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 #include "lib/util/util_process.h"
  35
  36 #include "ctdb_private.h"
  37 #include "ctdb_client.h"
  38
  39 #include "common/rb_tree.h"
  40 #include "common/system.h"
  41 #include "common/common.h"
  42 #include "common/logging.h"
  43
  44 #define TIMELIMIT() timeval_current_ofs(10, 0)
  45
  46 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  47
  48 struct ctdb_vacuum_child_context {
  49         struct ctdb_vacuum_child_context *next, *prev;
  50         struct ctdb_vacuum_handle *vacuum_handle;
  51         /* fd child writes status to */
  52         int fd[2];
  53         pid_t child_pid;
  54         enum vacuum_child_status status;
  55         struct timeval start_time;
  56 };
  57
  58 struct ctdb_vacuum_handle {
  59         struct ctdb_db_context *ctdb_db;
  60         struct ctdb_vacuum_child_context *child_ctx;
  61         uint32_t fast_path_count;
  62 };
  63
  64
  65 /*  a list of records to possibly delete */
  66 struct vacuum_data {
  67         struct ctdb_context *ctdb;
  68         struct ctdb_db_context *ctdb_db;
  69         struct tdb_context *dest_db;
  70         trbt_tree_t *delete_list;
  71         struct ctdb_marshall_buffer **vacuum_fetch_list;
  72         struct timeval start;
  73         bool traverse_error;
  74         bool vacuum;
  75         struct {
  76                 struct {
  77                         uint32_t added_to_vacuum_fetch_list;
  78                         uint32_t added_to_delete_list;
  79                         uint32_t deleted;
  80                         uint32_t skipped;
  81                         uint32_t error;
  82                         uint32_t total;
  83                 } delete_queue;
  84                 struct {
  85                         uint32_t scheduled;
  86                         uint32_t skipped;
  87                         uint32_t error;
  88                         uint32_t total;
  89                 } db_traverse;
  90                 struct {
  91                         uint32_t total;
  92                         uint32_t remote_error;
  93                         uint32_t local_error;
  94                         uint32_t deleted;
  95                         uint32_t skipped;
  96                         uint32_t left;
  97                 } delete_list;
  98                 struct {
  99                         uint32_t vacuumed;
 100                         uint32_t copied;
 101                 } repack;
 102         } count;
 103 };
 104
 105 /* this structure contains the information for one record to be deleted */
 106 struct delete_record_data {
 107         struct ctdb_context *ctdb;
 108         struct ctdb_db_context *ctdb_db;
 109         struct ctdb_ltdb_header hdr;
 110         TDB_DATA key;
 111         uint8_t keydata[1];
 112 };
 113
 114 struct delete_records_list {
 115         struct ctdb_marshall_buffer *records;
 116         struct vacuum_data *vdata;
 117 };
 118
 119 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
 120                                            const struct ctdb_ltdb_header *hdr,
 121                                            TDB_DATA key);
 122
 123 /**
 124  * Store key and header in a tree, indexed by the key hash.
 125  */
 126 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
 127                                                struct ctdb_db_context *ctdb_db,
 128                                                trbt_tree_t *tree,
 129                                                const struct ctdb_ltdb_header *hdr,
 130                                                TDB_DATA key)
 131 {
 132         struct delete_record_data *dd;
 133         uint32_t hash;
 134         size_t len;
 135
 136         len = offsetof(struct delete_record_data, keydata) + key.dsize;
 137
 138         dd = (struct delete_record_data *)talloc_size(tree, len);
 139         if (dd == NULL) {
 140                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 141                 return -1;
 142         }
 143         talloc_set_name_const(dd, "struct delete_record_data");
 144
 145         dd->ctdb      = ctdb;
 146         dd->ctdb_db   = ctdb_db;
 147         dd->key.dsize = key.dsize;
 148         dd->key.dptr  = dd->keydata;
 149         memcpy(dd->keydata, key.dptr, key.dsize);
 150
 151         dd->hdr = *hdr;
 152
 153         hash = ctdb_hash(&key);
 154
 155         trbt_insert32(tree, hash, dd);
 156
 157         return 0;
 158 }
 159
 160 static int add_record_to_delete_list(struct vacuum_data *vdata, TDB_DATA key,
 161                                      struct ctdb_ltdb_header *hdr)
 162 {
 163         struct ctdb_context *ctdb = vdata->ctdb;
 164         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 165         uint32_t hash;
 166         int ret;
 167
 168         hash = ctdb_hash(&key);
 169
 170         if (trbt_lookup32(vdata->delete_list, hash)) {
 171                 DEBUG(DEBUG_INFO, (__location__ " Hash collision when vacuuming, skipping this record.\n"));
 172                 return 0;
 173         }
 174
 175         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 176                                                   vdata->delete_list,
 177                                                   hdr, key);
 178         if (ret != 0) {
 179                 return -1;
 180         }
 181
 182         vdata->count.delete_list.total++;
 183
 184         return 0;
 185 }
 186
 187 /**
 188  * Add a record to the list of records to be sent
 189  * to their lmaster with VACUUM_FETCH.
 190  */
 191 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 192                                            TDB_DATA key)
 193 {
 194         struct ctdb_context *ctdb = vdata->ctdb;
 195         uint32_t lmaster;
 196         struct ctdb_marshall_buffer *vfl;
 197
 198         lmaster = ctdb_lmaster(ctdb, &key);
 199
 200         vfl = vdata->vacuum_fetch_list[lmaster];
 201
 202         vfl = ctdb_marshall_add(ctdb, vfl, vfl->db_id, ctdb->pnn,
 203                                 key, NULL, tdb_null);
 204         if (vfl == NULL) {
 205                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 206                 vdata->traverse_error = true;
 207                 return -1;
 208         }
 209
 210         vdata->vacuum_fetch_list[lmaster] = vfl;
 211
 212         return 0;
 213 }
 214
 215
 216 static void ctdb_vacuum_event(struct tevent_context *ev,
 217                               struct tevent_timer *te,
 218                               struct timeval t, void *private_data);
 219
 220 static int vacuum_record_parser(TDB_DATA key, TDB_DATA data, void *private_data)
 221 {
 222         struct ctdb_ltdb_header *header =
 223                 (struct ctdb_ltdb_header *)private_data;
 224
 225         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 226                 return -1;
 227         }
 228
 229         *header = *(struct ctdb_ltdb_header *)data.dptr;
 230
 231         return 0;
 232 }
 233
 234 /*
 235  * traverse function for gathering the records that can be deleted
 236  */
 237 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data,
 238                            void *private_data)
 239 {
 240         struct vacuum_data *vdata = talloc_get_type(private_data,
 241                                                     struct vacuum_data);
 242         struct ctdb_context *ctdb = vdata->ctdb;
 243         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 244         uint32_t lmaster;
 245         struct ctdb_ltdb_header *hdr;
 246         int res = 0;
 247
 248         vdata->count.db_traverse.total++;
 249
 250         lmaster = ctdb_lmaster(ctdb, &key);
 251         if (lmaster >= ctdb->num_nodes) {
 252                 vdata->count.db_traverse.error++;
 253                 DEBUG(DEBUG_CRIT, (__location__
 254                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 255                                    " with hash[%u]!\n",
 256                                    (unsigned)lmaster,
 257                                    (unsigned)ctdb->num_nodes,
 258                                    (unsigned)ctdb_hash(&key)));
 259                 return -1;
 260         }
 261
 262         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 263                 /* it is not a deleted record */
 264                 vdata->count.db_traverse.skipped++;
 265                 return 0;
 266         }
 267
 268         hdr = (struct ctdb_ltdb_header *)data.dptr;
 269
 270         if (hdr->dmaster != ctdb->pnn) {
 271                 vdata->count.db_traverse.skipped++;
 272                 return 0;
 273         }
 274
 275         /*
 276          * Add the record to this process's delete_queue for processing
 277          * in the subsequent traverse in the fast vacuum run.
 278          */
 279         res = insert_record_into_delete_queue(ctdb_db, hdr, key);
 280         if (res != 0) {
 281                 vdata->count.db_traverse.error++;
 282         } else {
 283                 vdata->count.db_traverse.scheduled++;
 284         }
 285
 286         return 0;
 287 }
 288
 289 /*
 290  * traverse the tree of records to delete and marshall them into
 291  * a blob
 292  */
 293 static int delete_marshall_traverse(void *param, void *data)
 294 {
 295         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 296         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 297         struct ctdb_marshall_buffer *m;
 298
 299         m = ctdb_marshall_add(recs, recs->records, recs->records->db_id,
 300                               recs->records->db_id,
 301                               dd->key, &dd->hdr, tdb_null);
 302         if (m == NULL) {
 303                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 304                 return -1;
 305         }
 306
 307         recs->records = m;
 308         return 0;
 309 }
 310
 311 /**
 312  * Variant of delete_marshall_traverse() that bumps the
 313  * RSN of each traversed record in the database.
 314  *
 315  * This is needed to ensure that when rolling out our
 316  * empty record copy before remote deletion, we as the
 317  * record's dmaster keep a higher RSN than the non-dmaster
 318  * nodes. This is needed to prevent old copies from
 319  * resurrection in recoveries.
 320  */
 321 static int delete_marshall_traverse_first(void *param, void *data)
 322 {
 323         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 324         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 325         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 326         struct ctdb_context *ctdb = ctdb_db->ctdb;
 327         struct ctdb_ltdb_header header;
 328         uint32_t lmaster;
 329         uint32_t hash = ctdb_hash(&(dd->key));
 330         int res;
 331
 332         res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
 333         if (res != 0) {
 334                 recs->vdata->count.delete_list.skipped++;
 335                 recs->vdata->count.delete_list.left--;
 336                 talloc_free(dd);
 337                 return 0;
 338         }
 339
 340         /*
 341          * Verify that the record is still empty, its RSN has not
 342          * changed and that we are still its lmaster and dmaster.
 343          */
 344
 345         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 346                                vacuum_record_parser, &header);
 347         if (res != 0) {
 348                 goto skip;
 349         }
 350
 351         if (header.flags & CTDB_REC_RO_FLAGS) {
 352                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 353                                    "on database db[%s] has read-only flags. "
 354                                    "skipping.\n",
 355                                    hash, ctdb_db->db_name));
 356                 goto skip;
 357         }
 358
 359         if (header.dmaster != ctdb->pnn) {
 360                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 361                                    "on database db[%s] has been migrated away. "
 362                                    "skipping.\n",
 363                                    hash, ctdb_db->db_name));
 364                 goto skip;
 365         }
 366
 367         if (header.rsn != dd->hdr.rsn) {
 368                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 369                                    "on database db[%s] seems to have been "
 370                                    "migrated away and back again (with empty "
 371                                    "data). skipping.\n",
 372                                    hash, ctdb_db->db_name));
 373                 goto skip;
 374         }
 375
 376         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 377
 378         if (lmaster != ctdb->pnn) {
 379                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 380                                    "delete list (key hash [0x%08x], db[%s]). "
 381                                    "Strange! skipping.\n",
 382                                    hash, ctdb_db->db_name));
 383                 goto skip;
 384         }
 385
 386         /*
 387          * Increment the record's RSN to ensure the dmaster (i.e. the current
 388          * node) has the highest RSN of the record in the cluster.
 389          * This is to prevent old record copies from resurrecting in recoveries
 390          * if something should fail during the deletion process.
 391          * Note that ctdb_ltdb_store_server() increments the RSN if called
 392          * on the record's dmaster.
 393          */
 394
 395         res = ctdb_ltdb_store(ctdb_db, dd->key, &header, tdb_null);
 396         if (res != 0) {
 397                 DEBUG(DEBUG_ERR, (__location__ ": Failed to store record with "
 398                                   "key hash [0x%08x] on database db[%s].\n",
 399                                   hash, ctdb_db->db_name));
 400                 goto skip;
 401         }
 402
 403         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 404
 405         goto done;
 406
 407 skip:
 408         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 409
 410         recs->vdata->count.delete_list.skipped++;
 411         recs->vdata->count.delete_list.left--;
 412         talloc_free(dd);
 413         dd = NULL;
 414
 415 done:
 416         if (dd == NULL) {
 417                 return 0;
 418         }
 419
 420         return delete_marshall_traverse(param, data);
 421 }
 422
 423 /**
 424  * traverse function for the traversal of the delete_queue,
 425  * the fast-path vacuuming list.
 426  *
 427  *  - If the record has been migrated off the node
 428  *    or has been revived (filled with data) on the node,
 429  *    then skip the record.
 430  *
 431  *  - If the current node is the record's lmaster and it is
 432  *    a record that has never been migrated with data, then
 433  *    delete the record from the local tdb.
 434  *
 435  *  - If the current node is the record's lmaster and it has
 436  *    been migrated with data, then schedule it for the normal
 437  *    vacuuming procedure (i.e. add it to the delete_list).
 438  *
 439  *  - If the current node is NOT the record's lmaster then
 440  *    add it to the list of records that are to be sent to
 441  *    the lmaster with the VACUUM_FETCH message.
 442  */
 443 static int delete_queue_traverse(void *param, void *data)
 444 {
 445         struct delete_record_data *dd =
 446                 talloc_get_type(data, struct delete_record_data);
 447         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 448         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 449         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 450         int res;
 451         struct ctdb_ltdb_header header;
 452         uint32_t lmaster;
 453         uint32_t hash = ctdb_hash(&(dd->key));
 454
 455         vdata->count.delete_queue.total++;
 456
 457         res = tdb_chainlock_nonblock(ctdb_db->ltdb->tdb, dd->key);
 458         if (res != 0) {
 459                 vdata->count.delete_queue.error++;
 460                 return 0;
 461         }
 462
 463         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 464                                vacuum_record_parser, &header);
 465         if (res != 0) {
 466                 goto skipped;
 467         }
 468
 469         if (header.dmaster != ctdb->pnn) {
 470                 /* The record has been migrated off the node. Skip. */
 471                 goto skipped;
 472         }
 473
 474         if (header.rsn != dd->hdr.rsn) {
 475                 /*
 476                  * The record has been migrated off the node and back again.
 477                  * But not requeued for deletion. Skip it.
 478                  */
 479                 goto skipped;
 480         }
 481
 482         /*
 483          * We are dmaster, and the record has no data, and it has
 484          * not been migrated after it has been queued for deletion.
 485          *
 486          * At this stage, the record could still have been revived locally
 487          * and last been written with empty data. This can only be
 488          * fixed with the addition of an active or delete flag. (TODO)
 489          */
 490
 491         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 492
 493         if (lmaster != ctdb->pnn) {
 494                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 495
 496                 if (res != 0) {
 497                         DEBUG(DEBUG_ERR,
 498                               (__location__ " Error adding record to list "
 499                                "of records to send to lmaster.\n"));
 500                         vdata->count.delete_queue.error++;
 501                 } else {
 502                         vdata->count.delete_queue.added_to_vacuum_fetch_list++;
 503                 }
 504                 goto done;
 505         }
 506
 507         /* use header->flags or dd->hdr.flags ?? */
 508         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 509                 res = add_record_to_delete_list(vdata, dd->key, &dd->hdr);
 510
 511                 if (res != 0) {
 512                         DEBUG(DEBUG_ERR,
 513                               (__location__ " Error adding record to list "
 514                                "of records for deletion on lmaster.\n"));
 515                         vdata->count.delete_queue.error++;
 516                 } else {
 517                         vdata->count.delete_queue.added_to_delete_list++;
 518                 }
 519         } else {
 520                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 521
 522                 if (res != 0) {
 523                         DEBUG(DEBUG_ERR,
 524                               (__location__ " Error deleting record with key "
 525                                "hash [0x%08x] from local data base db[%s].\n",
 526                                hash, ctdb_db->db_name));
 527                         vdata->count.delete_queue.error++;
 528                         goto done;
 529                 }
 530
 531                 DEBUG(DEBUG_DEBUG,
 532                       (__location__ " Deleted record with key hash "
 533                        "[0x%08x] from local data base db[%s].\n",
 534                        hash, ctdb_db->db_name));
 535                 vdata->count.delete_queue.deleted++;
 536         }
 537
 538         goto done;
 539
 540 skipped:
 541         vdata->count.delete_queue.skipped++;
 542
 543 done:
 544         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 545
 546         return 0;
 547 }
 548
 549 /**
 550  * Delete the records that we are lmaster and dmaster for and
 551  * that could be deleted on all other nodes via the TRY_DELETE_RECORDS
 552  * control.
 553  */
 554 static int delete_record_traverse(void *param, void *data)
 555 {
 556         struct delete_record_data *dd =
 557                 talloc_get_type(data, struct delete_record_data);
 558         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 559         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 560         struct ctdb_context *ctdb = ctdb_db->ctdb;
 561         int res;
 562         struct ctdb_ltdb_header header;
 563         uint32_t lmaster;
 564         uint32_t hash = ctdb_hash(&(dd->key));
 565
 566         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 567         if (res != 0) {
 568                 DEBUG(DEBUG_ERR,
 569                       (__location__ " Error getting chainlock on record with "
 570                        "key hash [0x%08x] on database db[%s].\n",
 571                        hash, ctdb_db->db_name));
 572                 vdata->count.delete_list.local_error++;
 573                 vdata->count.delete_list.left--;
 574                 talloc_free(dd);
 575                 return 0;
 576         }
 577
 578         /*
 579          * Verify that the record is still empty, its RSN has not
 580          * changed and that we are still its lmaster and dmaster.
 581          */
 582
 583         res = tdb_parse_record(ctdb_db->ltdb->tdb, dd->key,
 584                                vacuum_record_parser, &header);
 585         if (res != 0) {
 586                 goto skip;
 587         }
 588
 589         if (header.flags & CTDB_REC_RO_FLAGS) {
 590                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 591                                    "on database db[%s] has read-only flags. "
 592                                    "skipping.\n",
 593                                    hash, ctdb_db->db_name));
 594                 goto skip;
 595         }
 596
 597         if (header.dmaster != ctdb->pnn) {
 598                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 599                                    "on database db[%s] has been migrated away. "
 600                                    "skipping.\n",
 601                                    hash, ctdb_db->db_name));
 602                 goto skip;
 603         }
 604
 605         if (header.rsn != dd->hdr.rsn + 1) {
 606                 /*
 607                  * The record has been migrated off the node and back again.
 608                  * But not requeued for deletion. Skip it.
 609                  * (Note that the first marshall traverse has bumped the RSN
 610                  *  on disk.)
 611                  */
 612                 DEBUG(DEBUG_INFO, (__location__ ": record with hash [0x%08x] "
 613                                    "on database db[%s] seems to have been "
 614                                    "migrated away and back again (with empty "
 615                                    "data). skipping.\n",
 616                                    hash, ctdb_db->db_name));
 617                 goto skip;
 618         }
 619
 620         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 621
 622         if (lmaster != ctdb->pnn) {
 623                 DEBUG(DEBUG_INFO, (__location__ ": not lmaster for record in "
 624                                    "delete list (key hash [0x%08x], db[%s]). "
 625                                    "Strange! skipping.\n",
 626                                    hash, ctdb_db->db_name));
 627                 goto skip;
 628         }
 629
 630         res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 631
 632         if (res != 0) {
 633                 DEBUG(DEBUG_ERR,
 634                       (__location__ " Error deleting record with key hash "
 635                        "[0x%08x] from local data base db[%s].\n",
 636                        hash, ctdb_db->db_name));
 637                 vdata->count.delete_list.local_error++;
 638                 goto done;
 639         }
 640
 641         DEBUG(DEBUG_DEBUG,
 642               (__location__ " Deleted record with key hash [0x%08x] from "
 643                "local data base db[%s].\n", hash, ctdb_db->db_name));
 644
 645         vdata->count.delete_list.deleted++;
 646         goto done;
 647
 648 skip:
 649         vdata->count.delete_list.skipped++;
 650
 651 done:
 652         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 653
 654         talloc_free(dd);
 655         vdata->count.delete_list.left--;
 656
 657         return 0;
 658 }
 659
 660 /**
 661  * Traverse the delete_queue.
 662  * Records are either deleted directly or filled
 663  * into the delete list or the vacuum fetch lists
 664  * for further processing.
 665  */
 666 static void ctdb_process_delete_queue(struct ctdb_db_context *ctdb_db,
 667                                       struct vacuum_data *vdata)
 668 {
 669         uint32_t sum;
 670         int ret;
 671
 672         ret = trbt_traversearray32(ctdb_db->delete_queue, 1,
 673                                    delete_queue_traverse, vdata);
 674
 675         if (ret != 0) {
 676                 DEBUG(DEBUG_ERR, (__location__ " Error traversing "
 677                       "the delete queue.\n"));
 678         }
 679
 680         sum = vdata->count.delete_queue.deleted
 681             + vdata->count.delete_queue.skipped
 682             + vdata->count.delete_queue.error
 683             + vdata->count.delete_queue.added_to_delete_list
 684             + vdata->count.delete_queue.added_to_vacuum_fetch_list;
 685
 686         if (vdata->count.delete_queue.total != sum) {
 687                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in fast vacuum "
 688                       "counts for db[%s]: total[%u] != sum[%u]\n",
 689                       ctdb_db->db_name,
 690                       (unsigned)vdata->count.delete_queue.total,
 691                       (unsigned)sum));
 692         }
 693
 694         if (vdata->count.delete_queue.total > 0) {
 695                 DEBUG(DEBUG_INFO,
 696                       (__location__
 697                        " fast vacuuming delete_queue traverse statistics: "
 698                        "db[%s] "
 699                        "total[%u] "
 700                        "del[%u] "
 701                        "skp[%u] "
 702                        "err[%u] "
 703                        "adl[%u] "
 704                        "avf[%u]\n",
 705                        ctdb_db->db_name,
 706                        (unsigned)vdata->count.delete_queue.total,
 707                        (unsigned)vdata->count.delete_queue.deleted,
 708                        (unsigned)vdata->count.delete_queue.skipped,
 709                        (unsigned)vdata->count.delete_queue.error,
 710                        (unsigned)vdata->count.delete_queue.added_to_delete_list,
 711                        (unsigned)vdata->count.delete_queue.added_to_vacuum_fetch_list));
 712         }
 713
 714         return;
 715 }
 716
 717 /**
 718  * read-only traverse of the database, looking for records that
 719  * might be able to be vacuumed.
 720  *
 721  * This is not done each time but only every tunable
 722  * VacuumFastPathCount times.
 723  */
 724 static void ctdb_vacuum_traverse_db(struct ctdb_db_context *ctdb_db,
 725                                     struct vacuum_data *vdata)
 726 {
 727         int ret;
 728
 729         ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 730         if (ret == -1 || vdata->traverse_error) {
 731                 DEBUG(DEBUG_ERR, (__location__ " Traverse error in vacuuming "
 732                                   "'%s'\n", ctdb_db->db_name));
 733                 return;
 734         }
 735
 736         if (vdata->count.db_traverse.total > 0) {
 737                 DEBUG(DEBUG_INFO,
 738                       (__location__
 739                        " full vacuuming db traverse statistics: "
 740                        "db[%s] "
 741                        "total[%u] "
 742                        "skp[%u] "
 743                        "err[%u] "
 744                        "sched[%u]\n",
 745                        ctdb_db->db_name,
 746                        (unsigned)vdata->count.db_traverse.total,
 747                        (unsigned)vdata->count.db_traverse.skipped,
 748                        (unsigned)vdata->count.db_traverse.error,
 749                        (unsigned)vdata->count.db_traverse.scheduled));
 750         }
 751
 752         return;
 753 }
 754
 755 /**
 756  * Process the vacuum fetch lists:
 757  * For records for which we are not the lmaster, tell the lmaster to
 758  * fetch the record.
 759  */
 760 static void ctdb_process_vacuum_fetch_lists(struct ctdb_db_context *ctdb_db,
 761                                             struct vacuum_data *vdata)
 762 {
 763         int i;
 764         struct ctdb_context *ctdb = ctdb_db->ctdb;
 765
 766         for (i = 0; i < ctdb->num_nodes; i++) {
 767                 TDB_DATA data;
 768                 struct ctdb_marshall_buffer *vfl = vdata->vacuum_fetch_list[i];
 769
 770                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 771                         continue;
 772                 }
 773
 774                 if (vfl->count == 0) {
 775                         continue;
 776                 }
 777
 778                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 779                                    vfl->count, ctdb->nodes[i]->pnn,
 780                                    ctdb_db->db_name));
 781
 782                 data = ctdb_marshall_finish(vfl);
 783                 if (ctdb_client_send_message(ctdb, ctdb->nodes[i]->pnn,
 784                                              CTDB_SRVID_VACUUM_FETCH,
 785                                              data) != 0)
 786                 {
 787                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 788                                           "fetch message to %u\n",
 789                                           ctdb->nodes[i]->pnn));
 790                 }
 791         }
 792
 793         return;
 794 }
 795
 796 /**
 797  * Process the delete list:
 798  *
 799  * This is the last step of vacuuming that consistently deletes
 800  * those records that have been migrated with data and can hence
 801  * not be deleted when leaving a node.
 802  *
 803  * In this step, the lmaster does the final deletion of those empty
 804  * records that it is also dmaster for. It has ususally received
 805  * at least some of these records previously from the former dmasters
 806  * with the vacuum fetch message.
 807  *
 808  * This last step is implemented as a 3-phase process to protect from
 809  * races leading to data corruption:
 810  *
 811  *  1) Send the lmaster's copy to all other active nodes with the
 812  *     RECEIVE_RECORDS control: The remote nodes store the lmaster's copy.
 813  *  2) Send the records that could successfully be stored remotely
 814  *     in step #1 to all active nodes with the TRY_DELETE_RECORDS
 815  *     control. The remote notes delete their local copy.
 816  *  3) The lmaster locally deletes its copies of all records that
 817  *     could successfully be deleted remotely in step #2.
 818  */
 819 static void ctdb_process_delete_list(struct ctdb_db_context *ctdb_db,
 820                                      struct vacuum_data *vdata)
 821 {
 822         int ret, i;
 823         struct ctdb_context *ctdb = ctdb_db->ctdb;
 824         struct delete_records_list *recs;
 825         TDB_DATA indata;
 826         struct ctdb_node_map_old *nodemap;
 827         uint32_t *active_nodes;
 828         int num_active_nodes;
 829         TALLOC_CTX *tmp_ctx;
 830         uint32_t sum;
 831
 832         if (vdata->count.delete_list.total == 0) {
 833                 return;
 834         }
 835
 836         tmp_ctx = talloc_new(vdata);
 837         if (tmp_ctx == NULL) {
 838                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 839                 return;
 840         }
 841
 842         vdata->count.delete_list.left = vdata->count.delete_list.total;
 843
 844         /*
 845          * get the list of currently active nodes
 846          */
 847
 848         ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 849                                    CTDB_CURRENT_NODE,
 850                                    tmp_ctx,
 851                                    &nodemap);
 852         if (ret != 0) {
 853                 DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 854                 goto done;
 855         }
 856
 857         active_nodes = list_of_active_nodes(ctdb, nodemap,
 858                                             nodemap, /* talloc context */
 859                                             false /* include self */);
 860         /* yuck! ;-) */
 861         num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 862
 863         /*
 864          * Now delete the records all active nodes in a three-phase process:
 865          * 1) send all active remote nodes the current empty copy with this
 866          *    node as DMASTER
 867          * 2) if all nodes could store the new copy,
 868          *    tell all the active remote nodes to delete all their copy
 869          * 3) if all remote nodes deleted their record copy, delete it locally
 870          */
 871
 872         /*
 873          * Step 1:
 874          * Send currently empty record copy to all active nodes for storing.
 875          */
 876
 877         recs = talloc_zero(tmp_ctx, struct delete_records_list);
 878         if (recs == NULL) {
 879                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 880                 goto done;
 881         }
 882         recs->records = (struct ctdb_marshall_buffer *)
 883                 talloc_zero_size(recs,
 884                                  offsetof(struct ctdb_marshall_buffer, data));
 885         if (recs->records == NULL) {
 886                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 887                 goto done;
 888         }
 889         recs->records->db_id = ctdb_db->db_id;
 890         recs->vdata = vdata;
 891
 892         /*
 893          * traverse the tree of all records we want to delete and
 894          * create a blob we can send to the other nodes.
 895          *
 896          * We call delete_marshall_traverse_first() to bump the
 897          * records' RSNs in the database, to ensure we (as dmaster)
 898          * keep the highest RSN of the records in the cluster.
 899          */
 900         ret = trbt_traversearray32(vdata->delete_list, 1,
 901                                    delete_marshall_traverse_first, recs);
 902         if (ret != 0) {
 903                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
 904                       "delete list for first marshalling.\n"));
 905                 goto done;
 906         }
 907
 908         indata = ctdb_marshall_finish(recs->records);
 909
 910         for (i = 0; i < num_active_nodes; i++) {
 911                 struct ctdb_marshall_buffer *records;
 912                 struct ctdb_rec_data_old *rec;
 913                 int32_t res;
 914                 TDB_DATA outdata;
 915
 916                 ret = ctdb_control(ctdb, active_nodes[i], 0,
 917                                 CTDB_CONTROL_RECEIVE_RECORDS, 0,
 918                                 indata, recs, &outdata, &res,
 919                                 NULL, NULL);
 920                 if (ret != 0 || res != 0) {
 921                         DEBUG(DEBUG_ERR, ("Error storing record copies on "
 922                                           "node %u: ret[%d] res[%d]\n",
 923                                           active_nodes[i], ret, res));
 924                         goto done;
 925                 }
 926
 927                 /*
 928                  * outdata contains the list of records coming back
 929                  * from the node: These are the records that the
 930                  * remote node could not store. We remove these from
 931                  * the list to process further.
 932                  */
 933                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
 934                 rec = (struct ctdb_rec_data_old *)&records->data[0];
 935                 while (records->count-- > 1) {
 936                         TDB_DATA reckey, recdata;
 937                         struct ctdb_ltdb_header *rechdr;
 938                         struct delete_record_data *dd;
 939
 940                         reckey.dptr = &rec->data[0];
 941                         reckey.dsize = rec->keylen;
 942                         recdata.dptr = &rec->data[reckey.dsize];
 943                         recdata.dsize = rec->datalen;
 944
 945                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 946                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 947                                 goto done;
 948                         }
 949                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 950                         recdata.dptr += sizeof(*rechdr);
 951                         recdata.dsize -= sizeof(*rechdr);
 952
 953                         dd = (struct delete_record_data *)trbt_lookup32(
 954                                         vdata->delete_list,
 955                                         ctdb_hash(&reckey));
 956                         if (dd != NULL) {
 957                                 /*
 958                                  * The other node could not store the record
 959                                  * copy and it is the first node that failed.
 960                                  * So we should remove it from the tree and
 961                                  * update statistics.
 962                                  */
 963                                 talloc_free(dd);
 964                                 vdata->count.delete_list.remote_error++;
 965                                 vdata->count.delete_list.left--;
 966                         } else {
 967                                 DEBUG(DEBUG_ERR, (__location__ " Failed to "
 968                                       "find record with hash 0x%08x coming "
 969                                       "back from RECEIVE_RECORDS "
 970                                       "control in delete list.\n",
 971                                       ctdb_hash(&reckey)));
 972                                 vdata->count.delete_list.local_error++;
 973                                 vdata->count.delete_list.left--;
 974                         }
 975
 976                         rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
 977                 }
 978         }
 979
 980         if (vdata->count.delete_list.left == 0) {
 981                 goto success;
 982         }
 983
 984         /*
 985          * Step 2:
 986          * Send the remaining records to all active nodes for deletion.
 987          *
 988          * The lmaster's (i.e. our) copies of these records have been stored
 989          * successfully on the other nodes.
 990          */
 991
 992         /*
 993          * Create a marshall blob from the remaining list of records to delete.
 994          */
 995
 996         talloc_free(recs->records);
 997
 998         recs->records = (struct ctdb_marshall_buffer *)
 999                 talloc_zero_size(recs,
1000                                  offsetof(struct ctdb_marshall_buffer, data));
1001         if (recs->records == NULL) {
1002                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1003                 goto done;
1004         }
1005         recs->records->db_id = ctdb_db->db_id;
1006
1007         ret = trbt_traversearray32(vdata->delete_list, 1,
1008                                    delete_marshall_traverse, recs);
1009         if (ret != 0) {
1010                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1011                       "delete list for second marshalling.\n"));
1012                 goto done;
1013         }
1014
1015         indata = ctdb_marshall_finish(recs->records);
1016
1017         for (i = 0; i < num_active_nodes; i++) {
1018                 struct ctdb_marshall_buffer *records;
1019                 struct ctdb_rec_data_old *rec;
1020                 int32_t res;
1021                 TDB_DATA outdata;
1022
1023                 ret = ctdb_control(ctdb, active_nodes[i], 0,
1024                                 CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
1025                                 indata, recs, &outdata, &res,
1026                                 NULL, NULL);
1027                 if (ret != 0 || res != 0) {
1028                         DEBUG(DEBUG_ERR, ("Failed to delete records on "
1029                                           "node %u: ret[%d] res[%d]\n",
1030                                           active_nodes[i], ret, res));
1031                         goto done;
1032                 }
1033
1034                 /*
1035                  * outdata contains the list of records coming back
1036                  * from the node: These are the records that the
1037                  * remote node could not delete. We remove these from
1038                  * the list to delete locally.
1039                  */
1040                 records = (struct ctdb_marshall_buffer *)outdata.dptr;
1041                 rec = (struct ctdb_rec_data_old *)&records->data[0];
1042                 while (records->count-- > 1) {
1043                         TDB_DATA reckey, recdata;
1044                         struct ctdb_ltdb_header *rechdr;
1045                         struct delete_record_data *dd;
1046
1047                         reckey.dptr = &rec->data[0];
1048                         reckey.dsize = rec->keylen;
1049                         recdata.dptr = &rec->data[reckey.dsize];
1050                         recdata.dsize = rec->datalen;
1051
1052                         if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
1053                                 DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
1054                                 goto done;
1055                         }
1056                         rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
1057                         recdata.dptr += sizeof(*rechdr);
1058                         recdata.dsize -= sizeof(*rechdr);
1059
1060                         dd = (struct delete_record_data *)trbt_lookup32(
1061                                         vdata->delete_list,
1062                                         ctdb_hash(&reckey));
1063                         if (dd != NULL) {
1064                                 /*
1065                                  * The other node could not delete the
1066                                  * record and it is the first node that
1067                                  * failed. So we should remove it from
1068                                  * the tree and update statistics.
1069                                  */
1070                                 talloc_free(dd);
1071                                 vdata->count.delete_list.remote_error++;
1072                                 vdata->count.delete_list.left--;
1073                         } else {
1074                                 DEBUG(DEBUG_ERR, (__location__ " Failed to "
1075                                       "find record with hash 0x%08x coming "
1076                                       "back from TRY_DELETE_RECORDS "
1077                                       "control in delete list.\n",
1078                                       ctdb_hash(&reckey)));
1079                                 vdata->count.delete_list.local_error++;
1080                                 vdata->count.delete_list.left--;
1081                         }
1082
1083                         rec = (struct ctdb_rec_data_old *)(rec->length + (uint8_t *)rec);
1084                 }
1085         }
1086
1087         if (vdata->count.delete_list.left == 0) {
1088                 goto success;
1089         }
1090
1091         /*
1092          * Step 3:
1093          * Delete the remaining records locally.
1094          *
1095          * These records have successfully been deleted on all
1096          * active remote nodes.
1097          */
1098
1099         ret = trbt_traversearray32(vdata->delete_list, 1,
1100                                    delete_record_traverse, vdata);
1101         if (ret != 0) {
1102                 DEBUG(DEBUG_ERR, (__location__ " Error traversing the "
1103                       "delete list for deletion.\n"));
1104         }
1105
1106 success:
1107
1108         if (vdata->count.delete_list.left != 0) {
1109                 DEBUG(DEBUG_ERR, (__location__ " Vaccum db[%s] error: "
1110                       "there are %u records left for deletion after "
1111                       "processing delete list\n",
1112                       ctdb_db->db_name,
1113                       (unsigned)vdata->count.delete_list.left));
1114         }
1115
1116         sum = vdata->count.delete_list.deleted
1117             + vdata->count.delete_list.skipped
1118             + vdata->count.delete_list.remote_error
1119             + vdata->count.delete_list.local_error
1120             + vdata->count.delete_list.left;
1121
1122         if (vdata->count.delete_list.total != sum) {
1123                 DEBUG(DEBUG_ERR, (__location__ " Inconsistency in vacuum "
1124                       "delete list counts for db[%s]: total[%u] != sum[%u]\n",
1125                       ctdb_db->db_name,
1126                       (unsigned)vdata->count.delete_list.total,
1127                       (unsigned)sum));
1128         }
1129
1130         if (vdata->count.delete_list.total > 0) {
1131                 DEBUG(DEBUG_INFO,
1132                       (__location__
1133                        " vacuum delete list statistics: "
1134                        "db[%s] "
1135                        "total[%u] "
1136                        "del[%u] "
1137                        "skip[%u] "
1138                        "rem.err[%u] "
1139                        "loc.err[%u] "
1140                        "left[%u]\n",
1141                        ctdb_db->db_name,
1142                        (unsigned)vdata->count.delete_list.total,
1143                        (unsigned)vdata->count.delete_list.deleted,
1144                        (unsigned)vdata->count.delete_list.skipped,
1145                        (unsigned)vdata->count.delete_list.remote_error,
1146                        (unsigned)vdata->count.delete_list.local_error,
1147                        (unsigned)vdata->count.delete_list.left));
1148         }
1149
1150 done:
1151         talloc_free(tmp_ctx);
1152
1153         return;
1154 }
1155
1156 /**
1157  * initialize the vacuum_data
1158  */
1159 static struct vacuum_data *ctdb_vacuum_init_vacuum_data(
1160                                         struct ctdb_db_context *ctdb_db,
1161                                         TALLOC_CTX *mem_ctx)
1162 {
1163         int i;
1164         struct ctdb_context *ctdb = ctdb_db->ctdb;
1165         struct vacuum_data *vdata;
1166
1167         vdata = talloc_zero(mem_ctx, struct vacuum_data);
1168         if (vdata == NULL) {
1169                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1170                 return NULL;
1171         }
1172
1173         vdata->ctdb = ctdb_db->ctdb;
1174         vdata->ctdb_db = ctdb_db;
1175         vdata->delete_list = trbt_create(vdata, 0);
1176         if (vdata->delete_list == NULL) {
1177                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1178                 goto fail;
1179         }
1180
1181         vdata->start = timeval_current();
1182
1183         vdata->count.delete_queue.added_to_delete_list = 0;
1184         vdata->count.delete_queue.added_to_vacuum_fetch_list = 0;
1185         vdata->count.delete_queue.deleted = 0;
1186         vdata->count.delete_queue.skipped = 0;
1187         vdata->count.delete_queue.error = 0;
1188         vdata->count.delete_queue.total = 0;
1189         vdata->count.db_traverse.scheduled = 0;
1190         vdata->count.db_traverse.skipped = 0;
1191         vdata->count.db_traverse.error = 0;
1192         vdata->count.db_traverse.total = 0;
1193         vdata->count.delete_list.total = 0;
1194         vdata->count.delete_list.left = 0;
1195         vdata->count.delete_list.remote_error = 0;
1196         vdata->count.delete_list.local_error = 0;
1197         vdata->count.delete_list.skipped = 0;
1198         vdata->count.delete_list.deleted = 0;
1199
1200         /* the list needs to be of length num_nodes */
1201         vdata->vacuum_fetch_list = talloc_zero_array(vdata,
1202                                                 struct ctdb_marshall_buffer *,
1203                                                 ctdb->num_nodes);
1204         if (vdata->vacuum_fetch_list == NULL) {
1205                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1206                 goto fail;
1207         }
1208         for (i = 0; i < ctdb->num_nodes; i++) {
1209                 vdata->vacuum_fetch_list[i] = (struct ctdb_marshall_buffer *)
1210                         talloc_zero_size(vdata->vacuum_fetch_list,
1211                                          offsetof(struct ctdb_marshall_buffer, data));
1212                 if (vdata->vacuum_fetch_list[i] == NULL) {
1213                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
1214                         talloc_free(vdata);
1215                         return NULL;
1216                 }
1217                 vdata->vacuum_fetch_list[i]->db_id = ctdb_db->db_id;
1218         }
1219
1220         return vdata;
1221
1222 fail:
1223         talloc_free(vdata);
1224         return NULL;
1225 }
1226
1227 /**
1228  * Vacuum a DB:
1229  *  - Always do the fast vacuuming run, which traverses
1230  *    the in-memory delete queue: these records have been
1231  *    scheduled for deletion.
1232  *  - Only if explicitly requested, the database is traversed
1233  *    in order to use the traditional heuristics on empty records
1234  *    to trigger deletion.
1235  *    This is done only every VacuumFastPathCount'th vacuuming run.
1236  *
1237  * The traverse runs fill two lists:
1238  *
1239  * - The delete_list:
1240  *   This is the list of empty records the current
1241  *   node is lmaster and dmaster for. These records are later
1242  *   deleted first on other nodes and then locally.
1243  *
1244  *   The fast vacuuming run has a short cut for those records
1245  *   that have never been migrated with data: these records
1246  *   are immediately deleted locally, since they have left
1247  *   no trace on other nodes.
1248  *
1249  * - The vacuum_fetch lists
1250  *   (one for each other lmaster node):
1251  *   The records in this list are sent for deletion to
1252  *   their lmaster in a bulk VACUUM_FETCH message.
1253  *
1254  *   The lmaster then migrates all these records to itelf
1255  *   so that they can be vacuumed there.
1256  *
1257  * This executes in the child context.
1258  */
1259 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
1260                           bool full_vacuum_run)
1261 {
1262         struct ctdb_context *ctdb = ctdb_db->ctdb;
1263         int ret, pnn;
1264         struct vacuum_data *vdata;
1265         TALLOC_CTX *tmp_ctx;
1266
1267         DEBUG(DEBUG_INFO, (__location__ " Entering %s vacuum run for db "
1268                            "%s db_id[0x%08x]\n",
1269                            full_vacuum_run ? "full" : "fast",
1270                            ctdb_db->db_name, ctdb_db->db_id));
1271
1272         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
1273         if (ret != 0) {
1274                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
1275                 return ret;
1276         }
1277
1278         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1279         if (pnn == -1) {
1280                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
1281                 return -1;
1282         }
1283
1284         ctdb->pnn = pnn;
1285
1286         tmp_ctx = talloc_new(ctdb_db);
1287         if (tmp_ctx == NULL) {
1288                 DEBUG(DEBUG_ERR, ("Out of memory!\n"));
1289                 return -1;
1290         }
1291
1292         vdata = ctdb_vacuum_init_vacuum_data(ctdb_db, tmp_ctx);
1293         if (vdata == NULL) {
1294                 talloc_free(tmp_ctx);
1295                 return -1;
1296         }
1297
1298         if (full_vacuum_run) {
1299                 ctdb_vacuum_traverse_db(ctdb_db, vdata);
1300         }
1301
1302         ctdb_process_delete_queue(ctdb_db, vdata);
1303
1304         ctdb_process_vacuum_fetch_lists(ctdb_db, vdata);
1305
1306         ctdb_process_delete_list(ctdb_db, vdata);
1307
1308         talloc_free(tmp_ctx);
1309
1310         /* this ensures we run our event queue */
1311         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
1312
1313         return 0;
1314 }
1315
1316 /*
1317  * repack and vaccum a db
1318  * called from the child context
1319  */
1320 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
1321                                      bool full_vacuum_run)
1322 {
1323         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
1324         const char *name = ctdb_db->db_name;
1325         int freelist_size = 0;
1326         int ret;
1327
1328         if (ctdb_vacuum_db(ctdb_db, full_vacuum_run) != 0) {
1329                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
1330         }
1331
1332         freelist_size = tdb_freelist_size(ctdb_db->ltdb->tdb);
1333         if (freelist_size == -1) {
1334                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
1335                 return -1;
1336         }
1337
1338         /*
1339          * decide if a repack is necessary
1340          */
1341         if ((repack_limit == 0 || (uint32_t)freelist_size < repack_limit))
1342         {
1343                 return 0;
1344         }
1345
1346         DEBUG(DEBUG_INFO, ("Repacking %s with %u freelist entries\n",
1347                            name, freelist_size));
1348
1349         ret = tdb_repack(ctdb_db->ltdb->tdb);
1350         if (ret != 0) {
1351                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
1352                 return -1;
1353         }
1354
1355         return 0;
1356 }
1357
1358 static uint32_t get_vacuum_interval(struct ctdb_db_context *ctdb_db)
1359 {
1360         uint32_t interval = ctdb_db->ctdb->tunable.vacuum_interval;
1361
1362         return interval;
1363 }
1364
1365 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
1366 {
1367         double l = timeval_elapsed(&child_ctx->start_time);
1368         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
1369         struct ctdb_context *ctdb = ctdb_db->ctdb;
1370
1371         CTDB_UPDATE_DB_LATENCY(ctdb_db, "vacuum", vacuum.latency, l);
1372         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
1373
1374         if (child_ctx->child_pid != -1) {
1375                 ctdb_kill(ctdb, child_ctx->child_pid, SIGKILL);
1376         } else {
1377                 /* Bump the number of successful fast-path runs. */
1378                 child_ctx->vacuum_handle->fast_path_count++;
1379         }
1380
1381         DLIST_REMOVE(ctdb->vacuumers, child_ctx);
1382
1383         tevent_add_timer(ctdb->ev, child_ctx->vacuum_handle,
1384                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1385                          ctdb_vacuum_event, child_ctx->vacuum_handle);
1386
1387         return 0;
1388 }
1389
1390 /*
1391  * this event is generated when a vacuum child process times out
1392  */
1393 static void vacuum_child_timeout(struct tevent_context *ev,
1394                                  struct tevent_timer *te,
1395                                  struct timeval t, void *private_data)
1396 {
1397         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1398
1399         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
1400
1401         child_ctx->status = VACUUM_TIMEOUT;
1402
1403         talloc_free(child_ctx);
1404 }
1405
1406
1407 /*
1408  * this event is generated when a vacuum child process has completed
1409  */
1410 static void vacuum_child_handler(struct tevent_context *ev,
1411                                  struct tevent_fd *fde,
1412                                  uint16_t flags, void *private_data)
1413 {
1414         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1415         char c = 0;
1416         int ret;
1417
1418         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1419         child_ctx->child_pid = -1;
1420
1421         ret = sys_read(child_ctx->fd[0], &c, 1);
1422         if (ret != 1 || c != 0) {
1423                 child_ctx->status = VACUUM_ERROR;
1424                 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));
1425         } else {
1426                 child_ctx->status = VACUUM_OK;
1427         }
1428
1429         talloc_free(child_ctx);
1430 }
1431
1432 /*
1433  * this event is called every time we need to start a new vacuum process
1434  */
1435 static void ctdb_vacuum_event(struct tevent_context *ev,
1436                               struct tevent_timer *te,
1437                               struct timeval t, void *private_data)
1438 {
1439         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1440         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1441         struct ctdb_context *ctdb = ctdb_db->ctdb;
1442         struct ctdb_vacuum_child_context *child_ctx;
1443         struct tevent_fd *fde;
1444         int ret;
1445
1446         /* we don't vacuum if we are in recovery mode, or db frozen */
1447         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ||
1448             ctdb->freeze_mode[ctdb_db->priority] != CTDB_FREEZE_NONE) {
1449                 DEBUG(DEBUG_INFO, ("Not vacuuming %s (%s)\n", ctdb_db->db_name,
1450                                    ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE ? "in recovery"
1451                                    : ctdb->freeze_mode[ctdb_db->priority] == CTDB_FREEZE_PENDING
1452                                    ? "freeze pending"
1453                                    : "frozen"));
1454                 tevent_add_timer(ctdb->ev, vacuum_handle,
1455                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1456                                  ctdb_vacuum_event, vacuum_handle);
1457                 return;
1458         }
1459
1460         /* Do not allow multiple vacuuming child processes to be active at the
1461          * same time.  If there is vacuuming child process active, delay
1462          * new vacuuming event to stagger vacuuming events.
1463          */
1464         if (ctdb->vacuumers != NULL) {
1465                 tevent_add_timer(ctdb->ev, vacuum_handle,
1466                                  timeval_current_ofs(0, 500*1000),
1467                                  ctdb_vacuum_event, vacuum_handle);
1468                 return;
1469         }
1470
1471         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1472         if (child_ctx == NULL) {
1473                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1474                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1475         }
1476
1477
1478         ret = pipe(child_ctx->fd);
1479         if (ret != 0) {
1480                 talloc_free(child_ctx);
1481                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1482                 tevent_add_timer(ctdb->ev, vacuum_handle,
1483                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1484                                  ctdb_vacuum_event, vacuum_handle);
1485                 return;
1486         }
1487
1488         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1489                 vacuum_handle->fast_path_count = 0;
1490         }
1491
1492         child_ctx->child_pid = ctdb_fork(ctdb);
1493         if (child_ctx->child_pid == (pid_t)-1) {
1494                 close(child_ctx->fd[0]);
1495                 close(child_ctx->fd[1]);
1496                 talloc_free(child_ctx);
1497                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1498                 tevent_add_timer(ctdb->ev, vacuum_handle,
1499                                  timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1500                                  ctdb_vacuum_event, vacuum_handle);
1501                 return;
1502         }
1503
1504
1505         if (child_ctx->child_pid == 0) {
1506                 char cc = 0;
1507                 bool full_vacuum_run = false;
1508                 close(child_ctx->fd[0]);
1509
1510                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1511                 prctl_set_comment("ctdb_vacuum");
1512                 if (switch_from_server_to_client(ctdb, "vacuum-%s", ctdb_db->db_name) != 0) {
1513                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1514                         _exit(1);
1515                 }
1516
1517                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1518                     (vacuum_handle->fast_path_count == 0))
1519                 {
1520                         full_vacuum_run = true;
1521                 }
1522                 cc = ctdb_vacuum_and_repack_db(ctdb_db, full_vacuum_run);
1523
1524                 sys_write(child_ctx->fd[1], &cc, 1);
1525                 _exit(0);
1526         }
1527
1528         set_close_on_exec(child_ctx->fd[0]);
1529         close(child_ctx->fd[1]);
1530
1531         child_ctx->status = VACUUM_RUNNING;
1532         child_ctx->start_time = timeval_current();
1533
1534         DLIST_ADD(ctdb->vacuumers, child_ctx);
1535         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1536
1537         /*
1538          * Clear the fastpath vacuuming list in the parent.
1539          */
1540         talloc_free(ctdb_db->delete_queue);
1541         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1542         if (ctdb_db->delete_queue == NULL) {
1543                 /* fatal here? ... */
1544                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1545                                  "in parent context. Shutting down\n");
1546         }
1547
1548         tevent_add_timer(ctdb->ev, child_ctx,
1549                          timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1550                          vacuum_child_timeout, child_ctx);
1551
1552         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1553
1554         fde = tevent_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1555                             TEVENT_FD_READ, vacuum_child_handler, child_ctx);
1556         tevent_fd_set_auto_close(fde);
1557
1558         vacuum_handle->child_ctx = child_ctx;
1559         child_ctx->vacuum_handle = vacuum_handle;
1560 }
1561
1562 void ctdb_stop_vacuuming(struct ctdb_context *ctdb)
1563 {
1564         /* Simply free them all. */
1565         while (ctdb->vacuumers) {
1566                 DEBUG(DEBUG_INFO, ("Aborting vacuuming for %s (%i)\n",
1567                            ctdb->vacuumers->vacuum_handle->ctdb_db->db_name,
1568                            (int)ctdb->vacuumers->child_pid));
1569                 /* vacuum_child_destructor kills it, removes from list */
1570                 talloc_free(ctdb->vacuumers);
1571         }
1572 }
1573
1574 /* this function initializes the vacuuming context for a database
1575  * starts the vacuuming events
1576  */
1577 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1578 {
1579         if (ctdb_db->persistent != 0) {
1580                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1581                 return 0;
1582         }
1583
1584         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1585         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1586
1587         ctdb_db->vacuum_handle->ctdb_db         = ctdb_db;
1588         ctdb_db->vacuum_handle->fast_path_count = 0;
1589
1590         tevent_add_timer(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1591                          timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1592                          ctdb_vacuum_event, ctdb_db->vacuum_handle);
1593
1594         return 0;
1595 }
1596
1597 static void remove_record_from_delete_queue(struct ctdb_db_context *ctdb_db,
1598                                             const struct ctdb_ltdb_header *hdr,
1599                                             const TDB_DATA key)
1600 {
1601         struct delete_record_data *kd;
1602         uint32_t hash;
1603
1604         hash = (uint32_t)ctdb_hash(&key);
1605
1606         DEBUG(DEBUG_DEBUG, (__location__
1607                             " remove_record_from_delete_queue: "
1608                             "db[%s] "
1609                             "db_id[0x%08x] "
1610                             "key_hash[0x%08x] "
1611                             "lmaster[%u] "
1612                             "migrated_with_data[%s]\n",
1613                              ctdb_db->db_name, ctdb_db->db_id,
1614                              hash,
1615                              ctdb_lmaster(ctdb_db->ctdb, &key),
1616                              hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1617
1618         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1619         if (kd == NULL) {
1620                 DEBUG(DEBUG_DEBUG, (__location__
1621                                     " remove_record_from_delete_queue: "
1622                                     "record not in queue (hash[0x%08x])\n.",
1623                                     hash));
1624                 return;
1625         }
1626
1627         if ((kd->key.dsize != key.dsize) ||
1628             (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1629         {
1630                 DEBUG(DEBUG_DEBUG, (__location__
1631                                     " remove_record_from_delete_queue: "
1632                                     "hash collision for key with hash[0x%08x] "
1633                                     "in db[%s] - skipping\n",
1634                                     hash, ctdb_db->db_name));
1635                 return;
1636         }
1637
1638         DEBUG(DEBUG_DEBUG, (__location__
1639                             " remove_record_from_delete_queue: "
1640                             "removing key with hash[0x%08x]\n",
1641                              hash));
1642
1643         talloc_free(kd);
1644
1645         return;
1646 }
1647
1648 /**
1649  * Insert a record into the ctdb_db context's delete queue,
1650  * handling hash collisions.
1651  */
1652 static int insert_record_into_delete_queue(struct ctdb_db_context *ctdb_db,
1653                                            const struct ctdb_ltdb_header *hdr,
1654                                            TDB_DATA key)
1655 {
1656         struct delete_record_data *kd;
1657         uint32_t hash;
1658         int ret;
1659
1660         hash = (uint32_t)ctdb_hash(&key);
1661
1662         DEBUG(DEBUG_DEBUG, (__location__ " schedule for deletion: db[%s] "
1663                             "db_id[0x%08x] "
1664                             "key_hash[0x%08x] "
1665                             "lmaster[%u] "
1666                             "migrated_with_data[%s]\n",
1667                             ctdb_db->db_name, ctdb_db->db_id,
1668                             hash,
1669                             ctdb_lmaster(ctdb_db->ctdb, &key),
1670                             hdr->flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA ? "yes" : "no"));
1671
1672         kd = (struct delete_record_data *)trbt_lookup32(ctdb_db->delete_queue, hash);
1673         if (kd != NULL) {
1674                 if ((kd->key.dsize != key.dsize) ||
1675                     (memcmp(kd->key.dptr, key.dptr, key.dsize) != 0))
1676                 {
1677                         DEBUG(DEBUG_INFO,
1678                               (__location__ " schedule for deletion: "
1679                                "hash collision for key hash [0x%08x]. "
1680                                "Skipping the record.\n", hash));
1681                         return 0;
1682                 } else {
1683                         DEBUG(DEBUG_DEBUG,
1684                               (__location__ " schedule for deletion: "
1685                                "updating entry for key with hash [0x%08x].\n",
1686                                hash));
1687                 }
1688         }
1689
1690         ret = insert_delete_record_data_into_tree(ctdb_db->ctdb, ctdb_db,
1691                                                   ctdb_db->delete_queue,
1692                                                   hdr, key);
1693         if (ret != 0) {
1694                 DEBUG(DEBUG_INFO,
1695                       (__location__ " schedule for deletion: error "
1696                        "inserting key with hash [0x%08x] into delete queue\n",
1697                        hash));
1698                 return -1;
1699         }
1700
1701         return 0;
1702 }
1703
1704 /**
1705  * Schedule a record for deletetion.
1706  * Called from the parent context.
1707  */
1708 int32_t ctdb_control_schedule_for_deletion(struct ctdb_context *ctdb,
1709                                            TDB_DATA indata)
1710 {
1711         struct ctdb_control_schedule_for_deletion *dd;
1712         struct ctdb_db_context *ctdb_db;
1713         int ret;
1714         TDB_DATA key;
1715
1716         dd = (struct ctdb_control_schedule_for_deletion *)indata.dptr;
1717
1718         ctdb_db = find_ctdb_db(ctdb, dd->db_id);
1719         if (ctdb_db == NULL) {
1720                 DEBUG(DEBUG_ERR, (__location__ " Unknown db id 0x%08x\n",
1721                                   dd->db_id));
1722                 return -1;
1723         }
1724
1725         key.dsize = dd->keylen;
1726         key.dptr = dd->key;
1727
1728         ret = insert_record_into_delete_queue(ctdb_db, &dd->hdr, key);
1729
1730         return ret;
1731 }
1732
1733 int32_t ctdb_local_schedule_for_deletion(struct ctdb_db_context *ctdb_db,
1734                                          const struct ctdb_ltdb_header *hdr,
1735                                          TDB_DATA key)
1736 {
1737         int ret;
1738         struct ctdb_control_schedule_for_deletion *dd;
1739         TDB_DATA indata;
1740         int32_t status;
1741
1742         if (ctdb_db->ctdb->ctdbd_pid == getpid()) {
1743                 /* main daemon - directly queue */
1744                 ret = insert_record_into_delete_queue(ctdb_db, hdr, key);
1745
1746                 return ret;
1747         }
1748
1749         /* if we don't have a connection to the daemon we can not send
1750            a control. For example sometimes from update_record control child
1751            process.
1752         */
1753         if (!ctdb_db->ctdb->can_send_controls) {
1754                 return -1;
1755         }
1756
1757
1758         /* child process: send the main daemon a control */
1759         indata.dsize = offsetof(struct ctdb_control_schedule_for_deletion, key) + key.dsize;
1760         indata.dptr = talloc_zero_array(ctdb_db, uint8_t, indata.dsize);
1761         if (indata.dptr == NULL) {
1762                 DEBUG(DEBUG_ERR, (__location__ " out of memory\n"));
1763                 return -1;
1764         }
1765         dd = (struct ctdb_control_schedule_for_deletion *)(void *)indata.dptr;
1766         dd->db_id = ctdb_db->db_id;
1767         dd->hdr = *hdr;
1768         dd->keylen = key.dsize;
1769         memcpy(dd->key, key.dptr, key.dsize);
1770
1771         ret = ctdb_control(ctdb_db->ctdb,
1772                            CTDB_CURRENT_NODE,
1773                            ctdb_db->db_id,
1774                            CTDB_CONTROL_SCHEDULE_FOR_DELETION,
1775                            CTDB_CTRL_FLAG_NOREPLY, /* flags */
1776                            indata,
1777                            NULL, /* mem_ctx */
1778                            NULL, /* outdata */
1779                            &status,
1780                            NULL, /* timeout : NULL == wait forever */
1781                            NULL); /* error message */
1782
1783         talloc_free(indata.dptr);
1784
1785         if (ret != 0 || status != 0) {
1786                 DEBUG(DEBUG_ERR, (__location__ " Error sending "
1787                                   "SCHEDULE_FOR_DELETION "
1788                                   "control.\n"));
1789                 if (status != 0) {
1790                         ret = -1;
1791                 }
1792         }
1793
1794         return ret;
1795 }
1796
1797 void ctdb_local_remove_from_delete_queue(struct ctdb_db_context *ctdb_db,
1798                                          const struct ctdb_ltdb_header *hdr,
1799                                          const TDB_DATA key)
1800 {
1801         if (ctdb_db->ctdb->ctdbd_pid != getpid()) {
1802                 /*
1803                  * Only remove the record from the delete queue if called
1804                  * in the main daemon.
1805                  */
1806                 return;
1807         }
1808
1809         remove_record_from_delete_queue(ctdb_db, hdr, key);
1810
1811         return;
1812 }