ctdb/server/ctdb_lock.c

   1 /*
   2    ctdb lock handling
   3    provide API to do non-blocking locks for single or all databases
   4
   5    Copyright (C) Amitay Isaacs  2012
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program; if not, see <http://www.gnu.org/licenses/>.
  19 */
  20 #include "includes.h"
  21 #include "include/ctdb_private.h"
  22 #include "include/ctdb_protocol.h"
  23 #include "tevent.h"
  24 #include "tdb.h"
  25 #include "lib/tdb_wrap/tdb_wrap.h"
  26 #include "system/filesys.h"
  27 #include "lib/util/dlinklist.h"
  28
  29 /*
  30  * Non-blocking Locking API
  31  *
  32  * 1. Create a child process to do blocking locks.
  33  * 2. Once the locks are obtained, signal parent process via fd.
  34  * 3. Invoke registered callback routine with locking status.
  35  * 4. If the child process cannot get locks within certain time,
  36  *    execute an external script to debug.
  37  *
  38  * ctdb_lock_record()      - get a lock on a record
  39  * ctdb_lock_db()          - get a lock on a DB
  40  * ctdb_lock_alldb_prio()  - get a lock on all DBs with given priority
  41  * ctdb_lock_alldb()       - get a lock on all DBs
  42  *
  43  *  auto_mark              - whether to mark/unmark DBs in before/after callback
  44  */
  45
  46 enum lock_type {
  47         LOCK_RECORD,
  48         LOCK_DB,
  49         LOCK_ALLDB_PRIO,
  50         LOCK_ALLDB,
  51 };
  52
  53 static const char * const lock_type_str[] = {
  54         "lock_record",
  55         "lock_db",
  56         "lock_alldb_prio",
  57         "lock_alldb",
  58 };
  59
  60 struct lock_request;
  61
  62 /* lock_context is the common part for a lock request */
  63 struct lock_context {
  64         struct lock_context *next, *prev;
  65         enum lock_type type;
  66         struct ctdb_context *ctdb;
  67         struct ctdb_db_context *ctdb_db;
  68         TDB_DATA key;
  69         uint32_t priority;
  70         bool auto_mark;
  71         struct lock_request *request;
  72         pid_t child;
  73         int fd[2];
  74         struct tevent_fd *tfd;
  75         struct tevent_timer *ttimer;
  76         struct timeval start_time;
  77         uint32_t key_hash;
  78         bool can_schedule;
  79 };
  80
  81 /* lock_request is the client specific part for a lock request */
  82 struct lock_request {
  83         struct lock_context *lctx;
  84         void (*callback)(void *, bool);
  85         void *private_data;
  86 };
  87
  88
  89 /*
  90  * Support samba 3.6.x (and older) versions which do not set db priority.
  91  *
  92  * By default, all databases are set to priority 1. So only when priority
  93  * is set to 1, check for databases that need higher priority.
  94  */
  95 static bool later_db(struct ctdb_context *ctdb, const char *name)
  96 {
  97         if (ctdb->tunable.samba3_hack == 0) {
  98                 return false;
  99         }
 100
 101         if (strstr(name, "brlock") ||
 102             strstr(name, "g_lock") ||
 103             strstr(name, "notify_onelevel") ||
 104             strstr(name, "serverid") ||
 105             strstr(name, "xattr_tdb")) {
 106                 return true;
 107         }
 108
 109         return false;
 110 }
 111
 112 typedef int (*db_handler_t)(struct ctdb_db_context *ctdb_db,
 113                             uint32_t priority,
 114                             void *private_data);
 115
 116 static int ctdb_db_iterator(struct ctdb_context *ctdb, uint32_t priority,
 117                             db_handler_t handler, void *private_data)
 118 {
 119         struct ctdb_db_context *ctdb_db;
 120         int ret;
 121
 122         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 123                 if (ctdb_db->priority != priority) {
 124                         continue;
 125                 }
 126                 if (later_db(ctdb, ctdb_db->db_name)) {
 127                         continue;
 128                 }
 129                 ret = handler(ctdb_db, priority, private_data);
 130                 if (ret != 0) {
 131                         return -1;
 132                 }
 133         }
 134
 135         /* If priority != 1, later_db check is not required and can return */
 136         if (priority != 1) {
 137                 return 0;
 138         }
 139
 140         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 141                 if (!later_db(ctdb, ctdb_db->db_name)) {
 142                         continue;
 143                 }
 144                 ret = handler(ctdb_db, priority, private_data);
 145                 if (ret != 0) {
 146                         return -1;
 147                 }
 148         }
 149
 150         return 0;
 151 }
 152
 153
 154 /*
 155  * lock all databases - mark only
 156  */
 157 static int db_lock_mark_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 158                                 void *private_data)
 159 {
 160         int tdb_transaction_write_lock_mark(struct tdb_context *);
 161
 162         DEBUG(DEBUG_INFO, ("marking locked database %s, priority:%u\n",
 163                            ctdb_db->db_name, priority));
 164
 165         if (tdb_transaction_write_lock_mark(ctdb_db->ltdb->tdb) != 0) {
 166                 DEBUG(DEBUG_ERR, ("Failed to mark (transaction lock) database %s\n",
 167                                   ctdb_db->db_name));
 168                 return -1;
 169         }
 170
 171         if (tdb_lockall_mark(ctdb_db->ltdb->tdb) != 0) {
 172                 DEBUG(DEBUG_ERR, ("Failed to mark (all lock) database %s\n",
 173                                   ctdb_db->db_name));
 174                 return -1;
 175         }
 176
 177         return 0;
 178 }
 179
 180 int ctdb_lockall_mark_prio(struct ctdb_context *ctdb, uint32_t priority)
 181 {
 182         /*
 183          * This function is only used by the main dameon during recovery.
 184          * At this stage, the databases have already been locked, by a
 185          * dedicated child process. The freeze_mode variable is used to track
 186          * whether the actual locks are held by the child process or not.
 187          */
 188
 189         if (ctdb->freeze_mode[priority] != CTDB_FREEZE_FROZEN) {
 190                 DEBUG(DEBUG_ERR, ("Attempt to mark all databases locked when not frozen\n"));
 191                 return -1;
 192         }
 193
 194         return ctdb_db_iterator(ctdb, priority, db_lock_mark_handler, NULL);
 195 }
 196
 197 static int ctdb_lockall_mark(struct ctdb_context *ctdb)
 198 {
 199         uint32_t priority;
 200
 201         for (priority=1; priority<=NUM_DB_PRIORITIES; priority++) {
 202                 if (ctdb_db_iterator(ctdb, priority, db_lock_mark_handler, NULL) != 0) {
 203                         return -1;
 204                 }
 205         }
 206
 207         return 0;
 208 }
 209
 210
 211 /*
 212  * lock all databases - unmark only
 213  */
 214 static int db_lock_unmark_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 215                                   void *private_data)
 216 {
 217         int tdb_transaction_write_lock_unmark(struct tdb_context *);
 218
 219         DEBUG(DEBUG_INFO, ("unmarking locked database %s, priority:%u\n",
 220                            ctdb_db->db_name, priority));
 221
 222         if (tdb_transaction_write_lock_unmark(ctdb_db->ltdb->tdb) != 0) {
 223                 DEBUG(DEBUG_ERR, ("Failed to unmark (transaction lock) database %s\n",
 224                                   ctdb_db->db_name));
 225                 return -1;
 226         }
 227
 228         if (tdb_lockall_unmark(ctdb_db->ltdb->tdb) != 0) {
 229                 DEBUG(DEBUG_ERR, ("Failed to unmark (all lock) database %s\n",
 230                                   ctdb_db->db_name));
 231                 return -1;
 232         }
 233
 234         return 0;
 235 }
 236
 237 int ctdb_lockall_unmark_prio(struct ctdb_context *ctdb, uint32_t priority)
 238 {
 239         /*
 240          * This function is only used by the main daemon during recovery.
 241          * At this stage, the databases have already been locked, by a
 242          * dedicated child process. The freeze_mode variable is used to track
 243          * whether the actual locks are held by the child process or not.
 244          */
 245
 246         if (ctdb->freeze_mode[priority] != CTDB_FREEZE_FROZEN) {
 247                 DEBUG(DEBUG_ERR, ("Attempt to unmark all databases locked when not frozen\n"));
 248                 return -1;
 249         }
 250
 251         return ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL);
 252 }
 253
 254 static int ctdb_lockall_unmark(struct ctdb_context *ctdb)
 255 {
 256         uint32_t priority;
 257
 258         for (priority=NUM_DB_PRIORITIES; priority>0; priority--) {
 259                 if (ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL) != 0) {
 260                         return -1;
 261                 }
 262         }
 263
 264         return 0;
 265 }
 266
 267
 268 static void ctdb_lock_schedule(struct ctdb_context *ctdb);
 269
 270 /*
 271  * Destructor to kill the child locking process
 272  */
 273 static int ctdb_lock_context_destructor(struct lock_context *lock_ctx)
 274 {
 275         if (lock_ctx->request) {
 276                 lock_ctx->request->lctx = NULL;
 277         }
 278         if (lock_ctx->child > 0) {
 279                 ctdb_kill(lock_ctx->ctdb, lock_ctx->child, SIGKILL);
 280                 if (lock_ctx->type == LOCK_RECORD) {
 281                         DLIST_REMOVE(lock_ctx->ctdb_db->lock_current, lock_ctx);
 282                 } else {
 283                         DLIST_REMOVE(lock_ctx->ctdb->lock_current, lock_ctx);
 284                 }
 285                 if (lock_ctx->ctdb_db) {
 286                         lock_ctx->ctdb_db->lock_num_current--;
 287                 }
 288                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 289                 if (lock_ctx->ctdb_db) {
 290                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 291                 }
 292         } else {
 293                 if (lock_ctx->type == LOCK_RECORD) {
 294                         DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 295                 } else {
 296                         DLIST_REMOVE(lock_ctx->ctdb->lock_pending, lock_ctx);
 297                 }
 298                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 299                 if (lock_ctx->ctdb_db) {
 300                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 301                 }
 302         }
 303
 304         ctdb_lock_schedule(lock_ctx->ctdb);
 305
 306         return 0;
 307 }
 308
 309
 310 /*
 311  * Destructor to remove lock request
 312  */
 313 static int ctdb_lock_request_destructor(struct lock_request *lock_request)
 314 {
 315         TALLOC_FREE(lock_request->lctx);
 316         return 0;
 317 }
 318
 319 /*
 320  * Process all the callbacks waiting for lock
 321  *
 322  * If lock has failed, callback is executed with locked=false
 323  */
 324 static void process_callbacks(struct lock_context *lock_ctx, bool locked)
 325 {
 326         struct lock_request *request;
 327
 328         if (lock_ctx->auto_mark && locked) {
 329                 switch (lock_ctx->type) {
 330                 case LOCK_RECORD:
 331                         tdb_chainlock_mark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 332                         break;
 333
 334                 case LOCK_DB:
 335                         tdb_lockall_mark(lock_ctx->ctdb_db->ltdb->tdb);
 336                         break;
 337
 338                 case LOCK_ALLDB_PRIO:
 339                         ctdb_lockall_mark_prio(lock_ctx->ctdb, lock_ctx->priority);
 340                         break;
 341
 342                 case LOCK_ALLDB:
 343                         ctdb_lockall_mark(lock_ctx->ctdb);
 344                         break;
 345                 }
 346         }
 347
 348         request = lock_ctx->request;
 349         if (lock_ctx->auto_mark) {
 350                 /* Reset the destructor, so request is not removed from the list */
 351                 talloc_set_destructor(request, NULL);
 352         }
 353         request->callback(request->private_data, locked);
 354
 355         if (lock_ctx->auto_mark && locked) {
 356                 switch (lock_ctx->type) {
 357                 case LOCK_RECORD:
 358                         tdb_chainlock_unmark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 359                         break;
 360
 361                 case LOCK_DB:
 362                         tdb_lockall_unmark(lock_ctx->ctdb_db->ltdb->tdb);
 363                         break;
 364
 365                 case LOCK_ALLDB_PRIO:
 366                         ctdb_lockall_unmark_prio(lock_ctx->ctdb, lock_ctx->priority);
 367                         break;
 368
 369                 case LOCK_ALLDB:
 370                         ctdb_lockall_unmark(lock_ctx->ctdb);
 371                         break;
 372                 }
 373         }
 374 }
 375
 376
 377 static int lock_bucket_id(double t)
 378 {
 379         double ms = 1.e-3, s = 1;
 380         int id;
 381
 382         if (t < 1*ms) {
 383                 id = 0;
 384         } else if (t < 10*ms) {
 385                 id = 1;
 386         } else if (t < 100*ms) {
 387                 id = 2;
 388         } else if (t < 1*s) {
 389                 id = 3;
 390         } else if (t < 2*s) {
 391                 id = 4;
 392         } else if (t < 4*s) {
 393                 id = 5;
 394         } else if (t < 8*s) {
 395                 id = 6;
 396         } else if (t < 16*s) {
 397                 id = 7;
 398         } else if (t < 32*s) {
 399                 id = 8;
 400         } else if (t < 64*s) {
 401                 id = 9;
 402         } else {
 403                 id = 10;
 404         }
 405
 406         return id;
 407 }
 408
 409 /*
 410  * Callback routine when the required locks are obtained.
 411  * Called from parent context
 412  */
 413 static void ctdb_lock_handler(struct tevent_context *ev,
 414                             struct tevent_fd *tfd,
 415                             uint16_t flags,
 416                             void *private_data)
 417 {
 418         struct lock_context *lock_ctx;
 419         TALLOC_CTX *tmp_ctx = NULL;
 420         char c;
 421         bool locked;
 422         double t;
 423         int id;
 424
 425         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 426
 427         /* cancel the timeout event */
 428         TALLOC_FREE(lock_ctx->ttimer);
 429
 430         t = timeval_elapsed(&lock_ctx->start_time);
 431         id = lock_bucket_id(t);
 432
 433         if (lock_ctx->auto_mark) {
 434                 tmp_ctx = talloc_new(ev);
 435                 talloc_steal(tmp_ctx, lock_ctx);
 436         }
 437
 438         /* Read the status from the child process */
 439         if (sys_read(lock_ctx->fd[0], &c, 1) != 1) {
 440                 locked = false;
 441         } else {
 442                 locked = (c == 0 ? true : false);
 443         }
 444
 445         /* Update statistics */
 446         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_calls);
 447         if (lock_ctx->ctdb_db) {
 448                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_calls);
 449         }
 450
 451         if (locked) {
 452                 if (lock_ctx->ctdb_db) {
 453                         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.buckets[id]);
 454                         CTDB_UPDATE_LATENCY(lock_ctx->ctdb, lock_ctx->ctdb_db,
 455                                             lock_type_str[lock_ctx->type], locks.latency,
 456                                             lock_ctx->start_time);
 457
 458                         CTDB_UPDATE_DB_LATENCY(lock_ctx->ctdb_db, lock_type_str[lock_ctx->type], locks.latency, t);
 459                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.buckets[id]);
 460                 }
 461         } else {
 462                 CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_failed);
 463                 if (lock_ctx->ctdb_db) {
 464                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_failed);
 465                 }
 466         }
 467
 468         process_callbacks(lock_ctx, locked);
 469
 470         if (lock_ctx->auto_mark) {
 471                 talloc_free(tmp_ctx);
 472         }
 473 }
 474
 475
 476 /*
 477  * Callback routine when required locks are not obtained within timeout
 478  * Called from parent context
 479  */
 480 static void ctdb_lock_timeout_handler(struct tevent_context *ev,
 481                                     struct tevent_timer *ttimer,
 482                                     struct timeval current_time,
 483                                     void *private_data)
 484 {
 485         static const char * debug_locks = NULL;
 486         struct lock_context *lock_ctx;
 487         struct ctdb_context *ctdb;
 488         pid_t pid;
 489
 490         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 491         ctdb = lock_ctx->ctdb;
 492
 493         /* If a node stopped/banned, don't spam the logs */
 494         if (ctdb->nodes[ctdb->pnn]->flags & NODE_FLAGS_INACTIVE) {
 495                 lock_ctx->ttimer = NULL;
 496                 return;
 497         }
 498         if (lock_ctx->ctdb_db) {
 499                 DEBUG(DEBUG_WARNING,
 500                       ("Unable to get %s lock on database %s for %.0lf seconds\n",
 501                        (lock_ctx->type == LOCK_RECORD ? "RECORD" : "DB"),
 502                        lock_ctx->ctdb_db->db_name,
 503                        timeval_elapsed(&lock_ctx->start_time)));
 504         } else {
 505                 DEBUG(DEBUG_WARNING,
 506                       ("Unable to get ALLDB locks for %.0lf seconds\n",
 507                        timeval_elapsed(&lock_ctx->start_time)));
 508         }
 509
 510         /* Fire a child process to find the blocking process. */
 511         if (debug_locks == NULL) {
 512                 debug_locks = getenv("CTDB_DEBUG_LOCKS");
 513                 if (debug_locks == NULL) {
 514                         debug_locks = talloc_asprintf(ctdb,
 515                                                       "%s/debug_locks.sh",
 516                                                       getenv("CTDB_BASE"));
 517                 }
 518         }
 519         if (debug_locks != NULL) {
 520                 pid = vfork();
 521                 if (pid == 0) {
 522                         execl(debug_locks, debug_locks, NULL);
 523                         _exit(0);
 524                 }
 525                 ctdb_track_child(ctdb, pid);
 526         } else {
 527                 DEBUG(DEBUG_WARNING,
 528                       (__location__
 529                        " Unable to setup lock debugging - no memory?\n"));
 530         }
 531
 532         /* reset the timeout timer */
 533         // talloc_free(lock_ctx->ttimer);
 534         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 535                                             lock_ctx,
 536                                             timeval_current_ofs(10, 0),
 537                                             ctdb_lock_timeout_handler,
 538                                             (void *)lock_ctx);
 539 }
 540
 541
 542 static int db_count_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 543                             void *private_data)
 544 {
 545         int *count = (int *)private_data;
 546
 547         (*count)++;
 548
 549         return 0;
 550 }
 551
 552 struct db_namelist {
 553         const char **names;
 554         int n;
 555 };
 556
 557 static int db_name_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 558                            void *private_data)
 559 {
 560         struct db_namelist *list = (struct db_namelist *)private_data;
 561
 562         list->names[list->n] = talloc_strdup(list->names, ctdb_db->db_path);
 563         list->n++;
 564
 565         return 0;
 566 }
 567
 568 static bool lock_helper_args(TALLOC_CTX *mem_ctx,
 569                              struct lock_context *lock_ctx, int fd,
 570                              int *argc, const char ***argv)
 571 {
 572         struct ctdb_context *ctdb = lock_ctx->ctdb;
 573         const char **args = NULL;
 574         int nargs, i;
 575         int priority;
 576         struct db_namelist list;
 577
 578         switch (lock_ctx->type) {
 579         case LOCK_RECORD:
 580                 nargs = 5;
 581                 break;
 582
 583         case LOCK_DB:
 584                 nargs = 4;
 585                 break;
 586
 587         case LOCK_ALLDB_PRIO:
 588                 nargs = 3;
 589                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_count_handler, &nargs);
 590                 break;
 591
 592         case LOCK_ALLDB:
 593                 nargs = 3;
 594                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 595                         ctdb_db_iterator(ctdb, priority, db_count_handler, &nargs);
 596                 }
 597                 break;
 598         }
 599
 600         /* Add extra argument for null termination */
 601         nargs++;
 602
 603         args = talloc_array(mem_ctx, const char *, nargs);
 604         if (args == NULL) {
 605                 return false;
 606         }
 607
 608         args[0] = talloc_asprintf(args, "%d", getpid());
 609         args[1] = talloc_asprintf(args, "%d", fd);
 610
 611         switch (lock_ctx->type) {
 612         case LOCK_RECORD:
 613                 args[2] = talloc_strdup(args, "RECORD");
 614                 args[3] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 615                 if (lock_ctx->key.dsize == 0) {
 616                         args[4] = talloc_strdup(args, "NULL");
 617                 } else {
 618                         args[4] = hex_encode_talloc(args, lock_ctx->key.dptr, lock_ctx->key.dsize);
 619                 }
 620                 break;
 621
 622         case LOCK_DB:
 623                 args[2] = talloc_strdup(args, "DB");
 624                 args[3] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 625                 break;
 626
 627         case LOCK_ALLDB_PRIO:
 628                 args[2] = talloc_strdup(args, "DB");
 629                 list.names = args;
 630                 list.n = 3;
 631                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_name_handler, &list);
 632                 break;
 633
 634         case LOCK_ALLDB:
 635                 args[2] = talloc_strdup(args, "DB");
 636                 list.names = args;
 637                 list.n = 3;
 638                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 639                         ctdb_db_iterator(ctdb, priority, db_name_handler, &list);
 640                 }
 641                 break;
 642         }
 643
 644         /* Make sure last argument is NULL */
 645         args[nargs-1] = NULL;
 646
 647         for (i=0; i<nargs-1; i++) {
 648                 if (args[i] == NULL) {
 649                         talloc_free(args);
 650                         return false;
 651                 }
 652         }
 653
 654         *argc = nargs;
 655         *argv = args;
 656         return true;
 657 }
 658
 659 /*
 660  * Find a lock request that can be scheduled
 661  */
 662 static struct lock_context *ctdb_find_lock_context(struct ctdb_context *ctdb)
 663 {
 664         struct lock_context *lock_ctx, *next_ctx;
 665         struct ctdb_db_context *ctdb_db;
 666
 667         /* First check if there are database lock requests */
 668
 669         for (lock_ctx = ctdb->lock_pending; lock_ctx != NULL;
 670              lock_ctx = next_ctx) {
 671
 672                 if (lock_ctx->request != NULL) {
 673                         /* Found a lock context with a request */
 674                         return lock_ctx;
 675                 }
 676
 677                 next_ctx = lock_ctx->next;
 678
 679                 DEBUG(DEBUG_INFO, ("Removing lock context without lock "
 680                                    "request\n"));
 681                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 682                 CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 683                 if (lock_ctx->ctdb_db) {
 684                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db,
 685                                                locks.num_pending);
 686                 }
 687                 talloc_free(lock_ctx);
 688         }
 689
 690         /* Next check database queues */
 691         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 692                 if (ctdb_db->lock_num_current ==
 693                     ctdb->tunable.lock_processes_per_db) {
 694                         continue;
 695                 }
 696
 697                 for (lock_ctx = ctdb_db->lock_pending; lock_ctx != NULL;
 698                      lock_ctx = next_ctx) {
 699
 700                         next_ctx = lock_ctx->next;
 701
 702                         if (lock_ctx->request != NULL) {
 703                                 return lock_ctx;
 704                         }
 705
 706                         DEBUG(DEBUG_INFO, ("Removing lock context without "
 707                                            "lock request\n"));
 708                         DLIST_REMOVE(ctdb_db->lock_pending, lock_ctx);
 709                         CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 710                         CTDB_DECREMENT_DB_STAT(ctdb_db, locks.num_pending);
 711                         talloc_free(lock_ctx);
 712                 }
 713         }
 714
 715         return NULL;
 716 }
 717
 718 /*
 719  * Schedule a new lock child process
 720  * Set up callback handler and timeout handler
 721  */
 722 static void ctdb_lock_schedule(struct ctdb_context *ctdb)
 723 {
 724         struct lock_context *lock_ctx;
 725         int ret, argc;
 726         TALLOC_CTX *tmp_ctx;
 727         const char *helper = CTDB_HELPER_BINDIR "/ctdb_lock_helper";
 728         static const char *prog = NULL;
 729         const char **args;
 730
 731         if (prog == NULL) {
 732                 const char *t;
 733
 734                 t = getenv("CTDB_LOCK_HELPER");
 735                 if (t != NULL) {
 736                         prog = talloc_strdup(ctdb, t);
 737                 } else {
 738                         prog = talloc_strdup(ctdb, helper);
 739                 }
 740                 CTDB_NO_MEMORY_VOID(ctdb, prog);
 741         }
 742
 743         /* Find a lock context with requests */
 744         lock_ctx = ctdb_find_lock_context(ctdb);
 745         if (lock_ctx == NULL) {
 746                 return;
 747         }
 748
 749         lock_ctx->child = -1;
 750         ret = pipe(lock_ctx->fd);
 751         if (ret != 0) {
 752                 DEBUG(DEBUG_ERR, ("Failed to create pipe in ctdb_lock_schedule\n"));
 753                 return;
 754         }
 755
 756         set_close_on_exec(lock_ctx->fd[0]);
 757
 758         /* Create data for child process */
 759         tmp_ctx = talloc_new(lock_ctx);
 760         if (tmp_ctx == NULL) {
 761                 DEBUG(DEBUG_ERR, ("Failed to allocate memory for helper args\n"));
 762                 close(lock_ctx->fd[0]);
 763                 close(lock_ctx->fd[1]);
 764                 return;
 765         }
 766
 767         /* Create arguments for lock helper */
 768         if (!lock_helper_args(tmp_ctx, lock_ctx, lock_ctx->fd[1],
 769                               &argc, &args)) {
 770                 DEBUG(DEBUG_ERR, ("Failed to create lock helper args\n"));
 771                 close(lock_ctx->fd[0]);
 772                 close(lock_ctx->fd[1]);
 773                 talloc_free(tmp_ctx);
 774                 return;
 775         }
 776
 777         if (!ctdb_vfork_with_logging(lock_ctx, ctdb, "lock_helper",
 778                                      prog, argc, (const char **)args,
 779                                      NULL, NULL, &lock_ctx->child)) {
 780                 DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
 781                 close(lock_ctx->fd[0]);
 782                 close(lock_ctx->fd[1]);
 783                 talloc_free(tmp_ctx);
 784                 return;
 785         }
 786
 787         /* Parent process */
 788         close(lock_ctx->fd[1]);
 789
 790         talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
 791
 792         talloc_free(tmp_ctx);
 793
 794         /* Set up timeout handler */
 795         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 796                                             lock_ctx,
 797                                             timeval_current_ofs(10, 0),
 798                                             ctdb_lock_timeout_handler,
 799                                             (void *)lock_ctx);
 800         if (lock_ctx->ttimer == NULL) {
 801                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 802                 lock_ctx->child = -1;
 803                 talloc_set_destructor(lock_ctx, NULL);
 804                 close(lock_ctx->fd[0]);
 805                 return;
 806         }
 807
 808         /* Set up callback */
 809         lock_ctx->tfd = tevent_add_fd(ctdb->ev,
 810                                       lock_ctx,
 811                                       lock_ctx->fd[0],
 812                                       EVENT_FD_READ,
 813                                       ctdb_lock_handler,
 814                                       (void *)lock_ctx);
 815         if (lock_ctx->tfd == NULL) {
 816                 TALLOC_FREE(lock_ctx->ttimer);
 817                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 818                 lock_ctx->child = -1;
 819                 talloc_set_destructor(lock_ctx, NULL);
 820                 close(lock_ctx->fd[0]);
 821                 return;
 822         }
 823         tevent_fd_set_auto_close(lock_ctx->tfd);
 824
 825         /* Move the context from pending to current */
 826         if (lock_ctx->type == LOCK_RECORD) {
 827                 DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 828                 DLIST_ADD_END(lock_ctx->ctdb_db->lock_current, lock_ctx, NULL);
 829         } else {
 830                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 831                 DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
 832         }
 833         CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 834         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 835         if (lock_ctx->ctdb_db) {
 836                 lock_ctx->ctdb_db->lock_num_current++;
 837                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 838                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 839         }
 840 }
 841
 842
 843 /*
 844  * Lock record / db depending on type
 845  */
 846 static struct lock_request *ctdb_lock_internal(TALLOC_CTX *mem_ctx,
 847                                                struct ctdb_context *ctdb,
 848                                                struct ctdb_db_context *ctdb_db,
 849                                                TDB_DATA key,
 850                                                uint32_t priority,
 851                                                void (*callback)(void *, bool),
 852                                                void *private_data,
 853                                                enum lock_type type,
 854                                                bool auto_mark)
 855 {
 856         struct lock_context *lock_ctx = NULL;
 857         struct lock_request *request;
 858
 859         if (callback == NULL) {
 860                 DEBUG(DEBUG_WARNING, ("No callback function specified, not locking\n"));
 861                 return NULL;
 862         }
 863
 864         lock_ctx = talloc_zero(ctdb, struct lock_context);
 865         if (lock_ctx == NULL) {
 866                 DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
 867                 return NULL;
 868         }
 869
 870         if ((request = talloc_zero(mem_ctx, struct lock_request)) == NULL) {
 871                 talloc_free(lock_ctx);
 872                 return NULL;
 873         }
 874
 875         lock_ctx->type = type;
 876         lock_ctx->ctdb = ctdb;
 877         lock_ctx->ctdb_db = ctdb_db;
 878         lock_ctx->key.dsize = key.dsize;
 879         if (key.dsize > 0) {
 880                 lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
 881                 if (lock_ctx->key.dptr == NULL) {
 882                         DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
 883                         talloc_free(lock_ctx);
 884                         return NULL;
 885                 }
 886                 lock_ctx->key_hash = ctdb_hash(&key);
 887         } else {
 888                 lock_ctx->key.dptr = NULL;
 889         }
 890         lock_ctx->priority = priority;
 891         lock_ctx->auto_mark = auto_mark;
 892
 893         lock_ctx->request = request;
 894         lock_ctx->child = -1;
 895
 896         /* Non-record locks are required by recovery and should be scheduled
 897          * immediately, so keep them at the head of the pending queue.
 898          */
 899         if (lock_ctx->type == LOCK_RECORD) {
 900                 DLIST_ADD_END(ctdb_db->lock_pending, lock_ctx, NULL);
 901         } else {
 902                 DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
 903         }
 904         CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
 905         if (ctdb_db) {
 906                 CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
 907         }
 908
 909         /* Start the timer when we activate the context */
 910         lock_ctx->start_time = timeval_current();
 911
 912         request->lctx = lock_ctx;
 913         request->callback = callback;
 914         request->private_data = private_data;
 915
 916         talloc_set_destructor(request, ctdb_lock_request_destructor);
 917
 918         ctdb_lock_schedule(ctdb);
 919
 920         return request;
 921 }
 922
 923
 924 /*
 925  * obtain a lock on a record in a database
 926  */
 927 struct lock_request *ctdb_lock_record(TALLOC_CTX *mem_ctx,
 928                                       struct ctdb_db_context *ctdb_db,
 929                                       TDB_DATA key,
 930                                       bool auto_mark,
 931                                       void (*callback)(void *, bool),
 932                                       void *private_data)
 933 {
 934         return ctdb_lock_internal(mem_ctx,
 935                                   ctdb_db->ctdb,
 936                                   ctdb_db,
 937                                   key,
 938                                   0,
 939                                   callback,
 940                                   private_data,
 941                                   LOCK_RECORD,
 942                                   auto_mark);
 943 }
 944
 945
 946 /*
 947  * obtain a lock on a database
 948  */
 949 struct lock_request *ctdb_lock_db(TALLOC_CTX *mem_ctx,
 950                                   struct ctdb_db_context *ctdb_db,
 951                                   bool auto_mark,
 952                                   void (*callback)(void *, bool),
 953                                   void *private_data)
 954 {
 955         return ctdb_lock_internal(mem_ctx,
 956                                   ctdb_db->ctdb,
 957                                   ctdb_db,
 958                                   tdb_null,
 959                                   0,
 960                                   callback,
 961                                   private_data,
 962                                   LOCK_DB,
 963                                   auto_mark);
 964 }
 965
 966
 967 /*
 968  * obtain locks on all databases of specified priority
 969  */
 970 struct lock_request *ctdb_lock_alldb_prio(TALLOC_CTX *mem_ctx,
 971                                           struct ctdb_context *ctdb,
 972                                           uint32_t priority,
 973                                           bool auto_mark,
 974                                           void (*callback)(void *, bool),
 975                                           void *private_data)
 976 {
 977         if (priority < 1 || priority > NUM_DB_PRIORITIES) {
 978                 DEBUG(DEBUG_ERR, ("Invalid db priority: %u\n", priority));
 979                 return NULL;
 980         }
 981
 982         return ctdb_lock_internal(mem_ctx,
 983                                   ctdb,
 984                                   NULL,
 985                                   tdb_null,
 986                                   priority,
 987                                   callback,
 988                                   private_data,
 989                                   LOCK_ALLDB_PRIO,
 990                                   auto_mark);
 991 }
 992
 993
 994 /*
 995  * obtain locks on all databases
 996  */
 997 struct lock_request *ctdb_lock_alldb(TALLOC_CTX *mem_ctx,
 998                                      struct ctdb_context *ctdb,
 999                                      bool auto_mark,
1000                                      void (*callback)(void *, bool),
1001                                      void *private_data)
1002 {
1003         return ctdb_lock_internal(mem_ctx,
1004                                   ctdb,
1005                                   NULL,
1006                                   tdb_null,
1007                                   0,
1008                                   callback,
1009                                   private_data,
1010                                   LOCK_ALLDB,
1011                                   auto_mark);
1012 }
1013