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                 return;
 496         }
 497         if (lock_ctx->ctdb_db) {
 498                 DEBUG(DEBUG_WARNING,
 499                       ("Unable to get %s lock on database %s for %.0lf seconds\n",
 500                        (lock_ctx->type == LOCK_RECORD ? "RECORD" : "DB"),
 501                        lock_ctx->ctdb_db->db_name,
 502                        timeval_elapsed(&lock_ctx->start_time)));
 503         } else {
 504                 DEBUG(DEBUG_WARNING,
 505                       ("Unable to get ALLDB locks for %.0lf seconds\n",
 506                        timeval_elapsed(&lock_ctx->start_time)));
 507         }
 508
 509         /* Fire a child process to find the blocking process. */
 510         if (debug_locks == NULL) {
 511                 debug_locks = getenv("CTDB_DEBUG_LOCKS");
 512                 if (debug_locks == NULL) {
 513                         debug_locks = talloc_asprintf(ctdb,
 514                                                       "%s/debug_locks.sh",
 515                                                       getenv("CTDB_BASE"));
 516                 }
 517         }
 518         if (debug_locks != NULL) {
 519                 pid = vfork();
 520                 if (pid == 0) {
 521                         execl(debug_locks, debug_locks, NULL);
 522                         _exit(0);
 523                 }
 524                 ctdb_track_child(ctdb, pid);
 525         } else {
 526                 DEBUG(DEBUG_WARNING,
 527                       (__location__
 528                        " Unable to setup lock debugging - no memory?\n"));
 529         }
 530
 531         /* reset the timeout timer */
 532         // talloc_free(lock_ctx->ttimer);
 533         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 534                                             lock_ctx,
 535                                             timeval_current_ofs(10, 0),
 536                                             ctdb_lock_timeout_handler,
 537                                             (void *)lock_ctx);
 538 }
 539
 540
 541 static int db_count_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 542                             void *private_data)
 543 {
 544         int *count = (int *)private_data;
 545
 546         (*count)++;
 547
 548         return 0;
 549 }
 550
 551 struct db_namelist {
 552         char **names;
 553         int n;
 554 };
 555
 556 static int db_name_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 557                            void *private_data)
 558 {
 559         struct db_namelist *list = (struct db_namelist *)private_data;
 560
 561         list->names[list->n] = talloc_strdup(list->names, ctdb_db->db_path);
 562         list->n++;
 563
 564         return 0;
 565 }
 566
 567 static char **lock_helper_args(TALLOC_CTX *mem_ctx, struct lock_context *lock_ctx, int fd)
 568 {
 569         struct ctdb_context *ctdb = lock_ctx->ctdb;
 570         char **args = NULL;
 571         int nargs, i;
 572         int priority;
 573         struct db_namelist list;
 574
 575         switch (lock_ctx->type) {
 576         case LOCK_RECORD:
 577                 nargs = 6;
 578                 break;
 579
 580         case LOCK_DB:
 581                 nargs = 5;
 582                 break;
 583
 584         case LOCK_ALLDB_PRIO:
 585                 nargs = 4;
 586                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_count_handler, &nargs);
 587                 break;
 588
 589         case LOCK_ALLDB:
 590                 nargs = 4;
 591                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 592                         ctdb_db_iterator(ctdb, priority, db_count_handler, &nargs);
 593                 }
 594                 break;
 595         }
 596
 597         /* Add extra argument for null termination */
 598         nargs++;
 599
 600         args = talloc_array(mem_ctx, char *, nargs);
 601         if (args == NULL) {
 602                 return NULL;
 603         }
 604
 605         args[0] = talloc_strdup(args, "ctdb_lock_helper");
 606         args[1] = talloc_asprintf(args, "%d", getpid());
 607         args[2] = talloc_asprintf(args, "%d", fd);
 608
 609         switch (lock_ctx->type) {
 610         case LOCK_RECORD:
 611                 args[3] = talloc_strdup(args, "RECORD");
 612                 args[4] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 613                 if (lock_ctx->key.dsize == 0) {
 614                         args[5] = talloc_strdup(args, "NULL");
 615                 } else {
 616                         args[5] = hex_encode_talloc(args, lock_ctx->key.dptr, lock_ctx->key.dsize);
 617                 }
 618                 break;
 619
 620         case LOCK_DB:
 621                 args[3] = talloc_strdup(args, "DB");
 622                 args[4] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 623                 break;
 624
 625         case LOCK_ALLDB_PRIO:
 626                 args[3] = talloc_strdup(args, "DB");
 627                 list.names = args;
 628                 list.n = 4;
 629                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_name_handler, &list);
 630                 break;
 631
 632         case LOCK_ALLDB:
 633                 args[3] = talloc_strdup(args, "DB");
 634                 list.names = args;
 635                 list.n = 4;
 636                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 637                         ctdb_db_iterator(ctdb, priority, db_name_handler, &list);
 638                 }
 639                 break;
 640         }
 641
 642         /* Make sure last argument is NULL */
 643         args[nargs-1] = NULL;
 644
 645         for (i=0; i<nargs-1; i++) {
 646                 if (args[i] == NULL) {
 647                         talloc_free(args);
 648                         return NULL;
 649                 }
 650         }
 651
 652         return args;
 653 }
 654
 655 /*
 656  * Find a lock request that can be scheduled
 657  */
 658 static struct lock_context *ctdb_find_lock_context(struct ctdb_context *ctdb)
 659 {
 660         struct lock_context *lock_ctx, *next_ctx;
 661         struct ctdb_db_context *ctdb_db;
 662
 663         /* First check if there are database lock requests */
 664
 665         for (lock_ctx = ctdb->lock_pending; lock_ctx != NULL;
 666              lock_ctx = next_ctx) {
 667
 668                 if (lock_ctx->request != NULL) {
 669                         /* Found a lock context with a request */
 670                         return lock_ctx;
 671                 }
 672
 673                 next_ctx = lock_ctx->next;
 674
 675                 DEBUG(DEBUG_INFO, ("Removing lock context without lock "
 676                                    "request\n"));
 677                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 678                 CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 679                 if (lock_ctx->ctdb_db) {
 680                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db,
 681                                                locks.num_pending);
 682                 }
 683                 talloc_free(lock_ctx);
 684         }
 685
 686         /* Next check database queues */
 687         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 688                 if (ctdb_db->lock_num_current ==
 689                     ctdb->tunable.lock_processes_per_db) {
 690                         continue;
 691                 }
 692
 693                 for (lock_ctx = ctdb_db->lock_pending; lock_ctx != NULL;
 694                      lock_ctx = next_ctx) {
 695
 696                         next_ctx = lock_ctx->next;
 697
 698                         if (lock_ctx->request != NULL) {
 699                                 return lock_ctx;
 700                         }
 701
 702                         DEBUG(DEBUG_INFO, ("Removing lock context without "
 703                                            "lock request\n"));
 704                         DLIST_REMOVE(ctdb_db->lock_pending, lock_ctx);
 705                         CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 706                         CTDB_DECREMENT_DB_STAT(ctdb_db, locks.num_pending);
 707                         talloc_free(lock_ctx);
 708                 }
 709         }
 710
 711         return NULL;
 712 }
 713
 714 /*
 715  * Schedule a new lock child process
 716  * Set up callback handler and timeout handler
 717  */
 718 static void ctdb_lock_schedule(struct ctdb_context *ctdb)
 719 {
 720         struct lock_context *lock_ctx;
 721         int ret;
 722         TALLOC_CTX *tmp_ctx;
 723         const char *helper = BINDIR "/ctdb_lock_helper";
 724         static const char *prog = NULL;
 725         char **args;
 726
 727         if (prog == NULL) {
 728                 const char *t;
 729
 730                 t = getenv("CTDB_LOCK_HELPER");
 731                 if (t != NULL) {
 732                         prog = talloc_strdup(ctdb, t);
 733                 } else {
 734                         prog = talloc_strdup(ctdb, helper);
 735                 }
 736                 CTDB_NO_MEMORY_VOID(ctdb, prog);
 737         }
 738
 739         /* Find a lock context with requests */
 740         lock_ctx = ctdb_find_lock_context(ctdb);
 741         if (lock_ctx == NULL) {
 742                 return;
 743         }
 744
 745         lock_ctx->child = -1;
 746         ret = pipe(lock_ctx->fd);
 747         if (ret != 0) {
 748                 DEBUG(DEBUG_ERR, ("Failed to create pipe in ctdb_lock_schedule\n"));
 749                 return;
 750         }
 751
 752         set_close_on_exec(lock_ctx->fd[0]);
 753
 754         /* Create data for child process */
 755         tmp_ctx = talloc_new(lock_ctx);
 756         if (tmp_ctx == NULL) {
 757                 DEBUG(DEBUG_ERR, ("Failed to allocate memory for helper args\n"));
 758                 close(lock_ctx->fd[0]);
 759                 close(lock_ctx->fd[1]);
 760                 return;
 761         }
 762
 763         /* Create arguments for lock helper */
 764         args = lock_helper_args(tmp_ctx, lock_ctx, lock_ctx->fd[1]);
 765         if (args == NULL) {
 766                 DEBUG(DEBUG_ERR, ("Failed to create lock helper args\n"));
 767                 close(lock_ctx->fd[0]);
 768                 close(lock_ctx->fd[1]);
 769                 talloc_free(tmp_ctx);
 770                 return;
 771         }
 772
 773         lock_ctx->child = vfork();
 774
 775         if (lock_ctx->child == (pid_t)-1) {
 776                 DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
 777                 close(lock_ctx->fd[0]);
 778                 close(lock_ctx->fd[1]);
 779                 talloc_free(tmp_ctx);
 780                 return;
 781         }
 782
 783
 784         /* Child process */
 785         if (lock_ctx->child == 0) {
 786                 ret = execv(prog, args);
 787                 if (ret < 0) {
 788                         DEBUG(DEBUG_ERR, ("Failed to execute helper %s (%d, %s)\n",
 789                                           prog, errno, strerror(errno)));
 790                 }
 791                 _exit(1);
 792         }
 793
 794         /* Parent process */
 795         ctdb_track_child(ctdb, lock_ctx->child);
 796         close(lock_ctx->fd[1]);
 797
 798         talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
 799
 800         talloc_free(tmp_ctx);
 801
 802         /* Set up timeout handler */
 803         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 804                                             lock_ctx,
 805                                             timeval_current_ofs(10, 0),
 806                                             ctdb_lock_timeout_handler,
 807                                             (void *)lock_ctx);
 808         if (lock_ctx->ttimer == NULL) {
 809                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 810                 lock_ctx->child = -1;
 811                 talloc_set_destructor(lock_ctx, NULL);
 812                 close(lock_ctx->fd[0]);
 813                 return;
 814         }
 815
 816         /* Set up callback */
 817         lock_ctx->tfd = tevent_add_fd(ctdb->ev,
 818                                       lock_ctx,
 819                                       lock_ctx->fd[0],
 820                                       EVENT_FD_READ,
 821                                       ctdb_lock_handler,
 822                                       (void *)lock_ctx);
 823         if (lock_ctx->tfd == NULL) {
 824                 TALLOC_FREE(lock_ctx->ttimer);
 825                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 826                 lock_ctx->child = -1;
 827                 talloc_set_destructor(lock_ctx, NULL);
 828                 close(lock_ctx->fd[0]);
 829                 return;
 830         }
 831         tevent_fd_set_auto_close(lock_ctx->tfd);
 832
 833         /* Move the context from pending to current */
 834         if (lock_ctx->type == LOCK_RECORD) {
 835                 DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 836                 DLIST_ADD_END(lock_ctx->ctdb_db->lock_current, lock_ctx, NULL);
 837         } else {
 838                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 839                 DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
 840         }
 841         CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 842         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 843         if (lock_ctx->ctdb_db) {
 844                 lock_ctx->ctdb_db->lock_num_current++;
 845                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 846                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 847         }
 848 }
 849
 850
 851 /*
 852  * Lock record / db depending on type
 853  */
 854 static struct lock_request *ctdb_lock_internal(TALLOC_CTX *mem_ctx,
 855                                                struct ctdb_context *ctdb,
 856                                                struct ctdb_db_context *ctdb_db,
 857                                                TDB_DATA key,
 858                                                uint32_t priority,
 859                                                void (*callback)(void *, bool),
 860                                                void *private_data,
 861                                                enum lock_type type,
 862                                                bool auto_mark)
 863 {
 864         struct lock_context *lock_ctx = NULL;
 865         struct lock_request *request;
 866
 867         if (callback == NULL) {
 868                 DEBUG(DEBUG_WARNING, ("No callback function specified, not locking\n"));
 869                 return NULL;
 870         }
 871
 872         lock_ctx = talloc_zero(ctdb, struct lock_context);
 873         if (lock_ctx == NULL) {
 874                 DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
 875                 return NULL;
 876         }
 877
 878         if ((request = talloc_zero(mem_ctx, struct lock_request)) == NULL) {
 879                 talloc_free(lock_ctx);
 880                 return NULL;
 881         }
 882
 883         lock_ctx->type = type;
 884         lock_ctx->ctdb = ctdb;
 885         lock_ctx->ctdb_db = ctdb_db;
 886         lock_ctx->key.dsize = key.dsize;
 887         if (key.dsize > 0) {
 888                 lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
 889                 if (lock_ctx->key.dptr == NULL) {
 890                         DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
 891                         talloc_free(lock_ctx);
 892                         return NULL;
 893                 }
 894                 lock_ctx->key_hash = ctdb_hash(&key);
 895         } else {
 896                 lock_ctx->key.dptr = NULL;
 897         }
 898         lock_ctx->priority = priority;
 899         lock_ctx->auto_mark = auto_mark;
 900
 901         lock_ctx->request = request;
 902         lock_ctx->child = -1;
 903
 904         /* Non-record locks are required by recovery and should be scheduled
 905          * immediately, so keep them at the head of the pending queue.
 906          */
 907         if (lock_ctx->type == LOCK_RECORD) {
 908                 DLIST_ADD_END(ctdb_db->lock_pending, lock_ctx, NULL);
 909         } else {
 910                 DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
 911         }
 912         CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
 913         if (ctdb_db) {
 914                 CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
 915         }
 916
 917         /* Start the timer when we activate the context */
 918         lock_ctx->start_time = timeval_current();
 919
 920         request->lctx = lock_ctx;
 921         request->callback = callback;
 922         request->private_data = private_data;
 923
 924         talloc_set_destructor(request, ctdb_lock_request_destructor);
 925
 926         ctdb_lock_schedule(ctdb);
 927
 928         return request;
 929 }
 930
 931
 932 /*
 933  * obtain a lock on a record in a database
 934  */
 935 struct lock_request *ctdb_lock_record(TALLOC_CTX *mem_ctx,
 936                                       struct ctdb_db_context *ctdb_db,
 937                                       TDB_DATA key,
 938                                       bool auto_mark,
 939                                       void (*callback)(void *, bool),
 940                                       void *private_data)
 941 {
 942         return ctdb_lock_internal(mem_ctx,
 943                                   ctdb_db->ctdb,
 944                                   ctdb_db,
 945                                   key,
 946                                   0,
 947                                   callback,
 948                                   private_data,
 949                                   LOCK_RECORD,
 950                                   auto_mark);
 951 }
 952
 953
 954 /*
 955  * obtain a lock on a database
 956  */
 957 struct lock_request *ctdb_lock_db(TALLOC_CTX *mem_ctx,
 958                                   struct ctdb_db_context *ctdb_db,
 959                                   bool auto_mark,
 960                                   void (*callback)(void *, bool),
 961                                   void *private_data)
 962 {
 963         return ctdb_lock_internal(mem_ctx,
 964                                   ctdb_db->ctdb,
 965                                   ctdb_db,
 966                                   tdb_null,
 967                                   0,
 968                                   callback,
 969                                   private_data,
 970                                   LOCK_DB,
 971                                   auto_mark);
 972 }
 973
 974
 975 /*
 976  * obtain locks on all databases of specified priority
 977  */
 978 struct lock_request *ctdb_lock_alldb_prio(TALLOC_CTX *mem_ctx,
 979                                           struct ctdb_context *ctdb,
 980                                           uint32_t priority,
 981                                           bool auto_mark,
 982                                           void (*callback)(void *, bool),
 983                                           void *private_data)
 984 {
 985         if (priority < 1 || priority > NUM_DB_PRIORITIES) {
 986                 DEBUG(DEBUG_ERR, ("Invalid db priority: %u\n", priority));
 987                 return NULL;
 988         }
 989
 990         return ctdb_lock_internal(mem_ctx,
 991                                   ctdb,
 992                                   NULL,
 993                                   tdb_null,
 994                                   priority,
 995                                   callback,
 996                                   private_data,
 997                                   LOCK_ALLDB_PRIO,
 998                                   auto_mark);
 999 }
1000
1001
1002 /*
1003  * obtain locks on all databases
1004  */
1005 struct lock_request *ctdb_lock_alldb(TALLOC_CTX *mem_ctx,
1006                                      struct ctdb_context *ctdb,
1007                                      bool auto_mark,
1008                                      void (*callback)(void *, bool),
1009                                      void *private_data)
1010 {
1011         return ctdb_lock_internal(mem_ctx,
1012                                   ctdb,
1013                                   NULL,
1014                                   tdb_null,
1015                                   0,
1016                                   callback,
1017                                   private_data,
1018                                   LOCK_ALLDB,
1019                                   auto_mark);
1020 }
1021