27 #include "db_wrap.h"
  28 #include "lib/util/dlinklist.h"
  29 #include "lib/tevent/tevent.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34 #define TUNINGDBNAME "vactune.tdb"
  35
  36 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  37
  38 struct ctdb_vacuum_child_context {
  39         struct ctdb_vacuum_child_context *next, *prev;
  40         struct ctdb_vacuum_handle *vacuum_handle;
  41         /* fd child writes status to */
  42         int fd[2];
  43         pid_t child_pid;
  44         enum vacuum_child_status status;
  45         struct timeval start_time;
  46 };
  47
  48 struct ctdb_vacuum_handle {
  49         struct ctdb_db_context *ctdb_db;
  50         struct ctdb_vacuum_child_context *child_ctx;
  51         uint32_t fast_path_count;
  52 };
  53
  54
  55 /*  a list of records to possibly delete */
  56 struct vacuum_data {
  57         uint32_t vacuum_limit;
  58         uint32_t repack_limit;
  59         struct ctdb_context *ctdb;
  60         struct ctdb_db_context *ctdb_db;
  61         struct tdb_context *dest_db;
  62         trbt_tree_t *delete_tree;
  63         uint32_t delete_count;
  64         struct ctdb_marshall_buffer **list;
  65         struct timeval start;
  66         bool traverse_error;
  67         bool vacuum;
  68         uint32_t total;
  69         uint32_t vacuumed;
  70         uint32_t copied;
  71         uint32_t fast_added_to_vacuum_fetch_list;
  72         uint32_t fast_added_to_delete_tree;
  73         uint32_t fast_deleted;
  74         uint32_t fast_skipped;
  75         uint32_t fast_error;
  76         uint32_t fast_total;
  77         uint32_t full_added_to_vacuum_fetch_list;
  78         uint32_t full_added_to_delete_tree;
  79         uint32_t full_skipped;
  80         uint32_t full_error;
  81         uint32_t full_total;
  82 };
  83
  84 /* tuning information stored for every db */
  85 struct vacuum_tuning_data {
  86         uint32_t last_num_repack;
  87         uint32_t last_num_empty;
  88         uint32_t last_interval;
  89         uint32_t new_interval;
  90         struct timeval last_start;
  91         double   last_duration;
  92 };
  93
  94 /* this structure contains the information for one record to be deleted */
  95 struct delete_record_data {
  96         struct ctdb_context *ctdb;
  97         struct ctdb_db_context *ctdb_db;
  98         struct ctdb_ltdb_header hdr;
  99         TDB_DATA key;
 100 };
 101
 102 struct delete_records_list {
 103         struct ctdb_marshall_buffer *records;
 104 };
 105
 106 /**
 107  * Store key and header in a tree, indexed by the key hash.
 108  */
 109 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 110                                                struct ctdb_db_context *ctdb_db,
 111                                                trbt_tree_t *tree,
 112                                                const struct ctdb_ltdb_header *hdr,
 113                                                TDB_DATA key)
 114 {
 115         struct delete_record_data *dd;
 116         uint32_t hash;
 117
 118         dd = talloc_zero(tree, struct delete_record_data);
 119         if (dd == NULL) {
 120                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 121                 return -1;
 122         }
 123
 124         dd->ctdb      = ctdb;
 125         dd->ctdb_db   = ctdb_db;
 126         dd->key.dsize = key.dsize;
 127         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 128         if (dd->key.dptr == NULL) {
 129                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 130                 return -1;
 131         }
 132
 133         dd->hdr = *hdr;
 134
 135         hash = ctdb_hash(&key);
 136
 137         trbt_insert32(tree, hash, dd);
 138
 139         return 0;
 140 }
 141
 142 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
 143                                      struct ctdb_ltdb_header *hdr)
 144 {
 145         struct ctdb_context *ctdb = vdata->ctdb;
 146         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 147         uint32_t hash;
 148         int ret;
 149
 150         hash = ctdb_hash(&key);
 151
 152         if (trbt_lookup32(vdata->delete_tree, hash)) {
 153                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 154                 return 0;
 155         }
 156
 157         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 158                                                   vdata->delete_tree,
 159                                                   hdr, key);
 160         if (ret != 0) {
 161                 return -1;
 162         }
 163
 164         vdata->delete_count++;
 165
 166         return 0;
 167 }
 168
 169 /**
 170  * Add a record to the list of records to be sent
 171  * to their lmaster with VACUUM_FETCH.
 172  */
 173 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 174                                            TDB_DATA key)
 175 {
 176         struct ctdb_context *ctdb = vdata->ctdb;
 177         struct ctdb_rec_data *rec;
 178         uint32_t lmaster;
 179         size_t old_size;
 180
 181         lmaster = ctdb_lmaster(ctdb, &key);
 182
 183         rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
 184         if (rec == NULL) {
 185                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 186                 vdata->traverse_error = true;
 187                 return -1;
 188         }
 189
 190         old_size = talloc_get_size(vdata->list[lmaster]);
 191         vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
 192                                                    old_size + rec->length);
 193         if (vdata->list[lmaster] == NULL) {
 194                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 195                 vdata->traverse_error = true;
 196                 return -1;
 197         }
 198
 199         vdata->list[lmaster]->count++;
 200         memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
 201         talloc_free(rec);
 202
 203         vdata->total++;
 204
 205         return 0;
 206 }
 207
 208
 209 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 210                               struct timeval t, void *private_data);
 211
 212
 213 /*
 214  * traverse function for gathering the records that can be deleted
 215  */
 216 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 217 {
 218         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 219         struct ctdb_context *ctdb = vdata->ctdb;
 220         uint32_t lmaster;
 221         struct ctdb_ltdb_header *hdr;
 222         int res = 0;
 223
 224         vdata->full_total++;
 225
 226         lmaster = ctdb_lmaster(ctdb, &key);
 227         if (lmaster >= ctdb->num_nodes) {
 228                 vdata->full_error++;
 229                 DEBUG(DEBUG_CRIT, (__location__
 230                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 231                                    " with hash[%u]!\n",
 232                                    (unsigned)lmaster,
 233                                    (unsigned)ctdb->num_nodes,
 234                                    (unsigned)ctdb_hash(&key)));
 235                 return -1;
 236         }
 237
 238         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 239                 /* it is not a deleted record */
 240                 vdata->full_skipped++;
 241                 return 0;
 242         }
 243
 244         hdr = (struct ctdb_ltdb_header *)data.dptr;
 245
 246         if (hdr->dmaster != ctdb->pnn) {
 247                 vdata->full_skipped++;
 248                 return 0;
 249         }
 250
 251         if (lmaster == ctdb->pnn) {
 252                 /*
 253                  * We are both lmaster and dmaster, and the record is empty.
 254                  * So we should be able to delete it.
 255                  */
 256                 res = add_record_to_delete_tree(vdata, key, hdr);
 257                 if (res != 0) {
 258                         vdata->full_error++;
 259                 } else {
 260                         vdata->full_added_to_delete_tree++;
 261                 }
 262         } else {
 263                 /*
 264                  * We are not lmaster.
 265                  * Add the record to the blob ready to send to the nodes.
 266                  */
 267                 res = add_record_to_vacuum_fetch_list(vdata, key);
 268                 if (res != 0) {
 269                         vdata->full_error++;
 270                 } else {
 271                         vdata->full_added_to_vacuum_fetch_list++;
 272                 }
 273         }
 274
 275         return res;
 276 }
 277
 278 /*
 279  * traverse the tree of records to delete and marshall them into
 280  * a blob
 281  */
 282 static int delete_traverse(void *param, void *data)
 283 {
 284         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 285         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 286         struct ctdb_rec_data *rec;
 287         size_t old_size;
 288
 289         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 290         if (rec == NULL) {
 291                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 292                 return 0;
 293         }
 294
 295         old_size = talloc_get_size(recs->records);
 296         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 297         if (recs->records == NULL) {
 298                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 299                 return 0;
 300         }
 301         recs->records->count++;
 302         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 303         return 0;
 304 }
 305
 306 /**
 307  * traverse function for the traversal of the delete_queue,
 308  * the fast-path vacuuming list.
 309  *
 310  *  - If the record has been migrated off the node
 311  *    or has been revived (filled with data) on the node,
 312  *    then skip the record.
 313  *
 314  *  - If the current node is the record's lmaster and it is
 315  *    a record that has never been migrated with data, then
 316  *    delete the record from the local tdb.
 317  *
 318  *  - If the current node is the record's lmaster and it has
 319  *    been migrated with data, then schedule it for the normal
 320  *    vacuuming procedure (i.e. add it to the delete_list).
 321  *
 322  *  - If the current node is NOT the record's lmaster then
 323  *    add it to the list of records that are to be sent to
 324  *    the lmaster with the VACUUM_FETCH message.
 325  */
 326 static int delete_queue_traverse(void *param, void *data)
 327 {
 328         struct delete_record_data *dd =
 329                 talloc_get_type(data, struct delete_record_data);
 330         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 331         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 332         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 333         int res;
 334         struct ctdb_ltdb_header *header;
 335         TDB_DATA tdb_data;
 336         uint32_t lmaster;
 337
 338         vdata->fast_total++;
 339
 340         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 341         if (res != 0) {
 342                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 343                 vdata->fast_error++;
 344                 return 0;
 345         }
 346
 347         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 348         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 349                 /* Does not exist or not a ctdb record. Skip. */
 350                 goto skipped;
 351         }
 352
 353         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 354                 /* The record has been recycled (filled with data). Skip. */
 355                 goto skipped;
 356         }
 357
 358         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 359
 360         if (header->dmaster != ctdb->pnn) {
 361                 /* The record has been migrated off the node. Skip. */
 362                 goto skipped;
 363         }
 364
 365
 366         if (header->rsn != dd->hdr.rsn) {
 367                 /*
 368                  * The record has been migrated off the node and back again.
 369                  * But not requeued for deletion. Skip it.
 370                  */
 371                 goto skipped;
 372         }
 373
 374         /*
 375          * We are dmaster, and the record has no data, and it has
 376          * not been migrated after it has been queued for deletion.
 377          *
 378          * At this stage, the record could still have been revived locally
 379          * and last been written with empty data. This can only be
 380          * fixed with the addition of an active or delete flag. (TODO)
 381          */
 382
 383         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 384
 385         if (lmaster != ctdb->pnn) {
 386                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 387
 388                 if (res != 0) {
 389                         DEBUG(DEBUG_ERR,
 390                               (__location__ " Error adding record to list "
 391                                "of records to send to lmaster.\n"));
 392                         vdata->fast_error++;
 393                 } else {
 394                         vdata->fast_added_to_vacuum_fetch_list++;
 395                 }
 396                 goto done;
 397         }
 398
 399         /* use header->flags or dd->hdr.flags ?? */
 400         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 401                 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
 402
 403                 if (res != 0) {
 404                         DEBUG(DEBUG_ERR,
 405                               (__location__ " Error adding record to list "
 406                                "of records for deletion on lmaster.\n"));
 407                         vdata->fast_error++;
 408                 } else {
 409                         vdata->fast_added_to_delete_tree++;
 410                 }
 411         } else {
 412                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 413
 414                 if (res != 0) {
 415                         DEBUG(DEBUG_ERR,
 416                               (__location__ " Error deleting record from local "
 417                                "data base.\n"));
 418                         vdata->fast_error++;
 419                 } else {
 420                         vdata->fast_deleted++;
 421                 }
 422         }
 423
 424         goto done;
 425
 426 skipped:
 427         vdata->fast_skipped++;
 428
 429 done:
 430         if (tdb_data.dptr != NULL) {
 431                 free(tdb_data.dptr);
 432         }
 433         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 434
 435         return 0;
 436 }
 437
 438 /**
 439  * Vacuum a DB:
 440  *  - Always do the fast vacuuming run, which traverses
 441  *    the in-memory delete queue: these records have been
 442  *    scheduled for deletion.
 443  *  - Only if explicitly requested, the database is traversed
 444  *    in order to use the traditional heuristics on empty records
 445  *    to trigger deletion.
 446  *    This is done only every VacuumFastPathCount'th vacuuming run.
 447  *
 448  * The traverse runs fill two lists:
 449  *
 450  * - The delete_list:
 451  *   This is the list of empty records the current
 452  *   node is lmaster and dmaster for. These records are later
 453  *   deleted first on other nodes and then locally.
 454  *
 455  *   The fast vacuuming run has a short cut for those records
 456  *   that have never been migrated with data: these records
 457  *   are immediately deleted locally, since they have left
 458  *   no trace on other nodes.
 459  *
 460  * - The vacuum_fetch lists
 461  *   (one for each other lmaster node):
 462  *   The records in this list are sent for deletion to
 463  *   their lmaster in a bulk VACUUM_FETCH message.
 464  *
 465  *   The lmaster then migrates all these records to itelf
 466  *   so that they can be vacuumed there.
 467  *
 468  * This executes in the child context.
 469  */
 470 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 471                           struct vacuum_data *vdata,
 472                           bool full_vacuum_run)
 473 {
 474         struct ctdb_context *ctdb = ctdb_db->ctdb;
 475         const char *name = ctdb_db->db_name;
 476         int ret, i, pnn;
 477
 478         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 479                            "%s db_id[0x%08x]\n",
 480                            full_vacuum_run ? "full" : "fast",
 481                            ctdb_db->db_name, ctdb_db->db_id));
 482
 483         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 484         if (ret != 0) {
 485                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 486                 return ret;
 487         }
 488
 489         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 490         if (pnn == -1) {
 491                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 492                 return -1;
 493         }
 494
 495         ctdb->pnn = pnn;
 496
 497         vdata->fast_added_to_delete_tree = 0;
 498         vdata->fast_added_to_vacuum_fetch_list = 0;
 499         vdata->fast_deleted = 0;
 500         vdata->fast_skipped = 0;
 501         vdata->fast_error = 0;
 502         vdata->fast_total = 0;
 503         vdata->full_added_to_delete_tree = 0;
 504         vdata->full_added_to_vacuum_fetch_list = 0;
 505         vdata->full_skipped = 0;
 506         vdata->full_error = 0;
 507         vdata->full_total = 0;
 508
 509         /* the list needs to be of length num_nodes */
 510         vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
 511         if (vdata->list == NULL) {
 512                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 513                 return -1;
 514         }
 515         for (i = 0; i < ctdb->num_nodes; i++) {
 516                 vdata->list[i] = (struct ctdb_marshall_buffer *)
 517                         talloc_zero_size(vdata->list,
 518                                          offsetof(struct ctdb_marshall_buffer, data));
 519                 if (vdata->list[i] == NULL) {
 520                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 521                         return -1;
 522                 }
 523                 vdata->list[i]->db_id = ctdb_db->db_id;
 524         }
 525
 526         /*
 527          * Traverse the delete_queue.
 528          * This builds the same lists as the db traverse.
 529          */
 530         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 531
 532         if (vdata->fast_total > 0) {
 533                 DEBUG(DEBUG_INFO,
 534                       (__location__
 535                        " fast vacuuming delete_queue traverse statistics: "
 536                        "db[%s] "
 537                        "total[%u] "
 538                        "del[%u] "
 539                        "skp[%u] "
 540                        "err[%u] "
 541                        "adt[%u] "
 542                        "avf[%u]\n",
 543                        ctdb_db->db_name,
 544                        (unsigned)vdata->fast_total,
 545                        (unsigned)vdata->fast_deleted,
 546                        (unsigned)vdata->fast_skipped,
 547                        (unsigned)vdata->fast_error,
 548                        (unsigned)vdata->fast_added_to_delete_tree,
 549                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 550         }
 551
 552         /*
 553          * read-only traverse of the database, looking for records that
 554          * might be able to be vacuumed.
 555          *
 556          * This is not done each time but only every tunable
 557          * VacuumFastPathCount times.
 558          */
 559         if (full_vacuum_run) {
 560                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 561                 if (ret == -1 || vdata->traverse_error) {
 562                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 563                         return -1;
 564                 }
 565                 if (vdata->full_total > 0) {
 566                         DEBUG(DEBUG_INFO,
 567                               (__location__
 568                                " full vacuuming db traverse statistics: "
 569                                "db[%s] "
 570                                "total[%u] "
 571                                "skp[%u] "
 572                                "err[%u] "
 573                                "adt[%u] "
 574                                "avf[%u]\n",
 575                                ctdb_db->db_name,
 576                                (unsigned)vdata->full_total,
 577                                (unsigned)vdata->full_skipped,
 578                                (unsigned)vdata->full_error,
 579                                (unsigned)vdata->full_added_to_delete_tree,
 580                                (unsigned)vdata->full_added_to_vacuum_fetch_list));
 581                 }
 582         }
 583
 584         /*
 585          * For records where we are not the lmaster,
 586          * tell the lmaster to fetch the record.
 587          */
 588         for (i = 0; i < ctdb->num_nodes; i++) {
 589                 TDB_DATA data;
 590
 591                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 592                         continue;
 593                 }
 594
 595                 if (vdata->list[i]->count == 0) {
 596                         continue;
 597                 }
 598
 599                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 600                                    vdata->list[i]->count, ctdb->nodes[i]->pnn,
 601                                    name));
 602
 603                 data.dsize = talloc_get_size(vdata->list[i]);
 604                 data.dptr  = (void *)vdata->list[i];
 605                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 606                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 607                                           "fetch message to %u\n",
 608                                           ctdb->nodes[i]->pnn));
 609                         return -1;
 610                 }
 611         }
 612
 613         /* Process all records we can delete (if any) */
 614         if (vdata->delete_count > 0) {
 615                 struct delete_records_list *recs;
 616                 TDB_DATA indata, outdata;
 617                 int32_t res;
 618                 struct ctdb_node_map *nodemap;
 619                 uint32_t *active_nodes;
 620                 int num_active_nodes;
 621
 622                 recs = talloc_zero(vdata, struct delete_records_list);
 623                 if (recs == NULL) {
 624                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 625                         return -1;
 626                 }
 627                 recs->records = (struct ctdb_marshall_buffer *)
 628                         talloc_zero_size(vdata,
 629                                     offsetof(struct ctdb_marshall_buffer, data));
 630                 if (recs->records == NULL) {
 631                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 632                         return -1;
 633                 }
 634                 recs->records->db_id = ctdb_db->db_id;
 635
 636                 /*
 637                  * traverse the tree of all records we want to delete and
 638                  * create a blob we can send to the other nodes.
 639                  */
 640                 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
 641
 642                 indata.dsize = talloc_get_size(recs->records);
 643                 indata.dptr  = (void *)recs->records;
 644
 645                 /*
 646                  * now tell all the active nodes to delete all these records
 647                  * (if possible)
 648                  */
 649
 650                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 651                                            CTDB_CURRENT_NODE,
 652                                            recs, /* talloc context */
 653                                            &nodemap);
 654                 if (ret != 0) {
 655                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 656                         return -1;
 657                 }
 658
 659                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 660                                                     nodemap, /* talloc context */
 661                                                     false /* include self */);
 662                 /* yuck! ;-) */
 663                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 664
 665                 for (i = 0; i < num_active_nodes; i++) {
 666                         struct ctdb_marshall_buffer *records;
 667                         struct ctdb_rec_data *rec;
 668
 669                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 670                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 671                                         indata, recs, &outdata, &res,
 672                                         NULL, NULL);
 673                         if (ret != 0 || res != 0) {
 674                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 675                                                   "node %u: ret[%d] res[%d]\n",
 676                                                   active_nodes[i], ret, res));
 677                                 return -1;
 678                         }
 679
 680                         /*
 681                          * outdata countains the list of records coming back
 682                          * from the node which the node could not delete
 683                          */
 684                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 685                         rec = (struct ctdb_rec_data *)&records->data[0];
 686                         while (records->count-- > 1) {
 687                                 TDB_DATA reckey, recdata;
 688                                 struct ctdb_ltdb_header *rechdr;
 689
 690                                 reckey.dptr = &rec->data[0];
 691                                 reckey.dsize = rec->keylen;
 692                                 recdata.dptr = &rec->data[reckey.dsize];
 693                                 recdata.dsize = rec->datalen;
 694
 695                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 696                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 697                                         return -1;
 698                                 }
 699                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 700                                 recdata.dptr += sizeof(*rechdr);
 701                                 recdata.dsize -= sizeof(*rechdr);
 702
 703                                 /*
 704                                  * that other node couldnt delete the record
 705                                  * so we should delete it and thereby remove it from the tree
 706                                  */
 707                                 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
 708
 709                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 710                         }
 711                 }
 712
 713                 /* free nodemap and active_nodes */
 714                 talloc_free(nodemap);
 715
 716                 /*
 717                  * The only records remaining in the tree would be those
 718                  * records where all other nodes could successfully
 719                  * delete them, so we can safely delete them on the
 720                  * lmaster as well. Deletion implictely happens while
 721                  * we repack the database. The repack algorithm revisits
 722                  * the tree in order to find the records that don't need
 723                  * to be copied / repacked.
 724                  */
 725         }
 726
 727         /* this ensures we run our event queue */
 728         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 729
 730         return 0;
 731 }
 732
 733
 734 /*
 735  * traverse function for repacking
 736  */
 737 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 738 {
 739         struct vacuum_data *vdata = (struct vacuum_data *)private;
 740
 741         if (vdata->vacuum) {
 742                 uint32_t hash = ctdb_hash(&key);
 743                 struct delete_record_data *kd;
 744                 /*
 745                  * check if we can ignore this record because it's in the delete_tree
 746                  */
 747                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
 748                 /*
 749                  * there might be hash collisions so we have to compare the keys here to be sure
 750                  */
 751                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 752                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 753                         /*
 754                          * we have to check if the record hasn't changed in the meantime in order to
 755                          * savely remove it from the database
 756                          */
 757                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 758                                 hdr->dmaster == kd->ctdb->pnn &&
 759                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 760                                 kd->hdr.rsn == hdr->rsn) {
 761                                 vdata->vacuumed++;
 762                                 return 0;
 763                         }
 764                 }
 765         }
 766         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 767                 vdata->traverse_error = true;
 768                 return -1;
 769         }
 770         vdata->copied++;
 771         return 0;
 772 }
 773
 774 /*
 775  * repack a tdb
 776  */
 777 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 778 {
 779         struct tdb_context *tmp_db;
 780
 781         if (tdb_transaction_start(tdb) != 0) {
 782                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 783                 return -1;
 784         }
 785
 786         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 787                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 788                           O_RDWR|O_CREAT, 0);
 789         if (tmp_db == NULL) {
 790                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 791                 tdb_transaction_cancel(tdb);
 792                 return -1;
 793         }
 794
 795         vdata->traverse_error = false;
 796         vdata->dest_db = tmp_db;
 797         vdata->vacuum = true;
 798         vdata->vacuumed = 0;
 799         vdata->copied = 0;
 800
 801         /*
 802          * repack and vacuum on-the-fly by not writing the records that are
 803          * no longer needed
 804          */
 805         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 806                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 807                 tdb_transaction_cancel(tdb);
 808                 tdb_close(tmp_db);
 809                 return -1;
 810         }
 811
 812         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 813
 814         if (vdata->traverse_error) {
 815                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 816                 tdb_transaction_cancel(tdb);
 817                 tdb_close(tmp_db);
 818                 return -1;
 819         }
 820
 821         if (tdb_wipe_all(tdb) != 0) {
 822                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 823                 tdb_transaction_cancel(tdb);
 824                 tdb_close(tmp_db);
 825                 return -1;
 826         }
 827
 828         vdata->traverse_error = false;
 829         vdata->dest_db = tdb;
 830         vdata->vacuum = false;
 831         vdata->copied = 0;
 832
 833         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 834                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 835                 tdb_transaction_cancel(tdb);
 836                 tdb_close(tmp_db);
 837                 return -1;
 838         }
 839
 840         if (vdata->traverse_error) {
 841                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 842                 tdb_transaction_cancel(tdb);
 843                 tdb_close(tmp_db);
 844                 return -1;
 845         }
 846
 847         tdb_close(tmp_db);
 848
 849
 850         if (tdb_transaction_commit(tdb) != 0) {
 851                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 852                 return -1;
 853         }
 854         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 855
 856         return 0;
 857 }
 858
 859 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
 860 {
 861         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 862         TDB_CONTEXT *tune_tdb;
 863         TDB_DATA key, value;
 864         struct vacuum_tuning_data tdata;
 865         struct vacuum_tuning_data *tptr;
 866         char *vac_dbname;
 867         int flags;
 868
 869         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
 870                                      ctdb_db->ctdb->db_directory_state,
 871                                      TUNINGDBNAME, ctdb_db->ctdb->pnn);
 872         if (vac_dbname == NULL) {
 873                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
 874                 talloc_free(tmp_ctx);
 875                 return -1;
 876         }
 877
 878         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
 879         flags |= TDB_DISALLOW_NESTING;
 880         tune_tdb = tdb_open(vac_dbname, 0,
 881                             flags,
 882                             O_RDWR|O_CREAT, 0600);
 883         if (tune_tdb == NULL) {
 884                 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
 885                 talloc_free(tmp_ctx);
 886                 return -1;
 887         }
 888
 889         if (tdb_transaction_start(tune_tdb) != 0) {
 890                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 891                 tdb_close(tune_tdb);
 892                 return -1;
 893         }
 894         key.dptr = discard_const(ctdb_db->db_name);
 895         key.dsize = strlen(ctdb_db->db_name);
 896         value = tdb_fetch(tune_tdb, key);
 897
 898         if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
 899                 tptr = (struct vacuum_tuning_data *)value.dptr;
 900                 tdata = *tptr;
 901
 902                 /*
 903                  * re-calc new vacuum interval:
 904                  * in case no limit was reached we continuously increase the interval
 905                  * until vacuum_max_interval is reached
 906                  * in case a limit was reached we divide the current interval by 2
 907                  * unless vacuum_min_interval is reached
 908                  */
 909                 if (freelist < vdata->repack_limit &&
 910                     vdata->delete_count < vdata->vacuum_limit) {
 911                         if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
 912                                 tdata.new_interval = tdata.last_interval * 110 / 100;
 913                                 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
 914                                         tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 915                         }
 916                 } else {
 917                         tdata.new_interval = tdata.last_interval / 2;
 918                         if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
 919                                 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
 920                                 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
 921                         }
 922                         DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
 923                                          tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 924                 }
 925                 tdata.last_interval = tdata.new_interval;
 926         } else {
 927                 DEBUG(DEBUG_DEBUG,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
 928                 tdata.last_num_repack = freelist;
 929                 tdata.last_num_empty = vdata->delete_count;
 930                 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 931         }
 932
 933         if (value.dptr != NULL) {
 934                 free(value.dptr);
 935         }
 936
 937         tdata.last_start = vdata->start;
 938         tdata.last_duration = timeval_elapsed(&vdata->start);
 939
 940         value.dptr = (unsigned char *)&tdata;
 941         value.dsize = sizeof(tdata);
 942
 943         if (tdb_store(tune_tdb, key, value, 0) != 0) {
 944                 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
 945                 tdb_transaction_cancel(tune_tdb);
 946                 tdb_close(tune_tdb);
 947                 talloc_free(tmp_ctx);
 948                 return -1;
 949         }
 950         tdb_transaction_commit(tune_tdb);
 951         tdb_close(tune_tdb);
 952         talloc_free(tmp_ctx);
 953
 954         return 0;
 955 }
 956
 957 /*
 958  * repack and vaccum a db
 959  * called from the child context
 960  */
 961 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 962                                      TALLOC_CTX *mem_ctx,
 963                                      bool full_vacuum_run)
 964 {
 965         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
 966         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
 967         const char *name = ctdb_db->db_name;
 968         int freelist_size;
 969         struct vacuum_data *vdata;
 970
 971         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
 972         if (freelist_size == -1) {
 973                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
 974                 return -1;
 975         }
 976
 977         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 978         if (vdata == NULL) {
 979                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 980                 return -1;
 981         }
 982
 983         vdata->ctdb = ctdb_db->ctdb;
 984         vdata->vacuum_limit = vacuum_limit;
 985         vdata->repack_limit = repack_limit;
 986         vdata->delete_tree = trbt_create(vdata, 0);
 987         vdata->ctdb_db = ctdb_db;
 988         if (vdata->delete_tree == NULL) {
 989                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 990                 talloc_free(vdata);
 991                 return -1;
 992         }
 993
 994         vdata->start = timeval_current();
 995
 996         /*
 997          * gather all records that can be deleted in vdata
 998          */
 999         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1000                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1001         }
1002
1003         /*
1004          * decide if a repack is necessary
1005          */
1006         if (freelist_size < repack_limit && vdata->delete_count < vacuum_limit)
1007         {
1008                 update_tuning_db(ctdb_db, vdata, freelist_size);
1009                 talloc_free(vdata);
1010                 return 0;
1011         }
1012
1013         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1014                         name, freelist_size, vdata->delete_count));
1015
1016         /*
1017          * repack and implicitely get rid of the records we can delete
1018          */
1019         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1020                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1021                 update_tuning_db(ctdb_db, vdata, freelist_size);
1022                 talloc_free(vdata);
1023                 return -1;
1024         }
1025         update_tuning_db(ctdb_db, vdata, freelist_size);
1026         talloc_free(vdata);
1027
1028         return 0;
1029 }
1030
1031 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1032 {
1033         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1034         TDB_CONTEXT *tdb;
1035         TDB_DATA key, value;
1036         char *vac_dbname;
1037         uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
1038         struct ctdb_context *ctdb = ctdb_db->ctdb;
1039         int flags;
1040
1041         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
1042         if (vac_dbname == NULL) {
1043                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
1044                 talloc_free(tmp_ctx);
1045                 return interval;
1046         }
1047
1048         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
1049         flags |= TDB_DISALLOW_NESTING;
1050         tdb = tdb_open(vac_dbname, 0,
1051                        flags,
1052                        O_RDWR|O_CREAT, 0600);
1053         if (!tdb) {
1054                 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval. Errno : %s (%d)\n", vac_dbname, strerror(errno), errno));
1055                 talloc_free(tmp_ctx);
1056                 return interval;
1057         }
1058
1059         key.dptr = discard_const(ctdb_db->db_name);
1060         key.dsize = strlen(ctdb_db->db_name);
1061
1062         value = tdb_fetch(tdb, key);
1063
1064         if (value.dptr != NULL) {
1065                 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
1066                         struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
1067
1068                         interval = tptr->new_interval;
1069
1070                         if (interval < ctdb->tunable.vacuum_min_interval) {
1071                                 interval = ctdb->tunable.vacuum_min_interval;
1072                         }
1073                         if (interval > ctdb->tunable.vacuum_max_interval) {
1074                                 interval = ctdb->tunable.vacuum_max_interval;
1075                         }
1076                 }
1077                 free(value.dptr);
1078         }
1079         tdb_close(tdb);
1080
1081         talloc_free(tmp_ctx);
1082
1083         return interval;
1084 }
1085
1086 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1087 {
1088         double l = timeval_elapsed(&child_ctx->start_time);
1089         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1090         struct ctdb_context *ctdb = ctdb_db->ctdb;
1091
1092         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1093
1094         if (child_ctx->child_pid != -1) {
1095                 kill(child_ctx->child_pid, SIGKILL);
1096         } else {
1097                 /* Bump the number of successful fast-path runs. */
1098                 child_ctx->vacuum_handle->fast_path_count++;
1099         }
1100
1101         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1102
1103         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1104                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1105                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1106
1107         return 0;
1108 }
1109
1110 /*
1111  * this event is generated when a vacuum child process times out
1112  */
1113 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1114                                          struct timeval t, void *private_data)
1115 {
1116         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1117
1118         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1119
1120         child_ctx->status = VACUUM_TIMEOUT;
1121
1122         talloc_free(child_ctx);
1123 }
1124
1125
1126 /*
1127  * this event is generated when a vacuum child process has completed
1128  */
1129 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1130                              uint16_t flags, void *private_data)
1131 {
1132         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1133         char c = 0;
1134         int ret;
1135
1136         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1137         child_ctx->child_pid = -1;
1138
1139         ret = read(child_ctx->fd[0], &c, 1);
1140         if (ret != 1 || c != 0) {
1141                 child_ctx->status = VACUUM_ERROR;
1142                 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));
1143         } else {
1144                 child_ctx->status = VACUUM_OK;
1145         }
1146
1147         talloc_free(child_ctx);
1148 }
1149
1150 /*
1151  * this event is called every time we need to start a new vacuum process
1152  */
1153 static void
1154 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1155                                struct timeval t, void *private_data)
1156 {
1157         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1158         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1159         struct ctdb_context *ctdb = ctdb_db->ctdb;
1160         struct ctdb_vacuum_child_context *child_ctx;
1161         struct tevent_fd *fde;
1162         int ret;
1163
1164         /* we dont vacuum if we are in recovery mode, or db frozen */
1165         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1166             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1167                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1168                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1169                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1170                                    ? "freeze pending"
1171                                    : "frozen"));
1172                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1173                 return;
1174         }
1175
1176         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1177         if (child_ctx == NULL) {
1178                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1179                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1180         }
1181
1182
1183         ret = pipe(child_ctx->fd);
1184         if (ret != 0) {
1185                 talloc_free(child_ctx);
1186                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1187                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1188                 return;
1189         }
1190
1191         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1192                 vacuum_handle->fast_path_count = 0;
1193         }
1194
1195         child_ctx->child_pid = ctdb_fork(ctdb);
1196         if (child_ctx->child_pid == (pid_t)-1) {
1197                 close(child_ctx->fd[0]);
1198                 close(child_ctx->fd[1]);
1199                 talloc_free(child_ctx);
1200                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1201                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1202                 return;
1203         }
1204
1205
1206         if (child_ctx->child_pid == 0) {
1207                 char cc = 0;
1208                 bool full_vacuum_run = false;
1209                 close(child_ctx->fd[0]);
1210
1211                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1212
1213                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1214                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1215                         _exit(1);
1216                 }
1217
1218                 /*
1219                  * repack the db
1220                  */
1221                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1222                     (vacuum_handle->fast_path_count == 0))
1223                 {
1224                         full_vacuum_run = true;
1225                 }
1226                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1227                                                full_vacuum_run);
1228
1229                 write(child_ctx->fd[1], &cc, 1);
1230                 _exit(0);
1231         }
1232
1233         set_close_on_exec(child_ctx->fd[0]);
1234         close(child_ctx->fd[1]);
1235
1236         child_ctx->status = VACUUM_RUNNING;
1237         child_ctx->start_time = timeval_current();
1238
1239         DLIST_ADD(ctdb->vacuumers, child_ctx);
1240         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1241
1242         /*
1243          * Clear the fastpath vacuuming list in the parent.
1244          */
1245         talloc_free(ctdb_db->delete_queue);
1246         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1247         if (ctdb_db->delete_queue == NULL) {
1248                 /* fatal here? ... */
1249                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1250                                  "in parent context. Shutting down\n");
1251         }
1252
1253         event_add_timed(ctdb->ev, child_ctx,
1254                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1255                 vacuum_child_timeout, child_ctx);
1256
1257         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1258
1259         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1260                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1261         tevent_fd_set_auto_close(fde);
1262
1263         vacuum_handle->child_ctx = child_ctx;
1264         child_ctx->vacuum_handle = vacuum_handle;
1265 }
1266
1267 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1268 {
1269         /* Simply free them all. */
1270         while (ctdb->vacuumers) {
1271                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1272                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1273                            (int)ctdb->vacuumers->child_pid));
1274                 /* vacuum_child_destructor kills it, removes from list */
1275                 talloc_free(ctdb->vacuumers);
1276         }
1277 }
1278
1279 /* this function initializes the vacuuming context for a database
1280  * starts the vacuuming events
1281  */
1282 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1283 {
1284         if (ctdb_db->persistent != 0) {
1285                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1286                 return 0;
1287         }
1288
1289         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1290         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1291
1292         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1293         ctdb_db->vacuum_handle->fast_path_count = 0;
1294
1295         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1296                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1297                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1298
1299         return 0;
1300 }
1301
1302 /**
1303  * Insert a record into the ctdb_db context's delete queue,
1304  * handling hash collisions.
1305  */
1306 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1307                                            const struct ctdb_ltdb_header *hdr,
1308                                            TDB_DATA key)
1309 {
1310         struct delete_record_data *kd;
1311         uint32_t hash;
1312         int ret;
1313
1314         hash = (uint32_t)ctdb_hash(&key);
1315
1316         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1317                            "db_id[0x%08x] "
1318                            "key_hash[0x%08x] "
1319                            "lmaster[%u] "
1320                            "migrated_with_data[%s]\n",
1321                             ctdb_db->db_name, ctdb_db->db_id,
1322                             hash,
1323                             ctdb_lmaster(ctdb_db->ctdb, &key),
1324                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1325
1326         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1327         if (kd != NULL) {
1328                 if ((kd->key.dsize != key.dsize) ||
1329                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1330                 {
1331                         DEBUG(DEBUG_INFO,
1332                               ("schedule for deletion: Hash collision (0x%08x)."
1333                                " Skipping the record.\n", hash));
1334                         return 0;
1335                 } else {
1336                         DEBUG(DEBUG_DEBUG,
1337                               ("schedule for deletion: Overwriting entry for "
1338                                "key with hash 0x%08x.\n", hash));
1339                 }
1340         }
1341
1342         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1343                                                   ctdb_db->delete_queue,
1344                                                   hdr, key);
1345         if (ret != 0) {
1346                 return -1;
1347         }
1348
1349         return 0;
1350 }
1351
1352 /**
1353  * Schedule a record for deletetion.
1354  * Called from the parent context.
1355  */
1356 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1357                                            TDB_DATA indata)
1358 {
1359         struct ctdb_control_schedule_for_deletion *dd;
1360         struct ctdb_db_context *ctdb_db;
1361         int ret;
1362         TDB_DATA key;
1363
1364         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1365
1366         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1367         if (ctdb_db == NULL) {
1368                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1369                                   dd->db_id));
1370                 return -1;
1371         }
1372
1373         key.dsize = dd->keylen;
1374         key.dptr = dd->key;
1375
1376         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1377
1378         return ret;
1379 }
1380
1381 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1382                                          const struct ctdb_ltdb_header *hdr,
1383                                          TDB_DATA key)
1384 {
1385         int ret;
1386         struct ctdb_control_schedule_for_deletion *dd;
1387         TDB_DATA indata;
1388         int32_t status;
1389
1390         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1391                 /* main daemon - directly queue */
1392                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1393
1394                 return ret;
1395         }
1396
1397         /* child process: send the main daemon a control */
1398
1399         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1400         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1401         if (indata.dptr == NULL) {
1402                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1403                 return -1;
1404         }
1405         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1406         dd->db_id = ctdb_db->db_id;
1407         dd->hdr = *hdr;
1408         dd->keylen = key.dsize;
1409         memcpy(dd->key, key.dptr, key.dsize);
1410
1411         ret = ctdb_control(ctdb_db->ctdb,
1412                            CTDB_CURRENT_NODE,
1413                            ctdb_db->db_id,
1414                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1415                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1416                            indata,
1417                            NULL, /* mem_ctx */
1418                            NULL, /* outdata */
1419                            &status,
1420                            NULL, /* timeout : NULL == wait forever */
1421                            NULL); /* error message */
1422
1423         talloc_free(indata.dptr);
1424
1425         if (ret != 0 || status != 0) {
1426                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1427                                   "SCHEDULE_FOR_DELETION "
1428                                   "control.\n"));
1429                 if (status != 0) {
1430                         ret = -1;
1431                 }
1432         }
1433
1434         return ret;
1435 }