source3/lib/g_lock.c

   1 /*
   2    Unix SMB/CIFS implementation.
   3    global locks based on dbwrap and messaging
   4    Copyright (C) 2009 by Volker Lendecke
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18 */
  19
  20 #include "includes.h"
  21 #include "g_lock.h"
  22
  23 static NTSTATUS g_lock_force_unlock(struct g_lock_ctx *ctx, const char *name,
  24                                     struct server_id pid);
  25
  26 struct g_lock_ctx {
  27         struct db_context *db;
  28         struct messaging_context *msg;
  29 };
  30
  31 /*
  32  * The "g_lock.tdb" file contains records, indexed by the 0-terminated
  33  * lockname. The record contains an array of "struct g_lock_rec"
  34  * structures. Waiters have the lock_type with G_LOCK_PENDING or'ed.
  35  */
  36
  37 struct g_lock_rec {
  38         enum g_lock_type lock_type;
  39         struct server_id pid;
  40 };
  41
  42 struct g_lock_ctx *g_lock_ctx_init(TALLOC_CTX *mem_ctx,
  43                                    struct messaging_context *msg)
  44 {
  45         struct g_lock_ctx *result;
  46
  47         result = talloc(mem_ctx, struct g_lock_ctx);
  48         if (result == NULL) {
  49                 return NULL;
  50         }
  51         result->msg = msg;
  52
  53         result->db = db_open(result, lock_path("g_lock.tdb"), 0,
  54                              TDB_CLEAR_IF_FIRST, O_RDWR|O_CREAT, 0700);
  55         if (result->db == NULL) {
  56                 DEBUG(1, ("g_lock_init: Could not open g_lock.tdb"));
  57                 TALLOC_FREE(result);
  58                 return NULL;
  59         }
  60         return result;
  61 }
  62
  63 static bool g_lock_conflicts(enum g_lock_type lock_type,
  64                              const struct g_lock_rec *rec)
  65 {
  66         enum g_lock_type rec_lock = rec->lock_type;
  67
  68         if ((rec_lock & G_LOCK_PENDING) != 0) {
  69                 return false;
  70         }
  71
  72         /*
  73          * Only tested write locks so far. Very likely this routine
  74          * needs to be fixed for read locks....
  75          */
  76         if ((lock_type == G_LOCK_READ) && (rec_lock == G_LOCK_READ)) {
  77                 return false;
  78         }
  79         return true;
  80 }
  81
  82 static bool g_lock_parse(TALLOC_CTX *mem_ctx, TDB_DATA data,
  83                          int *pnum_locks, struct g_lock_rec **plocks)
  84 {
  85         int i, num_locks;
  86         struct g_lock_rec *locks;
  87
  88         if ((data.dsize % sizeof(struct g_lock_rec)) != 0) {
  89                 DEBUG(1, ("invalid lock record length %d\n", (int)data.dsize));
  90                 return false;
  91         }
  92
  93         num_locks = data.dsize / sizeof(struct g_lock_rec);
  94         locks = talloc_array(mem_ctx, struct g_lock_rec, num_locks);
  95         if (locks == NULL) {
  96                 DEBUG(1, ("talloc failed\n"));
  97                 return false;
  98         }
  99
 100         memcpy(locks, data.dptr, data.dsize);
 101
 102         DEBUG(10, ("locks:\n"));
 103         for (i=0; i<num_locks; i++) {
 104                 DEBUGADD(10, ("%s: %s %s\n",
 105                               procid_str(talloc_tos(), &locks[i].pid),
 106                               ((locks[i].lock_type & 1) == G_LOCK_READ) ?
 107                               "read" : "write",
 108                               (locks[i].lock_type & G_LOCK_PENDING) ?
 109                               "(pending)" : "(owner)"));
 110
 111                 if (process_exists(locks[i].pid)) {
 112                         continue;
 113                 }
 114                 DEBUGADD(10, ("%s does not exist -- discarding\n",
 115                               procid_str(talloc_tos(), &locks[i].pid)));
 116
 117                 if (i < (num_locks-1)) {
 118                         locks[i] = locks[num_locks-1];
 119                 }
 120                 num_locks -= 1;
 121         }
 122
 123         *plocks = locks;
 124         *pnum_locks = num_locks;
 125         return true;
 126 }
 127
 128 static struct g_lock_rec *g_lock_addrec(TALLOC_CTX *mem_ctx,
 129                                         struct g_lock_rec *locks,
 130                                         int num_locks,
 131                                         const struct server_id pid,
 132                                         enum g_lock_type lock_type)
 133 {
 134         struct g_lock_rec *result;
 135
 136         result = talloc_realloc(mem_ctx, locks, struct g_lock_rec,
 137                                 num_locks+1);
 138         if (result == NULL) {
 139                 return NULL;
 140         }
 141
 142         result[num_locks].pid = pid;
 143         result[num_locks].lock_type = lock_type;
 144         return result;
 145 }
 146
 147 static void g_lock_got_retry(struct messaging_context *msg,
 148                              void *private_data,
 149                              uint32_t msg_type,
 150                              struct server_id server_id,
 151                              DATA_BLOB *data);
 152 static void g_lock_timedout(struct tevent_context *ev,
 153                             struct tevent_timer *te,
 154                             struct timeval current_time,
 155                             void *private_data);
 156
 157 static NTSTATUS g_lock_trylock(struct g_lock_ctx *ctx, const char *name,
 158                                enum g_lock_type lock_type)
 159 {
 160         struct db_record *rec = NULL;
 161         struct g_lock_rec *locks = NULL;
 162         int i, num_locks;
 163         struct server_id self;
 164         int our_index;
 165         TDB_DATA data;
 166         NTSTATUS status = NT_STATUS_OK;
 167         NTSTATUS store_status;
 168
 169 again:
 170         rec = ctx->db->fetch_locked(ctx->db, talloc_tos(),
 171                                     string_term_tdb_data(name));
 172         if (rec == NULL) {
 173                 DEBUG(10, ("fetch_locked(\"%s\") failed\n", name));
 174                 status = NT_STATUS_LOCK_NOT_GRANTED;
 175                 goto done;
 176         }
 177
 178         if (!g_lock_parse(talloc_tos(), rec->value, &num_locks, &locks)) {
 179                 DEBUG(10, ("g_lock_parse for %s failed\n", name));
 180                 status = NT_STATUS_INTERNAL_ERROR;
 181                 goto done;
 182         }
 183
 184         self = procid_self();
 185         our_index = -1;
 186
 187         for (i=0; i<num_locks; i++) {
 188                 if (procid_equal(&self, &locks[i].pid)) {
 189                         if (our_index != -1) {
 190                                 DEBUG(1, ("g_lock_trylock: Added ourself "
 191                                           "twice!\n"));
 192                                 status = NT_STATUS_INTERNAL_ERROR;
 193                                 goto done;
 194                         }
 195                         if ((locks[i].lock_type & G_LOCK_PENDING) == 0) {
 196                                 DEBUG(1, ("g_lock_trylock: Found ourself not "
 197                                           "pending!\n"));
 198                                 status = NT_STATUS_INTERNAL_ERROR;
 199                                 goto done;
 200                         }
 201
 202                         our_index = i;
 203
 204                         /* never conflict with ourself */
 205                         continue;
 206                 }
 207                 if (g_lock_conflicts(lock_type, &locks[i])) {
 208                         struct server_id pid = locks[i].pid;
 209
 210                         if (!process_exists(pid)) {
 211                                 TALLOC_FREE(locks);
 212                                 TALLOC_FREE(rec);
 213                                 status = g_lock_force_unlock(ctx, name, pid);
 214                                 if (!NT_STATUS_IS_OK(status)) {
 215                                         DEBUG(1, ("Could not unlock dead lock "
 216                                                   "holder!\n"));
 217                                         goto done;
 218                                 }
 219                                 goto again;
 220                         }
 221                         lock_type |= G_LOCK_PENDING;
 222                 }
 223         }
 224
 225         if (our_index == -1) {
 226                 /* First round, add ourself */
 227
 228                 locks = g_lock_addrec(talloc_tos(), locks, num_locks,
 229                                       self, lock_type);
 230                 if (locks == NULL) {
 231                         DEBUG(10, ("g_lock_addrec failed\n"));
 232                         status = NT_STATUS_NO_MEMORY;
 233                         goto done;
 234                 }
 235         } else {
 236                 /*
 237                  * Retry. We were pending last time. Overwrite the
 238                  * stored lock_type with what we calculated, we might
 239                  * have acquired the lock this time.
 240                  */
 241                 locks[our_index].lock_type = lock_type;
 242         }
 243
 244         data = make_tdb_data((uint8_t *)locks, talloc_get_size(locks));
 245         store_status = rec->store(rec, data, 0);
 246         if (!NT_STATUS_IS_OK(store_status)) {
 247                 DEBUG(1, ("rec->store failed: %s\n",
 248                           nt_errstr(store_status)));
 249                 status = store_status;
 250         }
 251
 252 done:
 253         TALLOC_FREE(locks);
 254         TALLOC_FREE(rec);
 255
 256         if (NT_STATUS_IS_OK(status) && (lock_type & G_LOCK_PENDING) != 0) {
 257                 return STATUS_PENDING;
 258         }
 259
 260         return NT_STATUS_OK;
 261 }
 262
 263 NTSTATUS g_lock_lock(struct g_lock_ctx *ctx, const char *name,
 264                      enum g_lock_type lock_type, struct timeval timeout)
 265 {
 266         struct tevent_timer *te = NULL;
 267         NTSTATUS status;
 268         bool retry = false;
 269         bool timedout = false;
 270
 271         DEBUG(10, ("Trying to acquire lock %d for %s\n", (int)lock_type,
 272                    name));
 273
 274         if (lock_type & ~1) {
 275                 DEBUG(1, ("Got invalid lock type %d for %s\n",
 276                           (int)lock_type, name));
 277                 return NT_STATUS_INVALID_PARAMETER;
 278         }
 279
 280 #ifdef CLUSTER_SUPPORT
 281         if (lp_clustering()) {
 282                 status = ctdb_watch_us(messaging_ctdbd_connection());
 283                 if (!NT_STATUS_IS_OK(status)) {
 284                         DEBUG(10, ("could not register retry with ctdb: %s\n",
 285                                    nt_errstr(status)));
 286                         goto done;
 287                 }
 288         }
 289 #endif
 290
 291         status = messaging_register(ctx->msg, &retry, MSG_DBWRAP_G_LOCK_RETRY,
 292                                     g_lock_got_retry);
 293         if (!NT_STATUS_IS_OK(status)) {
 294                 DEBUG(10, ("messaging_register failed: %s\n",
 295                            nt_errstr(status)));
 296                 return status;
 297         }
 298 again:
 299         retry = false;
 300
 301         status = g_lock_trylock(ctx, name, lock_type);
 302         if (NT_STATUS_IS_OK(status)) {
 303                 DEBUG(10, ("Got lock %s\n", name));
 304                 goto done;
 305         }
 306         if (!NT_STATUS_EQUAL(status, STATUS_PENDING)) {
 307                 DEBUG(10, ("g_lock_trylock failed: %s\n",
 308                            nt_errstr(status)));
 309                 goto done;
 310         }
 311
 312         DEBUG(10, ("g_lock_trylock: Did not get lock, waiting...\n"));
 313
 314         if (te == NULL) {
 315                 te = tevent_add_timer(
 316                         ctx->msg->event_ctx, talloc_tos(),
 317                         timeval_current_ofs(timeout.tv_sec, timeout.tv_usec),
 318                         g_lock_timedout, &timedout);
 319                 if (te == NULL) {
 320                         DEBUG(10, ("tevent_add_timer failed\n"));
 321                         status = NT_STATUS_NO_MEMORY;
 322                         goto done;
 323                 }
 324         }
 325
 326         while (true) {
 327                 if (tevent_loop_once(ctx->msg->event_ctx) == -1) {
 328                         DEBUG(1, ("tevent_loop_once failed\n"));
 329                         status = NT_STATUS_INTERNAL_ERROR;
 330                         goto done;
 331                 }
 332                 if (retry) {
 333                         goto again;
 334                 }
 335                 if (timedout) {
 336                         DEBUG(10, ("g_lock_lock timed out\n"));
 337
 338                         te = NULL;
 339
 340                         status = NT_STATUS_LOCK_NOT_GRANTED;
 341                         goto done;
 342                 }
 343         }
 344 done:
 345
 346         if (!NT_STATUS_IS_OK(status)) {
 347                 NTSTATUS unlock_status;
 348
 349                 unlock_status = g_lock_unlock(ctx, name);
 350
 351                 if (!NT_STATUS_IS_OK(unlock_status)) {
 352                         DEBUG(1, ("Could not remove ourself from the locking "
 353                                   "db: %s\n", nt_errstr(status)));
 354                 }
 355         }
 356
 357         messaging_deregister(ctx->msg, MSG_DBWRAP_G_LOCK_RETRY, &retry);
 358         TALLOC_FREE(te);
 359
 360         return status;
 361 }
 362
 363 static void g_lock_got_retry(struct messaging_context *msg,
 364                              void *private_data,
 365                              uint32_t msg_type,
 366                              struct server_id server_id,
 367                              DATA_BLOB *data)
 368 {
 369         bool *pretry = (bool *)private_data;
 370
 371         DEBUG(10, ("Got retry message from pid %s\n",
 372                    procid_str(talloc_tos(), &server_id)));
 373
 374         *pretry = true;
 375 }
 376
 377 static void g_lock_timedout(struct tevent_context *ev,
 378                             struct tevent_timer *te,
 379                             struct timeval current_time,
 380                             void *private_data)
 381 {
 382         bool *ptimedout = (bool *)private_data;
 383         *ptimedout = true;
 384         TALLOC_FREE(te);
 385 }
 386
 387 static NTSTATUS g_lock_force_unlock(struct g_lock_ctx *ctx, const char *name,
 388                                     struct server_id pid)
 389 {
 390         struct db_record *rec = NULL;
 391         struct g_lock_rec *locks = NULL;
 392         int i, num_locks;
 393         enum g_lock_type lock_type;
 394         NTSTATUS status;
 395
 396         rec = ctx->db->fetch_locked(ctx->db, talloc_tos(),
 397                                     string_term_tdb_data(name));
 398         if (rec == NULL) {
 399                 DEBUG(10, ("fetch_locked(\"%s\") failed\n", name));
 400                 status = NT_STATUS_INTERNAL_ERROR;
 401                 goto done;
 402         }
 403
 404         if (!g_lock_parse(talloc_tos(), rec->value, &num_locks, &locks)) {
 405                 DEBUG(10, ("g_lock_parse for %s failed\n", name));
 406                 status = NT_STATUS_INTERNAL_ERROR;
 407                 goto done;
 408         }
 409
 410         for (i=0; i<num_locks; i++) {
 411                 if (procid_equal(&pid, &locks[i].pid)) {
 412                         break;
 413                 }
 414         }
 415
 416         if (i == num_locks) {
 417                 DEBUG(10, ("g_lock_force_unlock: Lock not found\n"));
 418                 status = NT_STATUS_INTERNAL_ERROR;
 419                 goto done;
 420         }
 421
 422         lock_type = locks[i].lock_type;
 423
 424         if (i < (num_locks-1)) {
 425                 locks[i] = locks[num_locks-1];
 426         }
 427         num_locks -= 1;
 428
 429         if (num_locks == 0) {
 430                 status = rec->delete_rec(rec);
 431         } else {
 432                 TDB_DATA data;
 433                 data = make_tdb_data((uint8_t *)locks,
 434                                      sizeof(struct g_lock_rec) * num_locks);
 435                 status = rec->store(rec, data, 0);
 436         }
 437
 438         if (!NT_STATUS_IS_OK(status)) {
 439                 DEBUG(1, ("g_lock_force_unlock: Could not store record: %s\n",
 440                           nt_errstr(status)));
 441                 goto done;
 442         }
 443
 444         if ((lock_type & G_LOCK_PENDING) == 0) {
 445                 /*
 446                  * We've been the lock holder. Tell all others to retry.
 447                  */
 448                 for (i=0; i<num_locks; i++) {
 449                         if ((locks[i].lock_type & G_LOCK_PENDING) == 0) {
 450                                 continue;
 451                         }
 452
 453                         /*
 454                          * Ping all waiters to retry
 455                          */
 456                         status = messaging_send(ctx->msg, locks[i].pid,
 457                                                 MSG_DBWRAP_G_LOCK_RETRY,
 458                                                 &data_blob_null);
 459                         if (!NT_STATUS_IS_OK(status)) {
 460                                 DEBUG(1, ("sending retry to %s failed: %s\n",
 461                                           procid_str(talloc_tos(),
 462                                                      &locks[i].pid),
 463                                           nt_errstr(status)));
 464                         }
 465                 }
 466         }
 467 done:
 468
 469         TALLOC_FREE(locks);
 470         TALLOC_FREE(rec);
 471         return status;
 472 }
 473
 474 NTSTATUS g_lock_unlock(struct g_lock_ctx *ctx, const char *name)
 475 {
 476         NTSTATUS status;
 477
 478         status = g_lock_force_unlock(ctx, name, procid_self());
 479
 480 #ifdef CLUSTER_SUPPORT
 481         if (lp_clustering()) {
 482                 ctdb_unwatch(messaging_ctdbd_connection());
 483         }
 484 #endif
 485         return status;
 486 }
 487
 488 struct g_lock_locks_state {
 489         int (*fn)(const char *name, void *private_data);
 490         void *private_data;
 491 };
 492
 493 static int g_lock_locks_fn(struct db_record *rec, void *priv)
 494 {
 495         struct g_lock_locks_state *state = (struct g_lock_locks_state *)priv;
 496
 497         if ((rec->key.dsize == 0) || (rec->key.dptr[rec->key.dsize-1] != 0)) {
 498                 DEBUG(1, ("invalid key in g_lock.tdb, ignoring\n"));
 499                 return 0;
 500         }
 501         return state->fn((char *)rec->key.dptr, state->private_data);
 502 }
 503
 504 int g_lock_locks(struct g_lock_ctx *ctx,
 505                  int (*fn)(const char *name, void *private_data),
 506                  void *private_data)
 507 {
 508         struct g_lock_locks_state state;
 509
 510         state.fn = fn;
 511         state.private_data = private_data;
 512
 513         return ctx->db->traverse_read(ctx->db, g_lock_locks_fn, &state);
 514 }
 515
 516 NTSTATUS g_lock_dump(struct g_lock_ctx *ctx, const char *name,
 517                      int (*fn)(struct server_id pid,
 518                                enum g_lock_type lock_type,
 519                                void *private_data),
 520                      void *private_data)
 521 {
 522         TDB_DATA data;
 523         int i, num_locks;
 524         struct g_lock_rec *locks = NULL;
 525         bool ret;
 526
 527         if (ctx->db->fetch(ctx->db, talloc_tos(), string_term_tdb_data(name),
 528                            &data) != 0) {
 529                 return NT_STATUS_NOT_FOUND;
 530         }
 531
 532         if ((data.dsize == 0) || (data.dptr == NULL)) {
 533                 return NT_STATUS_OK;
 534         }
 535
 536         ret = g_lock_parse(talloc_tos(), data, &num_locks, &locks);
 537
 538         TALLOC_FREE(data.dptr);
 539
 540         if (!ret) {
 541                 DEBUG(10, ("g_lock_parse for %s failed\n", name));
 542                 return NT_STATUS_INTERNAL_ERROR;
 543         }
 544
 545         for (i=0; i<num_locks; i++) {
 546                 if (fn(locks[i].pid, locks[i].lock_type, private_data) != 0) {
 547                         break;
 548                 }
 549         }
 550         TALLOC_FREE(locks);
 551         return NT_STATUS_OK;
 552 }
 553
 554 struct g_lock_get_state {
 555         bool found;
 556         struct server_id *pid;
 557 };
 558
 559 static int g_lock_get_fn(struct server_id pid, enum g_lock_type lock_type,
 560                          void *priv)
 561 {
 562         struct g_lock_get_state *state = (struct g_lock_get_state *)priv;
 563
 564         if ((lock_type & G_LOCK_PENDING) != 0) {
 565                 return 0;
 566         }
 567
 568         state->found = true;
 569         *state->pid = pid;
 570         return 1;
 571 }
 572
 573 NTSTATUS g_lock_get(struct g_lock_ctx *ctx, const char *name,
 574                     struct server_id *pid)
 575 {
 576         struct g_lock_get_state state;
 577         NTSTATUS status;
 578
 579         state.found = false;
 580         state.pid = pid;
 581
 582         status = g_lock_dump(ctx, name, g_lock_get_fn, &state);
 583         if (!NT_STATUS_IS_OK(status)) {
 584                 return status;
 585         }
 586         if (!state.found) {
 587                 return NT_STATUS_NOT_FOUND;
 588         }
 589         return NT_STATUS_OK;
 590 }