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