server/ctdb_vacuum.c

   1 /*
   2    ctdb vacuuming events
   3
   4    Copyright (C) Ronnie Sahlberg  2009
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program; if not, see <http://www.gnu.org/licenses/>.
  18 */
  19
  20 #include "includes.h"
  21 #include "lib/events/events.h"
  22 #include "lib/tdb/include/tdb.h"
  23 #include "system/network.h"
  24 #include "system/filesys.h"
  25 #include "system/dir.h"
  26 #include "../include/ctdb_private.h"
  27 #include "db_wrap.h"
  28 #include "lib/util/dlinklist.h"
  29 #include "lib/events/events.h"
  30 #include "../include/ctdb_private.h"
  31 #include "../common/rb_tree.h"
  32
  33 #define TIMELIMIT() timeval_current_ofs(10, 0)
  34 #define TUNINGDBNAME "vactune.tdb"
  35
  36 enum vacuum_child_status { VACUUM_RUNNING, VACUUM_OK, VACUUM_ERROR, VACUUM_TIMEOUT};
  37
  38 struct ctdb_vacuum_child_context {
  39         struct ctdb_vacuum_handle *vacuum_handle;
  40         int fd[2];
  41         pid_t child_pid;
  42         enum vacuum_child_status status;
  43         struct timeval start_time;
  44 };
  45
  46 struct ctdb_vacuum_handle {
  47         struct ctdb_db_context *ctdb_db;
  48         struct ctdb_vacuum_child_context *child_ctx;
  49         uint32_t fast_path_count;
  50 };
  51
  52
  53 /*  a list of records to possibly delete */
  54 struct vacuum_data {
  55         uint32_t vacuum_limit;
  56         uint32_t repack_limit;
  57         struct ctdb_context *ctdb;
  58         struct ctdb_db_context *ctdb_db;
  59         struct tdb_context *dest_db;
  60         trbt_tree_t *delete_tree;
  61         uint32_t delete_count;
  62         struct ctdb_marshall_buffer **list;
  63         struct timeval start;
  64         bool traverse_error;
  65         bool vacuum;
  66         uint32_t total;
  67         uint32_t vacuumed;
  68         uint32_t copied;
  69 };
  70
  71 /* tuning information stored for every db */
  72 struct vacuum_tuning_data {
  73         uint32_t last_num_repack;
  74         uint32_t last_num_empty;
  75         uint32_t last_interval;
  76         uint32_t new_interval;
  77         struct timeval last_start;
  78         double   last_duration;
  79 };
  80
  81 /* this structure contains the information for one record to be deleted */
  82 struct delete_record_data {
  83         struct ctdb_context *ctdb;
  84         struct ctdb_db_context *ctdb_db;
  85         struct ctdb_ltdb_header hdr;
  86         TDB_DATA key;
  87 };
  88
  89 struct delete_records_list {
  90         struct ctdb_marshall_buffer *records;
  91 };
  92
  93 /**
  94  * Store key and header in a tree, indexed by the key hash.
  95  */
  96 static int insert_delete_record_data_into_tree(struct ctdb_context *ctdb,
  97                                                struct ctdb_db_context *ctdb_db,
  98                                                trbt_tree_t *tree,
  99                                                const struct ctdb_ltdb_header *hdr,
 100                                                TDB_DATA key)
 101 {
 102         struct delete_record_data *dd;
 103         uint32_t hash;
 104
 105         dd = talloc_zero(tree, struct delete_record_data);
 106         if (dd == NULL) {
 107                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 108                 return -1;
 109         }
 110
 111         dd->ctdb      = ctdb;
 112         dd->ctdb_db   = ctdb_db;
 113         dd->key.dsize = key.dsize;
 114         dd->key.dptr  = talloc_memdup(dd, key.dptr, key.dsize);
 115         if (dd->key.dptr == NULL) {
 116                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 117                 return -1;
 118         }
 119
 120         dd->hdr = *hdr;
 121
 122         hash = ctdb_hash(&key);
 123
 124         trbt_insert32(tree, hash, dd);
 125
 126         return 0;
 127 }
 128
 129 static int add_record_to_delete_tree(struct vacuum_data *vdata, TDB_DATA key,
 130                                      struct ctdb_ltdb_header *hdr)
 131 {
 132         struct ctdb_context *ctdb = vdata->ctdb;
 133         struct ctdb_db_context *ctdb_db = vdata->ctdb_db;
 134         uint32_t hash;
 135         int ret;
 136
 137         hash = ctdb_hash(&key);
 138
 139         if (trbt_lookup32(vdata->delete_tree, hash)) {
 140                 DEBUG(DEBUG_DEBUG, (__location__ " Hash collission when vacuuming, skipping this record.\n"));
 141                 return 0;
 142         }
 143
 144         ret = insert_delete_record_data_into_tree(ctdb, ctdb_db,
 145                                                   vdata->delete_tree,
 146                                                   hdr, key);
 147         if (ret != 0) {
 148                 return -1;
 149         }
 150
 151         vdata->delete_count++;
 152
 153         return 0;
 154 }
 155
 156 /**
 157  * Add a record to the list of records to be sent
 158  * to their lmaster with VACUUM_FETCH.
 159  */
 160 static int add_record_to_vacuum_fetch_list(struct vacuum_data *vdata,
 161                                            TDB_DATA key)
 162 {
 163         struct ctdb_context *ctdb = vdata->ctdb;
 164         struct ctdb_rec_data *rec;
 165         uint32_t lmaster;
 166         size_t old_size;
 167
 168         lmaster = ctdb_lmaster(ctdb, &key);
 169
 170         rec = ctdb_marshall_record(vdata->list[lmaster], ctdb->pnn, key, NULL, tdb_null);
 171         if (rec == NULL) {
 172                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 173                 vdata->traverse_error = true;
 174                 return -1;
 175         }
 176
 177         old_size = talloc_get_size(vdata->list[lmaster]);
 178         vdata->list[lmaster] = talloc_realloc_size(NULL, vdata->list[lmaster],
 179                                                    old_size + rec->length);
 180         if (vdata->list[lmaster] == NULL) {
 181                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 182                 vdata->traverse_error = true;
 183                 return -1;
 184         }
 185
 186         vdata->list[lmaster]->count++;
 187         memcpy(old_size+(uint8_t *)vdata->list[lmaster], rec, rec->length);
 188         talloc_free(rec);
 189
 190         vdata->total++;
 191
 192         return 0;
 193 }
 194
 195
 196 static void ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
 197                                                           struct timeval t, void *private_data);
 198
 199
 200 /*
 201  * traverse function for gathering the records that can be deleted
 202  */
 203 static int vacuum_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 204 {
 205         struct vacuum_data *vdata = talloc_get_type(private, struct vacuum_data);
 206         struct ctdb_context *ctdb = vdata->ctdb;
 207         uint32_t lmaster;
 208         struct ctdb_ltdb_header *hdr;
 209         int res = 0;
 210
 211         lmaster = ctdb_lmaster(ctdb, &key);
 212         if (lmaster >= ctdb->num_nodes) {
 213                 DEBUG(DEBUG_CRIT, (__location__
 214                                    " lmaster[%u] >= ctdb->num_nodes[%u] for key"
 215                                    " with hash[%u]!\n",
 216                                    (unsigned)lmaster,
 217                                    (unsigned)ctdb->num_nodes,
 218                                    (unsigned)ctdb_hash(&key)));
 219                 return -1;
 220         }
 221
 222         if (data.dsize != sizeof(struct ctdb_ltdb_header)) {
 223                 /* its not a deleted record */
 224                 return 0;
 225         }
 226
 227         hdr = (struct ctdb_ltdb_header *)data.dptr;
 228
 229         if (hdr->dmaster != ctdb->pnn) {
 230                 return 0;
 231         }
 232
 233         if (lmaster == ctdb->pnn) {
 234                 /*
 235                  * We are both lmaster and dmaster, and the record * is empty.
 236                  * So we should be able to delete it.
 237                  */
 238                 res = add_record_to_delete_tree(vdata, key, hdr);
 239         } else {
 240                 /*
 241                  * We are not lmaster.
 242                  * Add the record to the blob ready to send to the nodes.
 243                  */
 244                 res = add_record_to_vacuum_fetch_list(vdata, key);
 245         }
 246
 247         return res;
 248 }
 249
 250 /*
 251  * traverse the tree of records to delete and marshall them into
 252  * a blob
 253  */
 254 static void delete_traverse(void *param, void *data)
 255 {
 256         struct delete_record_data *dd = talloc_get_type(data, struct delete_record_data);
 257         struct delete_records_list *recs = talloc_get_type(param, struct delete_records_list);
 258         struct ctdb_rec_data *rec;
 259         size_t old_size;
 260
 261         rec = ctdb_marshall_record(dd, recs->records->db_id, dd->key, &dd->hdr, tdb_null);
 262         if (rec == NULL) {
 263                 DEBUG(DEBUG_ERR, (__location__ " failed to marshall record\n"));
 264                 return;
 265         }
 266
 267         old_size = talloc_get_size(recs->records);
 268         recs->records = talloc_realloc_size(NULL, recs->records, old_size + rec->length);
 269         if (recs->records == NULL) {
 270                 DEBUG(DEBUG_ERR,(__location__ " Failed to expand\n"));
 271                 return;
 272         }
 273         recs->records->count++;
 274         memcpy(old_size+(uint8_t *)(recs->records), rec, rec->length);
 275 }
 276
 277 /**
 278  * traverse function for the traversal of the delete_queue,
 279  * the fast-path vacuuming list.
 280  *
 281  *  - If the record has been migrated off the node
 282  *    or has been revived (filled with data) on the node,
 283  *    then skip the record.
 284  *
 285  *  - If the current node is the record's lmaster and it is
 286  *    a record that has never been migrated with data, then
 287  *    delete the record from the local tdb.
 288  *
 289  *  - If the current node is the record's lmaster and it has
 290  *    been migrated with data, then schedule it for the normal
 291  *    vacuuming procedure (i.e. add it to the delete_list).
 292  *
 293  *  - If the current node is NOT the record's lmaster then
 294  *    add it to the list of records that are to be sent to
 295  *    the lmaster with the VACUUM_FETCH message.
 296  */
 297 static void delete_queue_traverse(void *param, void *data)
 298 {
 299         struct delete_record_data *dd =
 300                 talloc_get_type(data, struct delete_record_data);
 301         struct vacuum_data *vdata = talloc_get_type(param, struct vacuum_data);
 302         struct ctdb_db_context *ctdb_db = dd->ctdb_db;
 303         struct ctdb_context *ctdb = ctdb_db->ctdb; /* or dd->ctdb ??? */
 304         int res;
 305         struct ctdb_ltdb_header *header;
 306         TDB_DATA tdb_data;
 307         uint32_t lmaster;
 308
 309         res = tdb_chainlock(ctdb_db->ltdb->tdb, dd->key);
 310         if (res != 0) {
 311                 DEBUG(DEBUG_ERR, (__location__ " Error getting chainlock.\n"));
 312                 return;
 313         }
 314
 315         tdb_data = tdb_fetch(ctdb_db->ltdb->tdb, dd->key);
 316         if (tdb_data.dsize < sizeof(struct ctdb_ltdb_header)) {
 317                 /* Does not exist or not a ctdb record. Skip. */
 318                 goto done;
 319         }
 320
 321         if (tdb_data.dsize > sizeof(struct ctdb_ltdb_header)) {
 322                 /* The record has been recycled (filled with data). Skip. */
 323                 goto done;
 324         }
 325
 326         header = (struct ctdb_ltdb_header *)tdb_data.dptr;
 327
 328         if (header->dmaster != ctdb->pnn) {
 329                 /* The record has been migrated off the node. Skip. */
 330                 goto done;
 331         }
 332
 333
 334         if (header->rsn != dd->hdr.rsn) {
 335                 /*
 336                  * The record has been migrated off the node and back again.
 337                  * But not requeued for deletion. Skip it.
 338                  */
 339                 goto done;
 340         }
 341
 342         /*
 343          * We are dmaster, and the record has no data, and it has
 344          * not been migrated after it has been queued for deletion.
 345          *
 346          * At this stage, the record could still have been revived locally
 347          * and last been written with empty data. This can only be
 348          * fixed with the addition of an active or delete flag. (TODO)
 349          */
 350
 351         lmaster = ctdb_lmaster(ctdb_db->ctdb, &dd->key);
 352
 353         if (lmaster != ctdb->pnn) {
 354                 res = add_record_to_vacuum_fetch_list(vdata, dd->key);
 355
 356                 if (res != 0) {
 357                         DEBUG(DEBUG_ERR,
 358                               (__location__ " Error adding record to list "
 359                                "of records to send to lmaster.\n"));
 360                 }
 361
 362                 goto done;
 363         }
 364
 365         /* use header->flags or dd->hdr.flags ?? */
 366         if (dd->hdr.flags & CTDB_REC_FLAG_MIGRATED_WITH_DATA) {
 367                 res = add_record_to_delete_tree(vdata, dd->key, &dd->hdr);
 368
 369                 if (res != 0) {
 370                         DEBUG(DEBUG_ERR,
 371                               (__location__ " Error adding record to list "
 372                                "of records for deletion on lmaster.\n"));
 373                 }
 374         } else {
 375                 res = tdb_delete(ctdb_db->ltdb->tdb, dd->key);
 376
 377                 if (res != 0) {
 378                         DEBUG(DEBUG_ERR,
 379                               (__location__ " Error deleting record from local "
 380                                "data base.\n"));
 381                 }
 382         }
 383
 384 done:
 385         if (tdb_data.dptr != NULL) {
 386                 free(tdb_data.dptr);
 387         }
 388         tdb_chainunlock(ctdb_db->ltdb->tdb, dd->key);
 389
 390         return;
 391 }
 392
 393 /*
 394  * read-only traverse the database in order to find
 395  * records that can be deleted and try to delete these
 396  * records on the other nodes
 397  * this executes in the child context
 398  */
 399 static int ctdb_vacuum_db(struct ctdb_db_context *ctdb_db,
 400                           struct vacuum_data *vdata,
 401                           bool full_vacuum_run)
 402 {
 403         struct ctdb_context *ctdb = ctdb_db->ctdb;
 404         const char *name = ctdb_db->db_name;
 405         int ret, i, pnn;
 406
 407         ret = ctdb_ctrl_getvnnmap(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE, ctdb, &ctdb->vnn_map);
 408         if (ret != 0) {
 409                 DEBUG(DEBUG_ERR, ("Unable to get vnnmap from local node\n"));
 410                 return ret;
 411         }
 412
 413         pnn = ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 414         if (pnn == -1) {
 415                 DEBUG(DEBUG_ERR, ("Unable to get pnn from local node\n"));
 416                 return -1;
 417         }
 418
 419         ctdb->pnn = pnn;
 420         /* the list needs to be of length num_nodes */
 421         vdata->list = talloc_array(vdata, struct ctdb_marshall_buffer *, ctdb->num_nodes);
 422         if (vdata->list == NULL) {
 423                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 424                 return -1;
 425         }
 426         for (i = 0; i < ctdb->num_nodes; i++) {
 427                 vdata->list[i] = (struct ctdb_marshall_buffer *)
 428                         talloc_zero_size(vdata->list,
 429                                                          offsetof(struct ctdb_marshall_buffer, data));
 430                 if (vdata->list[i] == NULL) {
 431                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 432                         return -1;
 433                 }
 434                 vdata->list[i]->db_id = ctdb_db->db_id;
 435         }
 436
 437         /*
 438          * Traverse the delete_queue.
 439          * This builds the same lists as the db traverse.
 440          */
 441         trbt_traversearray32(ctdb_db->delete_queue, 1, delete_queue_traverse, vdata);
 442
 443         /*
 444          * read-only traverse of the database, looking for records that
 445          * might be able to be vacuumed.
 446          *
 447          * This is not done each time but only every tunable
 448          * VacuumFastPathCount times.
 449          */
 450         if (full_vacuum_run) {
 451                 ret = tdb_traverse_read(ctdb_db->ltdb->tdb, vacuum_traverse, vdata);
 452                 if (ret == -1 || vdata->traverse_error) {
 453                         DEBUG(DEBUG_ERR,(__location__ " Traverse error in vacuuming '%s'\n", name));
 454                         return -1;
 455                 }
 456         }
 457
 458         /*
 459          * For records where we are not the lmaster,
 460          * tell the lmaster to fetch the record.
 461          */
 462         for (i = 0; i < ctdb->num_nodes; i++) {
 463                 TDB_DATA data;
 464
 465                 if (ctdb->nodes[i]->pnn == ctdb->pnn) {
 466                         continue;
 467                 }
 468
 469                 if (vdata->list[i]->count == 0) {
 470                         continue;
 471                 }
 472
 473                 DEBUG(DEBUG_INFO, ("Found %u records for lmaster %u in '%s'\n",
 474                                    vdata->list[i]->count, ctdb->nodes[i]->pnn,
 475                                    name));
 476
 477                 data.dsize = talloc_get_size(vdata->list[i]);
 478                 data.dptr  = (void *)vdata->list[i];
 479                 if (ctdb_send_message(ctdb, ctdb->nodes[i]->pnn, CTDB_SRVID_VACUUM_FETCH, data) != 0) {
 480                         DEBUG(DEBUG_ERR, (__location__ " Failed to send vacuum "
 481                                           "fetch message to %u\n",
 482                                           ctdb->nodes[i]->pnn));
 483                         return -1;
 484                 }
 485         }
 486
 487         /* Process all records we can delete (if any) */
 488         if (vdata->delete_count > 0) {
 489                 struct delete_records_list *recs;
 490                 TDB_DATA indata, outdata;
 491                 int32_t res;
 492                 struct ctdb_node_map *nodemap;
 493                 uint32_t *active_nodes;
 494                 int num_active_nodes;
 495
 496                 recs = talloc_zero(vdata, struct delete_records_list);
 497                 if (recs == NULL) {
 498                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 499                         return -1;
 500                 }
 501                 recs->records = (struct ctdb_marshall_buffer *)
 502                         talloc_zero_size(vdata,
 503                                     offsetof(struct ctdb_marshall_buffer, data));
 504                 if (recs->records == NULL) {
 505                         DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 506                         return -1;
 507                 }
 508                 recs->records->db_id = ctdb_db->db_id;
 509
 510                 /*
 511                  * traverse the tree of all records we want to delete and
 512                  * create a blob we can send to the other nodes.
 513                  */
 514                 trbt_traversearray32(vdata->delete_tree, 1, delete_traverse, recs);
 515
 516                 indata.dsize = talloc_get_size(recs->records);
 517                 indata.dptr  = (void *)recs->records;
 518
 519                 /*
 520                  * now tell all the active nodes to delete all these records
 521                  * (if possible)
 522                  */
 523
 524                 ret = ctdb_ctrl_getnodemap(ctdb, TIMELIMIT(),
 525                                            CTDB_CURRENT_NODE,
 526                                            recs, /* talloc context */
 527                                            &nodemap);
 528                 if (ret != 0) {
 529                         DEBUG(DEBUG_ERR,(__location__ " unable to get node map\n"));
 530                         return -1;
 531                 }
 532
 533                 active_nodes = list_of_active_nodes(ctdb, nodemap,
 534                                                     nodemap, /* talloc context */
 535                                                     false /* include self */);
 536                 /* yuck! ;-) */
 537                 num_active_nodes = talloc_get_size(active_nodes)/sizeof(*active_nodes);
 538
 539                 for (i = 0; i < num_active_nodes; i++) {
 540                         struct ctdb_marshall_buffer *records;
 541                         struct ctdb_rec_data *rec;
 542
 543                         ret = ctdb_control(ctdb, active_nodes[i], 0,
 544                                         CTDB_CONTROL_TRY_DELETE_RECORDS, 0,
 545                                         indata, recs, &outdata, &res,
 546                                         NULL, NULL);
 547                         if (ret != 0 || res != 0) {
 548                                 DEBUG(DEBUG_ERR, ("Failed to delete records on "
 549                                                   "node %u: ret[%d] res[%d]\n",
 550                                                   active_nodes[i], ret, res));
 551                                 return -1;
 552                         }
 553
 554                         /*
 555                          * outdata countains the list of records coming back
 556                          * from the node which the node could not delete
 557                          */
 558                         records = (struct ctdb_marshall_buffer *)outdata.dptr;
 559                         rec = (struct ctdb_rec_data *)&records->data[0];
 560                         while (records->count-- > 1) {
 561                                 TDB_DATA reckey, recdata;
 562                                 struct ctdb_ltdb_header *rechdr;
 563
 564                                 reckey.dptr = &rec->data[0];
 565                                 reckey.dsize = rec->keylen;
 566                                 recdata.dptr = &rec->data[reckey.dsize];
 567                                 recdata.dsize = rec->datalen;
 568
 569                                 if (recdata.dsize < sizeof(struct ctdb_ltdb_header)) {
 570                                         DEBUG(DEBUG_CRIT,(__location__ " bad ltdb record\n"));
 571                                         return -1;
 572                                 }
 573                                 rechdr = (struct ctdb_ltdb_header *)recdata.dptr;
 574                                 recdata.dptr += sizeof(*rechdr);
 575                                 recdata.dsize -= sizeof(*rechdr);
 576
 577                                 /*
 578                                  * that other node couldnt delete the record
 579                                  * so we should delete it and thereby remove it from the tree
 580                                  */
 581                                 talloc_free(trbt_lookup32(vdata->delete_tree, ctdb_hash(&reckey)));
 582
 583                                 rec = (struct ctdb_rec_data *)(rec->length + (uint8_t *)rec);
 584                         }
 585                 }
 586
 587                 /* free nodemap and active_nodes */
 588                 talloc_free(nodemap);
 589
 590                 /*
 591                  * The only records remaining in the tree would be those
 592                  * records where all other nodes could successfully
 593                  * delete them, so we can safely delete them on the
 594                  * lmaster as well. Deletion implictely happens while
 595                  * we repack the database. The repack algorithm revisits
 596                  * the tree in order to find the records that don't need
 597                  * to be copied / repacked.
 598                  */
 599         }
 600
 601         /* this ensures we run our event queue */
 602         ctdb_ctrl_getpnn(ctdb, TIMELIMIT(), CTDB_CURRENT_NODE);
 603
 604         return 0;
 605 }
 606
 607
 608 /*
 609  * traverse function for repacking
 610  */
 611 static int repack_traverse(struct tdb_context *tdb, TDB_DATA key, TDB_DATA data, void *private)
 612 {
 613         struct vacuum_data *vdata = (struct vacuum_data *)private;
 614
 615         if (vdata->vacuum) {
 616                 uint32_t hash = ctdb_hash(&key);
 617                 struct delete_record_data *kd;
 618                 /*
 619                  * check if we can ignore this record because it's in the delete_tree
 620                  */
 621                 kd = (struct delete_record_data *)trbt_lookup32(vdata->delete_tree, hash);
 622                 /*
 623                  * there might be hash collisions so we have to compare the keys here to be sure
 624                  */
 625                 if (kd && kd->key.dsize == key.dsize && memcmp(kd->key.dptr, key.dptr, key.dsize) == 0) {
 626                         struct ctdb_ltdb_header *hdr = (struct ctdb_ltdb_header *)data.dptr;
 627                         /*
 628                          * we have to check if the record hasn't changed in the meantime in order to
 629                          * savely remove it from the database
 630                          */
 631                         if (data.dsize == sizeof(struct ctdb_ltdb_header) &&
 632                                 hdr->dmaster == kd->ctdb->pnn &&
 633                                 ctdb_lmaster(kd->ctdb, &(kd->key)) == kd->ctdb->pnn &&
 634                                 kd->hdr.rsn == hdr->rsn) {
 635                                 vdata->vacuumed++;
 636                                 return 0;
 637                         }
 638                 }
 639         }
 640         if (tdb_store(vdata->dest_db, key, data, TDB_INSERT) != 0) {
 641                 vdata->traverse_error = true;
 642                 return -1;
 643         }
 644         vdata->copied++;
 645         return 0;
 646 }
 647
 648 /*
 649  * repack a tdb
 650  */
 651 static int ctdb_repack_tdb(struct tdb_context *tdb, TALLOC_CTX *mem_ctx, struct vacuum_data *vdata)
 652 {
 653         struct tdb_context *tmp_db;
 654
 655         if (tdb_transaction_start(tdb) != 0) {
 656                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 657                 return -1;
 658         }
 659
 660         tmp_db = tdb_open("tmpdb", tdb_hash_size(tdb),
 661                           TDB_INTERNAL|TDB_DISALLOW_NESTING,
 662                           O_RDWR|O_CREAT, 0);
 663         if (tmp_db == NULL) {
 664                 DEBUG(DEBUG_ERR,(__location__ " Failed to create tmp_db\n"));
 665                 tdb_transaction_cancel(tdb);
 666                 return -1;
 667         }
 668
 669         vdata->traverse_error = false;
 670         vdata->dest_db = tmp_db;
 671         vdata->vacuum = true;
 672         vdata->vacuumed = 0;
 673         vdata->copied = 0;
 674
 675         /*
 676          * repack and vacuum on-the-fly by not writing the records that are
 677          * no longer needed
 678          */
 679         if (tdb_traverse_read(tdb, repack_traverse, vdata) == -1) {
 680                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying out\n"));
 681                 tdb_transaction_cancel(tdb);
 682                 tdb_close(tmp_db);
 683                 return -1;
 684         }
 685
 686         DEBUG(DEBUG_INFO,(__location__ " %u records vacuumed\n", vdata->vacuumed));
 687
 688         if (vdata->traverse_error) {
 689                 DEBUG(DEBUG_ERR,(__location__ " Error during traversal\n"));
 690                 tdb_transaction_cancel(tdb);
 691                 tdb_close(tmp_db);
 692                 return -1;
 693         }
 694
 695         if (tdb_wipe_all(tdb) != 0) {
 696                 DEBUG(DEBUG_ERR,(__location__ " Failed to wipe database\n"));
 697                 tdb_transaction_cancel(tdb);
 698                 tdb_close(tmp_db);
 699                 return -1;
 700         }
 701
 702         vdata->traverse_error = false;
 703         vdata->dest_db = tdb;
 704         vdata->vacuum = false;
 705         vdata->copied = 0;
 706
 707         if (tdb_traverse_read(tmp_db, repack_traverse, vdata) == -1) {
 708                 DEBUG(DEBUG_ERR,(__location__ " Failed to traverse copying back\n"));
 709                 tdb_transaction_cancel(tdb);
 710                 tdb_close(tmp_db);
 711                 return -1;
 712         }
 713
 714         if (vdata->traverse_error) {
 715                 DEBUG(DEBUG_ERR,(__location__ " Error during second traversal\n"));
 716                 tdb_transaction_cancel(tdb);
 717                 tdb_close(tmp_db);
 718                 return -1;
 719         }
 720
 721         tdb_close(tmp_db);
 722
 723
 724         if (tdb_transaction_commit(tdb) != 0) {
 725                 DEBUG(DEBUG_ERR,(__location__ " Failed to commit\n"));
 726                 return -1;
 727         }
 728         DEBUG(DEBUG_INFO,(__location__ " %u records copied\n", vdata->copied));
 729
 730         return 0;
 731 }
 732
 733 static int update_tuning_db(struct ctdb_db_context *ctdb_db, struct vacuum_data *vdata, uint32_t freelist)
 734 {
 735         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 736         TDB_CONTEXT *tune_tdb;
 737         TDB_DATA key, value;
 738         struct vacuum_tuning_data tdata;
 739         struct vacuum_tuning_data *tptr;
 740         char *vac_dbname;
 741         int flags;
 742
 743         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u",
 744                                      ctdb_db->ctdb->db_directory_state,
 745                                      TUNINGDBNAME, ctdb_db->ctdb->pnn);
 746         if (vac_dbname == NULL) {
 747                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
 748                 talloc_free(tmp_ctx);
 749                 return -1;
 750         }
 751
 752         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
 753         flags |= TDB_DISALLOW_NESTING;
 754         tune_tdb = tdb_open(vac_dbname, 0,
 755                             flags,
 756                             O_RDWR|O_CREAT, 0600);
 757         if (tune_tdb == NULL) {
 758                 DEBUG(DEBUG_ERR,(__location__ " Failed to create/open %s\n", TUNINGDBNAME));
 759                 talloc_free(tmp_ctx);
 760                 return -1;
 761         }
 762
 763         if (tdb_transaction_start(tune_tdb) != 0) {
 764                 DEBUG(DEBUG_ERR,(__location__ " Failed to start transaction\n"));
 765                 tdb_close(tune_tdb);
 766                 return -1;
 767         }
 768         key.dptr = discard_const(ctdb_db->db_name);
 769         key.dsize = strlen(ctdb_db->db_name);
 770         value = tdb_fetch(tune_tdb, key);
 771
 772         if (value.dptr != NULL && value.dsize == sizeof(struct vacuum_tuning_data)) {
 773                 tptr = (struct vacuum_tuning_data *)value.dptr;
 774                 tdata = *tptr;
 775
 776                 /*
 777                  * re-calc new vacuum interval:
 778                  * in case no limit was reached we continously increase the interval
 779                  * until vacuum_max_interval is reached
 780                  * in case a limit was reached we divide the current interval by 2
 781                  * unless vacuum_min_interval is reached
 782                  */
 783                 if (freelist < vdata->repack_limit &&
 784                     vdata->delete_count < vdata->vacuum_limit) {
 785                         if (tdata.last_interval < ctdb_db->ctdb->tunable.vacuum_max_interval) {
 786                                 tdata.new_interval = tdata.last_interval * 110 / 100;
 787                                 DEBUG(DEBUG_INFO,("Increasing vacuum interval %u -> %u for %s\n",
 788                                         tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 789                         }
 790                 } else {
 791                         tdata.new_interval = tdata.last_interval / 2;
 792                         if (tdata.new_interval < ctdb_db->ctdb->tunable.vacuum_min_interval ||
 793                                 tdata.new_interval > ctdb_db->ctdb->tunable.vacuum_max_interval) {
 794                                 tdata.new_interval = ctdb_db->ctdb->tunable.vacuum_min_interval;
 795                         }
 796                         DEBUG(DEBUG_INFO,("Decreasing vacuum interval %u -> %u for %s\n",
 797                                          tdata.last_interval, tdata.new_interval, ctdb_db->db_name));
 798                 }
 799                 tdata.last_interval = tdata.new_interval;
 800         } else {
 801                 DEBUG(DEBUG_ERR,(__location__ " Cannot find tunedb record for %s. Using default interval\n", ctdb_db->db_name));
 802                 tdata.last_num_repack = freelist;
 803                 tdata.last_num_empty = vdata->delete_count;
 804                 tdata.last_interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 805         }
 806
 807         if (value.dptr != NULL) {
 808                 free(value.dptr);
 809         }
 810
 811         tdata.last_start = vdata->start;
 812         tdata.last_duration = timeval_elapsed(&vdata->start);
 813
 814         value.dptr = (unsigned char *)&tdata;
 815         value.dsize = sizeof(tdata);
 816
 817         if (tdb_store(tune_tdb, key, value, 0) != 0) {
 818                 DEBUG(DEBUG_ERR,(__location__ " Unable to store tundb record for %s\n", ctdb_db->db_name));
 819                 tdb_transaction_cancel(tune_tdb);
 820                 tdb_close(tune_tdb);
 821                 talloc_free(tmp_ctx);
 822                 return -1;
 823         }
 824         tdb_transaction_commit(tune_tdb);
 825         tdb_close(tune_tdb);
 826         talloc_free(tmp_ctx);
 827
 828         return 0;
 829 }
 830
 831 /*
 832  * repack and vaccum a db
 833  * called from the child context
 834  */
 835 static int ctdb_vacuum_and_repack_db(struct ctdb_db_context *ctdb_db,
 836                                      TALLOC_CTX *mem_ctx,
 837                                      bool full_vacuum_run)
 838 {
 839         uint32_t repack_limit = ctdb_db->ctdb->tunable.repack_limit;
 840         uint32_t vacuum_limit = ctdb_db->ctdb->tunable.vacuum_limit;
 841         const char *name = ctdb_db->db_name;
 842         int size;
 843         struct vacuum_data *vdata;
 844
 845         size = tdb_freelist_size(ctdb_db->ltdb->tdb);
 846         if (size == -1) {
 847                 DEBUG(DEBUG_ERR,(__location__ " Failed to get freelist size for '%s'\n", name));
 848                 return -1;
 849         }
 850
 851         vdata = talloc_zero(mem_ctx, struct vacuum_data);
 852         if (vdata == NULL) {
 853                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 854                 return -1;
 855         }
 856
 857         vdata->ctdb = ctdb_db->ctdb;
 858         vdata->vacuum_limit = vacuum_limit;
 859         vdata->repack_limit = repack_limit;
 860         vdata->delete_tree = trbt_create(vdata, 0);
 861         if (vdata->delete_tree == NULL) {
 862                 DEBUG(DEBUG_ERR,(__location__ " Out of memory\n"));
 863                 talloc_free(vdata);
 864                 return -1;
 865         }
 866
 867         vdata->start = timeval_current();
 868
 869         /*
 870          * gather all records that can be deleted in vdata
 871          */
 872         if (ctdb_vacuum_db(ctdb_db, vdata, full_vacuum_run) != 0) {
 873                 DEBUG(DEBUG_ERR,(__location__ " Failed to vacuum '%s'\n", name));
 874         }
 875
 876         /*
 877          * decide if a repack is necessary
 878          */
 879         if (size < repack_limit && vdata->delete_count < vacuum_limit) {
 880                 update_tuning_db(ctdb_db, vdata, size);
 881                 talloc_free(vdata);
 882                 return 0;
 883         }
 884
 885         DEBUG(DEBUG_INFO,("Repacking %s with %u freelist entries and %u records to delete\n",
 886                         name, size, vdata->delete_count));
 887
 888         /*
 889          * repack and implicitely get rid of the records we can delete
 890          */
 891         if (ctdb_repack_tdb(ctdb_db->ltdb->tdb, mem_ctx, vdata) != 0) {
 892                 DEBUG(DEBUG_ERR,(__location__ " Failed to repack '%s'\n", name));
 893                 update_tuning_db(ctdb_db, vdata, size);
 894                 talloc_free(vdata);
 895                 return -1;
 896         }
 897         update_tuning_db(ctdb_db, vdata, size);
 898         talloc_free(vdata);
 899
 900         return 0;
 901 }
 902
 903 static int get_vacuum_interval(struct ctdb_db_context *ctdb_db)
 904 {
 905         TALLOC_CTX *tmp_ctx = talloc_new(NULL);
 906         TDB_CONTEXT *tdb;
 907         TDB_DATA key, value;
 908         char *vac_dbname;
 909         uint interval = ctdb_db->ctdb->tunable.vacuum_default_interval;
 910         struct ctdb_context *ctdb = ctdb_db->ctdb;
 911         int flags;
 912
 913         vac_dbname = talloc_asprintf(tmp_ctx, "%s/%s.%u", ctdb->db_directory, TUNINGDBNAME, ctdb->pnn);
 914         if (vac_dbname == NULL) {
 915                 DEBUG(DEBUG_CRIT,(__location__ " Out of memory error while allocating '%s'\n", vac_dbname));
 916                 talloc_free(tmp_ctx);
 917                 return interval;
 918         }
 919
 920         flags  = ctdb_db->ctdb->valgrinding ? TDB_NOMMAP : 0;
 921         flags |= TDB_DISALLOW_NESTING;
 922         tdb = tdb_open(vac_dbname, 0,
 923                        flags,
 924                        O_RDWR|O_CREAT, 0600);
 925         if (!tdb) {
 926                 DEBUG(DEBUG_ERR,("Unable to open/create database %s using default interval\n", vac_dbname));
 927                 talloc_free(tmp_ctx);
 928                 return interval;
 929         }
 930
 931         key.dptr = discard_const(ctdb_db->db_name);
 932         key.dsize = strlen(ctdb_db->db_name);
 933
 934         value = tdb_fetch(tdb, key);
 935
 936         if (value.dptr != NULL) {
 937                 if (value.dsize == sizeof(struct vacuum_tuning_data)) {
 938                         struct vacuum_tuning_data *tptr = (struct vacuum_tuning_data *)value.dptr;
 939
 940                         interval = tptr->new_interval;
 941
 942                         if (interval < ctdb->tunable.vacuum_min_interval) {
 943                                 interval = ctdb->tunable.vacuum_min_interval;
 944                         }
 945                         if (interval > ctdb->tunable.vacuum_max_interval) {
 946                                 interval = ctdb->tunable.vacuum_max_interval;
 947                         }
 948                 }
 949                 free(value.dptr);
 950         }
 951         tdb_close(tdb);
 952
 953         talloc_free(tmp_ctx);
 954
 955         return interval;
 956 }
 957
 958 static int vacuum_child_destructor(struct ctdb_vacuum_child_context *child_ctx)
 959 {
 960         double l = timeval_elapsed(&child_ctx->start_time);
 961         struct ctdb_db_context *ctdb_db = child_ctx->vacuum_handle->ctdb_db;
 962         struct ctdb_context *ctdb = ctdb_db->ctdb;
 963
 964         DEBUG(DEBUG_INFO,("Vacuuming took %.3f seconds for database %s\n", l, ctdb_db->db_name));
 965
 966         if (child_ctx->child_pid != -1) {
 967                 kill(child_ctx->child_pid, SIGKILL);
 968         } else {
 969                 /* Bump the number of successful fast-path runs. */
 970                 child_ctx->vacuum_handle->fast_path_count++;
 971         }
 972
 973         event_add_timed(ctdb->ev, child_ctx->vacuum_handle,
 974                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
 975                         ctdb_vacuum_event, child_ctx->vacuum_handle);
 976
 977         return 0;
 978 }
 979
 980 /*
 981  * this event is generated when a vacuum child process times out
 982  */
 983 static void vacuum_child_timeout(struct event_context *ev, struct timed_event *te,
 984                                          struct timeval t, void *private_data)
 985 {
 986         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
 987
 988         DEBUG(DEBUG_ERR,("Vacuuming child process timed out for db %s\n", child_ctx->vacuum_handle->ctdb_db->db_name));
 989
 990         child_ctx->status = VACUUM_TIMEOUT;
 991
 992         talloc_free(child_ctx);
 993 }
 994
 995
 996 /*
 997  * this event is generated when a vacuum child process has completed
 998  */
 999 static void vacuum_child_handler(struct event_context *ev, struct fd_event *fde,
1000                              uint16_t flags, void *private_data)
1001 {
1002         struct ctdb_vacuum_child_context *child_ctx = talloc_get_type(private_data, struct ctdb_vacuum_child_context);
1003         char c = 0;
1004         int ret;
1005
1006         DEBUG(DEBUG_INFO,("Vacuuming child process %d finished for db %s\n", child_ctx->child_pid, child_ctx->vacuum_handle->ctdb_db->db_name));
1007         child_ctx->child_pid = -1;
1008
1009         ret = read(child_ctx->fd[0], &c, 1);
1010         if (ret != 1 || c != 0) {
1011                 child_ctx->status = VACUUM_ERROR;
1012                 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));
1013         } else {
1014                 child_ctx->status = VACUUM_OK;
1015         }
1016
1017         talloc_free(child_ctx);
1018 }
1019
1020 /*
1021  * this event is called every time we need to start a new vacuum process
1022  */
1023 static void
1024 ctdb_vacuum_event(struct event_context *ev, struct timed_event *te,
1025                                struct timeval t, void *private_data)
1026 {
1027         struct ctdb_vacuum_handle *vacuum_handle = talloc_get_type(private_data, struct ctdb_vacuum_handle);
1028         struct ctdb_db_context *ctdb_db = vacuum_handle->ctdb_db;
1029         struct ctdb_context *ctdb = ctdb_db->ctdb;
1030         struct ctdb_vacuum_child_context *child_ctx;
1031         int ret;
1032
1033         /* we dont vacuum if we are in recovery mode */
1034         if (ctdb->recovery_mode == CTDB_RECOVERY_ACTIVE) {
1035                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1036                 return;
1037         }
1038
1039         child_ctx = talloc(vacuum_handle, struct ctdb_vacuum_child_context);
1040         if (child_ctx == NULL) {
1041                 DEBUG(DEBUG_CRIT, (__location__ " Failed to allocate child context for vacuuming of %s\n", ctdb_db->db_name));
1042                 ctdb_fatal(ctdb, "Out of memory when crating vacuum child context. Shutting down\n");
1043         }
1044
1045
1046         ret = pipe(child_ctx->fd);
1047         if (ret != 0) {
1048                 talloc_free(child_ctx);
1049                 DEBUG(DEBUG_ERR, ("Failed to create pipe for vacuum child process.\n"));
1050                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1051                 return;
1052         }
1053
1054         if (vacuum_handle->fast_path_count > ctdb->tunable.vacuum_fast_path_count) {
1055                 vacuum_handle->fast_path_count = 0;
1056         }
1057
1058         child_ctx->child_pid = fork();
1059         if (child_ctx->child_pid == (pid_t)-1) {
1060                 close(child_ctx->fd[0]);
1061                 close(child_ctx->fd[1]);
1062                 talloc_free(child_ctx);
1063                 DEBUG(DEBUG_ERR, ("Failed to fork vacuum child process.\n"));
1064                 event_add_timed(ctdb->ev, vacuum_handle, timeval_current_ofs(ctdb->tunable.vacuum_default_interval, 0), ctdb_vacuum_event, vacuum_handle);
1065                 return;
1066         }
1067
1068
1069         if (child_ctx->child_pid == 0) {
1070                 char cc = 0;
1071                 bool full_vacuum_run = false;
1072                 close(child_ctx->fd[0]);
1073
1074                 DEBUG(DEBUG_INFO,("Vacuuming child process %d for db %s started\n", getpid(), ctdb_db->db_name));
1075
1076                 if (switch_from_server_to_client(ctdb) != 0) {
1077                         DEBUG(DEBUG_CRIT, (__location__ "ERROR: failed to switch vacuum daemon into client mode. Shutting down.\n"));
1078                         _exit(1);
1079                 }
1080
1081                 /*
1082                  * repack the db
1083                  */
1084                 if ((ctdb->tunable.vacuum_fast_path_count > 0) &&
1085                     (vacuum_handle->fast_path_count == 0))
1086                 {
1087                         full_vacuum_run = true;
1088                 }
1089                 cc = ctdb_vacuum_and_repack_db(ctdb_db, child_ctx,
1090                                                full_vacuum_run);
1091
1092                 write(child_ctx->fd[1], &cc, 1);
1093                 _exit(0);
1094         }
1095
1096         set_close_on_exec(child_ctx->fd[0]);
1097         close(child_ctx->fd[1]);
1098
1099         child_ctx->status = VACUUM_RUNNING;
1100         child_ctx->start_time = timeval_current();
1101
1102         talloc_set_destructor(child_ctx, vacuum_child_destructor);
1103
1104         /*
1105          * Clear the fastpath vacuuming list in the parent.
1106          */
1107         talloc_free(ctdb_db->delete_queue);
1108         ctdb_db->delete_queue = trbt_create(ctdb_db, 0);
1109         if (ctdb_db->delete_queue == NULL) {
1110                 /* fatal here? ... */
1111                 ctdb_fatal(ctdb, "Out of memory when re-creating vacuum tree "
1112                                  "in parent context. Shutting down\n");
1113         }
1114
1115         event_add_timed(ctdb->ev, child_ctx,
1116                 timeval_current_ofs(ctdb->tunable.vacuum_max_run_time, 0),
1117                 vacuum_child_timeout, child_ctx);
1118
1119         DEBUG(DEBUG_DEBUG, (__location__ " Created PIPE FD:%d to child vacuum process\n", child_ctx->fd[0]));
1120
1121         event_add_fd(ctdb->ev, child_ctx, child_ctx->fd[0],
1122                 EVENT_FD_READ|EVENT_FD_AUTOCLOSE,
1123                 vacuum_child_handler,
1124                 child_ctx);
1125
1126         vacuum_handle->child_ctx = child_ctx;
1127         child_ctx->vacuum_handle = vacuum_handle;
1128 }
1129
1130
1131 /* this function initializes the vacuuming context for a database
1132  * starts the vacuuming events
1133  */
1134 int ctdb_vacuum_init(struct ctdb_db_context *ctdb_db)
1135 {
1136         if (ctdb_db->persistent != 0) {
1137                 DEBUG(DEBUG_ERR,("Vacuuming is disabled for persistent database %s\n", ctdb_db->db_name));
1138                 return 0;
1139         }
1140
1141         ctdb_db->vacuum_handle = talloc(ctdb_db, struct ctdb_vacuum_handle);
1142         CTDB_NO_MEMORY(ctdb_db->ctdb, ctdb_db->vacuum_handle);
1143
1144         ctdb_db->vacuum_handle->ctdb_db = ctdb_db;
1145
1146         event_add_timed(ctdb_db->ctdb->ev, ctdb_db->vacuum_handle,
1147                         timeval_current_ofs(get_vacuum_interval(ctdb_db), 0),
1148                         ctdb_vacuum_event, ctdb_db->vacuum_handle);
1149
1150         return 0;
1151 }