ctdb/server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5    Copyright (C) Michael Adam 2010-2013
   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 "replace.h"
  23 #include "system/network.h"
  24 #include "system/filesys.h"
  25 #include "system/time.h"
  26
  27 #include <talloc.h>
  28 #include <tevent.h>
  29
  30 #include "lib/tdb_wrap/tdb_wrap.h"
  31 #include "lib/util/dlinklist.h"
  32 #include "lib/util/debug.h"
  33 #include "lib/util/samba_util.h"
  34
  35 #include "ctdb_private.h"
  36 #include "ctdb_client.h"
  37 #include "ctdb_logging.h"
  38
  39 #include "common/rb_tree.h"
  40 #include "common/system.h"
  41 #include "common/common.h"
  42
  43 #define TIMELIMIT() timeval_current_ofs(10, 0)
  44
  45 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  46
  47 struct ctdb_vacuum_child_context {
  48         struct ctdb_vacuum_child_context *next, *prev;
  49         struct ctdb_vacuum_handle *vacuum_handle;
  50         /* fd child writes status to */
  51         int fd[2];
  52         pid_t child_pid;
  53         enum vacuum_child_status status;
  54         struct timeval start_time;
  55 };
  56
  57 struct ctdb_vacuum_handle {
  58         struct ctdb_db_context *ctdb_db;
  59         struct ctdb_vacuum_child_context *child_ctx;
  60         uint32_t fast_path_count;
  61 };
  62
  63
  64 /*  a list of records to possibly delete */
  65 struct vacuum_data {
  66         struct ctdb_context *ctdb;
  67         struct ctdb_db_context *ctdb_db;
  68         struct tdb_context *dest_db;
  69         trbt_tree_t *delete_list;
  70         struct ctdb_marshall_buffer **vacuum_fetch_list;
  71         struct timeval start;
  72         bool traverse_error;
  73         bool vacuum;
  74         struct {
  75                 struct {
  76                         uint32_t added_to_vacuum_fetch_list;
  77                         uint32_t added_to_delete_list;
  78                         uint32_t deleted;
  79                         uint32_t skipped;
  80                         uint32_t error;
  81                         uint32_t total;
  82                 } delete_queue;
  83                 struct {
  84                         uint32_t scheduled;
  85                         uint32_t skipped;
  86                         uint32_t error;
  87                         uint32_t total;
  88                 } db_traverse;
  89                 struct {
  90                         uint32_t total;
  91                         uint32_t remote_error;
  92                         uint32_t local_error;
  93                         uint32_t deleted;
  94                         uint32_t skipped;
  95                         uint32_t left;
  96                 } delete_list;
  97                 struct {
  98                         uint32_t vacuumed;
  99                         uint32_t copied;
 100                 } repack;
 101         } count;
 102 };
 103
 104 /* this structure contains the information for one record to be deleted */
 105 struct delete_record_data {
 106         struct ctdb_context *ctdb;
 107         struct ctdb_db_context *ctdb_db;
 108         struct ctdb_ltdb_header hdr;
 109         TDB_DATA key;
 110         uint8_t keydata[1];
 111 };
 112
 113 struct delete_records_list {
 114         struct ctdb_marshall_buffer *records;
 115         struct vacuum_data *vdata;
 116 };
 117
 118 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 119                                            const struct ctdb_ltdb_header *hdr,
 120                                            TDB_DATA key);
 121
 122 /**
 123  * Store key and header in a tree, indexed by the key hash.
 124  */
 125 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 126                                                struct ctdb_db_context *ctdb_db,
 127                                                trbt_tree_t *tree,
 128                                                const struct ctdb_ltdb_header *hdr,
 129                                                TDB_DATA key)
 130 {
 131         struct delete_record_data *dd;
 132         uint32_t hash;
 133         size_t len;
 134
 135         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 136
 137         dd = (struct delete_record_data *)talloc_size(tree, len);
 138         if (dd == NULL) {
 139                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 140                 return -1;
 141         }
 142         talloc_set_name_const(dd, "struct delete_record_data");
 143
 144         dd->ctdb      = ctdb;
 145         dd->ctdb_db   = ctdb_db;
 146         dd->key.dsize = key.dsize;
 147         dd->key.dptr  = dd->keydata;
 148         memcpy(dd->keydata, key.dptr, key.dsize);
 149
 150         dd->hdr = *hdr;
 151
 152         hash = ctdb_hash(&key);
 153
 154         trbt_insert32(tree, hash, dd);
 155
 156         return 0;
 157 }
 158
 159 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 160                                      struct ctdb_ltdb_header *hdr)
 161 {
 162         struct ctdb_context *ctdb = vdata->ctdb;
 163         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 164         uint32_t hash;
 165         int ret;
 166
 167         hash = ctdb_hash(&key);
 168
 169         if (trbt_lookup32(vdata->delete_list, hash)) {
 170                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 171                 return 0;
 172         }
 173
 174         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 175                                                   vdata->delete_list,
 176                                                   hdr, key);
 177         if (ret != 0) {
 178                 return -1;
 179         }
 180
 181         vdata->count.delete_list.total++;
 182
 183         return 0;
 184 }
 185
 186 /**
 187  * Add a record to the list of records to be sent
 188  * to their lmaster with VACUUM_FETCH.
 189  */
 190 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 191                                            TDB_DATA key)
 192 {
 193         struct ctdb_context *ctdb = vdata->ctdb;
 194         uint32_t lmaster;
 195         struct ctdb_marshall_buffer *vfl;
 196
 197         lmaster = ctdb_lmaster(ctdb, &key);
 198
 199         vfl = vdata->vacuum_fetch_list[lmaster];
 200
 201         vfl = ctdb_marshall_add(ctdb, vfl, vfl->db_id, ctdb->pnn,
 202                                 key, NULL, tdb_null);
 203         if (vfl == NULL) {
 204                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 205                 vdata->traverse_error = true;
 206                 return -1;
 207         }
 208
 209         vdata->vacuum_fetch_list[lmaster] = vfl;
 210
 211         return 0;
 212 }
 213
 214
 215 static void ctdb_vacuum_event(struct tevent_context *ev,
 216                               struct tevent_timer *te,
 217                               struct timeval t, void *private_data);
 218
 219 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 220 {
 221         struct ctdb_ltdb_header *header =
 222                 (struct ctdb_ltdb_header *)private_data;
 223
 224         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 225                 return -1;
 226         }
 227
 228         *header = *(struct ctdb_ltdb_header *)data.dptr;
 229
 230         return 0;
 231 }
 232
 233 /*
 234  * traverse function for gathering the records that can be deleted
 235  */
 236 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 237                            void *private_data)
 238 {
 239         struct vacuum_data *vdata = talloc_get_type(private_data,
 240                                                     struct vacuum_data);
 241         struct ctdb_context *ctdb = vdata->ctdb;
 242         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 243         uint32_t lmaster;
 244         struct ctdb_ltdb_header *hdr;
 245         int res = 0;
 246
 247         vdata->count.db_traverse.total++;
 248
 249         lmaster = ctdb_lmaster(ctdb, &key);
 250         if (lmaster >= ctdb->num_nodes) {
 251                 vdata->count.db_traverse.error++;
 252                 DEBUG(DEBUG_CRIT, (__location__
 253                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 254                                    " with hash[%u]!\n",
 255                                    (unsigned)lmaster,
 256                                    (unsigned)ctdb->num_nodes,
 257                                    (unsigned)ctdb_hash(&key)));
 258                 return -1;
 259         }
 260
 261         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 262                 /* it is not a deleted record */
 263                 vdata->count.db_traverse.skipped++;
 264                 return 0;
 265         }
 266
 267         hdr = (struct ctdb_ltdb_header *)data.dptr;
 268
 269         if (hdr->dmaster != ctdb->pnn) {
 270                 vdata->count.db_traverse.skipped++;
 271                 return 0;
 272         }
 273
 274         /*
 275          * Add the record to this process's delete_queue for processing
 276          * in the subsequent traverse in the fast vacuum run.
 277          */
 278         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 279         if (res != 0) {
 280                 vdata->count.db_traverse.error++;
 281         } else {
 282                 vdata->count.db_traverse.scheduled++;
 283         }
 284
 285         return 0;
 286 }
 287
 288 /*
 289  * traverse the tree of records to delete and marshall them into
 290  * a blob
 291  */
 292 static int delete_marshall_traverse(void *param, void *data)
 293 {
 294         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 295         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 296         struct ctdb_marshall_buffer *m;
 297
 298         m = ctdb_marshall_add(recs, recs->records, recs->records->db_id,
 299                               recs->records->db_id,
 300                               dd->key, &dd->hdr, tdb_null);
 301         if (m == NULL) {
 302                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 303                 return -1;
 304         }
 305
 306         recs->records = m;
 307         return 0;
 308 }
 309
 310 /**
 311  * Variant of delete_marshall_traverse() that bumps the
 312  * RSN of each traversed record in the database.
 313  *
 314  * This is needed to ensure that when rolling out our
 315  * empty record copy before remote deletion, we as the
 316  * record's dmaster keep a higher RSN than the non-dmaster
 317  * nodes. This is needed to prevent old copies from
 318  * resurrection in recoveries.
 319  */
 320 static int delete_marshall_traverse_first(void *param, void *data)
 321 {
 322         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 323         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 324         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 325         struct ctdb_context *ctdb = ctdb_db->ctdb;
 326         struct ctdb_ltdb_header header;
 327         uint32_t lmaster;
 328         uint32_t hash = ctdb_hash(&(dd->key));
 329         int res;
 330
 331         res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
 332         if (res != 0) {
 333                 recs->vdata->count.delete_list.skipped++;
 334                 recs->vdata->count.delete_list.left--;
 335                 talloc_free(dd);
 336                 return 0;
 337         }
 338
 339         /*
 340          * Verify that the record is still empty, its RSN has not
 341          * changed and that we are still its lmaster and dmaster.
 342          */
 343
 344         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 345                                vacuum_record_parser, &header);
 346         if (res != 0) {
 347                 goto skip;
 348         }
 349
 350         if (header.flags & CTDB_REC_RO_FLAGS) {
 351                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 352                                    "on database db[%s] has read-only flags. "
 353                                    "skipping.\n",
 354                                    hash, ctdb_db->db_name));
 355                 goto skip;
 356         }
 357
 358         if (header.dmaster != ctdb->pnn) {
 359                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 360                                    "on database db[%s] has been migrated away. "
 361                                    "skipping.\n",
 362                                    hash, ctdb_db->db_name));
 363                 goto skip;
 364         }
 365
 366         if (header.rsn != dd->hdr.rsn) {
 367                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 368                                    "on database db[%s] seems to have been "
 369                                    "migrated away and back again (with empty "
 370                                    "data). skipping.\n",
 371                                    hash, ctdb_db->db_name));
 372                 goto skip;
 373         }
 374
 375         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 376
 377         if (lmaster != ctdb->pnn) {
 378                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 379                                    "delete list (key hash [0x%08x], db[%s]). "
 380                                    "Strange! skipping.\n",
 381                                    hash, ctdb_db->db_name));
 382                 goto skip;
 383         }
 384
 385         /*
 386          * Increment the record's RSN to ensure the dmaster (i.e. the current
 387          * node) has the highest RSN of the record in the cluster.
 388          * This is to prevent old record copies from resurrecting in recoveries
 389          * if something should fail during the deletion process.
 390          * Note that ctdb_ltdb_store_server() increments the RSN if called
 391          * on the record's dmaster.
 392          */
 393
 394         res = ctdb_ltdb_store(ctdb_db, dd->key, &header, tdb_null);
 395         if (res != 0) {
 396                 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
 397                                   "key hash [0x%08x] on database db[%s].\n",
 398                                   hash, ctdb_db->db_name));
 399                 goto skip;
 400         }
 401
 402         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 403
 404         goto done;
 405
 406 skip:
 407         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 408
 409         recs->vdata->count.delete_list.skipped++;
 410         recs->vdata->count.delete_list.left--;
 411         talloc_free(dd);
 412         dd = NULL;
 413
 414 done:
 415         if (dd == NULL) {
 416                 return 0;
 417         }
 418
 419         return delete_marshall_traverse(param, data);
 420 }
 421
 422 /**
 423  * traverse function for the traversal of the delete_queue,
 424  * the fast-path vacuuming list.
 425  *
 426  *  - If the record has been migrated off the node
 427  *    or has been revived (filled with data) on the node,
 428  *    then skip the record.
 429  *
 430  *  - If the current node is the record's lmaster and it is
 431  *    a record that has never been migrated with data, then
 432  *    delete the record from the local tdb.
 433  *
 434  *  - If the current node is the record's lmaster and it has
 435  *    been migrated with data, then schedule it for the normal
 436  *    vacuuming procedure (i.e. add it to the delete_list).
 437  *
 438  *  - If the current node is NOT the record's lmaster then
 439  *    add it to the list of records that are to be sent to
 440  *    the lmaster with the VACUUM_FETCH message.
 441  */
 442 static int delete_queue_traverse(void *param, void *data)
 443 {
 444         struct delete_record_data *dd =
 445                 talloc_get_type(data, struct delete_record_data);
 446         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 447         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 448         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 449         int res;
 450         struct ctdb_ltdb_header header;
 451         uint32_t lmaster;
 452         uint32_t hash = ctdb_hash(&(dd->key));
 453
 454         vdata->count.delete_queue.total++;
 455
 456         res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
 457         if (res != 0) {
 458                 vdata->count.delete_queue.error++;
 459                 return 0;
 460         }
 461
 462         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 463                                vacuum_record_parser, &header);
 464         if (res != 0) {
 465                 goto skipped;
 466         }
 467
 468         if (header.dmaster != ctdb->pnn) {
 469                 /* The record has been migrated off the node. Skip. */
 470                 goto skipped;
 471         }
 472
 473         if (header.rsn != dd->hdr.rsn) {
 474                 /*
 475                  * The record has been migrated off the node and back again.
 476                  * But not requeued for deletion. Skip it.
 477                  */
 478                 goto skipped;
 479         }
 480
 481         /*
 482          * We are dmaster, and the record has no data, and it has
 483          * not been migrated after it has been queued for deletion.
 484          *
 485          * At this stage, the record could still have been revived locally
 486          * and last been written with empty data. This can only be
 487          * fixed with the addition of an active or delete flag. (TODO)
 488          */
 489
 490         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 491
 492         if (lmaster != ctdb->pnn) {
 493                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 494
 495                 if (res != 0) {
 496                         DEBUG(DEBUG_ERR,
 497                               (__location__ " Error adding record to list "
 498                                "of records to send to lmaster.\n"));
 499                         vdata->count.delete_queue.error++;
 500                 } else {
 501                         vdata->count.delete_queue.added_to_vacuum_fetch_list++;
 502                 }
 503                 goto done;
 504         }
 505
 506         /* use header->flags or dd->hdr.flags ?? */
 507         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 508                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 509
 510                 if (res != 0) {
 511                         DEBUG(DEBUG_ERR,
 512                               (__location__ " Error adding record to list "
 513                                "of records for deletion on lmaster.\n"));
 514                         vdata->count.delete_queue.error++;
 515                 } else {
 516                         vdata->count.delete_queue.added_to_delete_list++;
 517                 }
 518         } else {
 519                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 520
 521                 if (res != 0) {
 522                         DEBUG(DEBUG_ERR,
 523                               (__location__ " Error deleting record with key "
 524                                "hash [0x%08x] from local data base db[%s].\n",
 525                                hash, ctdb_db->db_name));
 526                         vdata->count.delete_queue.error++;
 527                         goto done;
 528                 }
 529
 530                 DEBUG(DEBUG_DEBUG,
 531                       (__location__ " Deleted record with key hash "
 532                        "[0x%08x] from local data base db[%s].\n",
 533                        hash, ctdb_db->db_name));
 534                 vdata->count.delete_queue.deleted++;
 535         }
 536
 537         goto done;
 538
 539 skipped:
 540         vdata->count.delete_queue.skipped++;
 541
 542 done:
 543         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 544
 545         return 0;
 546 }
 547
 548 /**
 549  * Delete the records that we are lmaster and dmaster for and
 550  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 551  * control.
 552  */
 553 static int delete_record_traverse(void *param, void *data)
 554 {
 555         struct delete_record_data *dd =
 556                 talloc_get_type(data, struct delete_record_data);
 557         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 558         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 559         struct ctdb_context *ctdb = ctdb_db->ctdb;
 560         int res;
 561         struct ctdb_ltdb_header header;
 562         uint32_t lmaster;
 563         uint32_t hash = ctdb_hash(&(dd->key));
 564
 565         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 566         if (res != 0) {
 567                 DEBUG(DEBUG_ERR,
 568                       (__location__ " Error getting chainlock on record with "
 569                        "key hash [0x%08x] on database db[%s].\n",
 570                        hash, ctdb_db->db_name));
 571                 vdata->count.delete_list.local_error++;
 572                 vdata->count.delete_list.left--;
 573                 talloc_free(dd);
 574                 return 0;
 575         }
 576
 577         /*
 578          * Verify that the record is still empty, its RSN has not
 579          * changed and that we are still its lmaster and dmaster.
 580          */
 581
 582         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 583                                vacuum_record_parser, &header);
 584         if (res != 0) {
 585                 goto skip;
 586         }
 587
 588         if (header.flags & CTDB_REC_RO_FLAGS) {
 589                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 590                                    "on database db[%s] has read-only flags. "
 591                                    "skipping.\n",
 592                                    hash, ctdb_db->db_name));
 593                 goto skip;
 594         }
 595
 596         if (header.dmaster != ctdb->pnn) {
 597                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 598                                    "on database db[%s] has been migrated away. "
 599                                    "skipping.\n",
 600                                    hash, ctdb_db->db_name));
 601                 goto skip;
 602         }
 603
 604         if (header.rsn != dd->hdr.rsn + 1) {
 605                 /*
 606                  * The record has been migrated off the node and back again.
 607                  * But not requeued for deletion. Skip it.
 608                  * (Note that the first marshall traverse has bumped the RSN
 609                  *  on disk.)
 610                  */
 611                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 612                                    "on database db[%s] seems to have been "
 613                                    "migrated away and back again (with empty "
 614                                    "data). skipping.\n",
 615                                    hash, ctdb_db->db_name));
 616                 goto skip;
 617         }
 618
 619         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 620
 621         if (lmaster != ctdb->pnn) {
 622                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 623                                    "delete list (key hash [0x%08x], db[%s]). "
 624                                    "Strange! skipping.\n",
 625                                    hash, ctdb_db->db_name));
 626                 goto skip;
 627         }
 628
 629         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 630
 631         if (res != 0) {
 632                 DEBUG(DEBUG_ERR,
 633                       (__location__ " Error deleting record with key hash "
 634                        "[0x%08x] from local data base db[%s].\n",
 635                        hash, ctdb_db->db_name));
 636                 vdata->count.delete_list.local_error++;
 637                 goto done;
 638         }
 639
 640         DEBUG(DEBUG_DEBUG,
 641               (__location__ " Deleted record with key hash [0x%08x] from "
 642                "local data base db[%s].\n", hash, ctdb_db->db_name));
 643
 644         vdata->count.delete_list.deleted++;
 645         goto done;
 646
 647 skip:
 648         vdata->count.delete_list.skipped++;
 649
 650 done:
 651         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 652
 653         talloc_free(dd);
 654         vdata->count.delete_list.left--;
 655
 656         return 0;
 657 }
 658
 659 /**
 660  * Traverse the delete_queue.
 661  * Records are either deleted directly or filled
 662  * into the delete list or the vacuum fetch lists
 663  * for further processing.
 664  */
 665 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 666                                       struct vacuum_data *vdata)
 667 {
 668         uint32_t sum;
 669         int ret;
 670
 671         ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
 672                                    delete_queue_traverse, vdata);
 673
 674         if (ret != 0) {
 675                 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
 676                       "the delete queue.\n"));
 677         }
 678
 679         sum = vdata->count.delete_queue.deleted
 680             + vdata->count.delete_queue.skipped
 681             + vdata->count.delete_queue.error
 682             + vdata->count.delete_queue.added_to_delete_list
 683             + vdata->count.delete_queue.added_to_vacuum_fetch_list;
 684
 685         if (vdata->count.delete_queue.total != sum) {
 686                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 687                       "counts for db[%s]: total[%u] != sum[%u]\n",
 688                       ctdb_db->db_name,
 689                       (unsigned)vdata->count.delete_queue.total,
 690                       (unsigned)sum));
 691         }
 692
 693         if (vdata->count.delete_queue.total > 0) {
 694                 DEBUG(DEBUG_INFO,
 695                       (__location__
 696                        " fast vacuuming delete_queue traverse statistics: "
 697                        "db[%s] "
 698                        "total[%u] "
 699                        "del[%u] "
 700                        "skp[%u] "
 701                        "err[%u] "
 702                        "adl[%u] "
 703                        "avf[%u]\n",
 704                        ctdb_db->db_name,
 705                        (unsigned)vdata->count.delete_queue.total,
 706                        (unsigned)vdata->count.delete_queue.deleted,
 707                        (unsigned)vdata->count.delete_queue.skipped,
 708                        (unsigned)vdata->count.delete_queue.error,
 709                        (unsigned)vdata->count.delete_queue.added_to_delete_list,
 710                        (unsigned)vdata->count.delete_queue.added_to_vacuum_fetch_list));
 711         }
 712
 713         return;
 714 }
 715
 716 /**
 717  * read-only traverse of the database, looking for records that
 718  * might be able to be vacuumed.
 719  *
 720  * This is not done each time but only every tunable
 721  * VacuumFastPathCount times.
 722  */
 723 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 724                                     struct vacuum_data *vdata)
 725 {
 726         int ret;
 727
 728         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 729         if (ret == -1 || vdata->traverse_error) {
 730                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 731                                   "'%s'\n", ctdb_db->db_name));
 732                 return;
 733         }
 734
 735         if (vdata->count.db_traverse.total > 0) {
 736                 DEBUG(DEBUG_INFO,
 737                       (__location__
 738                        " full vacuuming db traverse statistics: "
 739                        "db[%s] "
 740                        "total[%u] "
 741                        "skp[%u] "
 742                        "err[%u] "
 743                        "sched[%u]\n",
 744                        ctdb_db->db_name,
 745                        (unsigned)vdata->count.db_traverse.total,
 746                        (unsigned)vdata->count.db_traverse.skipped,
 747                        (unsigned)vdata->count.db_traverse.error,
 748                        (unsigned)vdata->count.db_traverse.scheduled));
 749         }
 750
 751         return;
 752 }
 753
 754 /**
 755  * Process the vacuum fetch lists:
 756  * For records for which we are not the lmaster, tell the lmaster to
 757  * fetch the record.
 758  */
 759 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 760                                             struct vacuum_data *vdata)
 761 {
 762         int i;
 763         struct ctdb_context *ctdb = ctdb_db->ctdb;
 764
 765         for (i = 0; i < ctdb->num_nodes; i++) {
 766                 TDB_DATA data;
 767                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 768
 769                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 770                         continue;
 771                 }
 772
 773                 if (vfl->count == 0) {
 774                         continue;
 775                 }
 776
 777                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 778                                    vfl->count, ctdb->nodes[i]->pnn,
 779                                    ctdb_db->db_name));
 780
 781                 data = ctdb_marshall_finish(vfl);
 782                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 783                                              CTDB_SRVID_VACUUM_FETCH,
 784                                              data) != 0)
 785                 {
 786                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 787                                           "fetch message to %u\n",
 788                                           ctdb->nodes[i]->pnn));
 789                 }
 790         }
 791
 792         return;
 793 }
 794
 795 /**
 796  * Process the delete list:
 797  *
 798  * This is the last step of vacuuming that consistently deletes
 799  * those records that have been migrated with data and can hence
 800  * not be deleted when leaving a node.
 801  *
 802  * In this step, the lmaster does the final deletion of those empty
 803  * records that it is also dmaster for. It has ususally received
 804  * at least some of these records previously from the former dmasters
 805  * with the vacuum fetch message.
 806  *
 807  * This last step is implemented as a 3-phase process to protect from
 808  * races leading to data corruption:
 809  *
 810  *  1) Send the lmaster's copy to all other active nodes with the
 811  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 812  *  2) Send the records that could successfully be stored remotely
 813  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 814  *     control. The remote notes delete their local copy.
 815  *  3) The lmaster locally deletes its copies of all records that
 816  *     could successfully be deleted remotely in step #2.
 817  */
 818 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 819                                      struct vacuum_data *vdata)
 820 {
 821         int ret, i;
 822         struct ctdb_context *ctdb = ctdb_db->ctdb;
 823         struct delete_records_list *recs;
 824         TDB_DATA indata;
 825         struct ctdb_node_map_old *nodemap;
 826         uint32_t *active_nodes;
 827         int num_active_nodes;
 828         TALLOC_CTX *tmp_ctx;
 829         uint32_t sum;
 830
 831         if (vdata->count.delete_list.total == 0) {
 832                 return;
 833         }
 834
 835         tmp_ctx = talloc_new(vdata);
 836         if (tmp_ctx == NULL) {
 837                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 838                 return;
 839         }
 840
 841         vdata->count.delete_list.left = vdata->count.delete_list.total;
 842
 843         /*
 844          * get the list of currently active nodes
 845          */
 846
 847         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 848                                    CTDB_CURRENT_NODE,
 849                                    tmp_ctx,
 850                                    &nodemap);
 851         if (ret != 0) {
 852                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 853                 goto done;
 854         }
 855
 856         active_nodes = list_of_active_nodes(ctdb, nodemap,
 857                                             nodemap, /* talloc context */
 858                                             false /* include self */);
 859         /* yuck! ;-) */
 860         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 861
 862         /*
 863          * Now delete the records all active nodes in a three-phase process:
 864          * 1) send all active remote nodes the current empty copy with this
 865          *    node as DMASTER
 866          * 2) if all nodes could store the new copy,
 867          *    tell all the active remote nodes to delete all their copy
 868          * 3) if all remote nodes deleted their record copy, delete it locally
 869          */
 870
 871         /*
 872          * Step 1:
 873          * Send currently empty record copy to all active nodes for storing.
 874          */
 875
 876         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 877         if (recs == NULL) {
 878                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 879                 goto done;
 880         }
 881         recs->records = (struct ctdb_marshall_buffer *)
 882                 talloc_zero_size(recs,
 883                                  offsetof(struct ctdb_marshall_buffer, data));
 884         if (recs->records == NULL) {
 885                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 886                 goto done;
 887         }
 888         recs->records->db_id = ctdb_db->db_id;
 889         recs->vdata = vdata;
 890
 891         /*
 892          * traverse the tree of all records we want to delete and
 893          * create a blob we can send to the other nodes.
 894          *
 895          * We call delete_marshall_traverse_first() to bump the
 896          * records' RSNs in the database, to ensure we (as dmaster)
 897          * keep the highest RSN of the records in the cluster.
 898          */
 899         ret = trbt_traversearray32(vdata->delete_list, 1,
 900                                    delete_marshall_traverse_first, recs);
 901         if (ret != 0) {
 902                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
 903                       "delete list for first marshalling.\n"));
 904                 goto done;
 905         }
 906
 907         indata = ctdb_marshall_finish(recs->records);
 908
 909         for (i = 0; i < num_active_nodes; i++) {
 910                 struct ctdb_marshall_buffer *records;
 911                 struct ctdb_rec_data_old *rec;
 912                 int32_t res;
 913                 TDB_DATA outdata;
 914
 915                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 916                                 CTDB_CONTROL_RECEIVE_RECORDS, 0,
 917                                 indata, recs, &outdata, &res,
 918                                 NULL, NULL);
 919                 if (ret != 0 || res != 0) {
 920                         DEBUG(DEBUG_ERR, ("Error storing record copies on "
 921                                           "node %u: ret[%d] res[%d]\n",
 922                                           active_nodes[i], ret, res));
 923                         goto done;
 924                 }
 925
 926                 /*
 927                  * outdata contains the list of records coming back
 928                  * from the node: These are the records that the
 929                  * remote node could not store. We remove these from
 930                  * the list to process further.
 931                  */
 932                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 933                 rec = (struct ctdb_rec_data_old *)&records->data[0];
 934                 while (records->count-- > 1) {
 935                         TDB_DATA reckey, recdata;
 936                         struct ctdb_ltdb_header *rechdr;
 937                         struct delete_record_data *dd;
 938
 939                         reckey.dptr = &rec->data[0];
 940                         reckey.dsize = rec->keylen;
 941                         recdata.dptr = &rec->data[reckey.dsize];
 942                         recdata.dsize = rec->datalen;
 943
 944                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 945                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 946                                 goto done;
 947                         }
 948                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 949                         recdata.dptr += sizeof(*rechdr);
 950                         recdata.dsize -= sizeof(*rechdr);
 951
 952                         dd = (struct delete_record_data *)trbt_lookup32(
 953                                         vdata->delete_list,
 954                                         ctdb_hash(&reckey));
 955                         if (dd != NULL) {
 956                                 /*
 957                                  * The other node could not store the record
 958                                  * copy and it is the first node that failed.
 959                                  * So we should remove it from the tree and
 960                                  * update statistics.
 961                                  */
 962                                 talloc_free(dd);
 963                                 vdata->count.delete_list.remote_error++;
 964                                 vdata->count.delete_list.left--;
 965                         } else {
 966                                 DEBUG(DEBUG_ERR, (__location__ " Failed to "
 967                                       "find record with hash 0x%08x coming "
 968                                       "back from RECEIVE_RECORDS "
 969                                       "control in delete list.\n",
 970                                       ctdb_hash(&reckey)));
 971                                 vdata->count.delete_list.local_error++;
 972                                 vdata->count.delete_list.left--;
 973                         }
 974
 975                         rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
 976                 }
 977         }
 978
 979         if (vdata->count.delete_list.left == 0) {
 980                 goto success;
 981         }
 982
 983         /*
 984          * Step 2:
 985          * Send the remaining records to all active nodes for deletion.
 986          *
 987          * The lmaster's (i.e. our) copies of these records have been stored
 988          * successfully on the other nodes.
 989          */
 990
 991         /*
 992          * Create a marshall blob from the remaining list of records to delete.
 993          */
 994
 995         talloc_free(recs->records);
 996
 997         recs->records = (struct ctdb_marshall_buffer *)
 998                 talloc_zero_size(recs,
 999                                  offsetof(struct ctdb_marshall_buffer, data));
1000         if (recs->records == NULL) {
1001                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1002                 goto done;
1003         }
1004         recs->records->db_id = ctdb_db->db_id;
1005
1006         ret = trbt_traversearray32(vdata->delete_list, 1,
1007                                    delete_marshall_traverse, recs);
1008         if (ret != 0) {
1009                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1010                       "delete list for second marshalling.\n"));
1011                 goto done;
1012         }
1013
1014         indata = ctdb_marshall_finish(recs->records);
1015
1016         for (i = 0; i < num_active_nodes; i++) {
1017                 struct ctdb_marshall_buffer *records;
1018                 struct ctdb_rec_data_old *rec;
1019                 int32_t res;
1020                 TDB_DATA outdata;
1021
1022                 ret = ctdb_control(ctdb, active_nodes[i], 0,
1023                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1024                                 indata, recs, &outdata, &res,
1025                                 NULL, NULL);
1026                 if (ret != 0 || res != 0) {
1027                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
1028                                           "node %u: ret[%d] res[%d]\n",
1029                                           active_nodes[i], ret, res));
1030                         goto done;
1031                 }
1032
1033                 /*
1034                  * outdata contains the list of records coming back
1035                  * from the node: These are the records that the
1036                  * remote node could not delete. We remove these from
1037                  * the list to delete locally.
1038                  */
1039                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1040                 rec = (struct ctdb_rec_data_old *)&records->data[0];
1041                 while (records->count-- > 1) {
1042                         TDB_DATA reckey, recdata;
1043                         struct ctdb_ltdb_header *rechdr;
1044                         struct delete_record_data *dd;
1045
1046                         reckey.dptr = &rec->data[0];
1047                         reckey.dsize = rec->keylen;
1048                         recdata.dptr = &rec->data[reckey.dsize];
1049                         recdata.dsize = rec->datalen;
1050
1051                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1052                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1053                                 goto done;
1054                         }
1055                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1056                         recdata.dptr += sizeof(*rechdr);
1057                         recdata.dsize -= sizeof(*rechdr);
1058
1059                         dd = (struct delete_record_data *)trbt_lookup32(
1060                                         vdata->delete_list,
1061                                         ctdb_hash(&reckey));
1062                         if (dd != NULL) {
1063                                 /*
1064                                  * The other node could not delete the
1065                                  * record and it is the first node that
1066                                  * failed. So we should remove it from
1067                                  * the tree and update statistics.
1068                                  */
1069                                 talloc_free(dd);
1070                                 vdata->count.delete_list.remote_error++;
1071                                 vdata->count.delete_list.left--;
1072                         } else {
1073                                 DEBUG(DEBUG_ERR, (__location__ " Failed to "
1074                                       "find record with hash 0x%08x coming "
1075                                       "back from TRY_DELETE_RECORDS "
1076                                       "control in delete list.\n",
1077                                       ctdb_hash(&reckey)));
1078                                 vdata->count.delete_list.local_error++;
1079                                 vdata->count.delete_list.left--;
1080                         }
1081
1082                         rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
1083                 }
1084         }
1085
1086         if (vdata->count.delete_list.left == 0) {
1087                 goto success;
1088         }
1089
1090         /*
1091          * Step 3:
1092          * Delete the remaining records locally.
1093          *
1094          * These records have successfully been deleted on all
1095          * active remote nodes.
1096          */
1097
1098         ret = trbt_traversearray32(vdata->delete_list, 1,
1099                                    delete_record_traverse, vdata);
1100         if (ret != 0) {
1101                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1102                       "delete list for deletion.\n"));
1103         }
1104
1105 success:
1106
1107         if (vdata->count.delete_list.left != 0) {
1108                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
1109                       "there are %u records left for deletion after "
1110                       "processing delete list\n",
1111                       ctdb_db->db_name,
1112                       (unsigned)vdata->count.delete_list.left));
1113         }
1114
1115         sum = vdata->count.delete_list.deleted
1116             + vdata->count.delete_list.skipped
1117             + vdata->count.delete_list.remote_error
1118             + vdata->count.delete_list.local_error
1119             + vdata->count.delete_list.left;
1120
1121         if (vdata->count.delete_list.total != sum) {
1122                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
1123                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1124                       ctdb_db->db_name,
1125                       (unsigned)vdata->count.delete_list.total,
1126                       (unsigned)sum));
1127         }
1128
1129         if (vdata->count.delete_list.total > 0) {
1130                 DEBUG(DEBUG_INFO,
1131                       (__location__
1132                        " vacuum delete list statistics: "
1133                        "db[%s] "
1134                        "total[%u] "
1135                        "del[%u] "
1136                        "skip[%u] "
1137                        "rem.err[%u] "
1138                        "loc.err[%u] "
1139                        "left[%u]\n",
1140                        ctdb_db->db_name,
1141                        (unsigned)vdata->count.delete_list.total,
1142                        (unsigned)vdata->count.delete_list.deleted,
1143                        (unsigned)vdata->count.delete_list.skipped,
1144                        (unsigned)vdata->count.delete_list.remote_error,
1145                        (unsigned)vdata->count.delete_list.local_error,
1146                        (unsigned)vdata->count.delete_list.left));
1147         }
1148
1149 done:
1150         talloc_free(tmp_ctx);
1151
1152         return;
1153 }
1154
1155 /**
1156  * initialize the vacuum_data
1157  */
1158 static struct vacuum_data *ctdb_vacuum_init_vacuum_data(
1159                                         struct ctdb_db_context *ctdb_db,
1160                                         TALLOC_CTX *mem_ctx)
1161 {
1162         int i;
1163         struct ctdb_context *ctdb = ctdb_db->ctdb;
1164         struct vacuum_data *vdata;
1165
1166         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1167         if (vdata == NULL) {
1168                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1169                 return NULL;
1170         }
1171
1172         vdata->ctdb = ctdb_db->ctdb;
1173         vdata->ctdb_db = ctdb_db;
1174         vdata->delete_list = trbt_create(vdata, 0);
1175         if (vdata->delete_list == NULL) {
1176                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1177                 goto fail;
1178         }
1179
1180         vdata->start = timeval_current();
1181
1182         vdata->count.delete_queue.added_to_delete_list = 0;
1183         vdata->count.delete_queue.added_to_vacuum_fetch_list = 0;
1184         vdata->count.delete_queue.deleted = 0;
1185         vdata->count.delete_queue.skipped = 0;
1186         vdata->count.delete_queue.error = 0;
1187         vdata->count.delete_queue.total = 0;
1188         vdata->count.db_traverse.scheduled = 0;
1189         vdata->count.db_traverse.skipped = 0;
1190         vdata->count.db_traverse.error = 0;
1191         vdata->count.db_traverse.total = 0;
1192         vdata->count.delete_list.total = 0;
1193         vdata->count.delete_list.left = 0;
1194         vdata->count.delete_list.remote_error = 0;
1195         vdata->count.delete_list.local_error = 0;
1196         vdata->count.delete_list.skipped = 0;
1197         vdata->count.delete_list.deleted = 0;
1198
1199         /* the list needs to be of length num_nodes */
1200         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1201                                                 struct ctdb_marshall_buffer *,
1202                                                 ctdb->num_nodes);
1203         if (vdata->vacuum_fetch_list == NULL) {
1204                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1205                 goto fail;
1206         }
1207         for (i = 0; i < ctdb->num_nodes; i++) {
1208                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1209                         talloc_zero_size(vdata->vacuum_fetch_list,
1210                                          offsetof(struct ctdb_marshall_buffer, data));
1211                 if (vdata->vacuum_fetch_list[i] == NULL) {
1212                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1213                         talloc_free(vdata);
1214                         return NULL;
1215                 }
1216                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1217         }
1218
1219         return vdata;
1220
1221 fail:
1222         talloc_free(vdata);
1223         return NULL;
1224 }
1225
1226 /**
1227  * Vacuum a DB:
1228  *  - Always do the fast vacuuming run, which traverses
1229  *    the in-memory delete queue: these records have been
1230  *    scheduled for deletion.
1231  *  - Only if explicitly requested, the database is traversed
1232  *    in order to use the traditional heuristics on empty records
1233  *    to trigger deletion.
1234  *    This is done only every VacuumFastPathCount'th vacuuming run.
1235  *
1236  * The traverse runs fill two lists:
1237  *
1238  * - The delete_list:
1239  *   This is the list of empty records the current
1240  *   node is lmaster and dmaster for. These records are later
1241  *   deleted first on other nodes and then locally.
1242  *
1243  *   The fast vacuuming run has a short cut for those records
1244  *   that have never been migrated with data: these records
1245  *   are immediately deleted locally, since they have left
1246  *   no trace on other nodes.
1247  *
1248  * - The vacuum_fetch lists
1249  *   (one for each other lmaster node):
1250  *   The records in this list are sent for deletion to
1251  *   their lmaster in a bulk VACUUM_FETCH message.
1252  *
1253  *   The lmaster then migrates all these records to itelf
1254  *   so that they can be vacuumed there.
1255  *
1256  * This executes in the child context.
1257  */
1258 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1259                           bool full_vacuum_run)
1260 {
1261         struct ctdb_context *ctdb = ctdb_db->ctdb;
1262         int ret, pnn;
1263         struct vacuum_data *vdata;
1264         TALLOC_CTX *tmp_ctx;
1265
1266         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1267                            "%s db_id[0x%08x]\n",
1268                            full_vacuum_run ? "full" : "fast",
1269                            ctdb_db->db_name, ctdb_db->db_id));
1270
1271         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1272         if (ret != 0) {
1273                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1274                 return ret;
1275         }
1276
1277         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1278         if (pnn == -1) {
1279                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1280                 return -1;
1281         }
1282
1283         ctdb->pnn = pnn;
1284
1285         tmp_ctx = talloc_new(ctdb_db);
1286         if (tmp_ctx == NULL) {
1287                 DEBUG(DEBUG_ERR, ("Out of memory!\n"));
1288                 return -1;
1289         }
1290
1291         vdata = ctdb_vacuum_init_vacuum_data(ctdb_db, tmp_ctx);
1292         if (vdata == NULL) {
1293                 talloc_free(tmp_ctx);
1294                 return -1;
1295         }
1296
1297         if (full_vacuum_run) {
1298                 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1299         }
1300
1301         ctdb_process_delete_queue(ctdb_db, vdata);
1302
1303         ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1304
1305         ctdb_process_delete_list(ctdb_db, vdata);
1306
1307         talloc_free(tmp_ctx);
1308
1309         /* this ensures we run our event queue */
1310         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1311
1312         return 0;
1313 }
1314
1315 /*
1316  * repack and vaccum a db
1317  * called from the child context
1318  */
1319 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1320                                      bool full_vacuum_run)
1321 {
1322         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1323         const char *name = ctdb_db->db_name;
1324         int freelist_size = 0;
1325         int ret;
1326
1327         if (ctdb_vacuum_db(ctdb_db, full_vacuum_run) != 0) {
1328                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1329         }
1330
1331         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1332         if (freelist_size == -1) {
1333                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1334                 return -1;
1335         }
1336
1337         /*
1338          * decide if a repack is necessary
1339          */
1340         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1341         {
1342                 return 0;
1343         }
1344
1345         DEBUG(DEBUG_INFO, ("Repacking %s with %u freelist entries\n",
1346                            name, freelist_size));
1347
1348         ret = tdb_repack(ctdb_db->ltdb->tdb);
1349         if (ret != 0) {
1350                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1351                 return -1;
1352         }
1353
1354         return 0;
1355 }
1356
1357 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1358 {
1359         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1360
1361         return interval;
1362 }
1363
1364 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1365 {
1366         double l = timeval_elapsed(&child_ctx->start_time);
1367         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1368         struct ctdb_context *ctdb = ctdb_db->ctdb;
1369
1370         CTDB_UPDATE_DB_LATENCY(ctdb_db, "vacuum", vacuum.latency, l);
1371         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1372
1373         if (child_ctx->child_pid != -1) {
1374                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1375         } else {
1376                 /* Bump the number of successful fast-path runs. */
1377                 child_ctx->vacuum_handle->fast_path_count++;
1378         }
1379
1380         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1381
1382         tevent_add_timer(ctdb->ev, child_ctx->vacuum_handle,
1383                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1384                          ctdb_vacuum_event, child_ctx->vacuum_handle);
1385
1386         return 0;
1387 }
1388
1389 /*
1390  * this event is generated when a vacuum child process times out
1391  */
1392 static void vacuum_child_timeout(struct tevent_context *ev,
1393                                  struct tevent_timer *te,
1394                                  struct timeval t, void *private_data)
1395 {
1396         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1397
1398         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1399
1400         child_ctx->status = VACUUM_TIMEOUT;
1401
1402         talloc_free(child_ctx);
1403 }
1404
1405
1406 /*
1407  * this event is generated when a vacuum child process has completed
1408  */
1409 static void vacuum_child_handler(struct tevent_context *ev,
1410                                  struct tevent_fd *fde,
1411                                  uint16_t flags, void *private_data)
1412 {
1413         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1414         char c = 0;
1415         int ret;
1416
1417         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1418         child_ctx->child_pid = -1;
1419
1420         ret = sys_read(child_ctx->fd[0], &c, 1);
1421         if (ret != 1 || c != 0) {
1422                 child_ctx->status = VACUUM_ERROR;
1423                 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));
1424         } else {
1425                 child_ctx->status = VACUUM_OK;
1426         }
1427
1428         talloc_free(child_ctx);
1429 }
1430
1431 /*
1432  * this event is called every time we need to start a new vacuum process
1433  */
1434 static void ctdb_vacuum_event(struct tevent_context *ev,
1435                               struct tevent_timer *te,
1436                               struct timeval t, void *private_data)
1437 {
1438         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1439         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1440         struct ctdb_context *ctdb = ctdb_db->ctdb;
1441         struct ctdb_vacuum_child_context *child_ctx;
1442         struct tevent_fd *fde;
1443         int ret;
1444
1445         /* we dont vacuum if we are in recovery mode, or db frozen */
1446         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1447             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1448                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1449                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1450                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1451                                    ? "freeze pending"
1452                                    : "frozen"));
1453                 tevent_add_timer(ctdb->ev, vacuum_handle,
1454                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1455                                  ctdb_vacuum_event, vacuum_handle);
1456                 return;
1457         }
1458
1459         /* Do not allow multiple vacuuming child processes to be active at the
1460          * same time.  If there is vacuuming child process active, delay
1461          * new vacuuming event to stagger vacuuming events.
1462          */
1463         if (ctdb->vacuumers != NULL) {
1464                 tevent_add_timer(ctdb->ev, vacuum_handle,
1465                                  timeval_current_ofs(0, 500*1000),
1466                                  ctdb_vacuum_event, vacuum_handle);
1467                 return;
1468         }
1469
1470         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1471         if (child_ctx == NULL) {
1472                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1473                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1474         }
1475
1476
1477         ret = pipe(child_ctx->fd);
1478         if (ret != 0) {
1479                 talloc_free(child_ctx);
1480                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1481                 tevent_add_timer(ctdb->ev, vacuum_handle,
1482                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1483                                  ctdb_vacuum_event, vacuum_handle);
1484                 return;
1485         }
1486
1487         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1488                 vacuum_handle->fast_path_count = 0;
1489         }
1490
1491         child_ctx->child_pid = ctdb_fork(ctdb);
1492         if (child_ctx->child_pid == (pid_t)-1) {
1493                 close(child_ctx->fd[0]);
1494                 close(child_ctx->fd[1]);
1495                 talloc_free(child_ctx);
1496                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1497                 tevent_add_timer(ctdb->ev, vacuum_handle,
1498                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1499                                  ctdb_vacuum_event, vacuum_handle);
1500                 return;
1501         }
1502
1503
1504         if (child_ctx->child_pid == 0) {
1505                 char cc = 0;
1506                 bool full_vacuum_run = false;
1507                 close(child_ctx->fd[0]);
1508
1509                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1510                 ctdb_set_process_name("ctdb_vacuum");
1511                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1512                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1513                         _exit(1);
1514                 }
1515
1516                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1517                     (vacuum_handle->fast_path_count == 0))
1518                 {
1519                         full_vacuum_run = true;
1520                 }
1521                 cc = ctdb_vacuum_and_repack_db(ctdb_db, full_vacuum_run);
1522
1523                 sys_write(child_ctx->fd[1], &cc, 1);
1524                 _exit(0);
1525         }
1526
1527         set_close_on_exec(child_ctx->fd[0]);
1528         close(child_ctx->fd[1]);
1529
1530         child_ctx->status = VACUUM_RUNNING;
1531         child_ctx->start_time = timeval_current();
1532
1533         DLIST_ADD(ctdb->vacuumers, child_ctx);
1534         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1535
1536         /*
1537          * Clear the fastpath vacuuming list in the parent.
1538          */
1539         talloc_free(ctdb_db->delete_queue);
1540         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1541         if (ctdb_db->delete_queue == NULL) {
1542                 /* fatal here? ... */
1543                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1544                                  "in parent context. Shutting down\n");
1545         }
1546
1547         tevent_add_timer(ctdb->ev, child_ctx,
1548                          timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1549                          vacuum_child_timeout, child_ctx);
1550
1551         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1552
1553         fde = tevent_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1554                             TEVENT_FD_READ, vacuum_child_handler, child_ctx);
1555         tevent_fd_set_auto_close(fde);
1556
1557         vacuum_handle->child_ctx = child_ctx;
1558         child_ctx->vacuum_handle = vacuum_handle;
1559 }
1560
1561 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1562 {
1563         /* Simply free them all. */
1564         while (ctdb->vacuumers) {
1565                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1566                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1567                            (int)ctdb->vacuumers->child_pid));
1568                 /* vacuum_child_destructor kills it, removes from list */
1569                 talloc_free(ctdb->vacuumers);
1570         }
1571 }
1572
1573 /* this function initializes the vacuuming context for a database
1574  * starts the vacuuming events
1575  */
1576 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1577 {
1578         if (ctdb_db->persistent != 0) {
1579                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1580                 return 0;
1581         }
1582
1583         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1584         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1585
1586         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1587         ctdb_db->vacuum_handle->fast_path_count = 0;
1588
1589         tevent_add_timer(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1590                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1591                          ctdb_vacuum_event, ctdb_db->vacuum_handle);
1592
1593         return 0;
1594 }
1595
1596 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1597                                             const struct ctdb_ltdb_header *hdr,
1598                                             const TDB_DATA key)
1599 {
1600         struct delete_record_data *kd;
1601         uint32_t hash;
1602
1603         hash = (uint32_t)ctdb_hash(&key);
1604
1605         DEBUG(DEBUG_DEBUG, (__location__
1606                             " remove_record_from_delete_queue: "
1607                             "db[%s] "
1608                             "db_id[0x%08x] "
1609                             "key_hash[0x%08x] "
1610                             "lmaster[%u] "
1611                             "migrated_with_data[%s]\n",
1612                              ctdb_db->db_name, ctdb_db->db_id,
1613                              hash,
1614                              ctdb_lmaster(ctdb_db->ctdb, &key),
1615                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1616
1617         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1618         if (kd == NULL) {
1619                 DEBUG(DEBUG_DEBUG, (__location__
1620                                     " remove_record_from_delete_queue: "
1621                                     "record not in queue (hash[0x%08x])\n.",
1622                                     hash));
1623                 return;
1624         }
1625
1626         if ((kd->key.dsize != key.dsize) ||
1627             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1628         {
1629                 DEBUG(DEBUG_DEBUG, (__location__
1630                                     " remove_record_from_delete_queue: "
1631                                     "hash collision for key with hash[0x%08x] "
1632                                     "in db[%s] - skipping\n",
1633                                     hash, ctdb_db->db_name));
1634                 return;
1635         }
1636
1637         DEBUG(DEBUG_DEBUG, (__location__
1638                             " remove_record_from_delete_queue: "
1639                             "removing key with hash[0x%08x]\n",
1640                              hash));
1641
1642         talloc_free(kd);
1643
1644         return;
1645 }
1646
1647 /**
1648  * Insert a record into the ctdb_db context's delete queue,
1649  * handling hash collisions.
1650  */
1651 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1652                                            const struct ctdb_ltdb_header *hdr,
1653                                            TDB_DATA key)
1654 {
1655         struct delete_record_data *kd;
1656         uint32_t hash;
1657         int ret;
1658
1659         hash = (uint32_t)ctdb_hash(&key);
1660
1661         DEBUG(DEBUG_INFO, (__location__ " schedule for deletion: db[%s] "
1662                            "db_id[0x%08x] "
1663                            "key_hash[0x%08x] "
1664                            "lmaster[%u] "
1665                            "migrated_with_data[%s]\n",
1666                             ctdb_db->db_name, ctdb_db->db_id,
1667                             hash,
1668                             ctdb_lmaster(ctdb_db->ctdb, &key),
1669                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1670
1671         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1672         if (kd != NULL) {
1673                 if ((kd->key.dsize != key.dsize) ||
1674                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1675                 {
1676                         DEBUG(DEBUG_INFO,
1677                               (__location__ " schedule for deletion: "
1678                                "hash collision for key hash [0x%08x]. "
1679                                "Skipping the record.\n", hash));
1680                         return 0;
1681                 } else {
1682                         DEBUG(DEBUG_DEBUG,
1683                               (__location__ " schedule for deletion: "
1684                                "updating entry for key with hash [0x%08x].\n",
1685                                hash));
1686                 }
1687         }
1688
1689         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1690                                                   ctdb_db->delete_queue,
1691                                                   hdr, key);
1692         if (ret != 0) {
1693                 DEBUG(DEBUG_INFO,
1694                       (__location__ " schedule for deletion: error "
1695                        "inserting key with hash [0x%08x] into delete queue\n",
1696                        hash));
1697                 return -1;
1698         }
1699
1700         return 0;
1701 }
1702
1703 /**
1704  * Schedule a record for deletetion.
1705  * Called from the parent context.
1706  */
1707 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1708                                            TDB_DATA indata)
1709 {
1710         struct ctdb_control_schedule_for_deletion *dd;
1711         struct ctdb_db_context *ctdb_db;
1712         int ret;
1713         TDB_DATA key;
1714
1715         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1716
1717         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1718         if (ctdb_db == NULL) {
1719                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1720                                   dd->db_id));
1721                 return -1;
1722         }
1723
1724         key.dsize = dd->keylen;
1725         key.dptr = dd->key;
1726
1727         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1728
1729         return ret;
1730 }
1731
1732 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1733                                          const struct ctdb_ltdb_header *hdr,
1734                                          TDB_DATA key)
1735 {
1736         int ret;
1737         struct ctdb_control_schedule_for_deletion *dd;
1738         TDB_DATA indata;
1739         int32_t status;
1740
1741         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1742                 /* main daemon - directly queue */
1743                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1744
1745                 return ret;
1746         }
1747
1748         /* if we dont have a connection to the daemon we can not send
1749            a control. For example sometimes from update_record control child
1750            process.
1751         */
1752         if (!ctdb_db->ctdb->can_send_controls) {
1753                 return -1;
1754         }
1755
1756
1757         /* child process: send the main daemon a control */
1758         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1759         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1760         if (indata.dptr == NULL) {
1761                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1762                 return -1;
1763         }
1764         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1765         dd->db_id = ctdb_db->db_id;
1766         dd->hdr = *hdr;
1767         dd->keylen = key.dsize;
1768         memcpy(dd->key, key.dptr, key.dsize);
1769
1770         ret = ctdb_control(ctdb_db->ctdb,
1771                            CTDB_CURRENT_NODE,
1772                            ctdb_db->db_id,
1773                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1774                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1775                            indata,
1776                            NULL, /* mem_ctx */
1777                            NULL, /* outdata */
1778                            &status,
1779                            NULL, /* timeout : NULL == wait forever */
1780                            NULL); /* error message */
1781
1782         talloc_free(indata.dptr);
1783
1784         if (ret != 0 || status != 0) {
1785                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1786                                   "SCHEDULE_FOR_DELETION "
1787                                   "control.\n"));
1788                 if (status != 0) {
1789                         ret = -1;
1790                 }
1791         }
1792
1793         return ret;
1794 }
1795
1796 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1797                                          const struct ctdb_ltdb_header *hdr,
1798                                          const TDB_DATA key)
1799 {
1800         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1801                 /*
1802                  * Only remove the record from the delete queue if called
1803                  * in the main daemon.
1804                  */
1805                 return;
1806         }
1807
1808         remove_record_from_delete_queue(ctdb_db, hdr, key);
1809
1810         return;
1811 }