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 "db_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->child > 0) {
 276                 ctdb_kill(lock_ctx->ctdb, lock_ctx->child, SIGKILL);
 277                 DLIST_REMOVE(lock_ctx->ctdb->lock_current, lock_ctx);
 278                 if (lock_ctx->ctdb_db) {
 279                         lock_ctx->ctdb_db->lock_num_current--;
 280                 }
 281                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 282                 if (lock_ctx->ctdb_db) {
 283                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 284                 }
 285         } else {
 286                 DLIST_REMOVE(lock_ctx->ctdb->lock_pending, lock_ctx);
 287                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 288                 if (lock_ctx->ctdb_db) {
 289                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 290                 }
 291         }
 292
 293         ctdb_lock_schedule(lock_ctx->ctdb);
 294
 295         return 0;
 296 }
 297
 298
 299 /*
 300  * Destructor to remove lock request
 301  */
 302 static int ctdb_lock_request_destructor(struct lock_request *lock_request)
 303 {
 304         lock_request->lctx->request = NULL;
 305         return 0;
 306 }
 307
 308 void ctdb_lock_free_request_context(struct lock_request *lock_req)
 309 {
 310         struct lock_context *lock_ctx;
 311
 312         lock_ctx = lock_req->lctx;
 313         talloc_free(lock_req);
 314         talloc_free(lock_ctx);
 315 }
 316
 317
 318 /*
 319  * Process all the callbacks waiting for lock
 320  *
 321  * If lock has failed, callback is executed with locked=false
 322  */
 323 static void process_callbacks(struct lock_context *lock_ctx, bool locked)
 324 {
 325         struct lock_request *request;
 326
 327         if (lock_ctx->auto_mark && locked) {
 328                 switch (lock_ctx->type) {
 329                 case LOCK_RECORD:
 330                         tdb_chainlock_mark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 331                         break;
 332
 333                 case LOCK_DB:
 334                         tdb_lockall_mark(lock_ctx->ctdb_db->ltdb->tdb);
 335                         break;
 336
 337                 case LOCK_ALLDB_PRIO:
 338                         ctdb_lockall_mark_prio(lock_ctx->ctdb, lock_ctx->priority);
 339                         break;
 340
 341                 case LOCK_ALLDB:
 342                         ctdb_lockall_mark(lock_ctx->ctdb);
 343                         break;
 344                 }
 345         }
 346
 347         request = lock_ctx->request;
 348         if (lock_ctx->auto_mark) {
 349                 /* Reset the destructor, so request is not removed from the list */
 350                 talloc_set_destructor(request, NULL);
 351         }
 352         request->callback(request->private_data, locked);
 353
 354         if (lock_ctx->auto_mark && locked) {
 355                 switch (lock_ctx->type) {
 356                 case LOCK_RECORD:
 357                         tdb_chainlock_unmark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 358                         break;
 359
 360                 case LOCK_DB:
 361                         tdb_lockall_unmark(lock_ctx->ctdb_db->ltdb->tdb);
 362                         break;
 363
 364                 case LOCK_ALLDB_PRIO:
 365                         ctdb_lockall_unmark_prio(lock_ctx->ctdb, lock_ctx->priority);
 366                         break;
 367
 368                 case LOCK_ALLDB:
 369                         ctdb_lockall_unmark(lock_ctx->ctdb);
 370                         break;
 371                 }
 372         }
 373 }
 374
 375
 376 static int lock_bucket_id(double t)
 377 {
 378         double ms = 1.e-3, s = 1;
 379         int id;
 380
 381         if (t < 1*ms) {
 382                 id = 0;
 383         } else if (t < 10*ms) {
 384                 id = 1;
 385         } else if (t < 100*ms) {
 386                 id = 2;
 387         } else if (t < 1*s) {
 388                 id = 3;
 389         } else if (t < 2*s) {
 390                 id = 4;
 391         } else if (t < 4*s) {
 392                 id = 5;
 393         } else if (t < 8*s) {
 394                 id = 6;
 395         } else if (t < 16*s) {
 396                 id = 7;
 397         } else if (t < 32*s) {
 398                 id = 8;
 399         } else if (t < 64*s) {
 400                 id = 9;
 401         } else {
 402                 id = 10;
 403         }
 404
 405         return id;
 406 }
 407
 408 /*
 409  * Callback routine when the required locks are obtained.
 410  * Called from parent context
 411  */
 412 static void ctdb_lock_handler(struct tevent_context *ev,
 413                             struct tevent_fd *tfd,
 414                             uint16_t flags,
 415                             void *private_data)
 416 {
 417         struct lock_context *lock_ctx;
 418         TALLOC_CTX *tmp_ctx = NULL;
 419         char c;
 420         bool locked;
 421         double t;
 422         int id;
 423
 424         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 425
 426         /* cancel the timeout event */
 427         if (lock_ctx->ttimer) {
 428                 TALLOC_FREE(lock_ctx->ttimer);
 429         }
 430
 431         t = timeval_elapsed(&lock_ctx->start_time);
 432         id = lock_bucket_id(t);
 433
 434         if (lock_ctx->auto_mark) {
 435                 tmp_ctx = talloc_new(ev);
 436                 talloc_steal(tmp_ctx, lock_ctx);
 437         }
 438
 439         /* Read the status from the child process */
 440         if (read(lock_ctx->fd[0], &c, 1) != 1) {
 441                 locked = false;
 442         } else {
 443                 locked = (c == 0 ? true : false);
 444         }
 445
 446         /* Update statistics */
 447         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_calls);
 448         if (lock_ctx->ctdb_db) {
 449                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_calls);
 450         }
 451
 452         if (locked) {
 453                 if (lock_ctx->ctdb_db) {
 454                         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.buckets[id]);
 455                         CTDB_UPDATE_LATENCY(lock_ctx->ctdb, lock_ctx->ctdb_db,
 456                                             lock_type_str[lock_ctx->type], locks.latency,
 457                                             lock_ctx->start_time);
 458
 459                         CTDB_UPDATE_DB_LATENCY(lock_ctx->ctdb_db, lock_type_str[lock_ctx->type], locks.latency, t);
 460                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.buckets[id]);
 461                 }
 462         } else {
 463                 CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_failed);
 464                 if (lock_ctx->ctdb_db) {
 465                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_failed);
 466                 }
 467         }
 468
 469         process_callbacks(lock_ctx, locked);
 470
 471         if (lock_ctx->auto_mark) {
 472                 talloc_free(tmp_ctx);
 473         }
 474 }
 475
 476
 477 /*
 478  * Callback routine when required locks are not obtained within timeout
 479  * Called from parent context
 480  */
 481 static void ctdb_lock_timeout_handler(struct tevent_context *ev,
 482                                     struct tevent_timer *ttimer,
 483                                     struct timeval current_time,
 484                                     void *private_data)
 485 {
 486         static const char * debug_locks = NULL;
 487         struct lock_context *lock_ctx;
 488         struct ctdb_context *ctdb;
 489         pid_t pid;
 490
 491         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 492         ctdb = lock_ctx->ctdb;
 493
 494         if (lock_ctx->ctdb_db) {
 495                 DEBUG(DEBUG_WARNING,
 496                       ("Unable to get %s lock on database %s for %.0lf seconds\n",
 497                        (lock_ctx->type == LOCK_RECORD ? "RECORD" : "DB"),
 498                        lock_ctx->ctdb_db->db_name,
 499                        timeval_elapsed(&lock_ctx->start_time)));
 500         } else {
 501                 DEBUG(DEBUG_WARNING,
 502                       ("Unable to get ALLDB locks for %.0lf seconds\n",
 503                        timeval_elapsed(&lock_ctx->start_time)));
 504         }
 505
 506         /* Fire a child process to find the blocking process. */
 507         if (debug_locks == NULL) {
 508                 debug_locks = getenv("CTDB_DEBUG_LOCKS");
 509                 if (debug_locks == NULL) {
 510                         debug_locks = talloc_asprintf(ctdb,
 511                                                       "%s/debug_locks.sh",
 512                                                       getenv("CTDB_BASE"));
 513                 }
 514         }
 515         if (debug_locks != NULL) {
 516                 pid = vfork();
 517                 if (pid == 0) {
 518                         execl(debug_locks, debug_locks, NULL);
 519                         _exit(0);
 520                 }
 521                 ctdb_track_child(ctdb, pid);
 522         } else {
 523                 DEBUG(DEBUG_WARNING,
 524                       (__location__
 525                        " Unable to setup lock debugging - no memory?\n"));
 526         }
 527
 528         /* reset the timeout timer */
 529         // talloc_free(lock_ctx->ttimer);
 530         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 531                                             lock_ctx,
 532                                             timeval_current_ofs(10, 0),
 533                                             ctdb_lock_timeout_handler,
 534                                             (void *)lock_ctx);
 535 }
 536
 537
 538 static int db_count_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 539                             void *private_data)
 540 {
 541         int *count = (int *)private_data;
 542
 543         (*count)++;
 544
 545         return 0;
 546 }
 547
 548 struct db_namelist {
 549         char **names;
 550         int n;
 551 };
 552
 553 static int db_name_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 554                            void *private_data)
 555 {
 556         struct db_namelist *list = (struct db_namelist *)private_data;
 557
 558         list->names[list->n] = talloc_strdup(list->names, ctdb_db->db_path);
 559         list->n++;
 560
 561         return 0;
 562 }
 563
 564 static char **lock_helper_args(TALLOC_CTX *mem_ctx, struct lock_context *lock_ctx, int fd)
 565 {
 566         struct ctdb_context *ctdb = lock_ctx->ctdb;
 567         char **args = NULL;
 568         int nargs, i;
 569         int priority;
 570         struct db_namelist list;
 571
 572         switch (lock_ctx->type) {
 573         case LOCK_RECORD:
 574                 nargs = 6;
 575                 break;
 576
 577         case LOCK_DB:
 578                 nargs = 5;
 579                 break;
 580
 581         case LOCK_ALLDB_PRIO:
 582                 nargs = 4;
 583                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_count_handler, &nargs);
 584                 break;
 585
 586         case LOCK_ALLDB:
 587                 nargs = 4;
 588                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 589                         ctdb_db_iterator(ctdb, priority, db_count_handler, &nargs);
 590                 }
 591                 break;
 592         }
 593
 594         /* Add extra argument for null termination */
 595         nargs++;
 596
 597         args = talloc_array(mem_ctx, char *, nargs);
 598         if (args == NULL) {
 599                 return NULL;
 600         }
 601
 602         args[0] = talloc_strdup(args, "ctdb_lock_helper");
 603         args[1] = talloc_asprintf(args, "%d", getpid());
 604         args[2] = talloc_asprintf(args, "%d", fd);
 605
 606         switch (lock_ctx->type) {
 607         case LOCK_RECORD:
 608                 args[3] = talloc_strdup(args, "RECORD");
 609                 args[4] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 610                 if (lock_ctx->key.dsize == 0) {
 611                         args[5] = talloc_strdup(args, "NULL");
 612                 } else {
 613                         args[5] = hex_encode_talloc(args, lock_ctx->key.dptr, lock_ctx->key.dsize);
 614                 }
 615                 break;
 616
 617         case LOCK_DB:
 618                 args[3] = talloc_strdup(args, "DB");
 619                 args[4] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 620                 break;
 621
 622         case LOCK_ALLDB_PRIO:
 623                 args[3] = talloc_strdup(args, "DB");
 624                 list.names = args;
 625                 list.n = 4;
 626                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_name_handler, &list);
 627                 break;
 628
 629         case LOCK_ALLDB:
 630                 args[3] = talloc_strdup(args, "DB");
 631                 list.names = args;
 632                 list.n = 4;
 633                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 634                         ctdb_db_iterator(ctdb, priority, db_name_handler, &list);
 635                 }
 636                 break;
 637         }
 638
 639         /* Make sure last argument is NULL */
 640         args[nargs-1] = NULL;
 641
 642         for (i=0; i<nargs-1; i++) {
 643                 if (args[i] == NULL) {
 644                         talloc_free(args);
 645                         return NULL;
 646                 }
 647         }
 648
 649         return args;
 650 }
 651
 652
 653 /*
 654  * Schedule a new lock child process
 655  * Set up callback handler and timeout handler
 656  */
 657 static void ctdb_lock_schedule(struct ctdb_context *ctdb)
 658 {
 659         struct lock_context *lock_ctx, *next_ctx;
 660         int ret;
 661         TALLOC_CTX *tmp_ctx;
 662         const char *helper = BINDIR "/ctdb_lock_helper";
 663         static const char *prog = NULL;
 664         char **args;
 665
 666         if (prog == NULL) {
 667                 const char *t;
 668
 669                 t = getenv("CTDB_LOCK_HELPER");
 670                 if (t != NULL) {
 671                         prog = talloc_strdup(ctdb, t);
 672                 } else {
 673                         prog = talloc_strdup(ctdb, helper);
 674                 }
 675                 CTDB_NO_MEMORY_VOID(ctdb, prog);
 676         }
 677
 678         if (ctdb->lock_pending == NULL) {
 679                 return;
 680         }
 681
 682         /* Find a lock context with requests */
 683         lock_ctx = ctdb->lock_pending;
 684         while (lock_ctx != NULL) {
 685                 next_ctx = lock_ctx->next;
 686                 if (! lock_ctx->request) {
 687                         DEBUG(DEBUG_INFO, ("Removing lock context without lock request\n"));
 688                         DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 689                         CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 690                         if (lock_ctx->ctdb_db) {
 691                                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 692                         }
 693                         talloc_free(lock_ctx);
 694                 } else {
 695                         if (lock_ctx->ctdb_db == NULL ||
 696                             lock_ctx->ctdb_db->lock_num_current < ctdb->tunable.lock_processes_per_db) {
 697                                 /* Found a lock context with lock requests */
 698                                 break;
 699                         }
 700                 }
 701                 lock_ctx = next_ctx;
 702         }
 703
 704         if (lock_ctx == NULL) {
 705                 return;
 706         }
 707
 708         lock_ctx->child = -1;
 709         ret = pipe(lock_ctx->fd);
 710         if (ret != 0) {
 711                 DEBUG(DEBUG_ERR, ("Failed to create pipe in ctdb_lock_schedule\n"));
 712                 return;
 713         }
 714
 715         set_close_on_exec(lock_ctx->fd[0]);
 716
 717         /* Create data for child process */
 718         tmp_ctx = talloc_new(lock_ctx);
 719         if (tmp_ctx == NULL) {
 720                 DEBUG(DEBUG_ERR, ("Failed to allocate memory for helper args\n"));
 721                 close(lock_ctx->fd[0]);
 722                 close(lock_ctx->fd[1]);
 723                 return;
 724         }
 725
 726         /* Create arguments for lock helper */
 727         args = lock_helper_args(tmp_ctx, lock_ctx, lock_ctx->fd[1]);
 728         if (args == NULL) {
 729                 DEBUG(DEBUG_ERR, ("Failed to create lock helper args\n"));
 730                 close(lock_ctx->fd[0]);
 731                 close(lock_ctx->fd[1]);
 732                 talloc_free(tmp_ctx);
 733                 return;
 734         }
 735
 736         lock_ctx->child = vfork();
 737
 738         if (lock_ctx->child == (pid_t)-1) {
 739                 DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
 740                 close(lock_ctx->fd[0]);
 741                 close(lock_ctx->fd[1]);
 742                 talloc_free(tmp_ctx);
 743                 return;
 744         }
 745
 746
 747         /* Child process */
 748         if (lock_ctx->child == 0) {
 749                 ret = execv(prog, args);
 750                 if (ret < 0) {
 751                         DEBUG(DEBUG_ERR, ("Failed to execute helper %s (%d, %s)\n",
 752                                           prog, errno, strerror(errno)));
 753                 }
 754                 _exit(1);
 755         }
 756
 757         /* Parent process */
 758         ctdb_track_child(ctdb, lock_ctx->child);
 759         close(lock_ctx->fd[1]);
 760
 761         talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
 762
 763         talloc_free(tmp_ctx);
 764
 765         /* Set up timeout handler */
 766         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 767                                             lock_ctx,
 768                                             timeval_current_ofs(10, 0),
 769                                             ctdb_lock_timeout_handler,
 770                                             (void *)lock_ctx);
 771         if (lock_ctx->ttimer == NULL) {
 772                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 773                 lock_ctx->child = -1;
 774                 talloc_set_destructor(lock_ctx, NULL);
 775                 close(lock_ctx->fd[0]);
 776                 return;
 777         }
 778
 779         /* Set up callback */
 780         lock_ctx->tfd = tevent_add_fd(ctdb->ev,
 781                                       lock_ctx,
 782                                       lock_ctx->fd[0],
 783                                       EVENT_FD_READ,
 784                                       ctdb_lock_handler,
 785                                       (void *)lock_ctx);
 786         if (lock_ctx->tfd == NULL) {
 787                 TALLOC_FREE(lock_ctx->ttimer);
 788                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 789                 lock_ctx->child = -1;
 790                 talloc_set_destructor(lock_ctx, NULL);
 791                 close(lock_ctx->fd[0]);
 792                 return;
 793         }
 794         tevent_fd_set_auto_close(lock_ctx->tfd);
 795
 796         /* Move the context from pending to current */
 797         DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 798         DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
 799         CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 800         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 801         if (lock_ctx->ctdb_db) {
 802                 lock_ctx->ctdb_db->lock_num_current++;
 803                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 804                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 805         }
 806 }
 807
 808
 809 /*
 810  * Lock record / db depending on type
 811  */
 812 static struct lock_request *ctdb_lock_internal(struct ctdb_context *ctdb,
 813                                                struct ctdb_db_context *ctdb_db,
 814                                                TDB_DATA key,
 815                                                uint32_t priority,
 816                                                void (*callback)(void *, bool),
 817                                                void *private_data,
 818                                                enum lock_type type,
 819                                                bool auto_mark)
 820 {
 821         struct lock_context *lock_ctx = NULL;
 822         struct lock_request *request;
 823
 824         if (callback == NULL) {
 825                 DEBUG(DEBUG_WARNING, ("No callback function specified, not locking\n"));
 826                 return NULL;
 827         }
 828
 829         lock_ctx = talloc_zero(ctdb, struct lock_context);
 830         if (lock_ctx == NULL) {
 831                 DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
 832                 return NULL;
 833         }
 834
 835         if ((request = talloc_zero(lock_ctx, struct lock_request)) == NULL) {
 836                 talloc_free(lock_ctx);
 837                 return NULL;
 838         }
 839
 840         lock_ctx->type = type;
 841         lock_ctx->ctdb = ctdb;
 842         lock_ctx->ctdb_db = ctdb_db;
 843         lock_ctx->key.dsize = key.dsize;
 844         if (key.dsize > 0) {
 845                 lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
 846                 if (lock_ctx->key.dptr == NULL) {
 847                         DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
 848                         talloc_free(lock_ctx);
 849                         return NULL;
 850                 }
 851                 lock_ctx->key_hash = ctdb_hash(&key);
 852         } else {
 853                 lock_ctx->key.dptr = NULL;
 854         }
 855         lock_ctx->priority = priority;
 856         lock_ctx->auto_mark = auto_mark;
 857
 858         lock_ctx->request = request;
 859         lock_ctx->child = -1;
 860
 861         /* Non-record locks are required by recovery and should be scheduled
 862          * immediately, so keep them at the head of the pending queue.
 863          */
 864         if (lock_ctx->type == LOCK_RECORD) {
 865                 DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
 866         } else {
 867                 DLIST_ADD(ctdb->lock_pending, lock_ctx);
 868         }
 869         CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
 870         if (ctdb_db) {
 871                 CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
 872         }
 873
 874         /* Start the timer when we activate the context */
 875         lock_ctx->start_time = timeval_current();
 876
 877         request->lctx = lock_ctx;
 878         request->callback = callback;
 879         request->private_data = private_data;
 880
 881         talloc_set_destructor(request, ctdb_lock_request_destructor);
 882
 883         ctdb_lock_schedule(ctdb);
 884
 885         return request;
 886 }
 887
 888
 889 /*
 890  * obtain a lock on a record in a database
 891  */
 892 struct lock_request *ctdb_lock_record(struct ctdb_db_context *ctdb_db,
 893                                       TDB_DATA key,
 894                                       bool auto_mark,
 895                                       void (*callback)(void *, bool),
 896                                       void *private_data)
 897 {
 898         return ctdb_lock_internal(ctdb_db->ctdb,
 899                                   ctdb_db,
 900                                   key,
 901                                   0,
 902                                   callback,
 903                                   private_data,
 904                                   LOCK_RECORD,
 905                                   auto_mark);
 906 }
 907
 908
 909 /*
 910  * obtain a lock on a database
 911  */
 912 struct lock_request *ctdb_lock_db(struct ctdb_db_context *ctdb_db,
 913                                   bool auto_mark,
 914                                   void (*callback)(void *, bool),
 915                                   void *private_data)
 916 {
 917         return ctdb_lock_internal(ctdb_db->ctdb,
 918                                   ctdb_db,
 919                                   tdb_null,
 920                                   0,
 921                                   callback,
 922                                   private_data,
 923                                   LOCK_DB,
 924                                   auto_mark);
 925 }
 926
 927
 928 /*
 929  * obtain locks on all databases of specified priority
 930  */
 931 struct lock_request *ctdb_lock_alldb_prio(struct ctdb_context *ctdb,
 932                                           uint32_t priority,
 933                                           bool auto_mark,
 934                                           void (*callback)(void *, bool),
 935                                           void *private_data)
 936 {
 937         if (priority < 1 || priority > NUM_DB_PRIORITIES) {
 938                 DEBUG(DEBUG_ERR, ("Invalid db priority: %u\n", priority));
 939                 return NULL;
 940         }
 941
 942         return ctdb_lock_internal(ctdb,
 943                                   NULL,
 944                                   tdb_null,
 945                                   priority,
 946                                   callback,
 947                                   private_data,
 948                                   LOCK_ALLDB_PRIO,
 949                                   auto_mark);
 950 }
 951
 952
 953 /*
 954  * obtain locks on all databases
 955  */
 956 struct lock_request *ctdb_lock_alldb(struct ctdb_context *ctdb,
 957                                      bool auto_mark,
 958                                      void (*callback)(void *, bool),
 959                                      void *private_data)
 960 {
 961         return ctdb_lock_internal(ctdb,
 962                                   NULL,
 963                                   tdb_null,
 964                                   0,
 965                                   callback,
 966                                   private_data,
 967                                   LOCK_ALLDB,
 968                                   auto_mark);
 969 }
 970