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  *    diagnose using /proc/locks and log warning message
  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 /* FIXME: Add a tunable max_lock_processes_per_db */
  47 #define MAX_LOCK_PROCESSES_PER_DB               (100)
  48
  49 enum lock_type {
  50         LOCK_RECORD,
  51         LOCK_DB,
  52         LOCK_ALLDB_PRIO,
  53         LOCK_ALLDB,
  54 };
  55
  56 static const char * const lock_type_str[] = {
  57         "lock_record",
  58         "lock_db",
  59         "lock_alldb_prio",
  60         "lock_alldb",
  61 };
  62
  63 struct lock_request;
  64
  65 /* lock_context is the common part for a lock request */
  66 struct lock_context {
  67         struct lock_context *next, *prev;
  68         enum lock_type type;
  69         struct ctdb_context *ctdb;
  70         struct ctdb_db_context *ctdb_db;
  71         TDB_DATA key;
  72         uint32_t priority;
  73         bool auto_mark;
  74         struct lock_request *request;
  75         pid_t child;
  76         int fd[2];
  77         struct tevent_fd *tfd;
  78         struct tevent_timer *ttimer;
  79         struct timeval start_time;
  80         uint32_t key_hash;
  81         bool can_schedule;
  82 };
  83
  84 /* lock_request is the client specific part for a lock request */
  85 struct lock_request {
  86         struct lock_context *lctx;
  87         void (*callback)(void *, bool);
  88         void *private_data;
  89 };
  90
  91
  92 /*
  93  * Support samba 3.6.x (and older) versions which do not set db priority.
  94  *
  95  * By default, all databases are set to priority 1. So only when priority
  96  * is set to 1, check for databases that need higher priority.
  97  */
  98 static bool later_db(struct ctdb_context *ctdb, const char *name)
  99 {
 100         if (ctdb->tunable.samba3_hack == 0) {
 101                 return false;
 102         }
 103
 104         if (strstr(name, "brlock") ||
 105             strstr(name, "g_lock") ||
 106             strstr(name, "notify_onelevel") ||
 107             strstr(name, "serverid") ||
 108             strstr(name, "xattr_tdb")) {
 109                 return true;
 110         }
 111
 112         return false;
 113 }
 114
 115 typedef int (*db_handler_t)(struct ctdb_db_context *ctdb_db,
 116                             uint32_t priority,
 117                             void *private_data);
 118
 119 static int ctdb_db_iterator(struct ctdb_context *ctdb, uint32_t priority,
 120                             db_handler_t handler, void *private_data)
 121 {
 122         struct ctdb_db_context *ctdb_db;
 123         int ret;
 124
 125         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 126                 if (ctdb_db->priority != priority) {
 127                         continue;
 128                 }
 129                 if (later_db(ctdb, ctdb_db->db_name)) {
 130                         continue;
 131                 }
 132                 ret = handler(ctdb_db, priority, private_data);
 133                 if (ret != 0) {
 134                         return -1;
 135                 }
 136         }
 137
 138         /* If priority != 1, later_db check is not required and can return */
 139         if (priority != 1) {
 140                 return 0;
 141         }
 142
 143         for (ctdb_db = ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
 144                 if (!later_db(ctdb, ctdb_db->db_name)) {
 145                         continue;
 146                 }
 147                 ret = handler(ctdb_db, priority, private_data);
 148                 if (ret != 0) {
 149                         return -1;
 150                 }
 151         }
 152
 153         return 0;
 154 }
 155
 156
 157 /*
 158  * lock all databases - mark only
 159  */
 160 static int db_lock_mark_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 161                                 void *private_data)
 162 {
 163         int tdb_transaction_write_lock_mark(struct tdb_context *);
 164
 165         DEBUG(DEBUG_INFO, ("marking locked database %s, priority:%u\n",
 166                            ctdb_db->db_name, priority));
 167
 168         if (tdb_transaction_write_lock_mark(ctdb_db->ltdb->tdb) != 0) {
 169                 DEBUG(DEBUG_ERR, ("Failed to mark (transaction lock) database %s\n",
 170                                   ctdb_db->db_name));
 171                 return -1;
 172         }
 173
 174         if (tdb_lockall_mark(ctdb_db->ltdb->tdb) != 0) {
 175                 DEBUG(DEBUG_ERR, ("Failed to mark (all lock) database %s\n",
 176                                   ctdb_db->db_name));
 177                 return -1;
 178         }
 179
 180         return 0;
 181 }
 182
 183 int ctdb_lockall_mark_prio(struct ctdb_context *ctdb, uint32_t priority)
 184 {
 185         /*
 186          * This function is only used by the main dameon during recovery.
 187          * At this stage, the databases have already been locked, by a
 188          * dedicated child process. The freeze_mode variable is used to track
 189          * whether the actual locks are held by the child process or not.
 190          */
 191
 192         if (ctdb->freeze_mode[priority] != CTDB_FREEZE_FROZEN) {
 193                 DEBUG(DEBUG_ERR, ("Attempt to mark all databases locked when not frozen\n"));
 194                 return -1;
 195         }
 196
 197         return ctdb_db_iterator(ctdb, priority, db_lock_mark_handler, NULL);
 198 }
 199
 200 static int ctdb_lockall_mark(struct ctdb_context *ctdb)
 201 {
 202         uint32_t priority;
 203
 204         for (priority=1; priority<=NUM_DB_PRIORITIES; priority++) {
 205                 if (ctdb_db_iterator(ctdb, priority, db_lock_mark_handler, NULL) != 0) {
 206                         return -1;
 207                 }
 208         }
 209
 210         return 0;
 211 }
 212
 213
 214 /*
 215  * lock all databases - unmark only
 216  */
 217 static int db_lock_unmark_handler(struct ctdb_db_context *ctdb_db, uint32_t priority,
 218                                   void *private_data)
 219 {
 220         int tdb_transaction_write_lock_unmark(struct tdb_context *);
 221
 222         DEBUG(DEBUG_INFO, ("unmarking locked database %s, priority:%u\n",
 223                            ctdb_db->db_name, priority));
 224
 225         if (tdb_transaction_write_lock_unmark(ctdb_db->ltdb->tdb) != 0) {
 226                 DEBUG(DEBUG_ERR, ("Failed to unmark (transaction lock) database %s\n",
 227                                   ctdb_db->db_name));
 228                 return -1;
 229         }
 230
 231         if (tdb_lockall_unmark(ctdb_db->ltdb->tdb) != 0) {
 232                 DEBUG(DEBUG_ERR, ("Failed to unmark (all lock) database %s\n",
 233                                   ctdb_db->db_name));
 234                 return -1;
 235         }
 236
 237         return 0;
 238 }
 239
 240 int ctdb_lockall_unmark_prio(struct ctdb_context *ctdb, uint32_t priority)
 241 {
 242         /*
 243          * This function is only used by the main daemon during recovery.
 244          * At this stage, the databases have already been locked, by a
 245          * dedicated child process. The freeze_mode variable is used to track
 246          * whether the actual locks are held by the child process or not.
 247          */
 248
 249         if (ctdb->freeze_mode[priority] != CTDB_FREEZE_FROZEN) {
 250                 DEBUG(DEBUG_ERR, ("Attempt to unmark all databases locked when not frozen\n"));
 251                 return -1;
 252         }
 253
 254         return ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL);
 255 }
 256
 257 static int ctdb_lockall_unmark(struct ctdb_context *ctdb)
 258 {
 259         uint32_t priority;
 260
 261         for (priority=NUM_DB_PRIORITIES; priority>0; priority--) {
 262                 if (ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL) != 0) {
 263                         return -1;
 264                 }
 265         }
 266
 267         return 0;
 268 }
 269
 270
 271 static void ctdb_lock_schedule(struct ctdb_context *ctdb);
 272
 273 /*
 274  * Destructor to kill the child locking process
 275  */
 276 static int ctdb_lock_context_destructor(struct lock_context *lock_ctx)
 277 {
 278         if (lock_ctx->child > 0) {
 279                 ctdb_kill(lock_ctx->ctdb, lock_ctx->child, SIGKILL);
 280                 DLIST_REMOVE(lock_ctx->ctdb->lock_current, lock_ctx);
 281                 if (lock_ctx->ctdb_db) {
 282                         lock_ctx->ctdb_db->lock_num_current--;
 283                 }
 284                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 285                 if (lock_ctx->type == LOCK_RECORD || lock_ctx->type == LOCK_DB) {
 286                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 287                 }
 288         } else {
 289                 DLIST_REMOVE(lock_ctx->ctdb->lock_pending, lock_ctx);
 290                 lock_ctx->ctdb->lock_num_pending--;
 291                 CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 292                 if (lock_ctx->type == LOCK_RECORD || lock_ctx->type == LOCK_DB) {
 293                         CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 294                 }
 295         }
 296
 297         ctdb_lock_schedule(lock_ctx->ctdb);
 298
 299         return 0;
 300 }
 301
 302
 303 /*
 304  * Destructor to remove lock request
 305  */
 306 static int ctdb_lock_request_destructor(struct lock_request *lock_request)
 307 {
 308         lock_request->lctx->request = NULL;
 309         return 0;
 310 }
 311
 312 void ctdb_lock_free_request_context(struct lock_request *lock_req)
 313 {
 314         struct lock_context *lock_ctx;
 315
 316         lock_ctx = lock_req->lctx;
 317         talloc_free(lock_req);
 318         talloc_free(lock_ctx);
 319 }
 320
 321
 322 /*
 323  * Process all the callbacks waiting for lock
 324  *
 325  * If lock has failed, callback is executed with locked=false
 326  */
 327 static void process_callbacks(struct lock_context *lock_ctx, bool locked)
 328 {
 329         struct lock_request *request;
 330
 331         if (lock_ctx->auto_mark && locked) {
 332                 switch (lock_ctx->type) {
 333                 case LOCK_RECORD:
 334                         tdb_chainlock_mark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 335                         break;
 336
 337                 case LOCK_DB:
 338                         tdb_lockall_mark(lock_ctx->ctdb_db->ltdb->tdb);
 339                         break;
 340
 341                 case LOCK_ALLDB_PRIO:
 342                         ctdb_lockall_mark_prio(lock_ctx->ctdb, lock_ctx->priority);
 343                         break;
 344
 345                 case LOCK_ALLDB:
 346                         ctdb_lockall_mark(lock_ctx->ctdb);
 347                         break;
 348                 }
 349         }
 350
 351         request = lock_ctx->request;
 352         if (lock_ctx->auto_mark) {
 353                 /* Reset the destructor, so request is not removed from the list */
 354                 talloc_set_destructor(request, NULL);
 355         }
 356         request->callback(request->private_data, locked);
 357
 358         if (lock_ctx->auto_mark && locked) {
 359                 switch (lock_ctx->type) {
 360                 case LOCK_RECORD:
 361                         tdb_chainlock_unmark(lock_ctx->ctdb_db->ltdb->tdb, lock_ctx->key);
 362                         break;
 363
 364                 case LOCK_DB:
 365                         tdb_lockall_unmark(lock_ctx->ctdb_db->ltdb->tdb);
 366                         break;
 367
 368                 case LOCK_ALLDB_PRIO:
 369                         ctdb_lockall_unmark_prio(lock_ctx->ctdb, lock_ctx->priority);
 370                         break;
 371
 372                 case LOCK_ALLDB:
 373                         ctdb_lockall_unmark(lock_ctx->ctdb);
 374                         break;
 375                 }
 376         }
 377 }
 378
 379
 380 static int lock_bucket_id(double t)
 381 {
 382         double ms = 1.e-3, s = 1;
 383         int id;
 384
 385         if (t < 1*ms) {
 386                 id = 0;
 387         } else if (t < 10*ms) {
 388                 id = 1;
 389         } else if (t < 100*ms) {
 390                 id = 2;
 391         } else if (t < 1*s) {
 392                 id = 3;
 393         } else if (t < 2*s) {
 394                 id = 4;
 395         } else if (t < 4*s) {
 396                 id = 5;
 397         } else if (t < 8*s) {
 398                 id = 6;
 399         } else if (t < 16*s) {
 400                 id = 7;
 401         } else if (t < 32*s) {
 402                 id = 8;
 403         } else if (t < 64*s) {
 404                 id = 9;
 405         } else {
 406                 id = 10;
 407         }
 408
 409         return id;
 410 }
 411
 412 /*
 413  * Callback routine when the required locks are obtained.
 414  * Called from parent context
 415  */
 416 static void ctdb_lock_handler(struct tevent_context *ev,
 417                             struct tevent_fd *tfd,
 418                             uint16_t flags,
 419                             void *private_data)
 420 {
 421         struct lock_context *lock_ctx;
 422         TALLOC_CTX *tmp_ctx = NULL;
 423         char c;
 424         bool locked;
 425         double t;
 426         int id;
 427
 428         lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
 429
 430         /* cancel the timeout event */
 431         if (lock_ctx->ttimer) {
 432                 TALLOC_FREE(lock_ctx->ttimer);
 433         }
 434
 435         t = timeval_elapsed(&lock_ctx->start_time);
 436         id = lock_bucket_id(t);
 437
 438         if (lock_ctx->auto_mark) {
 439                 tmp_ctx = talloc_new(ev);
 440                 talloc_steal(tmp_ctx, lock_ctx);
 441         }
 442
 443         /* Read the status from the child process */
 444         if (read(lock_ctx->fd[0], &c, 1) != 1) {
 445                 locked = false;
 446         } else {
 447                 locked = (c == 0 ? true : false);
 448         }
 449
 450         /* Update statistics */
 451         CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
 452         CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_calls);
 453         if (lock_ctx->ctdb_db) {
 454                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 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->type == LOCK_RECORD || lock_ctx->type == LOCK_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 /*
 660  * Schedule a new lock child process
 661  * Set up callback handler and timeout handler
 662  */
 663 static void ctdb_lock_schedule(struct ctdb_context *ctdb)
 664 {
 665         struct lock_context *lock_ctx, *next_ctx;
 666         int ret;
 667         TALLOC_CTX *tmp_ctx;
 668         const char *helper = BINDIR "/ctdb_lock_helper";
 669         static const char *prog = NULL;
 670         char **args;
 671
 672         if (prog == NULL) {
 673                 const char *t;
 674
 675                 t = getenv("CTDB_LOCK_HELPER");
 676                 if (t != NULL) {
 677                         prog = talloc_strdup(ctdb, t);
 678                 } else {
 679                         prog = talloc_strdup(ctdb, helper);
 680                 }
 681                 CTDB_NO_MEMORY_VOID(ctdb, prog);
 682         }
 683
 684         if (ctdb->lock_pending == NULL) {
 685                 return;
 686         }
 687
 688         /* Find a lock context with requests */
 689         lock_ctx = ctdb->lock_pending;
 690         while (lock_ctx != NULL) {
 691                 next_ctx = lock_ctx->next;
 692                 if (! lock_ctx->request) {
 693                         DEBUG(DEBUG_INFO, ("Removing lock context without lock request\n"));
 694                         DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 695                         ctdb->lock_num_pending--;
 696                         CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
 697                         if (lock_ctx->ctdb_db) {
 698                                 CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
 699                         }
 700                         talloc_free(lock_ctx);
 701                 } else {
 702                         if (lock_ctx->ctdb_db == NULL ||
 703                             lock_ctx->ctdb_db->lock_num_current < MAX_LOCK_PROCESSES_PER_DB) {
 704                                 /* Found a lock context with lock requests */
 705                                 break;
 706                         }
 707                 }
 708                 lock_ctx = next_ctx;
 709         }
 710
 711         if (lock_ctx == NULL) {
 712                 return;
 713         }
 714
 715         lock_ctx->child = -1;
 716         ret = pipe(lock_ctx->fd);
 717         if (ret != 0) {
 718                 DEBUG(DEBUG_ERR, ("Failed to create pipe in ctdb_lock_schedule\n"));
 719                 return;
 720         }
 721
 722         set_close_on_exec(lock_ctx->fd[0]);
 723
 724         /* Create data for child process */
 725         tmp_ctx = talloc_new(lock_ctx);
 726         if (tmp_ctx == NULL) {
 727                 DEBUG(DEBUG_ERR, ("Failed to allocate memory for helper args\n"));
 728                 close(lock_ctx->fd[0]);
 729                 close(lock_ctx->fd[1]);
 730                 return;
 731         }
 732
 733         /* Create arguments for lock helper */
 734         args = lock_helper_args(tmp_ctx, lock_ctx, lock_ctx->fd[1]);
 735         if (args == NULL) {
 736                 DEBUG(DEBUG_ERR, ("Failed to create lock helper args\n"));
 737                 close(lock_ctx->fd[0]);
 738                 close(lock_ctx->fd[1]);
 739                 talloc_free(tmp_ctx);
 740                 return;
 741         }
 742
 743         lock_ctx->child = vfork();
 744
 745         if (lock_ctx->child == (pid_t)-1) {
 746                 DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
 747                 close(lock_ctx->fd[0]);
 748                 close(lock_ctx->fd[1]);
 749                 talloc_free(tmp_ctx);
 750                 return;
 751         }
 752
 753
 754         /* Child process */
 755         if (lock_ctx->child == 0) {
 756                 ret = execv(prog, args);
 757                 if (ret < 0) {
 758                         DEBUG(DEBUG_ERR, ("Failed to execute helper %s (%d, %s)\n",
 759                                           prog, errno, strerror(errno)));
 760                 }
 761                 _exit(1);
 762         }
 763
 764         /* Parent process */
 765         ctdb_track_child(ctdb, lock_ctx->child);
 766         close(lock_ctx->fd[1]);
 767
 768         talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
 769
 770         talloc_free(tmp_ctx);
 771
 772         /* Set up timeout handler */
 773         lock_ctx->ttimer = tevent_add_timer(ctdb->ev,
 774                                             lock_ctx,
 775                                             timeval_current_ofs(10, 0),
 776                                             ctdb_lock_timeout_handler,
 777                                             (void *)lock_ctx);
 778         if (lock_ctx->ttimer == NULL) {
 779                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 780                 lock_ctx->child = -1;
 781                 talloc_set_destructor(lock_ctx, NULL);
 782                 close(lock_ctx->fd[0]);
 783                 return;
 784         }
 785
 786         /* Set up callback */
 787         lock_ctx->tfd = tevent_add_fd(ctdb->ev,
 788                                       lock_ctx,
 789                                       lock_ctx->fd[0],
 790                                       EVENT_FD_READ,
 791                                       ctdb_lock_handler,
 792                                       (void *)lock_ctx);
 793         if (lock_ctx->tfd == NULL) {
 794                 TALLOC_FREE(lock_ctx->ttimer);
 795                 ctdb_kill(ctdb, lock_ctx->child, SIGKILL);
 796                 lock_ctx->child = -1;
 797                 talloc_set_destructor(lock_ctx, NULL);
 798                 close(lock_ctx->fd[0]);
 799                 return;
 800         }
 801         tevent_fd_set_auto_close(lock_ctx->tfd);
 802
 803         /* Move the context from pending to current */
 804         DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
 805         ctdb->lock_num_pending--;
 806         DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
 807         if (lock_ctx->ctdb_db) {
 808                 lock_ctx->ctdb_db->lock_num_current++;
 809                 CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
 810                 CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
 811         }
 812 }
 813
 814
 815 /*
 816  * Lock record / db depending on type
 817  */
 818 static struct lock_request *ctdb_lock_internal(struct ctdb_context *ctdb,
 819                                                struct ctdb_db_context *ctdb_db,
 820                                                TDB_DATA key,
 821                                                uint32_t priority,
 822                                                void (*callback)(void *, bool),
 823                                                void *private_data,
 824                                                enum lock_type type,
 825                                                bool auto_mark)
 826 {
 827         struct lock_context *lock_ctx = NULL;
 828         struct lock_request *request;
 829
 830         if (callback == NULL) {
 831                 DEBUG(DEBUG_WARNING, ("No callback function specified, not locking\n"));
 832                 return NULL;
 833         }
 834
 835         lock_ctx = talloc_zero(ctdb, struct lock_context);
 836         if (lock_ctx == NULL) {
 837                 DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
 838                 return NULL;
 839         }
 840
 841         if ((request = talloc_zero(lock_ctx, struct lock_request)) == NULL) {
 842                 talloc_free(lock_ctx);
 843                 return NULL;
 844         }
 845
 846         lock_ctx->type = type;
 847         lock_ctx->ctdb = ctdb;
 848         lock_ctx->ctdb_db = ctdb_db;
 849         lock_ctx->key.dsize = key.dsize;
 850         if (key.dsize > 0) {
 851                 lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
 852                 if (lock_ctx->key.dptr == NULL) {
 853                         DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
 854                         talloc_free(lock_ctx);
 855                         return NULL;
 856                 }
 857                 lock_ctx->key_hash = ctdb_hash(&key);
 858         } else {
 859                 lock_ctx->key.dptr = NULL;
 860         }
 861         lock_ctx->priority = priority;
 862         lock_ctx->auto_mark = auto_mark;
 863
 864         lock_ctx->request = request;
 865         lock_ctx->child = -1;
 866
 867         DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
 868         ctdb->lock_num_pending++;
 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