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