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
 153 static NTSTATUS g_lock_trylock(struct g_lock_ctx *ctx, const char *name,
 154                                enum g_lock_type lock_type)
 155 {
 156         struct db_record *rec = NULL;
 157         struct g_lock_rec *locks = NULL;
 158         int i, num_locks;
 159         struct server_id self;
 160         int our_index;
 161         TDB_DATA data;
 162         NTSTATUS status = NT_STATUS_OK;
 163         NTSTATUS store_status;
 164
 165 again:
 166         rec = ctx->db->fetch_locked(ctx->db, talloc_tos(),
 167                                     string_term_tdb_data(name));
 168         if (rec == NULL) {
 169                 DEBUG(10, ("fetch_locked(\"%s\") failed\n", name));
 170                 status = NT_STATUS_LOCK_NOT_GRANTED;
 171                 goto done;
 172         }
 173
 174         if (!g_lock_parse(talloc_tos(), rec->value, &num_locks, &locks)) {
 175                 DEBUG(10, ("g_lock_parse for %s failed\n", name));
 176                 status = NT_STATUS_INTERNAL_ERROR;
 177                 goto done;
 178         }
 179
 180         self = procid_self();
 181         our_index = -1;
 182
 183         for (i=0; i<num_locks; i++) {
 184                 if (procid_equal(&self, &locks[i].pid)) {
 185                         if (our_index != -1) {
 186                                 DEBUG(1, ("g_lock_trylock: Added ourself "
 187                                           "twice!\n"));
 188                                 status = NT_STATUS_INTERNAL_ERROR;
 189                                 goto done;
 190                         }
 191                         if ((locks[i].lock_type & G_LOCK_PENDING) == 0) {
 192                                 DEBUG(1, ("g_lock_trylock: Found ourself not "
 193                                           "pending!\n"));
 194                                 status = NT_STATUS_INTERNAL_ERROR;
 195                                 goto done;
 196                         }
 197
 198                         our_index = i;
 199
 200                         /* never conflict with ourself */
 201                         continue;
 202                 }
 203                 if (g_lock_conflicts(lock_type, &locks[i])) {
 204                         struct server_id pid = locks[i].pid;
 205
 206                         if (!process_exists(pid)) {
 207                                 TALLOC_FREE(locks);
 208                                 TALLOC_FREE(rec);
 209                                 status = g_lock_force_unlock(ctx, name, pid);
 210                                 if (!NT_STATUS_IS_OK(status)) {
 211                                         DEBUG(1, ("Could not unlock dead lock "
 212                                                   "holder!\n"));
 213                                         goto done;
 214                                 }
 215                                 goto again;
 216                         }
 217                         lock_type |= G_LOCK_PENDING;
 218                 }
 219         }
 220
 221         if (our_index == -1) {
 222                 /* First round, add ourself */
 223
 224                 locks = g_lock_addrec(talloc_tos(), locks, num_locks,
 225                                       self, lock_type);
 226                 if (locks == NULL) {
 227                         DEBUG(10, ("g_lock_addrec failed\n"));
 228                         status = NT_STATUS_NO_MEMORY;
 229                         goto done;
 230                 }
 231         } else {
 232                 /*
 233                  * Retry. We were pending last time. Overwrite the
 234                  * stored lock_type with what we calculated, we might
 235                  * have acquired the lock this time.
 236                  */
 237                 locks[our_index].lock_type = lock_type;
 238         }
 239
 240         data = make_tdb_data((uint8_t *)locks, talloc_get_size(locks));
 241         store_status = rec->store(rec, data, 0);
 242         if (!NT_STATUS_IS_OK(store_status)) {
 243                 DEBUG(1, ("rec->store failed: %s\n",
 244                           nt_errstr(store_status)));
 245                 status = store_status;
 246         }
 247
 248 done:
 249         TALLOC_FREE(locks);
 250         TALLOC_FREE(rec);
 251
 252         if (NT_STATUS_IS_OK(status) && (lock_type & G_LOCK_PENDING) != 0) {
 253                 return STATUS_PENDING;
 254         }
 255
 256         return NT_STATUS_OK;
 257 }
 258
 259 NTSTATUS g_lock_lock(struct g_lock_ctx *ctx, const char *name,
 260                      enum g_lock_type lock_type, struct timeval timeout)
 261 {
 262         struct tevent_timer *te = NULL;
 263         NTSTATUS status;
 264         bool retry = false;
 265         struct timeval timeout_end;
 266         struct timeval timeout_remaining;
 267         struct timeval time_now;
 268
 269         DEBUG(10, ("Trying to acquire lock %d for %s\n", (int)lock_type,
 270                    name));
 271
 272         if (lock_type & ~1) {
 273                 DEBUG(1, ("Got invalid lock type %d for %s\n",
 274                           (int)lock_type, name));
 275                 return NT_STATUS_INVALID_PARAMETER;
 276         }
 277
 278 #ifdef CLUSTER_SUPPORT
 279         if (lp_clustering()) {
 280                 status = ctdb_watch_us(messaging_ctdbd_connection());
 281                 if (!NT_STATUS_IS_OK(status)) {
 282                         DEBUG(10, ("could not register retry with ctdb: %s\n",
 283                                    nt_errstr(status)));
 284                         goto done;
 285                 }
 286         }
 287 #endif
 288
 289         status = messaging_register(ctx->msg, &retry, MSG_DBWRAP_G_LOCK_RETRY,
 290                                     g_lock_got_retry);
 291         if (!NT_STATUS_IS_OK(status)) {
 292                 DEBUG(10, ("messaging_register failed: %s\n",
 293                            nt_errstr(status)));
 294                 return status;
 295         }
 296
 297         time_now = timeval_current();
 298         timeout_end = timeval_sum(&time_now, &timeout);
 299
 300         while (true) {
 301 #ifdef CLUSTER_SUPPORT
 302                 fd_set _r_fds;
 303 #endif
 304                 fd_set *r_fds = NULL;
 305                 int max_fd = 0;
 306                 int ret;
 307
 308                 status = g_lock_trylock(ctx, name, lock_type);
 309                 if (NT_STATUS_IS_OK(status)) {
 310                         DEBUG(10, ("Got lock %s\n", name));
 311                         break;
 312                 }
 313                 if (!NT_STATUS_EQUAL(status, STATUS_PENDING)) {
 314                         DEBUG(10, ("g_lock_trylock failed: %s\n",
 315                                    nt_errstr(status)));
 316                         break;
 317                 }
 318
 319                 DEBUG(10, ("g_lock_trylock: Did not get lock, waiting...\n"));
 320
 321                 /* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 322                  *             !!! HACK ALERT --- FIX ME !!!
 323                  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 324                  * What we really want to do here is to react to
 325                  * MSG_DBWRAP_G_LOCK_RETRY messages that are either sent
 326                  * by a client doing g_lock_unlock or by ourselves when
 327                  * we receive a CTDB_SRVID_SAMBA_NOTIFY or
 328                  * CTDB_SRVID_RECONFIGURE message from ctdbd, i.e. when
 329                  * either a client holding a lock or a complete node
 330                  * has died.
 331                  *
 332                  * Doing this properly involves calling tevent_loop_once(),
 333                  * but doing this here with the main ctdbd messaging context
 334                  * creates a nested event loop when g_lock_lock() is called
 335                  * from the main event loop, e.g. in a tcon_and_X where the
 336                  * share_info.tdb needs to be initialized and is locked by
 337                  * another process, or when the remore registry is accessed
 338                  * for writing and some other process already holds a lock
 339                  * on the registry.tdb.
 340                  *
 341                  * So as a quick fix, we act a little coarsely here: we do
 342                  * a select on the ctdb connection fd and when it is readable
 343                  * or we get EINTR, then we retry without actually parsing
 344                  * any ctdb packages or dispatching messages. This means that
 345                  * we retry more often than intended by design, but this does
 346                  * not harm and it is unobtrusive. When we have finished,
 347                  * the main loop will pick up all the messages and ctdb
 348                  * packets. The only extra twist is that we cannot use timed
 349                  * events here but have to handcode a timeout.
 350                  */
 351
 352 #ifdef CLUSTER_SUPPORT
 353                 if (lp_clustering()) {
 354                         struct ctdbd_connection *conn = messaging_ctdbd_connection();
 355
 356                         r_fds = &_r_fds;
 357                         FD_ZERO(r_fds);
 358                         max_fd = ctdbd_conn_get_fd(conn);
 359                         FD_SET(max_fd, r_fds);
 360                 }
 361 #endif
 362
 363                 time_now = timeval_current();
 364                 timeout_remaining = timeval_until(&time_now, &timeout_end);
 365
 366                 ret = sys_select(max_fd + 1, r_fds, NULL, NULL,
 367                                  &timeout_remaining);
 368
 369                 if (ret == -1) {
 370                         if (errno != EINTR) {
 371                                 DEBUG(1, ("error calling select: %s\n",
 372                                           strerror(errno)));
 373                                 status = NT_STATUS_INTERNAL_ERROR;
 374                                 break;
 375                         }
 376                         /*
 377                          * errno == EINTR:
 378                          * This means a signal was received.
 379                          * It might have been a MSG_DBWRAP_G_LOCK_RETRY message.
 380                          * ==> retry
 381                          */
 382                 } else if (ret == 0) {
 383                         if (timeval_expired(&timeout_end)) {
 384                                 DEBUG(10, ("g_lock_lock timed out\n"));
 385                                 status = NT_STATUS_LOCK_NOT_GRANTED;
 386                                 break;
 387                         } else {
 388                                 DEBUG(10, ("select returned 0 but timeout not "
 389                                            "not expired: strange - retrying\n"));
 390                         }
 391                 } else if (ret != 1) {
 392                         DEBUG(1, ("invalid return code of select: %d\n", ret));
 393                         status = NT_STATUS_INTERNAL_ERROR;
 394                         break;
 395                 }
 396                 /*
 397                  * ret == 1:
 398                  * This means ctdbd has sent us some data.
 399                  * Might be a CTDB_SRVID_RECONFIGURE or a
 400                  * CTDB_SRVID_SAMBA_NOTIFY message.
 401                  * ==> retry
 402                  */
 403         }
 404
 405 #ifdef CLUSTER_SUPPORT
 406 done:
 407 #endif
 408
 409         if (!NT_STATUS_IS_OK(status)) {
 410                 NTSTATUS unlock_status;
 411
 412                 unlock_status = g_lock_unlock(ctx, name);
 413
 414                 if (!NT_STATUS_IS_OK(unlock_status)) {
 415                         DEBUG(1, ("Could not remove ourself from the locking "
 416                                   "db: %s\n", nt_errstr(status)));
 417                 }
 418         }
 419
 420         messaging_deregister(ctx->msg, MSG_DBWRAP_G_LOCK_RETRY, &retry);
 421         TALLOC_FREE(te);
 422
 423         return status;
 424 }
 425
 426 static void g_lock_got_retry(struct messaging_context *msg,
 427                              void *private_data,
 428                              uint32_t msg_type,
 429                              struct server_id server_id,
 430                              DATA_BLOB *data)
 431 {
 432         bool *pretry = (bool *)private_data;
 433
 434         DEBUG(10, ("Got retry message from pid %s\n",
 435                    procid_str(talloc_tos(), &server_id)));
 436
 437         *pretry = true;
 438 }
 439
 440 static NTSTATUS g_lock_force_unlock(struct g_lock_ctx *ctx, const char *name,
 441                                     struct server_id pid)
 442 {
 443         struct db_record *rec = NULL;
 444         struct g_lock_rec *locks = NULL;
 445         int i, num_locks;
 446         enum g_lock_type lock_type;
 447         NTSTATUS status;
 448
 449         rec = ctx->db->fetch_locked(ctx->db, talloc_tos(),
 450                                     string_term_tdb_data(name));
 451         if (rec == NULL) {
 452                 DEBUG(10, ("fetch_locked(\"%s\") failed\n", name));
 453                 status = NT_STATUS_INTERNAL_ERROR;
 454                 goto done;
 455         }
 456
 457         if (!g_lock_parse(talloc_tos(), rec->value, &num_locks, &locks)) {
 458                 DEBUG(10, ("g_lock_parse for %s failed\n", name));
 459                 status = NT_STATUS_INTERNAL_ERROR;
 460                 goto done;
 461         }
 462
 463         for (i=0; i<num_locks; i++) {
 464                 if (procid_equal(&pid, &locks[i].pid)) {
 465                         break;
 466                 }
 467         }
 468
 469         if (i == num_locks) {
 470                 DEBUG(10, ("g_lock_force_unlock: Lock not found\n"));
 471                 status = NT_STATUS_INTERNAL_ERROR;
 472                 goto done;
 473         }
 474
 475         lock_type = locks[i].lock_type;
 476
 477         if (i < (num_locks-1)) {
 478                 locks[i] = locks[num_locks-1];
 479         }
 480         num_locks -= 1;
 481
 482         if (num_locks == 0) {
 483                 status = rec->delete_rec(rec);
 484         } else {
 485                 TDB_DATA data;
 486                 data = make_tdb_data((uint8_t *)locks,
 487                                      sizeof(struct g_lock_rec) * num_locks);
 488                 status = rec->store(rec, data, 0);
 489         }
 490
 491         if (!NT_STATUS_IS_OK(status)) {
 492                 DEBUG(1, ("g_lock_force_unlock: Could not store record: %s\n",
 493                           nt_errstr(status)));
 494                 goto done;
 495         }
 496
 497         if ((lock_type & G_LOCK_PENDING) == 0) {
 498                 /*
 499                  * We've been the lock holder. Tell all others to retry.
 500                  */
 501                 for (i=0; i<num_locks; i++) {
 502                         if ((locks[i].lock_type & G_LOCK_PENDING) == 0) {
 503                                 continue;
 504                         }
 505
 506                         /*
 507                          * Ping all waiters to retry
 508                          */
 509                         status = messaging_send(ctx->msg, locks[i].pid,
 510                                                 MSG_DBWRAP_G_LOCK_RETRY,
 511                                                 &data_blob_null);
 512                         if (!NT_STATUS_IS_OK(status)) {
 513                                 DEBUG(1, ("sending retry to %s failed: %s\n",
 514                                           procid_str(talloc_tos(),
 515                                                      &locks[i].pid),
 516                                           nt_errstr(status)));
 517                         }
 518                 }
 519         }
 520 done:
 521
 522         TALLOC_FREE(locks);
 523         TALLOC_FREE(rec);
 524         return status;
 525 }
 526
 527 NTSTATUS g_lock_unlock(struct g_lock_ctx *ctx, const char *name)
 528 {
 529         NTSTATUS status;
 530
 531         status = g_lock_force_unlock(ctx, name, procid_self());
 532
 533 #ifdef CLUSTER_SUPPORT
 534         if (lp_clustering()) {
 535                 ctdb_unwatch(messaging_ctdbd_connection());
 536         }
 537 #endif
 538         return status;
 539 }
 540
 541 struct g_lock_locks_state {
 542         int (*fn)(const char *name, void *private_data);
 543         void *private_data;
 544 };
 545
 546 static int g_lock_locks_fn(struct db_record *rec, void *priv)
 547 {
 548         struct g_lock_locks_state *state = (struct g_lock_locks_state *)priv;
 549
 550         if ((rec->key.dsize == 0) || (rec->key.dptr[rec->key.dsize-1] != 0)) {
 551                 DEBUG(1, ("invalid key in g_lock.tdb, ignoring\n"));
 552                 return 0;
 553         }
 554         return state->fn((char *)rec->key.dptr, state->private_data);
 555 }
 556
 557 int g_lock_locks(struct g_lock_ctx *ctx,
 558                  int (*fn)(const char *name, void *private_data),
 559                  void *private_data)
 560 {
 561         struct g_lock_locks_state state;
 562
 563         state.fn = fn;
 564         state.private_data = private_data;
 565
 566         return ctx->db->traverse_read(ctx->db, g_lock_locks_fn, &state);
 567 }
 568
 569 NTSTATUS g_lock_dump(struct g_lock_ctx *ctx, const char *name,
 570                      int (*fn)(struct server_id pid,
 571                                enum g_lock_type lock_type,
 572                                void *private_data),
 573                      void *private_data)
 574 {
 575         TDB_DATA data;
 576         int i, num_locks;
 577         struct g_lock_rec *locks = NULL;
 578         bool ret;
 579
 580         if (ctx->db->fetch(ctx->db, talloc_tos(), string_term_tdb_data(name),
 581                            &data) != 0) {
 582                 return NT_STATUS_NOT_FOUND;
 583         }
 584
 585         if ((data.dsize == 0) || (data.dptr == NULL)) {
 586                 return NT_STATUS_OK;
 587         }
 588
 589         ret = g_lock_parse(talloc_tos(), data, &num_locks, &locks);
 590
 591         TALLOC_FREE(data.dptr);
 592
 593         if (!ret) {
 594                 DEBUG(10, ("g_lock_parse for %s failed\n", name));
 595                 return NT_STATUS_INTERNAL_ERROR;
 596         }
 597
 598         for (i=0; i<num_locks; i++) {
 599                 if (fn(locks[i].pid, locks[i].lock_type, private_data) != 0) {
 600                         break;
 601                 }
 602         }
 603         TALLOC_FREE(locks);
 604         return NT_STATUS_OK;
 605 }
 606
 607 struct g_lock_get_state {
 608         bool found;
 609         struct server_id *pid;
 610 };
 611
 612 static int g_lock_get_fn(struct server_id pid, enum g_lock_type lock_type,
 613                          void *priv)
 614 {
 615         struct g_lock_get_state *state = (struct g_lock_get_state *)priv;
 616
 617         if ((lock_type & G_LOCK_PENDING) != 0) {
 618                 return 0;
 619         }
 620
 621         state->found = true;
 622         *state->pid = pid;
 623         return 1;
 624 }
 625
 626 NTSTATUS g_lock_get(struct g_lock_ctx *ctx, const char *name,
 627                     struct server_id *pid)
 628 {
 629         struct g_lock_get_state state;
 630         NTSTATUS status;
 631
 632         state.found = false;
 633         state.pid = pid;
 634
 635         status = g_lock_dump(ctx, name, g_lock_get_fn, &state);
 636         if (!NT_STATUS_IS_OK(status)) {
 637                 return status;
 638         }
 639         if (!state.found) {
 640                 return NT_STATUS_NOT_FOUND;
 641         }
 642         return NT_STATUS_OK;
 643 }