server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program; if not, see <http://www.gnu.org/licenses/>.
  18 */
  19
  20 #include "includes.h"
  21 #include "lib/events/events.h"
  22 #include "lib/tdb/include/tdb.h"
  23 #include "system/network.h"
  24 #include "system/filesys.h"
  25 #include "system/dir.h"
  26 #include "../include/ctdb_private.h"
  27 #include "db_wrap.h"
  28 #include "lib/util/dlinklist.h"
  29 #include "lib/events/events.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34 #define TUNINGDBNAME "vactune.tdb"
  35
  36 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  37
  38 struct ctdb_vacuum_child_context {
  39         struct ctdb_vacuum_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 = (uint32_t)tdb_jenkins_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 = (uint32_t)tdb_jenkins_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  * read-only traverse the database in order to find
 440  * records that can be deleted and try to delete these
 441  * records on the other nodes
 442  * this executes in the child context
 443  */
 444 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 445                           struct vacuum_data *vdata,
 446                           bool full_vacuum_run)
 447 {
 448         struct ctdb_context *ctdb = ctdb_db->ctdb;
 449         const char *name = ctdb_db->db_name;
 450         int ret, i, pnn;
 451
 452         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 453                            "%s db_id[0x%08x]\n",
 454                            full_vacuum_run ? "full" : "fast",
 455                            ctdb_db->db_name, ctdb_db->db_id));
 456
 457         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 458         if (ret != 0) {
 459                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 460                 return ret;
 461         }
 462
 463         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 464         if (pnn == -1) {
 465                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 466                 return -1;
 467         }
 468
 469         ctdb->pnn = pnn;
 470
 471         vdata->fast_added_to_delete_tree = 0;
 472         vdata->fast_added_to_vacuum_fetch_list = 0;
 473         vdata->fast_deleted = 0;
 474         vdata->fast_skipped = 0;
 475         vdata->fast_error = 0;
 476         vdata->fast_total = 0;
 477         vdata->full_added_to_delete_tree = 0;
 478         vdata->full_added_to_vacuum_fetch_list = 0;
 479         vdata->full_skipped = 0;
 480         vdata->full_error = 0;
 481         vdata->full_total = 0;
 482
 483         /* the list needs to be of length num_nodes */
 484         vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
 485         if (vdata->list == NULL) {
 486                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 487                 return -1;
 488         }
 489         for (i = 0; i < ctdb->num_nodes; i++) {
 490                 vdata->list[i] = (struct ctdb_marshall_buffer *)
 491                         talloc_zero_size(vdata->list,
 492                                          offsetof(struct ctdb_marshall_buffer, data));
 493                 if (vdata->list[i] == NULL) {
 494                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 495                         return -1;
 496                 }
 497                 vdata->list[i]->db_id = ctdb_db->db_id;
 498         }
 499
 500         /*
 501          * Traverse the delete_queue.
 502          * This builds the same lists as the db traverse.
 503          */
 504         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 505
 506         if (vdata->fast_total > 0) {
 507                 DEBUG(DEBUG_INFO,
 508                       (__location__
 509                        " fast vacuuming delete_queue traverse statistics: "
 510                        "db[%s] "
 511                        "total[%u] "
 512                        "del[%u] "
 513                        "skp[%u] "
 514                        "err[%u] "
 515                        "adt[%u] "
 516                        "avf[%u]\n",
 517                        ctdb_db->db_name,
 518                        (unsigned)vdata->fast_total,
 519                        (unsigned)vdata->fast_deleted,
 520                        (unsigned)vdata->fast_skipped,
 521                        (unsigned)vdata->fast_error,
 522                        (unsigned)vdata->fast_added_to_delete_tree,
 523                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 524         }
 525
 526         /*
 527          * read-only traverse of the database, looking for records that
 528          * might be able to be vacuumed.
 529          *
 530          * This is not done each time but only every tunable
 531          * VacuumFastPathCount times.
 532          */
 533         if (full_vacuum_run) {
 534                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 535                 if (ret == -1 || vdata->traverse_error) {
 536                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 537                         return -1;
 538                 }
 539                 if (vdata->full_total > 0) {
 540                         DEBUG(DEBUG_INFO,
 541                               (__location__
 542                                " full vacuuming db traverse statistics: "
 543                                "db[%s] "
 544                                "total[%u] "
 545                                "skp[%u] "
 546                                "err[%u] "
 547                                "adt[%u] "
 548                                "avf[%u]\n",
 549                                ctdb_db->db_name,
 550                                (unsigned)vdata->full_total,
 551                                (unsigned)vdata->full_skipped,
 552                                (unsigned)vdata->full_error,
 553                                (unsigned)vdata->full_added_to_delete_tree,
 554                                (unsigned)vdata->full_added_to_vacuum_fetch_list));
 555                 }
 556         }
 557
 558         /*
 559          * For records where we are not the lmaster,
 560          * tell the lmaster to fetch the record.
 561          */
 562         for (i = 0; i < ctdb->num_nodes; i++) {
 563                 TDB_DATA data;
 564
 565                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 566                         continue;
 567                 }
 568
 569                 if (vdata->list[i]->count == 0) {
 570                         continue;
 571                 }
 572
 573                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 574                                    vdata->list[i]->count, ctdb->nodes[i]->pnn,
 575                                    name));
 576
 577                 data.dsize = talloc_get_size(vdata->list[i]);
 578                 data.dptr  = (void *)vdata->list[i];
 579                 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 580                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 581                                           "fetch message to %u\n",
 582                                           ctdb->nodes[i]->pnn));
 583                         return -1;
 584                 }
 585         }
 586
 587         /* Process all records we can delete (if any) */
 588         if (vdata->delete_count > 0) {
 589                 struct delete_records_list *recs;
 590                 TDB_DATA indata, outdata;
 591                 int32_t res;
 592                 struct ctdb_node_map *nodemap;
 593                 uint32_t *active_nodes;
 594                 int num_active_nodes;
 595
 596                 recs = talloc_zero(vdata, struct delete_records_list);
 597                 if (recs == NULL) {
 598                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 599                         return -1;
 600                 }
 601                 recs->records = (struct ctdb_marshall_buffer *)
 602                         talloc_zero_size(vdata,
 603                                     offsetof(struct ctdb_marshall_buffer, data));
 604                 if (recs->records == NULL) {
 605                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 606                         return -1;
 607                 }
 608                 recs->records->db_id = ctdb_db->db_id;
 609
 610                 /*
 611                  * traverse the tree of all records we want to delete and
 612                  * create a blob we can send to the other nodes.
 613                  */
 614                 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
 615
 616                 indata.dsize = talloc_get_size(recs->records);
 617                 indata.dptr  = (void *)recs->records;
 618
 619                 /*
 620                  * now tell all the active nodes to delete all these records
 621                  * (if possible)
 622                  */
 623
 624                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 625                                            CTDB_CURRENT_NODE,
 626                                            recs, /* talloc context */
 627                                            &nodemap);
 628                 if (ret != 0) {
 629                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 630                         return -1;
 631                 }
 632
 633                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 634                                                     nodemap, /* talloc context */
 635                                                     false /* include self */);
 636                 /* yuck! ;-) */
 637                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 638
 639                 for (i = 0; i < num_active_nodes; i++) {
 640                         struct ctdb_marshall_buffer *records;
 641                         struct ctdb_rec_data *rec;
 642
 643                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 644                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 645                                         indata, recs, &outdata, &res,
 646                                         NULL, NULL);
 647                         if (ret != 0 || res != 0) {
 648                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 649                                                   "node %u: ret[%d] res[%d]\n",
 650                                                   active_nodes[i], ret, res));
 651                                 return -1;
 652                         }
 653
 654                         /*
 655                          * outdata countains the list of records coming back
 656                          * from the node which the node could not delete
 657                          */
 658                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 659                         rec = (struct ctdb_rec_data *)&records->data[0];
 660                         while (records->count-- > 1) {
 661                                 TDB_DATA reckey, recdata;
 662                                 struct ctdb_ltdb_header *rechdr;
 663
 664                                 reckey.dptr = &rec->data[0];
 665                                 reckey.dsize = rec->keylen;
 666                                 recdata.dptr = &rec->data[reckey.dsize];
 667                                 recdata.dsize = rec->datalen;
 668
 669                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 670                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 671                                         return -1;
 672                                 }
 673                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 674                                 recdata.dptr += sizeof(*rechdr);
 675                                 recdata.dsize -= sizeof(*rechdr);
 676
 677                                 /*
 678                                  * that other node couldnt delete the record
 679                                  * so we should delete it and thereby remove it from the tree
 680                                  */
 681                                 talloc_free(trbt_lookup32(vdata->delete_tree,
 682                                                           (uint32_t)tdb_jenkins_hash(&reckey)));
 683
 684                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 685                         }
 686                 }
 687
 688                 /* free nodemap and active_nodes */
 689                 talloc_free(nodemap);
 690
 691                 /*
 692                  * The only records remaining in the tree would be those
 693                  * records where all other nodes could successfully
 694                  * delete them, so we can safely delete them on the
 695                  * lmaster as well. Deletion implictely happens while
 696                  * we repack the database. The repack algorithm revisits
 697                  * the tree in order to find the records that don't need
 698                  * to be copied / repacked.
 699                  */
 700         }
 701
 702         /* this ensures we run our event queue */
 703         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 704
 705         return 0;
 706 }
 707
 708
 709 /*
 710  * traverse function for repacking
 711  */
 712 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 713 {
 714         struct vacuum_data *vdata = (struct vacuum_data *)private;
 715
 716         if (vdata->vacuum) {
 717                 uint32_t hash = (uint32_t)tdb_jenkins_hash(&key);
 718                 struct delete_record_data *kd;
 719                 /*
 720                  * check if we can ignore this record because it's in the delete_tree
 721                  */
 722                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
 723                 /*
 724                  * there might be hash collisions so we have to compare the keys here to be sure
 725                  */
 726                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 727                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 728                         /*
 729                          * we have to check if the record hasn't changed in the meantime in order to
 730                          * savely remove it from the database
 731                          */
 732                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 733                                 hdr->dmaster == kd->ctdb->pnn &&
 734                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 735                                 kd->hdr.rsn == hdr->rsn) {
 736                                 vdata->vacuumed++;
 737                                 return 0;
 738                         }
 739                 }
 740         }
 741         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 742                 vdata->traverse_error = true;
 743                 return -1;
 744         }
 745         vdata->copied++;
 746         return 0;
 747 }
 748
 749 /*
 750  * repack a tdb
 751  */
 752 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 753 {
 754         struct tdb_context *tmp_db;
 755
 756         if (tdb_transaction_start(tdb) != 0) {
 757                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 758                 return -1;
 759         }
 760
 761         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 762                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 763                           O_RDWR|O_CREAT, 0);
 764         if (tmp_db == NULL) {
 765                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 766                 tdb_transaction_cancel(tdb);
 767                 return -1;
 768         }
 769
 770         vdata->traverse_error = false;
 771         vdata->dest_db = tmp_db;
 772         vdata->vacuum = true;
 773         vdata->vacuumed = 0;
 774         vdata->copied = 0;
 775
 776         /*
 777          * repack and vacuum on-the-fly by not writing the records that are
 778          * no longer needed
 779          */
 780         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 781                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 782                 tdb_transaction_cancel(tdb);
 783                 tdb_close(tmp_db);
 784                 return -1;
 785         }
 786
 787         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 788
 789         if (vdata->traverse_error) {
 790                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 791                 tdb_transaction_cancel(tdb);
 792                 tdb_close(tmp_db);
 793                 return -1;
 794         }
 795
 796         if (tdb_wipe_all(tdb) != 0) {
 797                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 798                 tdb_transaction_cancel(tdb);
 799                 tdb_close(tmp_db);
 800                 return -1;
 801         }
 802
 803         vdata->traverse_error = false;
 804         vdata->dest_db = tdb;
 805         vdata->vacuum = false;
 806         vdata->copied = 0;
 807
 808         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 809                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 810                 tdb_transaction_cancel(tdb);
 811                 tdb_close(tmp_db);
 812                 return -1;
 813         }
 814
 815         if (vdata->traverse_error) {
 816                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 817                 tdb_transaction_cancel(tdb);
 818                 tdb_close(tmp_db);
 819                 return -1;
 820         }
 821
 822         tdb_close(tmp_db);
 823
 824
 825         if (tdb_transaction_commit(tdb) != 0) {
 826                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 827                 return -1;
 828         }
 829         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 830
 831         return 0;
 832 }
 833
 834 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
 835 {
 836         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 837         TDB_CONTEXT *tune_tdb;
 838         TDB_DATA key, value;
 839         struct vacuum_tuning_data tdata;
 840         struct vacuum_tuning_data *tptr;
 841         char *vac_dbname;
 842         int flags;
 843
 844         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
 845                                      ctdb_db->ctdb->db_directory_state,
 846                                      TUNINGDBNAME, ctdb_db->ctdb->pnn);
 847         if (vac_dbname == NULL) {
 848                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
 849                 talloc_free(tmp_ctx);
 850                 return -1;
 851         }
 852
 853         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
 854         flags |= TDB_DISALLOW_NESTING;
 855         tune_tdb = tdb_open(vac_dbname, 0,
 856                             flags,
 857                             O_RDWR|O_CREAT, 0600);
 858         if (tune_tdb == NULL) {
 859                 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
 860                 talloc_free(tmp_ctx);
 861                 return -1;
 862         }
 863
 864         if (tdb_transaction_start(tune_tdb) != 0) {
 865                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 866                 tdb_close(tune_tdb);
 867                 return -1;
 868         }
 869         key.dptr = discard_const(ctdb_db->db_name);
 870         key.dsize = strlen(ctdb_db->db_name);
 871         value = tdb_fetch(tune_tdb, key);
 872
 873         if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
 874                 tptr = (struct vacuum_tuning_data *)value.dptr;
 875                 tdata = *tptr;
 876
 877                 /*
 878                  * re-calc new vacuum interval:
 879                  * in case no limit was reached we continuously increase the interval
 880                  * until vacuum_max_interval is reached
 881                  * in case a limit was reached we divide the current interval by 2
 882                  * unless vacuum_min_interval is reached
 883                  */
 884                 if (freelist < vdata->repack_limit &&
 885                     vdata->delete_count < vdata->vacuum_limit) {
 886                         if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
 887                                 tdata.new_interval = tdata.last_interval * 110 / 100;
 888                                 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
 889                                         tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 890                         }
 891                 } else {
 892                         tdata.new_interval = tdata.last_interval / 2;
 893                         if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
 894                                 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
 895                                 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
 896                         }
 897                         DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
 898                                          tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 899                 }
 900                 tdata.last_interval = tdata.new_interval;
 901         } else {
 902                 DEBUG(DEBUG_ERR,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
 903                 tdata.last_num_repack = freelist;
 904                 tdata.last_num_empty = vdata->delete_count;
 905                 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 906         }
 907
 908         if (value.dptr != NULL) {
 909                 free(value.dptr);
 910         }
 911
 912         tdata.last_start = vdata->start;
 913         tdata.last_duration = timeval_elapsed(&vdata->start);
 914
 915         value.dptr = (unsigned char *)&tdata;
 916         value.dsize = sizeof(tdata);
 917
 918         if (tdb_store(tune_tdb, key, value, 0) != 0) {
 919                 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
 920                 tdb_transaction_cancel(tune_tdb);
 921                 tdb_close(tune_tdb);
 922                 talloc_free(tmp_ctx);
 923                 return -1;
 924         }
 925         tdb_transaction_commit(tune_tdb);
 926         tdb_close(tune_tdb);
 927         talloc_free(tmp_ctx);
 928
 929         return 0;
 930 }
 931
 932 /*
 933  * repack and vaccum a db
 934  * called from the child context
 935  */
 936 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 937                                      TALLOC_CTX *mem_ctx,
 938                                      bool full_vacuum_run)
 939 {
 940         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
 941         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
 942         const char *name = ctdb_db->db_name;
 943         int freelist_size;
 944         struct vacuum_data *vdata;
 945
 946         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
 947         if (freelist_size == -1) {
 948                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
 949                 return -1;
 950         }
 951
 952         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 953         if (vdata == NULL) {
 954                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 955                 return -1;
 956         }
 957
 958         vdata->ctdb = ctdb_db->ctdb;
 959         vdata->vacuum_limit = vacuum_limit;
 960         vdata->repack_limit = repack_limit;
 961         vdata->delete_tree = trbt_create(vdata, 0);
 962         vdata->ctdb_db = ctdb_db;
 963         if (vdata->delete_tree == NULL) {
 964                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 965                 talloc_free(vdata);
 966                 return -1;
 967         }
 968
 969         vdata->start = timeval_current();
 970
 971         /*
 972          * gather all records that can be deleted in vdata
 973          */
 974         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
 975                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
 976         }
 977
 978         /*
 979          * decide if a repack is necessary
 980          */
 981         if (freelist_size < repack_limit && vdata->delete_count < vacuum_limit)
 982         {
 983                 update_tuning_db(ctdb_db, vdata, freelist_size);
 984                 talloc_free(vdata);
 985                 return 0;
 986         }
 987
 988         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
 989                         name, freelist_size, vdata->delete_count));
 990
 991         /*
 992          * repack and implicitely get rid of the records we can delete
 993          */
 994         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
 995                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
 996                 update_tuning_db(ctdb_db, vdata, freelist_size);
 997                 talloc_free(vdata);
 998                 return -1;
 999         }
1000         update_tuning_db(ctdb_db, vdata, freelist_size);
1001         talloc_free(vdata);
1002
1003         return 0;
1004 }
1005
1006 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1007 {
1008         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
1009         TDB_CONTEXT *tdb;
1010         TDB_DATA key, value;
1011         char *vac_dbname;
1012         uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
1013         struct ctdb_context *ctdb = ctdb_db->ctdb;
1014         int flags;
1015
1016         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
1017         if (vac_dbname == NULL) {
1018                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
1019                 talloc_free(tmp_ctx);
1020                 return interval;
1021         }
1022
1023         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
1024         flags |= TDB_DISALLOW_NESTING;
1025         tdb = tdb_open(vac_dbname, 0,
1026                        flags,
1027                        O_RDWR|O_CREAT, 0600);
1028         if (!tdb) {
1029                 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval\n", vac_dbname));
1030                 talloc_free(tmp_ctx);
1031                 return interval;
1032         }
1033
1034         key.dptr = discard_const(ctdb_db->db_name);
1035         key.dsize = strlen(ctdb_db->db_name);
1036
1037         value = tdb_fetch(tdb, key);
1038
1039         if (value.dptr != NULL) {
1040                 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
1041                         struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
1042
1043                         interval = tptr->new_interval;
1044
1045                         if (interval < ctdb->tunable.vacuum_min_interval) {
1046                                 interval = ctdb->tunable.vacuum_min_interval;
1047                         }
1048                         if (interval > ctdb->tunable.vacuum_max_interval) {
1049                                 interval = ctdb->tunable.vacuum_max_interval;
1050                         }
1051                 }
1052                 free(value.dptr);
1053         }
1054         tdb_close(tdb);
1055
1056         talloc_free(tmp_ctx);
1057
1058         return interval;
1059 }
1060
1061 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1062 {
1063         double l = timeval_elapsed(&child_ctx->start_time);
1064         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1065         struct ctdb_context *ctdb = ctdb_db->ctdb;
1066
1067         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1068
1069         if (child_ctx->child_pid != -1) {
1070                 kill(child_ctx->child_pid, SIGKILL);
1071         } else {
1072                 /* Bump the number of successful fast-path runs. */
1073                 child_ctx->vacuum_handle->fast_path_count++;
1074         }
1075
1076         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1077
1078         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1079                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1080                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1081
1082         return 0;
1083 }
1084
1085 /*
1086  * this event is generated when a vacuum child process times out
1087  */
1088 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1089                                          struct timeval t, void *private_data)
1090 {
1091         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1092
1093         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1094
1095         child_ctx->status = VACUUM_TIMEOUT;
1096
1097         talloc_free(child_ctx);
1098 }
1099
1100
1101 /*
1102  * this event is generated when a vacuum child process has completed
1103  */
1104 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1105                              uint16_t flags, void *private_data)
1106 {
1107         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1108         char c = 0;
1109         int ret;
1110
1111         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1112         child_ctx->child_pid = -1;
1113
1114         ret = read(child_ctx->fd[0], &c, 1);
1115         if (ret != 1 || c != 0) {
1116                 child_ctx->status = VACUUM_ERROR;
1117                 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));
1118         } else {
1119                 child_ctx->status = VACUUM_OK;
1120         }
1121
1122         talloc_free(child_ctx);
1123 }
1124
1125 /*
1126  * this event is called every time we need to start a new vacuum process
1127  */
1128 static void
1129 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1130                                struct timeval t, void *private_data)
1131 {
1132         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1133         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1134         struct ctdb_context *ctdb = ctdb_db->ctdb;
1135         struct ctdb_vacuum_child_context *child_ctx;
1136         int ret;
1137
1138         /* we dont vacuum if we are in recovery mode, or db frozen */
1139         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1140             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1141                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1142                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1143                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1144                                    ? "freeze pending"
1145                                    : "frozen"));
1146                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1147                 return;
1148         }
1149
1150         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1151         if (child_ctx == NULL) {
1152                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1153                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1154         }
1155
1156
1157         ret = pipe(child_ctx->fd);
1158         if (ret != 0) {
1159                 talloc_free(child_ctx);
1160                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1161                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1162                 return;
1163         }
1164
1165         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1166                 vacuum_handle->fast_path_count = 0;
1167         }
1168
1169         child_ctx->child_pid = fork();
1170         if (child_ctx->child_pid == (pid_t)-1) {
1171                 close(child_ctx->fd[0]);
1172                 close(child_ctx->fd[1]);
1173                 talloc_free(child_ctx);
1174                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1175                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1176                 return;
1177         }
1178
1179
1180         if (child_ctx->child_pid == 0) {
1181                 char cc = 0;
1182                 bool full_vacuum_run = false;
1183                 close(child_ctx->fd[0]);
1184
1185                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1186
1187                 if (switch_from_server_to_client(ctdb) != 0) {
1188                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1189                         _exit(1);
1190                 }
1191
1192                 /*
1193                  * repack the db
1194                  */
1195                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1196                     (vacuum_handle->fast_path_count == 0))
1197                 {
1198                         full_vacuum_run = true;
1199                 }
1200                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1201                                                full_vacuum_run);
1202
1203                 write(child_ctx->fd[1], &cc, 1);
1204                 _exit(0);
1205         }
1206
1207         set_close_on_exec(child_ctx->fd[0]);
1208         close(child_ctx->fd[1]);
1209
1210         child_ctx->status = VACUUM_RUNNING;
1211         child_ctx->start_time = timeval_current();
1212
1213         DLIST_ADD(ctdb->vacuumers, child_ctx);
1214         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1215
1216         /*
1217          * Clear the fastpath vacuuming list in the parent.
1218          */
1219         talloc_free(ctdb_db->delete_queue);
1220         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1221         if (ctdb_db->delete_queue == NULL) {
1222                 /* fatal here? ... */
1223                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1224                                  "in parent context. Shutting down\n");
1225         }
1226
1227         event_add_timed(ctdb->ev, child_ctx,
1228                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1229                 vacuum_child_timeout, child_ctx);
1230
1231         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1232
1233         event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1234                 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1235                 vacuum_child_handler,
1236                 child_ctx);
1237
1238         vacuum_handle->child_ctx = child_ctx;
1239         child_ctx->vacuum_handle = vacuum_handle;
1240 }
1241
1242 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1243 {
1244         /* Simply free them all. */
1245         while (ctdb->vacuumers) {
1246                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1247                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1248                            (int)ctdb->vacuumers->child_pid));
1249                 /* vacuum_child_destructor kills it, removes from list */
1250                 talloc_free(ctdb->vacuumers);
1251         }
1252 }
1253
1254 /* this function initializes the vacuuming context for a database
1255  * starts the vacuuming events
1256  */
1257 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1258 {
1259         if (ctdb_db->persistent != 0) {
1260                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1261                 return 0;
1262         }
1263
1264         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1265         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1266
1267         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1268         ctdb_db->vacuum_handle->fast_path_count = 0;
1269
1270         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1271                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1272                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1273
1274         return 0;
1275 }
1276
1277 /**
1278  * Insert a record into the ctdb_db context's delete queue,
1279  * handling hash collisions.
1280  */
1281 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1282                                            const struct ctdb_ltdb_header *hdr,
1283                                            TDB_DATA key)
1284 {
1285         struct delete_record_data *kd;
1286         uint32_t hash;
1287         int ret;
1288
1289         hash = (uint32_t)tdb_jenkins_hash(&key);
1290
1291         DEBUG(DEBUG_INFO, (__location__ " Schedule for deletion: db[%s] "
1292                            "db_id[0x%08x] "
1293                            "key_hash[0x%08x] "
1294                            "lmaster[%u] "
1295                            "migrated_with_data[%s]\n",
1296                             ctdb_db->db_name, ctdb_db->db_id,
1297                             hash,
1298                             ctdb_lmaster(ctdb_db->ctdb, &key),
1299                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1300
1301         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1302         if (kd != NULL) {
1303                 if ((kd->key.dsize != key.dsize) ||
1304                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1305                 {
1306                         DEBUG(DEBUG_INFO,
1307                               ("schedule for deletion: Hash collision (0x%08x)."
1308                                " Skipping the record.\n", hash));
1309                         return 0;
1310                 } else {
1311                         DEBUG(DEBUG_DEBUG,
1312                               ("schedule for deletion: Overwriting entry for "
1313                                "key with hash 0x%08x.\n", hash));
1314                 }
1315         }
1316
1317         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1318                                                   ctdb_db->delete_queue,
1319                                                   hdr, key);
1320         if (ret != 0) {
1321                 return -1;
1322         }
1323
1324         return 0;
1325 }
1326
1327 /**
1328  * Schedule a record for deletetion.
1329  * Called from the parent context.
1330  */
1331 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1332                                            TDB_DATA indata)
1333 {
1334         struct ctdb_control_schedule_for_deletion *dd;
1335         struct ctdb_db_context *ctdb_db;
1336         int ret;
1337         TDB_DATA key;
1338
1339         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1340
1341         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1342         if (ctdb_db == NULL) {
1343                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1344                                   dd->db_id));
1345                 return -1;
1346         }
1347
1348         key.dsize = dd->keylen;
1349         key.dptr = dd->key;
1350
1351         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1352
1353         return ret;
1354 }
1355
1356 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1357                                          const struct ctdb_ltdb_header *hdr,
1358                                          TDB_DATA key)
1359 {
1360         int ret;
1361         struct ctdb_control_schedule_for_deletion *dd;
1362         TDB_DATA indata;
1363         int32_t status;
1364
1365         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1366                 /* main daemon - directly queue */
1367                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1368
1369                 return ret;
1370         }
1371
1372         /* child process: send the main daemon a control */
1373
1374         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1375         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1376         if (indata.dptr == NULL) {
1377                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1378                 return -1;
1379         }
1380         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1381         dd->db_id = ctdb_db->db_id;
1382         dd->hdr = *hdr;
1383         dd->keylen = key.dsize;
1384         memcpy(dd->key, key.dptr, key.dsize);
1385
1386         ret = ctdb_control(ctdb_db->ctdb,
1387                            CTDB_CURRENT_NODE,
1388                            ctdb_db->db_id,
1389                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1390                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1391                            indata,
1392                            NULL, /* mem_ctx */
1393                            NULL, /* outdata */
1394                            &status,
1395                            NULL, /* timeout : NULL == wait forever */
1396                            NULL); /* error message */
1397
1398         talloc_free(indata.dptr);
1399
1400         if (ret != 0 || status != 0) {
1401                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1402                                   "SCHEDULE_FOR_DELETION "
1403                                   "control.\n"));
1404                 if (status != 0) {
1405                         ret = -1;
1406                 }
1407         }
1408
1409         return ret;
1410 }