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