source3/lib/ctdbd_conn.c

   1 /*
   2    Unix SMB/CIFS implementation.
   3    Samba internal messaging functions
   4    Copyright (C) 2007 by Volker Lendecke
   5    Copyright (C) 2007 by Andrew Tridgell
   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
  21 #include "includes.h"
  22 #include "util_tdb.h"
  23
  24 #ifdef CLUSTER_SUPPORT
  25
  26 #include "ctdbd_conn.h"
  27 #include "ctdb_packet.h"
  28 #include "messages.h"
  29
  30 /*
  31  * It is not possible to include ctdb.h and tdb_compat.h (included via
  32  * some other include above) without warnings. This fixes those
  33  * warnings.
  34  */
  35
  36 #ifdef typesafe_cb
  37 #undef typesafe_cb
  38 #endif
  39
  40 #ifdef typesafe_cb_preargs
  41 #undef typesafe_cb_preargs
  42 #endif
  43
  44 #ifdef typesafe_cb_postargs
  45 #undef typesafe_cb_postargs
  46 #endif
  47
  48 /* paths to these include files come from --with-ctdb= in configure */
  49
  50 #include "ctdb.h"
  51 #include "ctdb_private.h"
  52
  53 struct ctdbd_connection {
  54         struct messaging_context *msg_ctx;
  55         uint32 reqid;
  56         uint32 our_vnn;
  57         uint64 rand_srvid;
  58         struct ctdb_packet_context *pkt;
  59         struct fd_event *fde;
  60
  61         void (*release_ip_handler)(const char *ip_addr, void *private_data);
  62         void *release_ip_priv;
  63 };
  64
  65 static uint32_t ctdbd_next_reqid(struct ctdbd_connection *conn)
  66 {
  67         conn->reqid += 1;
  68         if (conn->reqid == 0) {
  69                 conn->reqid += 1;
  70         }
  71         return conn->reqid;
  72 }
  73
  74 static NTSTATUS ctdbd_control(struct ctdbd_connection *conn,
  75                               uint32_t vnn, uint32 opcode,
  76                               uint64_t srvid, uint32_t flags, TDB_DATA data,
  77                               TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
  78                               int *cstatus);
  79
  80 /*
  81  * exit on fatal communications errors with the ctdbd daemon
  82  */
  83 static void cluster_fatal(const char *why)
  84 {
  85         DEBUG(0,("cluster fatal event: %s - exiting immediately\n", why));
  86         /* we don't use smb_panic() as we don't want to delay to write
  87            a core file. We need to release this process id immediately
  88            so that someone else can take over without getting sharing
  89            violations */
  90         _exit(1);
  91 }
  92
  93 /*
  94  *
  95  */
  96 static void ctdb_packet_dump(struct ctdb_req_header *hdr)
  97 {
  98         if (DEBUGLEVEL < 10) {
  99                 return;
 100         }
 101         DEBUGADD(10, ("len=%d, magic=%x, vers=%d, gen=%d, op=%d, reqid=%d\n",
 102                       (int)hdr->length, (int)hdr->ctdb_magic,
 103                       (int)hdr->ctdb_version, (int)hdr->generation,
 104                       (int)hdr->operation, (int)hdr->reqid));
 105 }
 106
 107 /*
 108  * Register a srvid with ctdbd
 109  */
 110 static NTSTATUS register_with_ctdbd(struct ctdbd_connection *conn,
 111                                     uint64_t srvid)
 112 {
 113
 114         int cstatus;
 115         return ctdbd_control(conn, CTDB_CURRENT_NODE,
 116                              CTDB_CONTROL_REGISTER_SRVID, srvid, 0,
 117                              tdb_null, NULL, NULL, &cstatus);
 118 }
 119
 120 /*
 121  * get our vnn from the cluster
 122  */
 123 static NTSTATUS get_cluster_vnn(struct ctdbd_connection *conn, uint32 *vnn)
 124 {
 125         int32_t cstatus=-1;
 126         NTSTATUS status;
 127         status = ctdbd_control(conn,
 128                                CTDB_CURRENT_NODE, CTDB_CONTROL_GET_PNN, 0, 0,
 129                                tdb_null, NULL, NULL, &cstatus);
 130         if (!NT_STATUS_IS_OK(status)) {
 131                 cluster_fatal("ctdbd_control failed\n");
 132         }
 133         *vnn = (uint32_t)cstatus;
 134         return status;
 135 }
 136
 137 /*
 138  * Are we active (i.e. not banned or stopped?)
 139  */
 140 static bool ctdbd_working(struct ctdbd_connection *conn, uint32_t vnn)
 141 {
 142         int32_t cstatus=-1;
 143         NTSTATUS status;
 144         TDB_DATA outdata;
 145         struct ctdb_node_map *m;
 146         uint32_t failure_flags;
 147         bool ret = false;
 148         int i;
 149
 150         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
 151                                CTDB_CONTROL_GET_NODEMAP, 0, 0,
 152                                tdb_null, talloc_tos(), &outdata, &cstatus);
 153         if (!NT_STATUS_IS_OK(status)) {
 154                 cluster_fatal("ctdbd_control failed\n");
 155         }
 156         if ((cstatus != 0) || (outdata.dptr == NULL)) {
 157                 DEBUG(2, ("Received invalid ctdb data\n"));
 158                 return false;
 159         }
 160
 161         m = (struct ctdb_node_map *)outdata.dptr;
 162
 163         for (i=0; i<m->num; i++) {
 164                 if (vnn == m->nodes[i].pnn) {
 165                         break;
 166                 }
 167         }
 168
 169         if (i == m->num) {
 170                 DEBUG(2, ("Did not find ourselves (node %d) in nodemap\n",
 171                           (int)vnn));
 172                 goto fail;
 173         }
 174
 175         failure_flags = NODE_FLAGS_BANNED | NODE_FLAGS_DISCONNECTED
 176                 | NODE_FLAGS_PERMANENTLY_DISABLED | NODE_FLAGS_STOPPED;
 177
 178         if ((m->nodes[i].flags & failure_flags) != 0) {
 179                 DEBUG(2, ("Node has status %x, not active\n",
 180                           (int)m->nodes[i].flags));
 181                 goto fail;
 182         }
 183
 184         ret = true;
 185 fail:
 186         TALLOC_FREE(outdata.dptr);
 187         return ret;
 188 }
 189
 190 uint32 ctdbd_vnn(const struct ctdbd_connection *conn)
 191 {
 192         return conn->our_vnn;
 193 }
 194
 195 /*
 196  * Get us a ctdb connection
 197  */
 198
 199 static NTSTATUS ctdbd_connect(TALLOC_CTX *mem_ctx,
 200                               struct ctdb_packet_context **presult)
 201 {
 202         struct ctdb_packet_context *result;
 203         const char *sockname = lp_ctdbd_socket();
 204         struct sockaddr_un addr;
 205         int fd;
 206
 207         if (!sockname || !*sockname) {
 208                 sockname = CTDB_PATH;
 209         }
 210
 211         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 212         if (fd == -1) {
 213                 DEBUG(3, ("Could not create socket: %s\n", strerror(errno)));
 214                 return map_nt_error_from_unix(errno);
 215         }
 216
 217         ZERO_STRUCT(addr);
 218         addr.sun_family = AF_UNIX;
 219         strncpy(addr.sun_path, sockname, sizeof(addr.sun_path));
 220
 221         if (sys_connect(fd, (struct sockaddr *)(void *)&addr) == -1) {
 222                 DEBUG(1, ("connect(%s) failed: %s\n", sockname,
 223                           strerror(errno)));
 224                 close(fd);
 225                 return map_nt_error_from_unix(errno);
 226         }
 227
 228         if (!(result = ctdb_packet_init(mem_ctx, fd))) {
 229                 close(fd);
 230                 return NT_STATUS_NO_MEMORY;
 231         }
 232
 233         *presult = result;
 234         return NT_STATUS_OK;
 235 }
 236
 237 /*
 238  * Do we have a complete ctdb packet in the queue?
 239  */
 240
 241 static bool ctdb_req_complete(const uint8_t *buf, size_t available,
 242                               size_t *length,
 243                               void *private_data)
 244 {
 245         uint32 msglen;
 246
 247         if (available < sizeof(msglen)) {
 248                 return False;
 249         }
 250
 251         msglen = *((uint32 *)buf);
 252
 253         DEBUG(10, ("msglen = %d\n", msglen));
 254
 255         if (msglen < sizeof(struct ctdb_req_header)) {
 256                 DEBUG(0, ("Got invalid msglen: %d, expected at least %d for "
 257                           "the req_header\n", (int)msglen,
 258                           (int)sizeof(struct ctdb_req_header)));
 259                 cluster_fatal("ctdbd protocol error\n");
 260         }
 261
 262         if (available < msglen) {
 263                 return false;
 264         }
 265
 266         *length = msglen;
 267         return true;
 268 }
 269
 270 /*
 271  * State necessary to defer an incoming message while we are waiting for a
 272  * ctdb reply.
 273  */
 274
 275 struct deferred_msg_state {
 276         struct messaging_context *msg_ctx;
 277         struct messaging_rec *rec;
 278 };
 279
 280 /*
 281  * Timed event handler for the deferred message
 282  */
 283
 284 static void deferred_message_dispatch(struct event_context *event_ctx,
 285                                       struct timed_event *te,
 286                                       struct timeval now,
 287                                       void *private_data)
 288 {
 289         struct deferred_msg_state *state = talloc_get_type_abort(
 290                 private_data, struct deferred_msg_state);
 291
 292         messaging_dispatch_rec(state->msg_ctx, state->rec);
 293         TALLOC_FREE(state);
 294         TALLOC_FREE(te);
 295 }
 296
 297 struct req_pull_state {
 298         TALLOC_CTX *mem_ctx;
 299         DATA_BLOB req;
 300 };
 301
 302 /*
 303  * Pull a ctdb request out of the incoming ctdb_packet queue
 304  */
 305
 306 static NTSTATUS ctdb_req_pull(uint8_t *buf, size_t length,
 307                               void *private_data)
 308 {
 309         struct req_pull_state *state = (struct req_pull_state *)private_data;
 310
 311         state->req.data = talloc_move(state->mem_ctx, &buf);
 312         state->req.length = length;
 313         return NT_STATUS_OK;
 314 }
 315
 316 /*
 317  * Fetch a messaging_rec from an incoming ctdb style message
 318  */
 319
 320 static struct messaging_rec *ctdb_pull_messaging_rec(TALLOC_CTX *mem_ctx,
 321                                                      size_t overall_length,
 322                                                      struct ctdb_req_message *msg)
 323 {
 324         struct messaging_rec *result;
 325         DATA_BLOB blob;
 326         enum ndr_err_code ndr_err;
 327
 328         if ((overall_length < offsetof(struct ctdb_req_message, data))
 329             || (overall_length
 330                 < offsetof(struct ctdb_req_message, data) + msg->datalen)) {
 331
 332                 cluster_fatal("got invalid msg length");
 333         }
 334
 335         if (!(result = talloc(mem_ctx, struct messaging_rec))) {
 336                 DEBUG(0, ("talloc failed\n"));
 337                 return NULL;
 338         }
 339
 340         blob = data_blob_const(msg->data, msg->datalen);
 341
 342         ndr_err = ndr_pull_struct_blob(
 343                 &blob, result, result,
 344                 (ndr_pull_flags_fn_t)ndr_pull_messaging_rec);
 345
 346         if (!NDR_ERR_CODE_IS_SUCCESS(ndr_err)) {
 347                 DEBUG(0, ("ndr_pull_struct_blob failed: %s\n",
 348                           ndr_errstr(ndr_err)));
 349                 TALLOC_FREE(result);
 350                 return NULL;
 351         }
 352
 353         if (DEBUGLEVEL >= 10) {
 354                 DEBUG(10, ("ctdb_pull_messaging_rec:\n"));
 355                 NDR_PRINT_DEBUG(messaging_rec, result);
 356         }
 357
 358         return result;
 359 }
 360
 361 static NTSTATUS ctdb_packet_fd_read_sync(struct ctdb_packet_context *ctx)
 362 {
 363         int timeout = lp_ctdb_timeout();
 364
 365         if (timeout == 0) {
 366                 timeout = -1;
 367         }
 368         return ctdb_packet_fd_read_sync_timeout(ctx, timeout);
 369 }
 370
 371 /*
 372  * Read a full ctdbd request. If we have a messaging context, defer incoming
 373  * messages that might come in between.
 374  */
 375
 376 static NTSTATUS ctdb_read_req(struct ctdbd_connection *conn, uint32 reqid,
 377                               TALLOC_CTX *mem_ctx, void *result)
 378 {
 379         struct ctdb_req_header *hdr;
 380         struct req_pull_state state;
 381         NTSTATUS status;
 382
 383  again:
 384
 385         status = ctdb_packet_fd_read_sync(conn->pkt);
 386
 387         if (NT_STATUS_EQUAL(status, NT_STATUS_NETWORK_BUSY)) {
 388                 /* EAGAIN */
 389                 goto again;
 390         } else if (NT_STATUS_EQUAL(status, NT_STATUS_RETRY)) {
 391                 /* EAGAIN */
 392                 goto again;
 393         }
 394
 395         if (!NT_STATUS_IS_OK(status)) {
 396                 DEBUG(0, ("ctdb_packet_fd_read failed: %s\n", nt_errstr(status)));
 397                 cluster_fatal("ctdbd died\n");
 398         }
 399
 400  next_pkt:
 401
 402         ZERO_STRUCT(state);
 403         state.mem_ctx = mem_ctx;
 404
 405         if (!ctdb_packet_handler(conn->pkt, ctdb_req_complete, ctdb_req_pull,
 406                             &state, &status)) {
 407                 /*
 408                  * Not enough data
 409                  */
 410                 DEBUG(10, ("not enough data from ctdb socket, retrying\n"));
 411                 goto again;
 412         }
 413
 414         if (!NT_STATUS_IS_OK(status)) {
 415                 DEBUG(0, ("Could not read ctdb_packet: %s\n", nt_errstr(status)));
 416                 cluster_fatal("ctdbd died\n");
 417         }
 418
 419         hdr = (struct ctdb_req_header *)state.req.data;
 420
 421         DEBUG(10, ("Received ctdb packet\n"));
 422         ctdb_packet_dump(hdr);
 423
 424         if (hdr->operation == CTDB_REQ_MESSAGE) {
 425                 struct timed_event *evt;
 426                 struct deferred_msg_state *msg_state;
 427                 struct ctdb_req_message *msg = (struct ctdb_req_message *)hdr;
 428
 429                 if (conn->msg_ctx == NULL) {
 430                         DEBUG(1, ("Got a message without having a msg ctx, "
 431                                   "dropping msg %llu\n",
 432                                   (long long unsigned)msg->srvid));
 433                         goto next_pkt;
 434                 }
 435
 436                 if ((conn->release_ip_handler != NULL)
 437                     && (msg->srvid == CTDB_SRVID_RELEASE_IP)) {
 438                         /* must be dispatched immediately */
 439                         DEBUG(10, ("received CTDB_SRVID_RELEASE_IP\n"));
 440                         conn->release_ip_handler((const char *)msg->data,
 441                                                  conn->release_ip_priv);
 442                         TALLOC_FREE(hdr);
 443                         goto next_pkt;
 444                 }
 445
 446                 if ((msg->srvid == CTDB_SRVID_RECONFIGURE)
 447                     || (msg->srvid == CTDB_SRVID_SAMBA_NOTIFY)) {
 448
 449                         DEBUG(1, ("ctdb_read_req: Got %s message\n",
 450                                   (msg->srvid == CTDB_SRVID_RECONFIGURE)
 451                                   ? "cluster reconfigure" : "SAMBA_NOTIFY"));
 452
 453                         messaging_send(conn->msg_ctx,
 454                                        messaging_server_id(conn->msg_ctx),
 455                                        MSG_SMB_BRL_VALIDATE, &data_blob_null);
 456                         messaging_send(conn->msg_ctx,
 457                                        messaging_server_id(conn->msg_ctx),
 458                                        MSG_DBWRAP_G_LOCK_RETRY,
 459                                        &data_blob_null);
 460                         TALLOC_FREE(hdr);
 461                         goto next_pkt;
 462                 }
 463
 464                 msg_state = talloc(NULL, struct deferred_msg_state);
 465                 if (msg_state == NULL) {
 466                         DEBUG(0, ("talloc failed\n"));
 467                         TALLOC_FREE(hdr);
 468                         goto next_pkt;
 469                 }
 470
 471                 if (!(msg_state->rec = ctdb_pull_messaging_rec(
 472                               msg_state, state.req.length, msg))) {
 473                         DEBUG(0, ("ctdbd_pull_messaging_rec failed\n"));
 474                         TALLOC_FREE(msg_state);
 475                         TALLOC_FREE(hdr);
 476                         goto next_pkt;
 477                 }
 478
 479                 TALLOC_FREE(hdr);
 480
 481                 msg_state->msg_ctx = conn->msg_ctx;
 482
 483                 /*
 484                  * We're waiting for a call reply, but an async message has
 485                  * crossed. Defer dispatching to the toplevel event loop.
 486                  */
 487                 evt = event_add_timed(conn->msg_ctx->event_ctx,
 488                                       conn->msg_ctx->event_ctx,
 489                                       timeval_zero(),
 490                                       deferred_message_dispatch,
 491                                       msg_state);
 492                 if (evt == NULL) {
 493                         DEBUG(0, ("event_add_timed failed\n"));
 494                         TALLOC_FREE(msg_state);
 495                         TALLOC_FREE(hdr);
 496                         goto next_pkt;
 497                 }
 498
 499                 goto next_pkt;
 500         }
 501
 502         if ((reqid != 0) && (hdr->reqid != reqid)) {
 503                 /* we got the wrong reply */
 504                 DEBUG(0,("Discarding mismatched ctdb reqid %u should have "
 505                          "been %u\n", hdr->reqid, reqid));
 506                 TALLOC_FREE(hdr);
 507                 goto again;
 508         }
 509
 510         *((void **)result) = talloc_move(mem_ctx, &hdr);
 511
 512         return NT_STATUS_OK;
 513 }
 514
 515 /*
 516  * Get us a ctdbd connection
 517  */
 518
 519 static NTSTATUS ctdbd_init_connection(TALLOC_CTX *mem_ctx,
 520                                       struct ctdbd_connection **pconn)
 521 {
 522         struct ctdbd_connection *conn;
 523         NTSTATUS status;
 524
 525         if (!(conn = talloc_zero(mem_ctx, struct ctdbd_connection))) {
 526                 DEBUG(0, ("talloc failed\n"));
 527                 return NT_STATUS_NO_MEMORY;
 528         }
 529
 530         status = ctdbd_connect(conn, &conn->pkt);
 531
 532         if (!NT_STATUS_IS_OK(status)) {
 533                 DEBUG(10, ("ctdbd_connect failed: %s\n", nt_errstr(status)));
 534                 goto fail;
 535         }
 536
 537         status = get_cluster_vnn(conn, &conn->our_vnn);
 538
 539         if (!NT_STATUS_IS_OK(status)) {
 540                 DEBUG(10, ("get_cluster_vnn failed: %s\n", nt_errstr(status)));
 541                 goto fail;
 542         }
 543
 544         if (!ctdbd_working(conn, conn->our_vnn)) {
 545                 DEBUG(2, ("Node is not working, can not connect\n"));
 546                 status = NT_STATUS_INTERNAL_DB_ERROR;
 547                 goto fail;
 548         }
 549
 550         generate_random_buffer((unsigned char *)&conn->rand_srvid,
 551                                sizeof(conn->rand_srvid));
 552
 553         status = register_with_ctdbd(conn, conn->rand_srvid);
 554
 555         if (!NT_STATUS_IS_OK(status)) {
 556                 DEBUG(5, ("Could not register random srvid: %s\n",
 557                           nt_errstr(status)));
 558                 goto fail;
 559         }
 560
 561         *pconn = conn;
 562         return NT_STATUS_OK;
 563
 564  fail:
 565         TALLOC_FREE(conn);
 566         return status;
 567 }
 568
 569 /*
 570  * Get us a ctdbd connection and register us as a process
 571  */
 572
 573 NTSTATUS ctdbd_messaging_connection(TALLOC_CTX *mem_ctx,
 574                                     struct ctdbd_connection **pconn)
 575 {
 576         struct ctdbd_connection *conn;
 577         NTSTATUS status;
 578
 579         status = ctdbd_init_connection(mem_ctx, &conn);
 580
 581         if (!NT_STATUS_IS_OK(status)) {
 582                 return status;
 583         }
 584
 585         status = register_with_ctdbd(conn, (uint64_t)sys_getpid());
 586         if (!NT_STATUS_IS_OK(status)) {
 587                 goto fail;
 588         }
 589
 590         status = register_with_ctdbd(conn, MSG_SRVID_SAMBA);
 591         if (!NT_STATUS_IS_OK(status)) {
 592                 goto fail;
 593         }
 594
 595         status = register_with_ctdbd(conn, CTDB_SRVID_SAMBA_NOTIFY);
 596         if (!NT_STATUS_IS_OK(status)) {
 597                 goto fail;
 598         }
 599
 600         *pconn = conn;
 601         return NT_STATUS_OK;
 602
 603  fail:
 604         TALLOC_FREE(conn);
 605         return status;
 606 }
 607
 608 struct messaging_context *ctdb_conn_msg_ctx(struct ctdbd_connection *conn)
 609 {
 610         return conn->msg_ctx;
 611 }
 612
 613 int ctdbd_conn_get_fd(struct ctdbd_connection *conn)
 614 {
 615         return ctdb_packet_get_fd(conn->pkt);
 616 }
 617
 618 /*
 619  * Packet handler to receive and handle a ctdb message
 620  */
 621 static NTSTATUS ctdb_handle_message(uint8_t *buf, size_t length,
 622                                     void *private_data)
 623 {
 624         struct ctdbd_connection *conn = talloc_get_type_abort(
 625                 private_data, struct ctdbd_connection);
 626         struct ctdb_req_message *msg;
 627         struct messaging_rec *msg_rec;
 628
 629         msg = (struct ctdb_req_message *)buf;
 630
 631         if (msg->hdr.operation != CTDB_REQ_MESSAGE) {
 632                 DEBUG(0, ("Received async msg of type %u, discarding\n",
 633                           msg->hdr.operation));
 634                 TALLOC_FREE(buf);
 635                 return NT_STATUS_INVALID_PARAMETER;
 636         }
 637
 638         if ((conn->release_ip_handler != NULL)
 639             && (msg->srvid == CTDB_SRVID_RELEASE_IP)) {
 640                 /* must be dispatched immediately */
 641                 DEBUG(10, ("received CTDB_SRVID_RELEASE_IP\n"));
 642                 conn->release_ip_handler((const char *)msg->data,
 643                                          conn->release_ip_priv);
 644                 TALLOC_FREE(buf);
 645                 return NT_STATUS_OK;
 646         }
 647
 648         SMB_ASSERT(conn->msg_ctx != NULL);
 649
 650         if ((msg->srvid == CTDB_SRVID_RECONFIGURE)
 651             || (msg->srvid == CTDB_SRVID_SAMBA_NOTIFY)){
 652                 DEBUG(0,("Got cluster reconfigure message\n"));
 653                 /*
 654                  * when the cluster is reconfigured or someone of the
 655                  * family has passed away (SAMBA_NOTIFY), we need to
 656                  * clean the brl database
 657                  */
 658                 messaging_send(conn->msg_ctx,
 659                                messaging_server_id(conn->msg_ctx),
 660                                MSG_SMB_BRL_VALIDATE, &data_blob_null);
 661
 662                 messaging_send(conn->msg_ctx,
 663                                messaging_server_id(conn->msg_ctx),
 664                                MSG_DBWRAP_G_LOCK_RETRY,
 665                                &data_blob_null);
 666
 667                 TALLOC_FREE(buf);
 668                 return NT_STATUS_OK;
 669         }
 670
 671         /* only messages to our pid or the broadcast are valid here */
 672         if (msg->srvid != sys_getpid() && msg->srvid != MSG_SRVID_SAMBA) {
 673                 DEBUG(0,("Got unexpected message with srvid=%llu\n",
 674                          (unsigned long long)msg->srvid));
 675                 TALLOC_FREE(buf);
 676                 return NT_STATUS_OK;
 677         }
 678
 679         if (!(msg_rec = ctdb_pull_messaging_rec(NULL, length, msg))) {
 680                 DEBUG(10, ("ctdb_pull_messaging_rec failed\n"));
 681                 TALLOC_FREE(buf);
 682                 return NT_STATUS_NO_MEMORY;
 683         }
 684
 685         messaging_dispatch_rec(conn->msg_ctx, msg_rec);
 686
 687         TALLOC_FREE(msg_rec);
 688         TALLOC_FREE(buf);
 689         return NT_STATUS_OK;
 690 }
 691
 692 /*
 693  * The ctdbd socket is readable asynchronuously
 694  */
 695
 696 static void ctdbd_socket_handler(struct event_context *event_ctx,
 697                                  struct fd_event *event,
 698                                  uint16 flags,
 699                                  void *private_data)
 700 {
 701         struct ctdbd_connection *conn = talloc_get_type_abort(
 702                 private_data, struct ctdbd_connection);
 703
 704         NTSTATUS status;
 705
 706         status = ctdb_packet_fd_read(conn->pkt);
 707
 708         if (!NT_STATUS_IS_OK(status)) {
 709                 DEBUG(0, ("packet_fd_read failed: %s\n", nt_errstr(status)));
 710                 cluster_fatal("ctdbd died\n");
 711         }
 712
 713         while (ctdb_packet_handler(conn->pkt, ctdb_req_complete,
 714                               ctdb_handle_message, conn, &status)) {
 715                 if (!NT_STATUS_IS_OK(status)) {
 716                         DEBUG(10, ("could not handle incoming message: %s\n",
 717                                    nt_errstr(status)));
 718                 }
 719         }
 720 }
 721
 722 /*
 723  * Prepare a ctdbd connection to receive messages
 724  */
 725
 726 NTSTATUS ctdbd_register_msg_ctx(struct ctdbd_connection *conn,
 727                                 struct messaging_context *msg_ctx)
 728 {
 729         SMB_ASSERT(conn->msg_ctx == NULL);
 730         SMB_ASSERT(conn->fde == NULL);
 731
 732         if (!(conn->fde = event_add_fd(msg_ctx->event_ctx, conn,
 733                                        ctdb_packet_get_fd(conn->pkt),
 734                                        EVENT_FD_READ,
 735                                        ctdbd_socket_handler,
 736                                        conn))) {
 737                 DEBUG(0, ("event_add_fd failed\n"));
 738                 return NT_STATUS_NO_MEMORY;
 739         }
 740
 741         conn->msg_ctx = msg_ctx;
 742
 743         return NT_STATUS_OK;
 744 }
 745
 746 /*
 747  * Send a messaging message across a ctdbd
 748  */
 749
 750 NTSTATUS ctdbd_messaging_send(struct ctdbd_connection *conn,
 751                               uint32 dst_vnn, uint64 dst_srvid,
 752                               struct messaging_rec *msg)
 753 {
 754         struct ctdb_req_message r;
 755         TALLOC_CTX *mem_ctx;
 756         DATA_BLOB blob;
 757         NTSTATUS status;
 758         enum ndr_err_code ndr_err;
 759
 760         if (!(mem_ctx = talloc_init("ctdbd_messaging_send"))) {
 761                 DEBUG(0, ("talloc failed\n"));
 762                 return NT_STATUS_NO_MEMORY;
 763         }
 764
 765         ndr_err = ndr_push_struct_blob(
 766                 &blob, mem_ctx, msg,
 767                 (ndr_push_flags_fn_t)ndr_push_messaging_rec);
 768
 769         if (!NDR_ERR_CODE_IS_SUCCESS(ndr_err)) {
 770                 DEBUG(0, ("ndr_push_struct_blob failed: %s\n",
 771                           ndr_errstr(ndr_err)));
 772                 status = ndr_map_error2ntstatus(ndr_err);
 773                 goto fail;
 774         }
 775
 776         r.hdr.length = offsetof(struct ctdb_req_message, data) + blob.length;
 777         r.hdr.ctdb_magic = CTDB_MAGIC;
 778         r.hdr.ctdb_version = CTDB_VERSION;
 779         r.hdr.generation = 1;
 780         r.hdr.operation  = CTDB_REQ_MESSAGE;
 781         r.hdr.destnode   = dst_vnn;
 782         r.hdr.srcnode    = conn->our_vnn;
 783         r.hdr.reqid      = 0;
 784         r.srvid          = dst_srvid;
 785         r.datalen        = blob.length;
 786
 787         DEBUG(10, ("ctdbd_messaging_send: Sending ctdb packet\n"));
 788         ctdb_packet_dump(&r.hdr);
 789
 790         status = ctdb_packet_send(
 791                 conn->pkt, 2,
 792                 data_blob_const(&r, offsetof(struct ctdb_req_message, data)),
 793                 blob);
 794
 795         if (!NT_STATUS_IS_OK(status)) {
 796                 DEBUG(0, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
 797                 goto fail;
 798         }
 799
 800         status = ctdb_packet_flush(conn->pkt);
 801
 802         if (!NT_STATUS_IS_OK(status)) {
 803                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
 804                 cluster_fatal("cluster dispatch daemon msg write error\n");
 805         }
 806
 807         status = NT_STATUS_OK;
 808  fail:
 809         TALLOC_FREE(mem_ctx);
 810         return status;
 811 }
 812
 813 /*
 814  * send/recv a generic ctdb control message
 815  */
 816 static NTSTATUS ctdbd_control(struct ctdbd_connection *conn,
 817                               uint32_t vnn, uint32 opcode,
 818                               uint64_t srvid, uint32_t flags,
 819                               TDB_DATA data,
 820                               TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
 821                               int *cstatus)
 822 {
 823         struct ctdb_req_control req;
 824         struct ctdb_reply_control *reply = NULL;
 825         struct ctdbd_connection *new_conn = NULL;
 826         NTSTATUS status;
 827
 828         if (conn == NULL) {
 829                 status = ctdbd_init_connection(NULL, &new_conn);
 830
 831                 if (!NT_STATUS_IS_OK(status)) {
 832                         DEBUG(10, ("Could not init temp connection: %s\n",
 833                                    nt_errstr(status)));
 834                         goto fail;
 835                 }
 836
 837                 conn = new_conn;
 838         }
 839
 840         ZERO_STRUCT(req);
 841         req.hdr.length = offsetof(struct ctdb_req_control, data) + data.dsize;
 842         req.hdr.ctdb_magic   = CTDB_MAGIC;
 843         req.hdr.ctdb_version = CTDB_VERSION;
 844         req.hdr.operation    = CTDB_REQ_CONTROL;
 845         req.hdr.reqid        = ctdbd_next_reqid(conn);
 846         req.hdr.destnode     = vnn;
 847         req.opcode           = opcode;
 848         req.srvid            = srvid;
 849         req.datalen          = data.dsize;
 850         req.flags            = flags;
 851
 852         DEBUG(10, ("ctdbd_control: Sending ctdb packet\n"));
 853         ctdb_packet_dump(&req.hdr);
 854
 855         status = ctdb_packet_send(
 856                 conn->pkt, 2,
 857                 data_blob_const(&req, offsetof(struct ctdb_req_control, data)),
 858                 data_blob_const(data.dptr, data.dsize));
 859
 860         if (!NT_STATUS_IS_OK(status)) {
 861                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
 862                 goto fail;
 863         }
 864
 865         status = ctdb_packet_flush(conn->pkt);
 866
 867         if (!NT_STATUS_IS_OK(status)) {
 868                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
 869                 cluster_fatal("cluster dispatch daemon control write error\n");
 870         }
 871
 872         if (flags & CTDB_CTRL_FLAG_NOREPLY) {
 873                 TALLOC_FREE(new_conn);
 874                 if (cstatus) {
 875                         *cstatus = 0;
 876                 }
 877                 return NT_STATUS_OK;
 878         }
 879
 880         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
 881
 882         if (!NT_STATUS_IS_OK(status)) {
 883                 DEBUG(10, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
 884                 goto fail;
 885         }
 886
 887         if (reply->hdr.operation != CTDB_REPLY_CONTROL) {
 888                 DEBUG(0, ("received invalid reply\n"));
 889                 goto fail;
 890         }
 891
 892         if (outdata) {
 893                 if (!(outdata->dptr = (uint8 *)talloc_memdup(
 894                               mem_ctx, reply->data, reply->datalen))) {
 895                         TALLOC_FREE(reply);
 896                         return NT_STATUS_NO_MEMORY;
 897                 }
 898                 outdata->dsize = reply->datalen;
 899         }
 900         if (cstatus) {
 901                 (*cstatus) = reply->status;
 902         }
 903
 904         status = NT_STATUS_OK;
 905
 906  fail:
 907         TALLOC_FREE(new_conn);
 908         TALLOC_FREE(reply);
 909         return status;
 910 }
 911
 912 /*
 913  * see if a remote process exists
 914  */
 915 bool ctdbd_process_exists(struct ctdbd_connection *conn, uint32 vnn, pid_t pid)
 916 {
 917         NTSTATUS status;
 918         TDB_DATA data;
 919         int32_t cstatus;
 920
 921         data.dptr = (uint8_t*)&pid;
 922         data.dsize = sizeof(pid);
 923
 924         status = ctdbd_control(conn, vnn, CTDB_CONTROL_PROCESS_EXISTS, 0, 0,
 925                                data, NULL, NULL, &cstatus);
 926         if (!NT_STATUS_IS_OK(status)) {
 927                 DEBUG(0, (__location__ " ctdb_control for process_exists "
 928                           "failed\n"));
 929                 return False;
 930         }
 931
 932         return cstatus == 0;
 933 }
 934
 935 bool ctdb_processes_exist(struct ctdbd_connection *conn,
 936                           const struct server_id *pids, int num_pids,
 937                           bool *results)
 938 {
 939         TALLOC_CTX *frame = talloc_stackframe();
 940         int i, num_received;
 941         NTSTATUS status;
 942         uint32_t *reqids;
 943         bool result = false;
 944
 945         reqids = talloc_array(talloc_tos(), uint32_t, num_pids);
 946         if (reqids == NULL) {
 947                 goto fail;
 948         }
 949
 950         for (i=0; i<num_pids; i++) {
 951                 struct ctdb_req_control req;
 952
 953                 results[i] = false;
 954                 reqids[i] = ctdbd_next_reqid(conn);
 955
 956                 ZERO_STRUCT(req);
 957
 958                 req.hdr.length = offsetof(struct ctdb_req_control, data);
 959                 req.hdr.length += sizeof(pid_t);
 960                 req.hdr.ctdb_magic   = CTDB_MAGIC;
 961                 req.hdr.ctdb_version = CTDB_VERSION;
 962                 req.hdr.operation    = CTDB_REQ_CONTROL;
 963                 req.hdr.reqid        = reqids[i];
 964                 req.hdr.destnode     = pids[i].vnn;
 965                 req.opcode           = CTDB_CONTROL_PROCESS_EXISTS;
 966                 req.srvid            = 0;
 967                 req.datalen          = sizeof(pids[i].pid);
 968                 req.flags            = 0;
 969
 970                 DEBUG(10, ("ctdbd_control: Sending ctdb packet\n"));
 971                 ctdb_packet_dump(&req.hdr);
 972
 973                 status = ctdb_packet_send(
 974                         conn->pkt, 2,
 975                         data_blob_const(
 976                                 &req, offsetof(struct ctdb_req_control, data)),
 977                         data_blob_const(&pids[i].pid, sizeof(pids[i].pid)));
 978                 if (!NT_STATUS_IS_OK(status)) {
 979                         DEBUG(10, ("ctdb_packet_send failed: %s\n",
 980                                    nt_errstr(status)));
 981                         goto fail;
 982                 }
 983         }
 984
 985         status = ctdb_packet_flush(conn->pkt);
 986         if (!NT_STATUS_IS_OK(status)) {
 987                 DEBUG(10, ("ctdb_packet_flush failed: %s\n",
 988                            nt_errstr(status)));
 989                 goto fail;
 990         }
 991
 992         num_received = 0;
 993
 994         while (num_received < num_pids) {
 995                 struct ctdb_reply_control *reply = NULL;
 996                 uint32_t reqid;
 997
 998                 status = ctdb_read_req(conn, 0, talloc_tos(), (void *)&reply);
 999                 if (!NT_STATUS_IS_OK(status)) {
1000                         DEBUG(10, ("ctdb_read_req failed: %s\n",
1001                                    nt_errstr(status)));
1002                         goto fail;
1003                 }
1004
1005                 if (reply->hdr.operation != CTDB_REPLY_CONTROL) {
1006                         DEBUG(10, ("Received invalid reply\n"));
1007                         goto fail;
1008                 }
1009
1010                 reqid = reply->hdr.reqid;
1011
1012                 for (i=0; i<num_pids; i++) {
1013                         if (reqid == reqids[i]) {
1014                                 break;
1015                         }
1016                 }
1017                 if (i == num_pids) {
1018                         DEBUG(10, ("Received unknown record number %u\n",
1019                                    (unsigned)reqid));
1020                         goto fail;
1021                 }
1022                 results[i] = ((reply->status) == 0);
1023                 TALLOC_FREE(reply);
1024                 num_received += 1;
1025         }
1026
1027         result = true;
1028 fail:
1029         TALLOC_FREE(frame);
1030         return result;
1031 }
1032
1033 /*
1034  * Get a db path
1035  */
1036 char *ctdbd_dbpath(struct ctdbd_connection *conn,
1037                    TALLOC_CTX *mem_ctx, uint32_t db_id)
1038 {
1039         NTSTATUS status;
1040         TDB_DATA data;
1041         int32_t cstatus;
1042
1043         data.dptr = (uint8_t*)&db_id;
1044         data.dsize = sizeof(db_id);
1045
1046         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1047                                CTDB_CONTROL_GETDBPATH, 0, 0, data,
1048                                mem_ctx, &data, &cstatus);
1049         if (!NT_STATUS_IS_OK(status) || cstatus != 0) {
1050                 DEBUG(0,(__location__ " ctdb_control for getdbpath failed\n"));
1051                 return NULL;
1052         }
1053
1054         return (char *)data.dptr;
1055 }
1056
1057 /*
1058  * attach to a ctdb database
1059  */
1060 NTSTATUS ctdbd_db_attach(struct ctdbd_connection *conn,
1061                          const char *name, uint32_t *db_id, int tdb_flags)
1062 {
1063         NTSTATUS status;
1064         TDB_DATA data;
1065         int32_t cstatus;
1066         bool persistent = (tdb_flags & TDB_CLEAR_IF_FIRST) == 0;
1067
1068         data.dptr = (uint8_t*)name;
1069         data.dsize = strlen(name)+1;
1070
1071         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1072                                persistent
1073                                ? CTDB_CONTROL_DB_ATTACH_PERSISTENT
1074                                : CTDB_CONTROL_DB_ATTACH,
1075                                tdb_flags, 0, data, NULL, &data, &cstatus);
1076         if (!NT_STATUS_IS_OK(status)) {
1077                 DEBUG(0, (__location__ " ctdb_control for db_attach "
1078                           "failed: %s\n", nt_errstr(status)));
1079                 return status;
1080         }
1081
1082         if (cstatus != 0 || data.dsize != sizeof(uint32_t)) {
1083                 DEBUG(0,(__location__ " ctdb_control for db_attach failed\n"));
1084                 return NT_STATUS_INTERNAL_ERROR;
1085         }
1086
1087         *db_id = *(uint32_t *)data.dptr;
1088         talloc_free(data.dptr);
1089
1090         if (!(tdb_flags & TDB_SEQNUM)) {
1091                 return NT_STATUS_OK;
1092         }
1093
1094         data.dptr = (uint8_t *)db_id;
1095         data.dsize = sizeof(*db_id);
1096
1097         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1098                                CTDB_CONTROL_ENABLE_SEQNUM, 0, 0, data,
1099                                NULL, NULL, &cstatus);
1100         if (!NT_STATUS_IS_OK(status) || cstatus != 0) {
1101                 DEBUG(0,(__location__ " ctdb_control for enable seqnum "
1102                          "failed\n"));
1103                 return NT_STATUS_IS_OK(status) ? NT_STATUS_INTERNAL_ERROR :
1104                         status;
1105         }
1106
1107         return NT_STATUS_OK;
1108 }
1109
1110 /*
1111  * force the migration of a record to this node
1112  */
1113 NTSTATUS ctdbd_migrate(struct ctdbd_connection *conn, uint32 db_id,
1114                        TDB_DATA key)
1115 {
1116         struct ctdb_req_call req;
1117         struct ctdb_reply_call *reply;
1118         NTSTATUS status;
1119
1120         ZERO_STRUCT(req);
1121
1122         req.hdr.length = offsetof(struct ctdb_req_call, data) + key.dsize;
1123         req.hdr.ctdb_magic   = CTDB_MAGIC;
1124         req.hdr.ctdb_version = CTDB_VERSION;
1125         req.hdr.operation    = CTDB_REQ_CALL;
1126         req.hdr.reqid        = ctdbd_next_reqid(conn);
1127         req.flags            = CTDB_IMMEDIATE_MIGRATION;
1128         req.callid           = CTDB_NULL_FUNC;
1129         req.db_id            = db_id;
1130         req.keylen           = key.dsize;
1131
1132         DEBUG(10, ("ctdbd_migrate: Sending ctdb packet\n"));
1133         ctdb_packet_dump(&req.hdr);
1134
1135         status = ctdb_packet_send(
1136                 conn->pkt, 2,
1137                 data_blob_const(&req, offsetof(struct ctdb_req_call, data)),
1138                 data_blob_const(key.dptr, key.dsize));
1139
1140         if (!NT_STATUS_IS_OK(status)) {
1141                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
1142                 return status;
1143         }
1144
1145         status = ctdb_packet_flush(conn->pkt);
1146
1147         if (!NT_STATUS_IS_OK(status)) {
1148                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
1149                 cluster_fatal("cluster dispatch daemon control write error\n");
1150         }
1151
1152         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
1153
1154         if (!NT_STATUS_IS_OK(status)) {
1155                 DEBUG(0, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
1156                 goto fail;
1157         }
1158
1159         if (reply->hdr.operation != CTDB_REPLY_CALL) {
1160                 DEBUG(0, ("received invalid reply\n"));
1161                 status = NT_STATUS_INTERNAL_ERROR;
1162                 goto fail;
1163         }
1164
1165         status = NT_STATUS_OK;
1166  fail:
1167
1168         TALLOC_FREE(reply);
1169         return status;
1170 }
1171
1172 /*
1173  * remotely fetch a record without locking it or forcing a migration
1174  */
1175 NTSTATUS ctdbd_fetch(struct ctdbd_connection *conn, uint32 db_id,
1176                      TDB_DATA key, TALLOC_CTX *mem_ctx, TDB_DATA *data)
1177 {
1178         struct ctdb_req_call req;
1179         struct ctdb_reply_call *reply;
1180         NTSTATUS status;
1181
1182         ZERO_STRUCT(req);
1183
1184         req.hdr.length = offsetof(struct ctdb_req_call, data) + key.dsize;
1185         req.hdr.ctdb_magic   = CTDB_MAGIC;
1186         req.hdr.ctdb_version = CTDB_VERSION;
1187         req.hdr.operation    = CTDB_REQ_CALL;
1188         req.hdr.reqid        = ctdbd_next_reqid(conn);
1189         req.flags            = 0;
1190         req.callid           = CTDB_FETCH_FUNC;
1191         req.db_id            = db_id;
1192         req.keylen           = key.dsize;
1193
1194         status = ctdb_packet_send(
1195                 conn->pkt, 2,
1196                 data_blob_const(&req, offsetof(struct ctdb_req_call, data)),
1197                 data_blob_const(key.dptr, key.dsize));
1198
1199         if (!NT_STATUS_IS_OK(status)) {
1200                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
1201                 return status;
1202         }
1203
1204         status = ctdb_packet_flush(conn->pkt);
1205
1206         if (!NT_STATUS_IS_OK(status)) {
1207                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
1208                 cluster_fatal("cluster dispatch daemon control write error\n");
1209         }
1210
1211         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
1212
1213         if (!NT_STATUS_IS_OK(status)) {
1214                 DEBUG(0, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
1215                 goto fail;
1216         }
1217
1218         if (reply->hdr.operation != CTDB_REPLY_CALL) {
1219                 DEBUG(0, ("received invalid reply\n"));
1220                 status = NT_STATUS_INTERNAL_ERROR;
1221                 goto fail;
1222         }
1223
1224         data->dsize = reply->datalen;
1225         if (data->dsize == 0) {
1226                 data->dptr = NULL;
1227                 goto done;
1228         }
1229
1230         data->dptr = (uint8 *)talloc_memdup(mem_ctx, &reply->data[0],
1231                                             reply->datalen);
1232         if (data->dptr == NULL) {
1233                 DEBUG(0, ("talloc failed\n"));
1234                 status = NT_STATUS_NO_MEMORY;
1235                 goto fail;
1236         }
1237
1238  done:
1239         status = NT_STATUS_OK;
1240  fail:
1241         TALLOC_FREE(reply);
1242         return status;
1243 }
1244
1245 struct ctdbd_traverse_state {
1246         void (*fn)(TDB_DATA key, TDB_DATA data, void *private_data);
1247         void *private_data;
1248 };
1249
1250 /*
1251  * Handle a traverse record coming in on the ctdbd connection
1252  */
1253
1254 static NTSTATUS ctdb_traverse_handler(uint8_t *buf, size_t length,
1255                                       void *private_data)
1256 {
1257         struct ctdbd_traverse_state *state =
1258                 (struct ctdbd_traverse_state *)private_data;
1259
1260         struct ctdb_req_message *m;
1261         struct ctdb_rec_data *d;
1262         TDB_DATA key, data;
1263
1264         m = (struct ctdb_req_message *)buf;
1265
1266         if (length < sizeof(*m) || m->hdr.length != length) {
1267                 DEBUG(0, ("Got invalid message of length %d\n", (int)length));
1268                 TALLOC_FREE(buf);
1269                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1270         }
1271
1272         d = (struct ctdb_rec_data *)&m->data[0];
1273         if (m->datalen < sizeof(uint32_t) || m->datalen != d->length) {
1274                 DEBUG(0, ("Got invalid traverse data of length %d\n",
1275                           (int)m->datalen));
1276                 TALLOC_FREE(buf);
1277                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1278         }
1279
1280         key.dsize = d->keylen;
1281         key.dptr  = &d->data[0];
1282         data.dsize = d->datalen;
1283         data.dptr = &d->data[d->keylen];
1284
1285         if (key.dsize == 0 && data.dsize == 0) {
1286                 /* end of traverse */
1287                 return NT_STATUS_END_OF_FILE;
1288         }
1289
1290         if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
1291                 DEBUG(0, ("Got invalid ltdb header length %d\n",
1292                           (int)data.dsize));
1293                 TALLOC_FREE(buf);
1294                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1295         }
1296         data.dsize -= sizeof(struct ctdb_ltdb_header);
1297         data.dptr += sizeof(struct ctdb_ltdb_header);
1298
1299         if (state->fn) {
1300                 state->fn(key, data, state->private_data);
1301         }
1302
1303         TALLOC_FREE(buf);
1304         return NT_STATUS_OK;
1305 }
1306
1307 /*
1308   Traverse a ctdb database. This uses a kind-of hackish way to open a second
1309   connection to ctdbd to avoid the hairy recursive and async problems with
1310   everything in-line.
1311 */
1312
1313 NTSTATUS ctdbd_traverse(uint32 db_id,
1314                         void (*fn)(TDB_DATA key, TDB_DATA data,
1315                                    void *private_data),
1316                         void *private_data)
1317 {
1318         struct ctdbd_connection *conn;
1319         NTSTATUS status;
1320
1321         TDB_DATA data;
1322         struct ctdb_traverse_start t;
1323         int cstatus;
1324         struct ctdbd_traverse_state state;
1325
1326         status = ctdbd_init_connection(NULL, &conn);
1327         if (!NT_STATUS_IS_OK(status)) {
1328                 DEBUG(0, ("ctdbd_init_connection failed: %s\n",
1329                           nt_errstr(status)));
1330                 return status;
1331         }
1332
1333         t.db_id = db_id;
1334         t.srvid = conn->rand_srvid;
1335         t.reqid = ctdbd_next_reqid(conn);
1336
1337         data.dptr = (uint8_t *)&t;
1338         data.dsize = sizeof(t);
1339
1340         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1341                                CTDB_CONTROL_TRAVERSE_START, conn->rand_srvid, 0,
1342                                data, NULL, NULL, &cstatus);
1343
1344         if (!NT_STATUS_IS_OK(status) || (cstatus != 0)) {
1345
1346                 DEBUG(0,("ctdbd_control failed: %s, %d\n", nt_errstr(status),
1347                          cstatus));
1348
1349                 if (NT_STATUS_IS_OK(status)) {
1350                         /*
1351                          * We need a mapping here
1352                          */
1353                         status = NT_STATUS_UNSUCCESSFUL;
1354                 }
1355                 goto done;
1356         }
1357
1358         state.fn = fn;
1359         state.private_data = private_data;
1360
1361         while (True) {
1362
1363                 status = NT_STATUS_OK;
1364
1365                 if (ctdb_packet_handler(conn->pkt, ctdb_req_complete,
1366                                    ctdb_traverse_handler, &state, &status)) {
1367
1368                         if (NT_STATUS_EQUAL(status, NT_STATUS_END_OF_FILE)) {
1369                                 status = NT_STATUS_OK;
1370                                 break;
1371                         }
1372
1373                         /*
1374                          * There might be more in the queue
1375                          */
1376                         continue;
1377                 }
1378
1379                 if (!NT_STATUS_IS_OK(status)) {
1380                         break;
1381                 }
1382
1383                 status = ctdb_packet_fd_read_sync(conn->pkt);
1384
1385                 if (NT_STATUS_EQUAL(status, NT_STATUS_RETRY)) {
1386                         /*
1387                          * There might be more in the queue
1388                          */
1389                         continue;
1390                 }
1391
1392                 if (NT_STATUS_EQUAL(status, NT_STATUS_END_OF_FILE)) {
1393                         status = NT_STATUS_OK;
1394                         break;
1395                 }
1396
1397                 if (!NT_STATUS_IS_OK(status)) {
1398                         DEBUG(0, ("ctdb_packet_fd_read_sync failed: %s\n", nt_errstr(status)));
1399                         cluster_fatal("ctdbd died\n");
1400                 }
1401         }
1402
1403  done:
1404         TALLOC_FREE(conn);
1405         return status;
1406 }
1407
1408 /*
1409    This is used to canonicalize a ctdb_sock_addr structure.
1410 */
1411 static void smbd_ctdb_canonicalize_ip(const struct sockaddr_storage *in,
1412                                       struct sockaddr_storage *out)
1413 {
1414         memcpy(out, in, sizeof (*out));
1415
1416 #ifdef HAVE_IPV6
1417         if (in->ss_family == AF_INET6) {
1418                 const char prefix[12] = { 0,0,0,0,0,0,0,0,0,0,0xff,0xff };
1419                 const struct sockaddr_in6 *in6 = (struct sockaddr_in6 *)in;
1420                 struct sockaddr_in *out4 = (struct sockaddr_in *)out;
1421                 if (memcmp(&in6->sin6_addr, prefix, 12) == 0) {
1422                         memset(out, 0, sizeof(*out));
1423 #ifdef HAVE_SOCK_SIN_LEN
1424                         out4->sin_len = sizeof(*out);
1425 #endif
1426                         out4->sin_family = AF_INET;
1427                         out4->sin_port   = in6->sin6_port;
1428                         memcpy(&out4->sin_addr, &in6->sin6_addr.s6_addr32[3], 4);
1429                 }
1430         }
1431 #endif
1432 }
1433
1434 /*
1435  * Register us as a server for a particular tcp connection
1436  */
1437
1438 NTSTATUS ctdbd_register_ips(struct ctdbd_connection *conn,
1439                             const struct sockaddr_storage *_server,
1440                             const struct sockaddr_storage *_client,
1441                             void (*release_ip_handler)(const char *ip_addr,
1442                                                        void *private_data),
1443                             void *private_data)
1444 {
1445         /*
1446          * we still use ctdb_control_tcp for ipv4
1447          * because we want to work against older ctdb
1448          * versions at runtime
1449          */
1450         struct ctdb_control_tcp p4;
1451 #ifdef HAVE_STRUCT_CTDB_CONTROL_TCP_ADDR
1452         struct ctdb_control_tcp_addr p;
1453 #endif
1454         TDB_DATA data;
1455         NTSTATUS status;
1456         struct sockaddr_storage client;
1457         struct sockaddr_storage server;
1458
1459         /*
1460          * Only one connection so far
1461          */
1462         SMB_ASSERT(conn->release_ip_handler == NULL);
1463
1464         smbd_ctdb_canonicalize_ip(_client, &client);
1465         smbd_ctdb_canonicalize_ip(_server, &server);
1466
1467         switch (client.ss_family) {
1468         case AF_INET:
1469                 memcpy(&p4.dest, &server, sizeof(p4.dest));
1470                 memcpy(&p4.src, &client, sizeof(p4.src));
1471                 data.dptr = (uint8_t *)&p4;
1472                 data.dsize = sizeof(p4);
1473                 break;
1474 #ifdef HAVE_STRUCT_CTDB_CONTROL_TCP_ADDR
1475         case AF_INET6:
1476                 memcpy(&p.dest.ip6, &server, sizeof(p.dest.ip6));
1477                 memcpy(&p.src.ip6, &client, sizeof(p.src.ip6));
1478                 data.dptr = (uint8_t *)&p;
1479                 data.dsize = sizeof(p);
1480                 break;
1481 #endif
1482         default:
1483                 return NT_STATUS_INTERNAL_ERROR;
1484         }
1485
1486         conn->release_ip_handler = release_ip_handler;
1487         conn->release_ip_priv = private_data;
1488
1489         /*
1490          * We want to be told about IP releases
1491          */
1492
1493         status = register_with_ctdbd(conn, CTDB_SRVID_RELEASE_IP);
1494         if (!NT_STATUS_IS_OK(status)) {
1495                 return status;
1496         }
1497
1498         /*
1499          * inform ctdb of our tcp connection, so if IP takeover happens ctdb
1500          * can send an extra ack to trigger a reset for our client, so it
1501          * immediately reconnects
1502          */
1503         return ctdbd_control(conn, CTDB_CURRENT_NODE,
1504                              CTDB_CONTROL_TCP_CLIENT, 0,
1505                              CTDB_CTRL_FLAG_NOREPLY, data, NULL, NULL, NULL);
1506 }
1507
1508 /*
1509  * We want to handle reconfigure events
1510  */
1511 NTSTATUS ctdbd_register_reconfigure(struct ctdbd_connection *conn)
1512 {
1513         return register_with_ctdbd(conn, CTDB_SRVID_RECONFIGURE);
1514 }
1515
1516 /*
1517   call a control on the local node
1518  */
1519 NTSTATUS ctdbd_control_local(struct ctdbd_connection *conn, uint32 opcode,
1520                              uint64_t srvid, uint32_t flags, TDB_DATA data,
1521                              TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
1522                              int *cstatus)
1523 {
1524         return ctdbd_control(conn, CTDB_CURRENT_NODE, opcode, srvid, flags, data, mem_ctx, outdata, cstatus);
1525 }
1526
1527 NTSTATUS ctdb_watch_us(struct ctdbd_connection *conn)
1528 {
1529         struct ctdb_client_notify_register reg_data;
1530         size_t struct_len;
1531         NTSTATUS status;
1532         int cstatus;
1533
1534         reg_data.srvid = CTDB_SRVID_SAMBA_NOTIFY;
1535         reg_data.len = 1;
1536         reg_data.notify_data[0] = 0;
1537
1538         struct_len = offsetof(struct ctdb_client_notify_register,
1539                               notify_data) + reg_data.len;
1540
1541         status = ctdbd_control_local(
1542                 conn, CTDB_CONTROL_REGISTER_NOTIFY, conn->rand_srvid, 0,
1543                 make_tdb_data((uint8_t *)&reg_data, struct_len),
1544                 NULL, NULL, &cstatus);
1545         if (!NT_STATUS_IS_OK(status)) {
1546                 DEBUG(1, ("ctdbd_control_local failed: %s\n",
1547                           nt_errstr(status)));
1548         }
1549         return status;
1550 }
1551
1552 NTSTATUS ctdb_unwatch(struct ctdbd_connection *conn)
1553 {
1554         struct ctdb_client_notify_deregister dereg_data;
1555         NTSTATUS status;
1556         int cstatus;
1557
1558         dereg_data.srvid = CTDB_SRVID_SAMBA_NOTIFY;
1559
1560         status = ctdbd_control_local(
1561                 conn, CTDB_CONTROL_DEREGISTER_NOTIFY, conn->rand_srvid, 0,
1562                 make_tdb_data((uint8_t *)&dereg_data, sizeof(dereg_data)),
1563                 NULL, NULL, &cstatus);
1564         if (!NT_STATUS_IS_OK(status)) {
1565                 DEBUG(1, ("ctdbd_control_local failed: %s\n",
1566                           nt_errstr(status)));
1567         }
1568         return status;
1569 }
1570
1571 #endif