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                 if (lock_ctx->type == LOCK_RECORD) {
 278                         DLIST_REMOVE(lock_ctx->ctdb_db->lock_current, lock_ctx);
 279                 } else {
 280                         DLIST_REMOVE(lock_ctx->ctdb->lock_current, lock_ctx);
 281                 }
 282                 if (lock_ctx->ctdb_db) {
 283                         lock_ctx->ctdb_db->lock_num_current--;
 284                 }
 285                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 286                 if (lock_ctx->ctdb_db) {
 287                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 288                 }
 289         } else {
 290                 if (lock_ctx->type == LOCK_RECORD) {
 291                         DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 292                 } else {
 293                         DLIST_REMOVE(lock_ctx->ctdb->lock_pending, lock_ctx);
 294                 }
 295                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 296                 if (lock_ctx->ctdb_db) {
 297                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 298                 }
 299         }
 300
 301         ctdb_lock_schedule(lock_ctx->ctdb);
 302
 303         return 0;
 304 }
 305
 306
 307 /*
 308  * Destructor to remove lock request
 309  */
 310 static int ctdb_lock_request_destructor(struct lock_request *lock_request)
 311 {
 312         lock_request->lctx->request = NULL;
 313         return 0;
 314 }
 315
 316 void ctdb_lock_free_request_context(struct lock_request *lock_req)
 317 {
 318         struct lock_context *lock_ctx;
 319
 320         lock_ctx = lock_req->lctx;
 321         talloc_free(lock_req);
 322         talloc_free(lock_ctx);
 323 }
 324
 325
 326 /*
 327  * Process all the callbacks waiting for lock
 328  *
 329  * If lock has failed, callback is executed with locked=false
 330  */
 331 static void process_callbacks(struct lock_context *lock_ctx, bool locked)
 332 {
 333         struct lock_request *request;
 334
 335         if (lock_ctx->auto_mark && locked) {
 336                 switch (lock_ctx->type) {
 337                 case LOCK_RECORD:
 338                         tdb_chainlock_mark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 339                         break;
 340
 341                 case LOCK_DB:
 342                         tdb_lockall_mark(lock_ctx->ctdb_db->ltdb->tdb);
 343                         break;
 344
 345                 case LOCK_ALLDB_PRIO:
 346                         ctdb_lockall_mark_prio(lock_ctx->ctdb, lock_ctx->priority);
 347                         break;
 348
 349                 case LOCK_ALLDB:
 350                         ctdb_lockall_mark(lock_ctx->ctdb);
 351                         break;
 352                 }
 353         }
 354
 355         request = lock_ctx->request;
 356         if (lock_ctx->auto_mark) {
 357                 /* Reset the destructor, so request is not removed from the list */
 358                 talloc_set_destructor(request, NULL);
 359         }
 360         request->callback(request->private_data, locked);
 361
 362         if (lock_ctx->auto_mark && locked) {
 363                 switch (lock_ctx->type) {
 364                 case LOCK_RECORD:
 365                         tdb_chainlock_unmark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 366                         break;
 367
 368                 case LOCK_DB:
 369                         tdb_lockall_unmark(lock_ctx->ctdb_db->ltdb->tdb);
 370                         break;
 371
 372                 case LOCK_ALLDB_PRIO:
 373                         ctdb_lockall_unmark_prio(lock_ctx->ctdb, lock_ctx->priority);
 374                         break;
 375
 376                 case LOCK_ALLDB:
 377                         ctdb_lockall_unmark(lock_ctx->ctdb);
 378                         break;
 379                 }
 380         }
 381 }
 382
 383
 384 static int lock_bucket_id(double t)
 385 {
 386         double ms = 1.e-3, s = 1;
 387         int id;
 388
 389         if (t < 1*ms) {
 390                 id = 0;
 391         } else if (t < 10*ms) {
 392                 id = 1;
 393         } else if (t < 100*ms) {
 394                 id = 2;
 395         } else if (t < 1*s) {
 396                 id = 3;
 397         } else if (t < 2*s) {
 398                 id = 4;
 399         } else if (t < 4*s) {
 400                 id = 5;
 401         } else if (t < 8*s) {
 402                 id = 6;
 403         } else if (t < 16*s) {
 404                 id = 7;
 405         } else if (t < 32*s) {
 406                 id = 8;
 407         } else if (t < 64*s) {
 408                 id = 9;
 409         } else {
 410                 id = 10;
 411         }
 412
 413         return id;
 414 }
 415
 416 /*
 417  * Callback routine when the required locks are obtained.
 418  * Called from parent context
 419  */
 420 static void ctdb_lock_handler(struct tevent_context *ev,
 421                             struct tevent_fd *tfd,
 422                             uint16_t flags,
 423                             void *private_data)
 424 {
 425         struct lock_context *lock_ctx;
 426         TALLOC_CTX *tmp_ctx = NULL;
 427         char c;
 428         bool locked;
 429         double t;
 430         int id;
 431
 432         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 433
 434         /* cancel the timeout event */
 435         TALLOC_FREE(lock_ctx->ttimer);
 436
 437         t = timeval_elapsed(&lock_ctx->start_time);
 438         id = lock_bucket_id(t);
 439
 440         if (lock_ctx->auto_mark) {
 441                 tmp_ctx = talloc_new(ev);
 442                 talloc_steal(tmp_ctx, lock_ctx);
 443         }
 444
 445         /* Read the status from the child process */
 446         if (sys_read(lock_ctx->fd[0], &c, 1) != 1) {
 447                 locked = false;
 448         } else {
 449                 locked = (c == 0 ? true : false);
 450         }
 451
 452         /* Update statistics */
 453         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_calls);
 454         if (lock_ctx->ctdb_db) {
 455                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_calls);
 456         }
 457
 458         if (locked) {
 459                 if (lock_ctx->ctdb_db) {
 460                         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.buckets[id]);
 461                         CTDB_UPDATE_LATENCY(lock_ctx->ctdb, lock_ctx->ctdb_db,
 462                                             lock_type_str[lock_ctx->type], locks.latency,
 463                                             lock_ctx->start_time);
 464
 465                         CTDB_UPDATE_DB_LATENCY(lock_ctx->ctdb_db, lock_type_str[lock_ctx->type], locks.latency, t);
 466                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.buckets[id]);
 467                 }
 468         } else {
 469                 CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_failed);
 470                 if (lock_ctx->ctdb_db) {
 471                         CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_failed);
 472                 }
 473         }
 474
 475         process_callbacks(lock_ctx, locked);
 476
 477         if (lock_ctx->auto_mark) {
 478                 talloc_free(tmp_ctx);
 479         }
 480 }
 481
 482
 483 /*
 484  * Callback routine when required locks are not obtained within timeout
 485  * Called from parent context
 486  */
 487 static void ctdb_lock_timeout_handler(struct tevent_context *ev,
 488                                     struct tevent_timer *ttimer,
 489                                     struct timeval current_time,
 490                                     void *private_data)
 491 {
 492         static const char * debug_locks = NULL;
 493         struct lock_context *lock_ctx;
 494         struct ctdb_context *ctdb;
 495         pid_t pid;
 496
 497         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 498         ctdb = lock_ctx->ctdb;
 499
 500         if (lock_ctx->ctdb_db) {
 501                 DEBUG(DEBUG_WARNING,
 502                       ("Unable to get %s lock on database %s for %.0lf seconds\n",
 503                        (lock_ctx->type == LOCK_RECORD ? "RECORD" : "DB"),
 504                        lock_ctx->ctdb_db->db_name,
 505                        timeval_elapsed(&lock_ctx->start_time)));
 506         } else {
 507                 DEBUG(DEBUG_WARNING,
 508                       ("Unable to get ALLDB locks for %.0lf seconds\n",
 509                        timeval_elapsed(&lock_ctx->start_time)));
 510         }
 511
 512         /* Fire a child process to find the blocking process. */
 513         if (debug_locks == NULL) {
 514                 debug_locks = getenv("CTDB_DEBUG_LOCKS");
 515                 if (debug_locks == NULL) {
 516                         debug_locks = talloc_asprintf(ctdb,
 517                                                       "%s/debug_locks.sh",
 518                                                       getenv("CTDB_BASE"));
 519                 }
 520         }
 521         if (debug_locks != NULL) {
 522                 pid = vfork();
 523                 if (pid == 0) {
 524                         execl(debug_locks, debug_locks, NULL);
 525                         _exit(0);
 526                 }
 527                 ctdb_track_child(ctdb, pid);
 528         } else {
 529                 DEBUG(DEBUG_WARNING,
 530                       (__location__
 531                        " Unable to setup lock debugging - no memory?\n"));
 532         }
 533
 534         /* reset the timeout timer */
 535         // talloc_free(lock_ctx->ttimer);
 536         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 537                                             lock_ctx,
 538                                             timeval_current_ofs(10, 0),
 539                                             ctdb_lock_timeout_handler,
 540                                             (void *)lock_ctx);
 541 }
 542
 543
 544 static int db_count_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 545                             void *private_data)
 546 {
 547         int *count = (int *)private_data;
 548
 549         (*count)++;
 550
 551         return 0;
 552 }
 553
 554 struct db_namelist {
 555         char **names;
 556         int n;
 557 };
 558
 559 static int db_name_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 560                            void *private_data)
 561 {
 562         struct db_namelist *list = (struct db_namelist *)private_data;
 563
 564         list->names[list->n] = talloc_strdup(list->names, ctdb_db->db_path);
 565         list->n++;
 566
 567         return 0;
 568 }
 569
 570 static char **lock_helper_args(TALLOC_CTX *mem_ctx, struct lock_context *lock_ctx, int fd)
 571 {
 572         struct ctdb_context *ctdb = lock_ctx->ctdb;
 573         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 = 6;
 581                 break;
 582
 583         case LOCK_DB:
 584                 nargs = 5;
 585                 break;
 586
 587         case LOCK_ALLDB_PRIO:
 588                 nargs = 4;
 589                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_count_handler, &nargs);
 590                 break;
 591
 592         case LOCK_ALLDB:
 593                 nargs = 4;
 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, char *, nargs);
 604         if (args == NULL) {
 605                 return NULL;
 606         }
 607
 608         args[0] = talloc_strdup(args, "ctdb_lock_helper");
 609         args[1] = talloc_asprintf(args, "%d", getpid());
 610         args[2] = talloc_asprintf(args, "%d", fd);
 611
 612         switch (lock_ctx->type) {
 613         case LOCK_RECORD:
 614                 args[3] = talloc_strdup(args, "RECORD");
 615                 args[4] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 616                 if (lock_ctx->key.dsize == 0) {
 617                         args[5] = talloc_strdup(args, "NULL");
 618                 } else {
 619                         args[5] = hex_encode_talloc(args, lock_ctx->key.dptr, lock_ctx->key.dsize);
 620                 }
 621                 break;
 622
 623         case LOCK_DB:
 624                 args[3] = talloc_strdup(args, "DB");
 625                 args[4] = talloc_strdup(args, lock_ctx->ctdb_db->db_path);
 626                 break;
 627
 628         case LOCK_ALLDB_PRIO:
 629                 args[3] = talloc_strdup(args, "DB");
 630                 list.names = args;
 631                 list.n = 4;
 632                 ctdb_db_iterator(ctdb, lock_ctx->priority, db_name_handler, &list);
 633                 break;
 634
 635         case LOCK_ALLDB:
 636                 args[3] = talloc_strdup(args, "DB");
 637                 list.names = args;
 638                 list.n = 4;
 639                 for (priority=1; priority<NUM_DB_PRIORITIES; priority++) {
 640                         ctdb_db_iterator(ctdb, priority, db_name_handler, &list);
 641                 }
 642                 break;
 643         }
 644
 645         /* Make sure last argument is NULL */
 646         args[nargs-1] = NULL;
 647
 648         for (i=0; i<nargs-1; i++) {
 649                 if (args[i] == NULL) {
 650                         talloc_free(args);
 651                         return NULL;
 652                 }
 653         }
 654
 655         return args;
 656 }
 657
 658 /*
 659  * Find a lock request that can be scheduled
 660  */
 661 struct lock_context *ctdb_find_lock_context(struct ctdb_context *ctdb)
 662 {
 663         struct lock_context *lock_ctx, *next_ctx;
 664         struct ctdb_db_context *ctdb_db;
 665
 666         /* First check if there are database lock requests */
 667
 668         for (lock_ctx = ctdb->lock_pending; lock_ctx != NULL;
 669              lock_ctx = next_ctx) {
 670
 671                 if (lock_ctx->request != NULL) {
 672                         /* Found a lock context with a request */
 673                         return lock_ctx;
 674                 }
 675
 676                 next_ctx = lock_ctx->next;
 677
 678                 DEBUG(DEBUG_INFO, ("Removing lock context without lock "
 679                                    "request\n"));
 680                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 681                 CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 682                 if (lock_ctx->ctdb_db) {
 683                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db,
 684                                                locks.num_pending);
 685                 }
 686                 talloc_free(lock_ctx);
 687         }
 688
 689         /* Next check database queues */
 690         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 691                 if (ctdb_db->lock_num_current ==
 692                     ctdb->tunable.lock_processes_per_db) {
 693                         continue;
 694                 }
 695
 696                 for (lock_ctx = ctdb_db->lock_pending; lock_ctx != NULL;
 697                      lock_ctx = next_ctx) {
 698
 699                         next_ctx = lock_ctx->next;
 700
 701                         if (lock_ctx->request != NULL) {
 702                                 return lock_ctx;
 703                         }
 704
 705                         DEBUG(DEBUG_INFO, ("Removing lock context without "
 706                                            "lock request\n"));
 707                         DLIST_REMOVE(ctdb_db->lock_pending, lock_ctx);
 708                         CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 709                         CTDB_DECREMENT_DB_STAT(ctdb_db, locks.num_pending);
 710                         talloc_free(lock_ctx);
 711                 }
 712         }
 713
 714         return NULL;
 715 }
 716
 717 /*
 718  * Schedule a new lock child process
 719  * Set up callback handler and timeout handler
 720  */
 721 static void ctdb_lock_schedule(struct ctdb_context *ctdb)
 722 {
 723         struct lock_context *lock_ctx;
 724         int ret;
 725         TALLOC_CTX *tmp_ctx;
 726         const char *helper = BINDIR "/ctdb_lock_helper";
 727         static const char *prog = NULL;
 728         char **args;
 729
 730         if (prog == NULL) {
 731                 const char *t;
 732
 733                 t = getenv("CTDB_LOCK_HELPER");
 734                 if (t != NULL) {
 735                         prog = talloc_strdup(ctdb, t);
 736                 } else {
 737                         prog = talloc_strdup(ctdb, helper);
 738                 }
 739                 CTDB_NO_MEMORY_VOID(ctdb, prog);
 740         }
 741
 742         /* Find a lock context with requests */
 743         lock_ctx = ctdb_find_lock_context(ctdb);
 744         if (lock_ctx == NULL) {
 745                 return;
 746         }
 747
 748         lock_ctx->child = -1;
 749         ret = pipe(lock_ctx->fd);
 750         if (ret != 0) {
 751                 DEBUG(DEBUG_ERR, ("Failed to create pipe in ctdb_lock_schedule\n"));
 752                 return;
 753         }
 754
 755         set_close_on_exec(lock_ctx->fd[0]);
 756
 757         /* Create data for child process */
 758         tmp_ctx = talloc_new(lock_ctx);
 759         if (tmp_ctx == NULL) {
 760                 DEBUG(DEBUG_ERR, ("Failed to allocate memory for helper args\n"));
 761                 close(lock_ctx->fd[0]);
 762                 close(lock_ctx->fd[1]);
 763                 return;
 764         }
 765
 766         /* Create arguments for lock helper */
 767         args = lock_helper_args(tmp_ctx, lock_ctx, lock_ctx->fd[1]);
 768         if (args == NULL) {
 769                 DEBUG(DEBUG_ERR, ("Failed to create lock helper args\n"));
 770                 close(lock_ctx->fd[0]);
 771                 close(lock_ctx->fd[1]);
 772                 talloc_free(tmp_ctx);
 773                 return;
 774         }
 775
 776         lock_ctx->child = vfork();
 777
 778         if (lock_ctx->child == (pid_t)-1) {
 779                 DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
 780                 close(lock_ctx->fd[0]);
 781                 close(lock_ctx->fd[1]);
 782                 talloc_free(tmp_ctx);
 783                 return;
 784         }
 785
 786
 787         /* Child process */
 788         if (lock_ctx->child == 0) {
 789                 ret = execv(prog, args);
 790                 if (ret < 0) {
 791                         DEBUG(DEBUG_ERR, ("Failed to execute helper %s (%d, %s)\n",
 792                                           prog, errno, strerror(errno)));
 793                 }
 794                 _exit(1);
 795         }
 796
 797         /* Parent process */
 798         ctdb_track_child(ctdb, lock_ctx->child);
 799         close(lock_ctx->fd[1]);
 800
 801         talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
 802
 803         talloc_free(tmp_ctx);
 804
 805         /* Set up timeout handler */
 806         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 807                                             lock_ctx,
 808                                             timeval_current_ofs(10, 0),
 809                                             ctdb_lock_timeout_handler,
 810                                             (void *)lock_ctx);
 811         if (lock_ctx->ttimer == NULL) {
 812                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 813                 lock_ctx->child = -1;
 814                 talloc_set_destructor(lock_ctx, NULL);
 815                 close(lock_ctx->fd[0]);
 816                 return;
 817         }
 818
 819         /* Set up callback */
 820         lock_ctx->tfd = tevent_add_fd(ctdb->ev,
 821                                       lock_ctx,
 822                                       lock_ctx->fd[0],
 823                                       EVENT_FD_READ,
 824                                       ctdb_lock_handler,
 825                                       (void *)lock_ctx);
 826         if (lock_ctx->tfd == NULL) {
 827                 TALLOC_FREE(lock_ctx->ttimer);
 828                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 829                 lock_ctx->child = -1;
 830                 talloc_set_destructor(lock_ctx, NULL);
 831                 close(lock_ctx->fd[0]);
 832                 return;
 833         }
 834         tevent_fd_set_auto_close(lock_ctx->tfd);
 835
 836         /* Move the context from pending to current */
 837         if (lock_ctx->type == LOCK_RECORD) {
 838                 DLIST_REMOVE(lock_ctx->ctdb_db->lock_pending, lock_ctx);
 839                 DLIST_ADD_END(lock_ctx->ctdb_db->lock_current, lock_ctx, NULL);
 840         } else {
 841                 DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 842                 DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
 843         }
 844         CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 845         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 846         if (lock_ctx->ctdb_db) {
 847                 lock_ctx->ctdb_db->lock_num_current++;
 848                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 849                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 850         }
 851 }
 852
 853
 854 /*
 855  * Lock record / db depending on type
 856  */
 857 static struct lock_request *ctdb_lock_internal(struct ctdb_context *ctdb,
 858                                                struct ctdb_db_context *ctdb_db,
 859                                                TDB_DATA key,
 860                                                uint32_t priority,
 861                                                void (*callback)(void *, bool),
 862                                                void *private_data,
 863                                                enum lock_type type,
 864                                                bool auto_mark)
 865 {
 866         struct lock_context *lock_ctx = NULL;
 867         struct lock_request *request;
 868
 869         if (callback == NULL) {
 870                 DEBUG(DEBUG_WARNING, ("No callback function specified, not locking\n"));
 871                 return NULL;
 872         }
 873
 874         lock_ctx = talloc_zero(ctdb, struct lock_context);
 875         if (lock_ctx == NULL) {
 876                 DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
 877                 return NULL;
 878         }
 879
 880         if ((request = talloc_zero(lock_ctx, struct lock_request)) == NULL) {
 881                 talloc_free(lock_ctx);
 882                 return NULL;
 883         }
 884
 885         lock_ctx->type = type;
 886         lock_ctx->ctdb = ctdb;
 887         lock_ctx->ctdb_db = ctdb_db;
 888         lock_ctx->key.dsize = key.dsize;
 889         if (key.dsize > 0) {
 890                 lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
 891                 if (lock_ctx->key.dptr == NULL) {
 892                         DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
 893                         talloc_free(lock_ctx);
 894                         return NULL;
 895                 }
 896                 lock_ctx->key_hash = ctdb_hash(&key);
 897         } else {
 898                 lock_ctx->key.dptr = NULL;
 899         }
 900         lock_ctx->priority = priority;
 901         lock_ctx->auto_mark = auto_mark;
 902
 903         lock_ctx->request = request;
 904         lock_ctx->child = -1;
 905
 906         /* Non-record locks are required by recovery and should be scheduled
 907          * immediately, so keep them at the head of the pending queue.
 908          */
 909         if (lock_ctx->type == LOCK_RECORD) {
 910                 DLIST_ADD_END(ctdb_db->lock_pending, lock_ctx, NULL);
 911         } else {
 912                 DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
 913         }
 914         CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
 915         if (ctdb_db) {
 916                 CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
 917         }
 918
 919         /* Start the timer when we activate the context */
 920         lock_ctx->start_time = timeval_current();
 921
 922         request->lctx = lock_ctx;
 923         request->callback = callback;
 924         request->private_data = private_data;
 925
 926         talloc_set_destructor(request, ctdb_lock_request_destructor);
 927
 928         ctdb_lock_schedule(ctdb);
 929
 930         return request;
 931 }
 932
 933
 934 /*
 935  * obtain a lock on a record in a database
 936  */
 937 struct lock_request *ctdb_lock_record(struct ctdb_db_context *ctdb_db,
 938                                       TDB_DATA key,
 939                                       bool auto_mark,
 940                                       void (*callback)(void *, bool),
 941                                       void *private_data)
 942 {
 943         return ctdb_lock_internal(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(struct ctdb_db_context *ctdb_db,
 958                                   bool auto_mark,
 959                                   void (*callback)(void *, bool),
 960                                   void *private_data)
 961 {
 962         return ctdb_lock_internal(ctdb_db->ctdb,
 963                                   ctdb_db,
 964                                   tdb_null,
 965                                   0,
 966                                   callback,
 967                                   private_data,
 968                                   LOCK_DB,
 969                                   auto_mark);
 970 }
 971
 972
 973 /*
 974  * obtain locks on all databases of specified priority
 975  */
 976 struct lock_request *ctdb_lock_alldb_prio(struct ctdb_context *ctdb,
 977                                           uint32_t priority,
 978                                           bool auto_mark,
 979                                           void (*callback)(void *, bool),
 980                                           void *private_data)
 981 {
 982         if (priority < 1 || priority > NUM_DB_PRIORITIES) {
 983                 DEBUG(DEBUG_ERR, ("Invalid db priority: %u\n", priority));
 984                 return NULL;
 985         }
 986
 987         return ctdb_lock_internal(ctdb,
 988                                   NULL,
 989                                   tdb_null,
 990                                   priority,
 991                                   callback,
 992                                   private_data,
 993                                   LOCK_ALLDB_PRIO,
 994                                   auto_mark);
 995 }
 996
 997
 998 /*
 999  * obtain locks on all databases
1000  */
1001 struct lock_request *ctdb_lock_alldb(struct ctdb_context *ctdb,
1002                                      bool auto_mark,
1003                                      void (*callback)(void *, bool),
1004                                      void *private_data)
1005 {
1006         return ctdb_lock_internal(ctdb,
1007                                   NULL,
1008                                   tdb_null,
1009                                   0,
1010                                   callback,
1011                                   private_data,
1012                                   LOCK_ALLDB,
1013                                   auto_mark);
1014 }
1015