server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5    Copyright (C) Michael Adam 2010-2011
   6    Copyright (C) Stefan Metzmacher 2010-2011
   7
   8    This program is free software; you can redistribute it and/or modify
   9    it under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3 of the License, or
  11    (at your option) any later version.
  12
  13    This program is distributed in the hope that it will be useful,
  14    but WITHOUT ANY WARRANTY; without even the implied warranty of
  15    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16    GNU General Public License for more details.
  17
  18    You should have received a copy of the GNU General Public License
  19    along with this program; if not, see <http://www.gnu.org/licenses/>.
  20 */
  21
  22 #include "includes.h"
  23 #include "lib/tdb/include/tdb.h"
  24 #include "system/network.h"
  25 #include "system/filesys.h"
  26 #include "system/dir.h"
  27 #include "../include/ctdb_private.h"
  28 #include "db_wrap.h"
  29 #include "lib/util/dlinklist.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34
  35 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  36
  37 struct ctdb_vacuum_child_context {
  38         struct ctdb_vacuum_child_context *next, *prev;
  39         struct ctdb_vacuum_handle *vacuum_handle;
  40         /* fd child writes status to */
  41         int fd[2];
  42         pid_t child_pid;
  43         enum vacuum_child_status status;
  44         struct timeval start_time;
  45 };
  46
  47 struct ctdb_vacuum_handle {
  48         struct ctdb_db_context *ctdb_db;
  49         struct ctdb_vacuum_child_context *child_ctx;
  50         uint32_t fast_path_count;
  51 };
  52
  53
  54 /*  a list of records to possibly delete */
  55 struct vacuum_data {
  56         uint32_t vacuum_limit;
  57         uint32_t repack_limit;
  58         struct ctdb_context *ctdb;
  59         struct ctdb_db_context *ctdb_db;
  60         struct tdb_context *dest_db;
  61         trbt_tree_t *delete_list;
  62         uint32_t delete_count;
  63         struct ctdb_marshall_buffer **vacuum_fetch_list;
  64         struct timeval start;
  65         bool traverse_error;
  66         bool vacuum;
  67         uint32_t total;
  68         uint32_t vacuumed;
  69         uint32_t copied;
  70         uint32_t fast_added_to_vacuum_fetch_list;
  71         uint32_t fast_added_to_delete_list;
  72         uint32_t fast_deleted;
  73         uint32_t fast_skipped;
  74         uint32_t fast_error;
  75         uint32_t fast_total;
  76         uint32_t full_added_to_vacuum_fetch_list;
  77         uint32_t full_added_to_delete_list;
  78         uint32_t full_skipped;
  79         uint32_t full_error;
  80         uint32_t full_total;
  81         uint32_t delete_left;
  82         uint32_t delete_remote_error;
  83         uint32_t delete_local_error;
  84         uint32_t delete_deleted;
  85         uint32_t delete_skipped;
  86 };
  87
  88 /* this structure contains the information for one record to be deleted */
  89 struct delete_record_data {
  90         struct ctdb_context *ctdb;
  91         struct ctdb_db_context *ctdb_db;
  92         struct ctdb_ltdb_header hdr;
  93         TDB_DATA key;
  94         uint8_t keydata[1];
  95 };
  96
  97 struct delete_records_list {
  98         struct ctdb_marshall_buffer *records;
  99 };
 100
 101 /**
 102  * Store key and header in a tree, indexed by the key hash.
 103  */
 104 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 105                                                struct ctdb_db_context *ctdb_db,
 106                                                trbt_tree_t *tree,
 107                                                const struct ctdb_ltdb_header *hdr,
 108                                                TDB_DATA key)
 109 {
 110         struct delete_record_data *dd;
 111         uint32_t hash;
 112         size_t len;
 113
 114         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 115
 116         dd = (struct delete_record_data *)talloc_size(tree, len);
 117         if (dd == NULL) {
 118                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 119                 return -1;
 120         }
 121         talloc_set_name_const(dd, "struct delete_record_data");
 122
 123         dd->ctdb      = ctdb;
 124         dd->ctdb_db   = ctdb_db;
 125         dd->key.dsize = key.dsize;
 126         dd->key.dptr  = dd->keydata;
 127         memcpy(dd->keydata, key.dptr, key.dsize);
 128
 129         dd->hdr = *hdr;
 130
 131         hash = ctdb_hash(&key);
 132
 133         trbt_insert32(tree, hash, dd);
 134
 135         return 0;
 136 }
 137
 138 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 139                                      struct ctdb_ltdb_header *hdr)
 140 {
 141         struct ctdb_context *ctdb = vdata->ctdb;
 142         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 143         uint32_t hash;
 144         int ret;
 145
 146         hash = ctdb_hash(&key);
 147
 148         if (trbt_lookup32(vdata->delete_list, hash)) {
 149                 DEBUG(DEBUG_INFO, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 150                 return 0;
 151         }
 152
 153         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 154                                                   vdata->delete_list,
 155                                                   hdr, key);
 156         if (ret != 0) {
 157                 return -1;
 158         }
 159
 160         vdata->delete_count++;
 161
 162         return 0;
 163 }
 164
 165 /**
 166  * Add a record to the list of records to be sent
 167  * to their lmaster with VACUUM_FETCH.
 168  */
 169 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 170                                            TDB_DATA key)
 171 {
 172         struct ctdb_context *ctdb = vdata->ctdb;
 173         struct ctdb_rec_data *rec;
 174         uint32_t lmaster;
 175         size_t old_size;
 176         struct ctdb_marshall_buffer *vfl;
 177
 178         lmaster = ctdb_lmaster(ctdb, &key);
 179
 180         vfl = vdata->vacuum_fetch_list[lmaster];
 181
 182         rec = ctdb_marshall_record(vfl, ctdb->pnn, key, NULL, tdb_null);
 183         if (rec == NULL) {
 184                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 185                 vdata->traverse_error = true;
 186                 return -1;
 187         }
 188
 189         old_size = talloc_get_size(vfl);
 190         vfl = talloc_realloc_size(NULL, vfl, old_size + rec->length);
 191         if (vfl == NULL) {
 192                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 193                 vdata->traverse_error = true;
 194                 return -1;
 195         }
 196         vdata->vacuum_fetch_list[lmaster] = vfl;
 197
 198         vfl->count++;
 199         memcpy(old_size+(uint8_t *)vfl, rec, rec->length);
 200         talloc_free(rec);
 201
 202         vdata->total++;
 203
 204         return 0;
 205 }
 206
 207
 208 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 209                               struct timeval t, void *private_data);
 210
 211
 212 /*
 213  * traverse function for gathering the records that can be deleted
 214  */
 215 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 216 {
 217         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 218         struct ctdb_context *ctdb = vdata->ctdb;
 219         uint32_t lmaster;
 220         struct ctdb_ltdb_header *hdr;
 221         int res = 0;
 222
 223         vdata->full_total++;
 224
 225         lmaster = ctdb_lmaster(ctdb, &key);
 226         if (lmaster >= ctdb->num_nodes) {
 227                 vdata->full_error++;
 228                 DEBUG(DEBUG_CRIT, (__location__
 229                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 230                                    " with hash[%u]!\n",
 231                                    (unsigned)lmaster,
 232                                    (unsigned)ctdb->num_nodes,
 233                                    (unsigned)ctdb_hash(&key)));
 234                 return -1;
 235         }
 236
 237         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 238                 /* it is not a deleted record */
 239                 vdata->full_skipped++;
 240                 return 0;
 241         }
 242
 243         hdr = (struct ctdb_ltdb_header *)data.dptr;
 244
 245         if (hdr->dmaster != ctdb->pnn) {
 246                 vdata->full_skipped++;
 247                 return 0;
 248         }
 249
 250         if (lmaster == ctdb->pnn) {
 251                 /*
 252                  * We are both lmaster and dmaster, and the record is empty.
 253                  * So we should be able to delete it.
 254                  */
 255                 res = add_record_to_delete_list(vdata, key, hdr);
 256                 if (res != 0) {
 257                         vdata->full_error++;
 258                 } else {
 259                         vdata->full_added_to_delete_list++;
 260                 }
 261         } else {
 262                 /*
 263                  * We are not lmaster.
 264                  * Add the record to the blob ready to send to the nodes.
 265                  */
 266                 res = add_record_to_vacuum_fetch_list(vdata, key);
 267                 if (res != 0) {
 268                         vdata->full_error++;
 269                 } else {
 270                         vdata->full_added_to_vacuum_fetch_list++;
 271                 }
 272         }
 273
 274         return res;
 275 }
 276
 277 /*
 278  * traverse the tree of records to delete and marshall them into
 279  * a blob
 280  */
 281 static int delete_marshall_traverse(void *param, void *data)
 282 {
 283         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 284         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 285         struct ctdb_rec_data *rec;
 286         size_t old_size;
 287
 288         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 289         if (rec == NULL) {
 290                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 291                 return 0;
 292         }
 293
 294         old_size = talloc_get_size(recs->records);
 295         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 296         if (recs->records == NULL) {
 297                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 298                 return 0;
 299         }
 300         recs->records->count++;
 301         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 302         return 0;
 303 }
 304
 305 /**
 306  * traverse function for the traversal of the delete_queue,
 307  * the fast-path vacuuming list.
 308  *
 309  *  - If the record has been migrated off the node
 310  *    or has been revived (filled with data) on the node,
 311  *    then skip the record.
 312  *
 313  *  - If the current node is the record's lmaster and it is
 314  *    a record that has never been migrated with data, then
 315  *    delete the record from the local tdb.
 316  *
 317  *  - If the current node is the record's lmaster and it has
 318  *    been migrated with data, then schedule it for the normal
 319  *    vacuuming procedure (i.e. add it to the delete_list).
 320  *
 321  *  - If the current node is NOT the record's lmaster then
 322  *    add it to the list of records that are to be sent to
 323  *    the lmaster with the VACUUM_FETCH message.
 324  */
 325 static int delete_queue_traverse(void *param, void *data)
 326 {
 327         struct delete_record_data *dd =
 328                 talloc_get_type(data, struct delete_record_data);
 329         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 330         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 331         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 332         int res;
 333         struct ctdb_ltdb_header *header;
 334         TDB_DATA tdb_data;
 335         uint32_t lmaster;
 336         uint32_t hash = ctdb_hash(&(dd->key));
 337
 338         vdata->fast_total++;
 339
 340         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 341         if (res != 0) {
 342                 DEBUG(DEBUG_ERR,
 343                       (__location__ " Error getting chainlock on record with "
 344                        "key hash [0x%08x] on database db[%s].\n",
 345                        hash, ctdb_db->db_name));
 346                 vdata->fast_error++;
 347                 return 0;
 348         }
 349
 350         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 351         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 352                 /* Does not exist or not a ctdb record. Skip. */
 353                 goto skipped;
 354         }
 355
 356         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 357                 /* The record has been recycled (filled with data). Skip. */
 358                 goto skipped;
 359         }
 360
 361         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 362
 363         if (header->dmaster != ctdb->pnn) {
 364                 /* The record has been migrated off the node. Skip. */
 365                 goto skipped;
 366         }
 367
 368         if (header->rsn != dd->hdr.rsn) {
 369                 /*
 370                  * The record has been migrated off the node and back again.
 371                  * But not requeued for deletion. Skip it.
 372                  */
 373                 goto skipped;
 374         }
 375
 376         /*
 377          * We are dmaster, and the record has no data, and it has
 378          * not been migrated after it has been queued for deletion.
 379          *
 380          * At this stage, the record could still have been revived locally
 381          * and last been written with empty data. This can only be
 382          * fixed with the addition of an active or delete flag. (TODO)
 383          */
 384
 385         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 386
 387         if (lmaster != ctdb->pnn) {
 388                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 389
 390                 if (res != 0) {
 391                         DEBUG(DEBUG_ERR,
 392                               (__location__ " Error adding record to list "
 393                                "of records to send to lmaster.\n"));
 394                         vdata->fast_error++;
 395                 } else {
 396                         vdata->fast_added_to_vacuum_fetch_list++;
 397                 }
 398                 goto done;
 399         }
 400
 401         /* use header->flags or dd->hdr.flags ?? */
 402         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 403                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 404
 405                 if (res != 0) {
 406                         DEBUG(DEBUG_ERR,
 407                               (__location__ " Error adding record to list "
 408                                "of records for deletion on lmaster.\n"));
 409                         vdata->fast_error++;
 410                 } else {
 411                         vdata->fast_added_to_delete_list++;
 412                 }
 413         } else {
 414                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 415
 416                 if (res != 0) {
 417                         DEBUG(DEBUG_ERR,
 418                               (__location__ " Error deleting record with key "
 419                                "hash [0x%08x] from local data base db[%s].\n",
 420                                hash, ctdb_db->db_name));
 421                         vdata->fast_error++;
 422                 } else {
 423                         DEBUG(DEBUG_DEBUG,
 424                               (__location__ " Deleted record with key hash "
 425                                "[0x%08x] from local data base db[%s].\n",
 426                                hash, ctdb_db->db_name));
 427                         vdata->fast_deleted++;
 428                 }
 429         }
 430
 431         goto done;
 432
 433 skipped:
 434         vdata->fast_skipped++;
 435
 436 done:
 437         if (tdb_data.dptr != NULL) {
 438                 free(tdb_data.dptr);
 439         }
 440         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 441
 442         return 0;
 443 }
 444
 445 /**
 446  * Delete the records that we are lmaster and dmaster for and
 447  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 448  * control.
 449  */
 450 static int delete_record_traverse(void *param, void *data)
 451 {
 452         struct delete_record_data *dd =
 453                 talloc_get_type(data, struct delete_record_data);
 454         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 455         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 456         struct ctdb_context *ctdb = ctdb_db->ctdb;
 457         int res;
 458         struct ctdb_ltdb_header *header;
 459         TDB_DATA tdb_data;
 460         uint32_t lmaster;
 461         bool deleted = false;
 462         uint32_t hash = ctdb_hash(&(dd->key));
 463
 464         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 465         if (res != 0) {
 466                 DEBUG(DEBUG_ERR,
 467                       (__location__ " Error getting chainlock on record with "
 468                        "key hash [0x%08x] on database db[%s].\n",
 469                        hash, ctdb_db->db_name));
 470                 vdata->delete_local_error++;
 471                 return 0;
 472         }
 473
 474         /*
 475          * Verify that the record is still empty, its RSN has not
 476          * changed and that we are still its lmaster and dmaster.
 477          */
 478
 479         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 480         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 481                 DEBUG(DEBUG_ERR, (__location__ ": record with hash [0x%08x] "
 482                                   "on database db[%s] does not exist or is not "
 483                                   "a ctdb-record.  skipping.\n",
 484                                   hash, ctdb_db->db_name));
 485                 vdata->delete_skipped++;
 486                 goto done;
 487         }
 488
 489         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 490                 DEBUG(DEBUG_ERR, (__location__ ": record with hash [0x%08x] "
 491                                   "on database db[%s] has been recycled. "
 492                                   "skipping.\n",
 493                                   hash, ctdb_db->db_name));
 494                 vdata->delete_skipped++;
 495                 goto done;
 496         }
 497
 498         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 499
 500         if (header->flags & (CTDB_REC_RO_HAVE_DELEGATIONS|
 501                              CTDB_REC_RO_HAVE_READONLY|
 502                              CTDB_REC_RO_REVOKING_READONLY|
 503                              CTDB_REC_RO_REVOKE_COMPLETE))
 504         {
 505                 DEBUG(DEBUG_ERR, (__location__ ": record with hash [0x%08x] "
 506                                   "on database db[%s] has read-only flags. "
 507                                   "skipping.\n",
 508                                   hash, ctdb_db->db_name));
 509                 vdata->delete_skipped++;
 510                 goto done;
 511         }
 512
 513         if (header->dmaster != ctdb->pnn) {
 514                 DEBUG(DEBUG_ERR, (__location__ ": record with hash [0x%08x] "
 515                                   "on database db[%s] has been migrated away. "
 516                                   "skipping.\n",
 517                                   hash, ctdb_db->db_name));
 518                 vdata->delete_skipped++;
 519                 goto done;
 520         }
 521
 522
 523         if (header->rsn != dd->hdr.rsn) {
 524                 /*
 525                  * The record has been migrated off the node and back again.
 526                  * But not requeued for deletion. Skip it.
 527                  */
 528                 DEBUG(DEBUG_ERR, (__location__ ": record with hash [0x%08x] "
 529                                   "on database db[%s] seems to have been "
 530                                   "migrated away and back again (with empty "
 531                                   "data). skipping.\n",
 532                                   hash, ctdb_db->db_name));
 533                 vdata->delete_skipped++;
 534                 goto done;
 535         }
 536
 537         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 538
 539         if (lmaster != ctdb->pnn) {
 540                 DEBUG(DEBUG_ERR, (__location__ ": not lmaster for record in "
 541                                   "delete list (key hash [0x%08x], db[%s]). "
 542                                   "Strange! skipping.\n",
 543                                   hash, ctdb_db->db_name));
 544                 vdata->delete_skipped++;
 545                 goto done;
 546         }
 547
 548         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 549
 550         if (res != 0) {
 551                 DEBUG(DEBUG_ERR,
 552                       (__location__ " Error deleting record with key hash "
 553                        "[0x%08x] from local data base db[%s].\n",
 554                        hash, ctdb_db->db_name));
 555                 vdata->delete_local_error++;
 556                 goto done;
 557         }
 558
 559         deleted = true;
 560
 561         DEBUG(DEBUG_DEBUG,
 562               (__location__ " Deleted record with key hash [0x%08x] from "
 563                "local data base db[%s].\n", hash, ctdb_db->db_name));
 564
 565 done:
 566         if (tdb_data.dptr != NULL) {
 567                 free(tdb_data.dptr);
 568         }
 569
 570         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 571
 572         if (deleted) {
 573                 /*
 574                  * successfully deleted the record locally.
 575                  * remove it from the list and update statistics.
 576                  */
 577                 talloc_free(dd);
 578                 vdata->delete_deleted++;
 579                 vdata->delete_left--;
 580         }
 581
 582         return 0;
 583 }
 584
 585 /**
 586  * Fast vacuuming run:
 587  * Traverse the delete_queue.
 588  * This fills the same lists as the database traverse.
 589  */
 590 static void ctdb_vacuum_db_fast(struct ctdb_db_context *ctdb_db,
 591                                 struct vacuum_data *vdata)
 592 {
 593         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 594
 595         if (vdata->fast_total > 0) {
 596                 DEBUG(DEBUG_INFO,
 597                       (__location__
 598                        " fast vacuuming delete_queue traverse statistics: "
 599                        "db[%s] "
 600                        "total[%u] "
 601                        "del[%u] "
 602                        "skp[%u] "
 603                        "err[%u] "
 604                        "adl[%u] "
 605                        "avf[%u]\n",
 606                        ctdb_db->db_name,
 607                        (unsigned)vdata->fast_total,
 608                        (unsigned)vdata->fast_deleted,
 609                        (unsigned)vdata->fast_skipped,
 610                        (unsigned)vdata->fast_error,
 611                        (unsigned)vdata->fast_added_to_delete_list,
 612                        (unsigned)vdata->fast_added_to_vacuum_fetch_list));
 613         }
 614
 615         return;
 616 }
 617
 618 /**
 619  * Full vacuum run:
 620  * read-only traverse of the database, looking for records that
 621  * might be able to be vacuumed.
 622  *
 623  * This is not done each time but only every tunable
 624  * VacuumFastPathCount times.
 625  */
 626 static int ctdb_vacuum_db_full(struct ctdb_db_context *ctdb_db,
 627                                struct vacuum_data *vdata,
 628                                bool full_vacuum_run)
 629 {
 630         int ret;
 631
 632         if (!full_vacuum_run) {
 633                 return 0;
 634         }
 635
 636         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 637         if (ret == -1 || vdata->traverse_error) {
 638                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 639                                   "'%s'\n", ctdb_db->db_name));
 640                 return -1;
 641         }
 642
 643         if (vdata->full_total > 0) {
 644                 DEBUG(DEBUG_INFO,
 645                       (__location__
 646                        " full vacuuming db traverse statistics: "
 647                        "db[%s] "
 648                        "total[%u] "
 649                        "skp[%u] "
 650                        "err[%u] "
 651                        "adl[%u] "
 652                        "avf[%u]\n",
 653                        ctdb_db->db_name,
 654                        (unsigned)vdata->full_total,
 655                        (unsigned)vdata->full_skipped,
 656                        (unsigned)vdata->full_error,
 657                        (unsigned)vdata->full_added_to_delete_list,
 658                        (unsigned)vdata->full_added_to_vacuum_fetch_list));
 659         }
 660
 661         return 0;
 662 }
 663
 664 /**
 665  * Process the vacuum fetch lists:
 666  * For records for which we are not the lmaster, tell the lmaster to
 667  * fetch the record.
 668  */
 669 static int ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 670                                            struct vacuum_data *vdata)
 671 {
 672         int i;
 673         struct ctdb_context *ctdb = ctdb_db->ctdb;
 674
 675         for (i = 0; i < ctdb->num_nodes; i++) {
 676                 TDB_DATA data;
 677                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 678
 679                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 680                         continue;
 681                 }
 682
 683                 if (vfl->count == 0) {
 684                         continue;
 685                 }
 686
 687                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 688                                    vfl->count, ctdb->nodes[i]->pnn,
 689                                    ctdb_db->db_name));
 690
 691                 data.dsize = talloc_get_size(vfl);
 692                 data.dptr  = (void *)vfl;
 693                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 694                                              CTDB_SRVID_VACUUM_FETCH,
 695                                              data) != 0)
 696                 {
 697                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 698                                           "fetch message to %u\n",
 699                                           ctdb->nodes[i]->pnn));
 700                         return -1;
 701                 }
 702         }
 703
 704         return 0;
 705 }
 706
 707 /**
 708  * Process the delete list:
 709  * Send the records to delete to all other nodes with the
 710  * try_delete_records control.
 711  */
 712 static int ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 713                                     struct vacuum_data *vdata)
 714 {
 715         int ret, i;
 716         struct ctdb_context *ctdb = ctdb_db->ctdb;
 717         struct delete_records_list *recs;
 718         TDB_DATA indata;
 719         struct ctdb_node_map *nodemap;
 720         uint32_t *active_nodes;
 721         int num_active_nodes;
 722         TALLOC_CTX *tmp_ctx;
 723
 724         if (vdata->delete_count == 0) {
 725                 return 0;
 726         }
 727
 728         tmp_ctx = talloc_new(vdata);
 729         if (tmp_ctx == NULL) {
 730                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 731                 return 0;
 732         }
 733
 734         vdata->delete_left = vdata->delete_count;
 735
 736         /*
 737          * now tell all the active nodes to delete all these records
 738          * (if possible)
 739          */
 740
 741         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 742                                    CTDB_CURRENT_NODE,
 743                                    tmp_ctx,
 744                                    &nodemap);
 745         if (ret != 0) {
 746                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 747                 ret = -1;
 748                 goto done;
 749         }
 750
 751         active_nodes = list_of_active_nodes(ctdb, nodemap,
 752                                             nodemap, /* talloc context */
 753                                             false /* include self */);
 754         /* yuck! ;-) */
 755         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 756
 757         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 758         if (recs == NULL) {
 759                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 760                 ret = -1;
 761                 goto done;
 762         }
 763         recs->records = (struct ctdb_marshall_buffer *)
 764                 talloc_zero_size(recs,
 765                                  offsetof(struct ctdb_marshall_buffer, data));
 766         if (recs->records == NULL) {
 767                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 768                 ret = -1;
 769                 goto done;
 770         }
 771         recs->records->db_id = ctdb_db->db_id;
 772
 773         /*
 774          * traverse the tree of all records we want to delete and
 775          * create a blob we can send to the other nodes.
 776          */
 777         trbt_traversearray32(vdata->delete_list, 1,
 778                              delete_marshall_traverse, recs);
 779
 780         indata.dsize = talloc_get_size(recs->records);
 781         indata.dptr  = (void *)recs->records;
 782
 783         for (i = 0; i < num_active_nodes; i++) {
 784                 struct ctdb_marshall_buffer *records;
 785                 struct ctdb_rec_data *rec;
 786                 int32_t res;
 787                 TDB_DATA outdata;
 788
 789                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 790                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 791                                 indata, recs, &outdata, &res,
 792                                 NULL, NULL);
 793                 if (ret != 0 || res != 0) {
 794                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
 795                                           "node %u: ret[%d] res[%d]\n",
 796                                           active_nodes[i], ret, res));
 797                         ret = -1;
 798                         goto done;
 799                 }
 800
 801                 /*
 802                  * outdata contains the list of records coming back
 803                  * from the node: These are the records that the
 804                  * remote node could not delete.
 805                  *
 806                  * NOTE: There is a problem here:
 807                  *
 808                  * When a node failed to delete the record, but
 809                  * others succeeded, we may have created gaps in the
 810                  * history of the record. Hence when a node dies, an
 811                  * closed file handle might be resurrected or an open
 812                  * file handle might be lost, leading to blocked access
 813                  * or data corruption.
 814                  *
 815                  * TODO: This needs to be fixed!
 816                  */
 817                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 818                 rec = (struct ctdb_rec_data *)&records->data[0];
 819                 while (records->count-- > 1) {
 820                         TDB_DATA reckey, recdata;
 821                         struct ctdb_ltdb_header *rechdr;
 822                         struct delete_record_data *dd;
 823
 824                         reckey.dptr = &rec->data[0];
 825                         reckey.dsize = rec->keylen;
 826                         recdata.dptr = &rec->data[reckey.dsize];
 827                         recdata.dsize = rec->datalen;
 828
 829                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 830                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 831                                 ret = -1;
 832                                 goto done;
 833                         }
 834                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 835                         recdata.dptr += sizeof(*rechdr);
 836                         recdata.dsize -= sizeof(*rechdr);
 837
 838                         dd = (struct delete_record_data *)trbt_lookup32(
 839                                         vdata->delete_list,
 840                                         ctdb_hash(&reckey));
 841                         if (dd != NULL) {
 842                                 /*
 843                                  * The other node could not delete the
 844                                  * record and it is the first node that
 845                                  * failed. So we should remove it from
 846                                  * the tree and update statistics.
 847                                  */
 848                                 talloc_free(dd);
 849                                 vdata->delete_remote_error++;
 850                                 vdata->delete_left--;
 851                         }
 852
 853                         rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 854                 }
 855         }
 856
 857         if (vdata->delete_left > 0) {
 858                 /*
 859                  * The only records remaining in the tree are those
 860                  * records which all other nodes could successfully
 861                  * delete, so we can safely delete them on the
 862                  * lmaster as well.
 863                  */
 864                 trbt_traversearray32(vdata->delete_list, 1,
 865                                      delete_record_traverse, vdata);
 866         }
 867
 868         if (vdata->delete_count > 0) {
 869                 DEBUG(DEBUG_INFO,
 870                       (__location__
 871                        " vacuum delete list statistics: "
 872                        "db[%s] "
 873                        "coll[%u] "
 874                        "rem.err[%u] "
 875                        "loc.err[%u] "
 876                        "skip[%u] "
 877                        "del[%u] "
 878                        "left[%u]\n",
 879                        ctdb_db->db_name,
 880                        (unsigned)vdata->delete_count,
 881                        (unsigned)vdata->delete_remote_error,
 882                        (unsigned)vdata->delete_local_error,
 883                        (unsigned)vdata->delete_skipped,
 884                        (unsigned)vdata->delete_deleted,
 885                        (unsigned)vdata->delete_left));
 886         }
 887
 888         ret = 0;
 889
 890 done:
 891         talloc_free(tmp_ctx);
 892
 893         return ret;
 894 }
 895
 896 /**
 897  * initialize the vacuum_data
 898  */
 899 static int ctdb_vacuum_init_vacuum_data(struct ctdb_db_context *ctdb_db,
 900                                         struct vacuum_data *vdata)
 901 {
 902         int i;
 903         struct ctdb_context *ctdb = ctdb_db->ctdb;
 904
 905         vdata->fast_added_to_delete_list = 0;
 906         vdata->fast_added_to_vacuum_fetch_list = 0;
 907         vdata->fast_deleted = 0;
 908         vdata->fast_skipped = 0;
 909         vdata->fast_error = 0;
 910         vdata->fast_total = 0;
 911         vdata->full_added_to_delete_list = 0;
 912         vdata->full_added_to_vacuum_fetch_list = 0;
 913         vdata->full_skipped = 0;
 914         vdata->full_error = 0;
 915         vdata->full_total = 0;
 916         vdata->delete_count = 0;
 917         vdata->delete_left = 0;
 918         vdata->delete_remote_error = 0;
 919         vdata->delete_local_error = 0;
 920         vdata->delete_skipped = 0;
 921         vdata->delete_deleted = 0;
 922
 923         /* the list needs to be of length num_nodes */
 924         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
 925                                                 struct ctdb_marshall_buffer *,
 926                                                 ctdb->num_nodes);
 927         if (vdata->vacuum_fetch_list == NULL) {
 928                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 929                 return -1;
 930         }
 931         for (i = 0; i < ctdb->num_nodes; i++) {
 932                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
 933                         talloc_zero_size(vdata->vacuum_fetch_list,
 934                                          offsetof(struct ctdb_marshall_buffer, data));
 935                 if (vdata->vacuum_fetch_list[i] == NULL) {
 936                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 937                         return -1;
 938                 }
 939                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
 940         }
 941
 942         return 0;
 943 }
 944
 945 /**
 946  * Vacuum a DB:
 947  *  - Always do the fast vacuuming run, which traverses
 948  *    the in-memory delete queue: these records have been
 949  *    scheduled for deletion.
 950  *  - Only if explicitly requested, the database is traversed
 951  *    in order to use the traditional heuristics on empty records
 952  *    to trigger deletion.
 953  *    This is done only every VacuumFastPathCount'th vacuuming run.
 954  *
 955  * The traverse runs fill two lists:
 956  *
 957  * - The delete_list:
 958  *   This is the list of empty records the current
 959  *   node is lmaster and dmaster for. These records are later
 960  *   deleted first on other nodes and then locally.
 961  *
 962  *   The fast vacuuming run has a short cut for those records
 963  *   that have never been migrated with data: these records
 964  *   are immediately deleted locally, since they have left
 965  *   no trace on other nodes.
 966  *
 967  * - The vacuum_fetch lists
 968  *   (one for each other lmaster node):
 969  *   The records in this list are sent for deletion to
 970  *   their lmaster in a bulk VACUUM_FETCH message.
 971  *
 972  *   The lmaster then migrates all these records to itelf
 973  *   so that they can be vacuumed there.
 974  *
 975  * This executes in the child context.
 976  */
 977 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 978                           struct vacuum_data *vdata,
 979                           bool full_vacuum_run)
 980 {
 981         struct ctdb_context *ctdb = ctdb_db->ctdb;
 982         int ret, pnn;
 983
 984         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
 985                            "%s db_id[0x%08x]\n",
 986                            full_vacuum_run ? "full" : "fast",
 987                            ctdb_db->db_name, ctdb_db->db_id));
 988
 989         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 990         if (ret != 0) {
 991                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 992                 return ret;
 993         }
 994
 995         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 996         if (pnn == -1) {
 997                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 998                 return -1;
 999         }
1000
1001         ctdb->pnn = pnn;
1002
1003         ret = ctdb_vacuum_init_vacuum_data(ctdb_db, vdata);
1004         if (ret != 0) {
1005                 return ret;
1006         }
1007
1008         ctdb_vacuum_db_fast(ctdb_db, vdata);
1009
1010         ret = ctdb_vacuum_db_full(ctdb_db, vdata, full_vacuum_run);
1011         if (ret != 0) {
1012                 return ret;
1013         }
1014
1015         ret = ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1016         if (ret != 0) {
1017                 return ret;
1018         }
1019
1020         ret = ctdb_process_delete_list(ctdb_db, vdata);
1021         if (ret != 0) {
1022                 return ret;
1023         }
1024
1025         /* this ensures we run our event queue */
1026         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1027
1028         return 0;
1029 }
1030
1031
1032 /*
1033  * traverse function for repacking
1034  */
1035 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
1036 {
1037         struct vacuum_data *vdata = (struct vacuum_data *)private;
1038
1039         if (vdata->vacuum) {
1040                 uint32_t hash = ctdb_hash(&key);
1041                 struct delete_record_data *kd;
1042                 /*
1043                  * check if we can ignore this record because it's in the delete_list
1044                  */
1045                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_list, hash);
1046                 /*
1047                  * there might be hash collisions so we have to compare the keys here to be sure
1048                  */
1049                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
1050                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
1051                         /*
1052                          * we have to check if the record hasn't changed in the meantime in order to
1053                          * savely remove it from the database
1054                          */
1055                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
1056                                 hdr->dmaster == kd->ctdb->pnn &&
1057                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
1058                                 kd->hdr.rsn == hdr->rsn) {
1059                                 vdata->vacuumed++;
1060                                 return 0;
1061                         }
1062                 }
1063         }
1064         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
1065                 vdata->traverse_error = true;
1066                 return -1;
1067         }
1068         vdata->copied++;
1069         return 0;
1070 }
1071
1072 /*
1073  * repack a tdb
1074  */
1075 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
1076 {
1077         struct tdb_context *tmp_db;
1078
1079         if (tdb_transaction_start(tdb) != 0) {
1080                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
1081                 return -1;
1082         }
1083
1084         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
1085                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
1086                           O_RDWR|O_CREAT, 0);
1087         if (tmp_db == NULL) {
1088                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
1089                 tdb_transaction_cancel(tdb);
1090                 return -1;
1091         }
1092
1093         vdata->traverse_error = false;
1094         vdata->dest_db = tmp_db;
1095         vdata->vacuum = true;
1096         vdata->vacuumed = 0;
1097         vdata->copied = 0;
1098
1099         /*
1100          * repack and vacuum on-the-fly by not writing the records that are
1101          * no longer needed
1102          */
1103         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
1104                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
1105                 tdb_transaction_cancel(tdb);
1106                 tdb_close(tmp_db);
1107                 return -1;
1108         }
1109
1110         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
1111
1112         if (vdata->traverse_error) {
1113                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
1114                 tdb_transaction_cancel(tdb);
1115                 tdb_close(tmp_db);
1116                 return -1;
1117         }
1118
1119         if (tdb_wipe_all(tdb) != 0) {
1120                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
1121                 tdb_transaction_cancel(tdb);
1122                 tdb_close(tmp_db);
1123                 return -1;
1124         }
1125
1126         vdata->traverse_error = false;
1127         vdata->dest_db = tdb;
1128         vdata->vacuum = false;
1129         vdata->copied = 0;
1130
1131         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
1132                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
1133                 tdb_transaction_cancel(tdb);
1134                 tdb_close(tmp_db);
1135                 return -1;
1136         }
1137
1138         if (vdata->traverse_error) {
1139                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
1140                 tdb_transaction_cancel(tdb);
1141                 tdb_close(tmp_db);
1142                 return -1;
1143         }
1144
1145         tdb_close(tmp_db);
1146
1147
1148         if (tdb_transaction_commit(tdb) != 0) {
1149                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
1150                 return -1;
1151         }
1152         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
1153
1154         return 0;
1155 }
1156
1157 /*
1158  * repack and vaccum a db
1159  * called from the child context
1160  */
1161 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1162                                      TALLOC_CTX *mem_ctx,
1163                                      bool full_vacuum_run)
1164 {
1165         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1166         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
1167         const char *name = ctdb_db->db_name;
1168         int freelist_size;
1169         struct vacuum_data *vdata;
1170
1171         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1172         if (freelist_size == -1) {
1173                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1174                 return -1;
1175         }
1176
1177         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1178         if (vdata == NULL) {
1179                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1180                 return -1;
1181         }
1182
1183         vdata->ctdb = ctdb_db->ctdb;
1184         vdata->vacuum_limit = vacuum_limit;
1185         vdata->repack_limit = repack_limit;
1186         vdata->delete_list = trbt_create(vdata, 0);
1187         vdata->ctdb_db = ctdb_db;
1188         if (vdata->delete_list == NULL) {
1189                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1190                 talloc_free(vdata);
1191                 return -1;
1192         }
1193
1194         vdata->start = timeval_current();
1195
1196         /*
1197          * gather all records that can be deleted in vdata
1198          */
1199         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
1200                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1201         }
1202
1203         /*
1204          * decide if a repack is necessary
1205          */
1206         if (freelist_size < repack_limit && vdata->delete_left < vacuum_limit)
1207         {
1208                 talloc_free(vdata);
1209                 return 0;
1210         }
1211
1212         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
1213                         name, freelist_size, vdata->delete_left));
1214
1215         /*
1216          * repack and implicitely get rid of the records we can delete
1217          */
1218         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
1219                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1220                 talloc_free(vdata);
1221                 return -1;
1222         }
1223         talloc_free(vdata);
1224
1225         return 0;
1226 }
1227
1228 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1229 {
1230         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1231
1232         return interval;
1233 }
1234
1235 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1236 {
1237         double l = timeval_elapsed(&child_ctx->start_time);
1238         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1239         struct ctdb_context *ctdb = ctdb_db->ctdb;
1240
1241         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1242
1243         if (child_ctx->child_pid != -1) {
1244                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1245         } else {
1246                 /* Bump the number of successful fast-path runs. */
1247                 child_ctx->vacuum_handle->fast_path_count++;
1248         }
1249
1250         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1251
1252         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
1253                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1254                         ctdb_vacuum_event, child_ctx->vacuum_handle);
1255
1256         return 0;
1257 }
1258
1259 /*
1260  * this event is generated when a vacuum child process times out
1261  */
1262 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
1263                                          struct timeval t, void *private_data)
1264 {
1265         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1266
1267         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1268
1269         child_ctx->status = VACUUM_TIMEOUT;
1270
1271         talloc_free(child_ctx);
1272 }
1273
1274
1275 /*
1276  * this event is generated when a vacuum child process has completed
1277  */
1278 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1279                              uint16_t flags, void *private_data)
1280 {
1281         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1282         char c = 0;
1283         int ret;
1284
1285         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1286         child_ctx->child_pid = -1;
1287
1288         ret = read(child_ctx->fd[0], &c, 1);
1289         if (ret != 1 || c != 0) {
1290                 child_ctx->status = VACUUM_ERROR;
1291                 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));
1292         } else {
1293                 child_ctx->status = VACUUM_OK;
1294         }
1295
1296         talloc_free(child_ctx);
1297 }
1298
1299 /*
1300  * this event is called every time we need to start a new vacuum process
1301  */
1302 static void
1303 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1304                                struct timeval t, void *private_data)
1305 {
1306         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1307         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1308         struct ctdb_context *ctdb = ctdb_db->ctdb;
1309         struct ctdb_vacuum_child_context *child_ctx;
1310         struct tevent_fd *fde;
1311         int ret;
1312
1313         /* we dont vacuum if we are in recovery mode, or db frozen */
1314         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1315             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1316                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1317                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1318                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1319                                    ? "freeze pending"
1320                                    : "frozen"));
1321                 event_add_timed(ctdb->ev, vacuum_handle,
1322                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1323                         ctdb_vacuum_event, vacuum_handle);
1324                 return;
1325         }
1326
1327         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1328         if (child_ctx == NULL) {
1329                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1330                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1331         }
1332
1333
1334         ret = pipe(child_ctx->fd);
1335         if (ret != 0) {
1336                 talloc_free(child_ctx);
1337                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1338                 event_add_timed(ctdb->ev, vacuum_handle,
1339                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1340                         ctdb_vacuum_event, vacuum_handle);
1341                 return;
1342         }
1343
1344         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1345                 vacuum_handle->fast_path_count = 0;
1346         }
1347
1348         child_ctx->child_pid = ctdb_fork(ctdb);
1349         if (child_ctx->child_pid == (pid_t)-1) {
1350                 close(child_ctx->fd[0]);
1351                 close(child_ctx->fd[1]);
1352                 talloc_free(child_ctx);
1353                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1354                 event_add_timed(ctdb->ev, vacuum_handle,
1355                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1356                         ctdb_vacuum_event, vacuum_handle);
1357                 return;
1358         }
1359
1360
1361         if (child_ctx->child_pid == 0) {
1362                 char cc = 0;
1363                 bool full_vacuum_run = false;
1364                 close(child_ctx->fd[0]);
1365
1366                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1367
1368                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1369                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1370                         _exit(1);
1371                 }
1372
1373                 /*
1374                  * repack the db
1375                  */
1376                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1377                     (vacuum_handle->fast_path_count == 0))
1378                 {
1379                         full_vacuum_run = true;
1380                 }
1381                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1382                                                full_vacuum_run);
1383
1384                 write(child_ctx->fd[1], &cc, 1);
1385                 _exit(0);
1386         }
1387
1388         set_close_on_exec(child_ctx->fd[0]);
1389         close(child_ctx->fd[1]);
1390
1391         child_ctx->status = VACUUM_RUNNING;
1392         child_ctx->start_time = timeval_current();
1393
1394         DLIST_ADD(ctdb->vacuumers, child_ctx);
1395         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1396
1397         /*
1398          * Clear the fastpath vacuuming list in the parent.
1399          */
1400         talloc_free(ctdb_db->delete_queue);
1401         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1402         if (ctdb_db->delete_queue == NULL) {
1403                 /* fatal here? ... */
1404                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1405                                  "in parent context. Shutting down\n");
1406         }
1407
1408         event_add_timed(ctdb->ev, child_ctx,
1409                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1410                 vacuum_child_timeout, child_ctx);
1411
1412         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1413
1414         fde = event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1415                            EVENT_FD_READ, vacuum_child_handler, child_ctx);
1416         tevent_fd_set_auto_close(fde);
1417
1418         vacuum_handle->child_ctx = child_ctx;
1419         child_ctx->vacuum_handle = vacuum_handle;
1420 }
1421
1422 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1423 {
1424         /* Simply free them all. */
1425         while (ctdb->vacuumers) {
1426                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1427                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1428                            (int)ctdb->vacuumers->child_pid));
1429                 /* vacuum_child_destructor kills it, removes from list */
1430                 talloc_free(ctdb->vacuumers);
1431         }
1432 }
1433
1434 /* this function initializes the vacuuming context for a database
1435  * starts the vacuuming events
1436  */
1437 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1438 {
1439         if (ctdb_db->persistent != 0) {
1440                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1441                 return 0;
1442         }
1443
1444         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1445         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1446
1447         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1448         ctdb_db->vacuum_handle->fast_path_count = 0;
1449
1450         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1451                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1452                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1453
1454         return 0;
1455 }
1456
1457 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1458                                             const struct ctdb_ltdb_header *hdr,
1459                                             const TDB_DATA key)
1460 {
1461         struct delete_record_data *kd;
1462         uint32_t hash;
1463
1464         hash = (uint32_t)ctdb_hash(&key);
1465
1466         DEBUG(DEBUG_DEBUG, (__location__
1467                             " remove_record_from_delete_queue: "
1468                             "db[%s] "
1469                             "db_id[0x%08x] "
1470                             "key_hash[0x%08x] "
1471                             "lmaster[%u] "
1472                             "migrated_with_data[%s]\n",
1473                              ctdb_db->db_name, ctdb_db->db_id,
1474                              hash,
1475                              ctdb_lmaster(ctdb_db->ctdb, &key),
1476                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1477
1478         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1479         if (kd == NULL) {
1480                 DEBUG(DEBUG_DEBUG, (__location__
1481                                     " remove_record_from_delete_queue: "
1482                                     "record not in queue (hash[0x%08x])\n.",
1483                                     hash));
1484                 return;
1485         }
1486
1487         if ((kd->key.dsize != key.dsize) ||
1488             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1489         {
1490                 DEBUG(DEBUG_DEBUG, (__location__
1491                                     " remove_record_from_delete_queue: "
1492                                     "hash collision for key with hash[0x%08x] "
1493                                     "in db[%s] - skipping\n",
1494                                     hash, ctdb_db->db_name));
1495                 return;
1496         }
1497
1498         DEBUG(DEBUG_DEBUG, (__location__
1499                             " remove_record_from_delete_queue: "
1500                             "removing key with hash[0x%08x]\n",
1501                              hash));
1502
1503         talloc_free(kd);
1504
1505         return;
1506 }
1507
1508 /**
1509  * Insert a record into the ctdb_db context's delete queue,
1510  * handling hash collisions.
1511  */
1512 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1513                                            const struct ctdb_ltdb_header *hdr,
1514                                            TDB_DATA key)
1515 {
1516         struct delete_record_data *kd;
1517         uint32_t hash;
1518         int ret;
1519
1520         hash = (uint32_t)ctdb_hash(&key);
1521
1522         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1523                            "db_id[0x%08x] "
1524                            "key_hash[0x%08x] "
1525                            "lmaster[%u] "
1526                            "migrated_with_data[%s]\n",
1527                             ctdb_db->db_name, ctdb_db->db_id,
1528                             hash,
1529                             ctdb_lmaster(ctdb_db->ctdb, &key),
1530                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1531
1532         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1533         if (kd != NULL) {
1534                 if ((kd->key.dsize != key.dsize) ||
1535                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1536                 {
1537                         DEBUG(DEBUG_INFO,
1538                               (__location__ " schedule for deletion: "
1539                                "hash collision for key hash [0x%08x]. "
1540                                "Skipping the record.\n", hash));
1541                         return 0;
1542                 } else {
1543                         DEBUG(DEBUG_DEBUG,
1544                               (__location__ " schedule for deletion: "
1545                                "updating entry for key with hash [0x%08x].\n",
1546                                hash));
1547                 }
1548         }
1549
1550         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1551                                                   ctdb_db->delete_queue,
1552                                                   hdr, key);
1553         if (ret != 0) {
1554                 DEBUG(DEBUG_INFO,
1555                       (__location__ " schedule for deletion: error "
1556                        "inserting key with hash [0x%08x] into delete queue\n",
1557                        hash));
1558                 return -1;
1559         }
1560
1561         return 0;
1562 }
1563
1564 /**
1565  * Schedule a record for deletetion.
1566  * Called from the parent context.
1567  */
1568 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1569                                            TDB_DATA indata)
1570 {
1571         struct ctdb_control_schedule_for_deletion *dd;
1572         struct ctdb_db_context *ctdb_db;
1573         int ret;
1574         TDB_DATA key;
1575
1576         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1577
1578         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1579         if (ctdb_db == NULL) {
1580                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1581                                   dd->db_id));
1582                 return -1;
1583         }
1584
1585         key.dsize = dd->keylen;
1586         key.dptr = dd->key;
1587
1588         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1589
1590         return ret;
1591 }
1592
1593 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1594                                          const struct ctdb_ltdb_header *hdr,
1595                                          TDB_DATA key)
1596 {
1597         int ret;
1598         struct ctdb_control_schedule_for_deletion *dd;
1599         TDB_DATA indata;
1600         int32_t status;
1601
1602         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1603                 /* main daemon - directly queue */
1604                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1605
1606                 return ret;
1607         }
1608
1609         /* if we dont have a connection to the daemon we can not send
1610            a control. For example sometimes from update_record control child
1611            process.
1612         */
1613         if (!ctdb_db->ctdb->can_send_controls) {
1614                 return -1;
1615         }
1616
1617
1618         /* child process: send the main daemon a control */
1619         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1620         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1621         if (indata.dptr == NULL) {
1622                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1623                 return -1;
1624         }
1625         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1626         dd->db_id = ctdb_db->db_id;
1627         dd->hdr = *hdr;
1628         dd->keylen = key.dsize;
1629         memcpy(dd->key, key.dptr, key.dsize);
1630
1631         ret = ctdb_control(ctdb_db->ctdb,
1632                            CTDB_CURRENT_NODE,
1633                            ctdb_db->db_id,
1634                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1635                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1636                            indata,
1637                            NULL, /* mem_ctx */
1638                            NULL, /* outdata */
1639                            &status,
1640                            NULL, /* timeout : NULL == wait forever */
1641                            NULL); /* error message */
1642
1643         talloc_free(indata.dptr);
1644
1645         if (ret != 0 || status != 0) {
1646                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1647                                   "SCHEDULE_FOR_DELETION "
1648                                   "control.\n"));
1649                 if (status != 0) {
1650                         ret = -1;
1651                 }
1652         }
1653
1654         return ret;
1655 }
1656
1657 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1658                                          const struct ctdb_ltdb_header *hdr,
1659                                          const TDB_DATA key)
1660 {
1661         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1662                 /*
1663                  * Only remove the record from the delete queue if called
1664                  * in the main daemon.
1665                  */
1666                 return;
1667         }
1668
1669         remove_record_from_delete_queue(ctdb_db, hdr, key);
1670
1671         return;
1672 }