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 #include "serverid.h"
  24 #include "ctdbd_conn.h"
  25
  26 #ifdef CLUSTER_SUPPORT
  27
  28 #include "ctdb_packet.h"
  29 #include "messages.h"
  30
  31 /*
  32  * It is not possible to include ctdb.h and tdb_compat.h (included via
  33  * some other include above) without warnings. This fixes those
  34  * warnings.
  35  */
  36
  37 #ifdef typesafe_cb
  38 #undef typesafe_cb
  39 #endif
  40
  41 #ifdef typesafe_cb_preargs
  42 #undef typesafe_cb_preargs
  43 #endif
  44
  45 #ifdef typesafe_cb_postargs
  46 #undef typesafe_cb_postargs
  47 #endif
  48
  49 /* paths to these include files come from --with-ctdb= in configure */
  50
  51 #include "ctdb.h"
  52 #include "ctdb_private.h"
  53
  54 struct ctdbd_connection {
  55         struct messaging_context *msg_ctx;
  56         uint32_t reqid;
  57         uint32_t our_vnn;
  58         uint64_t rand_srvid;
  59         struct ctdb_packet_context *pkt;
  60         struct tevent_fd *fde;
  61
  62         void (*release_ip_handler)(const char *ip_addr, void *private_data);
  63         void *release_ip_priv;
  64 };
  65
  66 static uint32_t ctdbd_next_reqid(struct ctdbd_connection *conn)
  67 {
  68         conn->reqid += 1;
  69         if (conn->reqid == 0) {
  70                 conn->reqid += 1;
  71         }
  72         return conn->reqid;
  73 }
  74
  75 static NTSTATUS ctdbd_control(struct ctdbd_connection *conn,
  76                               uint32_t vnn, uint32_t opcode,
  77                               uint64_t srvid, uint32_t flags, TDB_DATA data,
  78                               TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
  79                               int *cstatus);
  80
  81 /*
  82  * exit on fatal communications errors with the ctdbd daemon
  83  */
  84 static void cluster_fatal(const char *why)
  85 {
  86         DEBUG(0,("cluster fatal event: %s - exiting immediately\n", why));
  87         /* we don't use smb_panic() as we don't want to delay to write
  88            a core file. We need to release this process id immediately
  89            so that someone else can take over without getting sharing
  90            violations */
  91         _exit(1);
  92 }
  93
  94 /*
  95  *
  96  */
  97 static void ctdb_packet_dump(struct ctdb_req_header *hdr)
  98 {
  99         if (DEBUGLEVEL < 11) {
 100                 return;
 101         }
 102         DEBUGADD(11, ("len=%d, magic=%x, vers=%d, gen=%d, op=%d, reqid=%d\n",
 103                       (int)hdr->length, (int)hdr->ctdb_magic,
 104                       (int)hdr->ctdb_version, (int)hdr->generation,
 105                       (int)hdr->operation, (int)hdr->reqid));
 106 }
 107
 108 /*
 109  * Register a srvid with ctdbd
 110  */
 111 NTSTATUS register_with_ctdbd(struct ctdbd_connection *conn, 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_t *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                 DEBUG(1, ("ctdbd_control failed: %s\n", nt_errstr(status)));
 132                 return status;
 133         }
 134         *vnn = (uint32_t)cstatus;
 135         return status;
 136 }
 137
 138 /*
 139  * Are we active (i.e. not banned or stopped?)
 140  */
 141 static bool ctdbd_working(struct ctdbd_connection *conn, uint32_t vnn)
 142 {
 143         int32_t cstatus=-1;
 144         NTSTATUS status;
 145         TDB_DATA outdata;
 146         struct ctdb_node_map *m;
 147         uint32_t failure_flags;
 148         bool ret = false;
 149         int i;
 150
 151         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
 152                                CTDB_CONTROL_GET_NODEMAP, 0, 0,
 153                                tdb_null, talloc_tos(), &outdata, &cstatus);
 154         if (!NT_STATUS_IS_OK(status)) {
 155                 DEBUG(1, ("ctdbd_control failed: %s\n", nt_errstr(status)));
 156                 return false;
 157         }
 158         if ((cstatus != 0) || (outdata.dptr == NULL)) {
 159                 DEBUG(2, ("Received invalid ctdb data\n"));
 160                 return false;
 161         }
 162
 163         m = (struct ctdb_node_map *)outdata.dptr;
 164
 165         for (i=0; i<m->num; i++) {
 166                 if (vnn == m->nodes[i].pnn) {
 167                         break;
 168                 }
 169         }
 170
 171         if (i == m->num) {
 172                 DEBUG(2, ("Did not find ourselves (node %d) in nodemap\n",
 173                           (int)vnn));
 174                 goto fail;
 175         }
 176
 177         failure_flags = NODE_FLAGS_BANNED | NODE_FLAGS_DISCONNECTED
 178                 | NODE_FLAGS_PERMANENTLY_DISABLED | NODE_FLAGS_STOPPED;
 179
 180         if ((m->nodes[i].flags & failure_flags) != 0) {
 181                 DEBUG(2, ("Node has status %x, not active\n",
 182                           (int)m->nodes[i].flags));
 183                 goto fail;
 184         }
 185
 186         ret = true;
 187 fail:
 188         TALLOC_FREE(outdata.dptr);
 189         return ret;
 190 }
 191
 192 uint32_t ctdbd_vnn(const struct ctdbd_connection *conn)
 193 {
 194         return conn->our_vnn;
 195 }
 196
 197 /*
 198  * Get us a ctdb connection
 199  */
 200
 201 static NTSTATUS ctdbd_connect(TALLOC_CTX *mem_ctx,
 202                               struct ctdb_packet_context **presult)
 203 {
 204         struct ctdb_packet_context *result;
 205         const char *sockname = lp_ctdbd_socket();
 206         struct sockaddr_un addr = { 0, };
 207         int fd;
 208         socklen_t salen;
 209
 210         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 211         if (fd == -1) {
 212                 DEBUG(3, ("Could not create socket: %s\n", strerror(errno)));
 213                 return map_nt_error_from_unix(errno);
 214         }
 215
 216         addr.sun_family = AF_UNIX;
 217         snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", sockname);
 218
 219         salen = sizeof(struct sockaddr_un);
 220         if (connect(fd, (struct sockaddr *)(void *)&addr, salen) == -1) {
 221                 DEBUG(1, ("connect(%s) failed: %s\n", sockname,
 222                           strerror(errno)));
 223                 close(fd);
 224                 return map_nt_error_from_unix(errno);
 225         }
 226
 227         if (!(result = ctdb_packet_init(mem_ctx, fd))) {
 228                 close(fd);
 229                 return NT_STATUS_NO_MEMORY;
 230         }
 231
 232         *presult = result;
 233         return NT_STATUS_OK;
 234 }
 235
 236 /*
 237  * Do we have a complete ctdb packet in the queue?
 238  */
 239
 240 static bool ctdb_req_complete(const uint8_t *buf, size_t available,
 241                               size_t *length,
 242                               void *private_data)
 243 {
 244         uint32_t msglen;
 245
 246         if (available < sizeof(msglen)) {
 247                 return False;
 248         }
 249
 250         msglen = *((const uint32_t *)buf);
 251
 252         DEBUG(11, ("msglen = %d\n", msglen));
 253
 254         if (msglen < sizeof(struct ctdb_req_header)) {
 255                 DEBUG(0, ("Got invalid msglen: %d, expected at least %d for "
 256                           "the req_header\n", (int)msglen,
 257                           (int)sizeof(struct ctdb_req_header)));
 258                 cluster_fatal("ctdbd protocol error\n");
 259         }
 260
 261         if (available < msglen) {
 262                 return false;
 263         }
 264
 265         *length = msglen;
 266         return true;
 267 }
 268
 269 /*
 270  * State necessary to defer an incoming message while we are waiting for a
 271  * ctdb reply.
 272  */
 273
 274 struct deferred_msg_state {
 275         struct messaging_context *msg_ctx;
 276         struct messaging_rec *rec;
 277 };
 278
 279 /*
 280  * Timed event handler for the deferred message
 281  */
 282
 283 static void deferred_message_dispatch(struct tevent_context *event_ctx,
 284                                       struct tevent_timer *te,
 285                                       struct timeval now,
 286                                       void *private_data)
 287 {
 288         struct deferred_msg_state *state = talloc_get_type_abort(
 289                 private_data, struct deferred_msg_state);
 290
 291         messaging_dispatch_rec(state->msg_ctx, state->rec);
 292         TALLOC_FREE(state);
 293         TALLOC_FREE(te);
 294 }
 295
 296 struct req_pull_state {
 297         TALLOC_CTX *mem_ctx;
 298         DATA_BLOB req;
 299 };
 300
 301 /*
 302  * Pull a ctdb request out of the incoming ctdb_packet queue
 303  */
 304
 305 static NTSTATUS ctdb_req_pull(uint8_t *buf, size_t length,
 306                               void *private_data)
 307 {
 308         struct req_pull_state *state = (struct req_pull_state *)private_data;
 309
 310         state->req.data = talloc_move(state->mem_ctx, &buf);
 311         state->req.length = length;
 312         return NT_STATUS_OK;
 313 }
 314
 315 /*
 316  * Fetch a messaging_rec from an incoming ctdb style message
 317  */
 318
 319 static struct messaging_rec *ctdb_pull_messaging_rec(TALLOC_CTX *mem_ctx,
 320                                                      size_t overall_length,
 321                                                      struct ctdb_req_message *msg)
 322 {
 323         struct messaging_rec *result;
 324         DATA_BLOB blob;
 325         enum ndr_err_code ndr_err;
 326
 327         if ((overall_length < offsetof(struct ctdb_req_message, data))
 328             || (overall_length
 329                 < offsetof(struct ctdb_req_message, data) + msg->datalen)) {
 330
 331                 cluster_fatal("got invalid msg length");
 332         }
 333
 334         if (!(result = talloc(mem_ctx, struct messaging_rec))) {
 335                 DEBUG(0, ("talloc failed\n"));
 336                 return NULL;
 337         }
 338
 339         blob = data_blob_const(msg->data, msg->datalen);
 340
 341         ndr_err = ndr_pull_struct_blob(
 342                 &blob, result, result,
 343                 (ndr_pull_flags_fn_t)ndr_pull_messaging_rec);
 344
 345         if (!NDR_ERR_CODE_IS_SUCCESS(ndr_err)) {
 346                 DEBUG(0, ("ndr_pull_struct_blob failed: %s\n",
 347                           ndr_errstr(ndr_err)));
 348                 TALLOC_FREE(result);
 349                 return NULL;
 350         }
 351
 352         if (DEBUGLEVEL >= 11) {
 353                 DEBUG(11, ("ctdb_pull_messaging_rec:\n"));
 354                 NDR_PRINT_DEBUG(messaging_rec, result);
 355         }
 356
 357         return result;
 358 }
 359
 360 static NTSTATUS ctdb_packet_fd_read_sync(struct ctdb_packet_context *ctx)
 361 {
 362         int timeout = lp_ctdb_timeout();
 363
 364         if (timeout == 0) {
 365                 timeout = -1;
 366         }
 367         return ctdb_packet_fd_read_sync_timeout(ctx, timeout);
 368 }
 369
 370 /*
 371  * Read a full ctdbd request. If we have a messaging context, defer incoming
 372  * messages that might come in between.
 373  */
 374
 375 static NTSTATUS ctdb_read_req(struct ctdbd_connection *conn, uint32_t reqid,
 376                               TALLOC_CTX *mem_ctx, void *result)
 377 {
 378         struct ctdb_req_header *hdr;
 379         struct req_pull_state state;
 380         NTSTATUS status;
 381
 382  next_pkt:
 383         ZERO_STRUCT(state);
 384         state.mem_ctx = mem_ctx;
 385
 386         while (!ctdb_packet_handler(conn->pkt, ctdb_req_complete,
 387                                     ctdb_req_pull, &state, &status)) {
 388                 /*
 389                  * Not enough data
 390                  */
 391                 status = ctdb_packet_fd_read_sync(conn->pkt);
 392
 393                 if (NT_STATUS_EQUAL(status, NT_STATUS_NETWORK_BUSY)) {
 394                         /* EAGAIN */
 395                         continue;
 396                 } else if (NT_STATUS_EQUAL(status, NT_STATUS_RETRY)) {
 397                         /* EAGAIN */
 398                         continue;
 399                 }
 400
 401                 if (!NT_STATUS_IS_OK(status)) {
 402                         DEBUG(0, ("packet_fd_read failed: %s\n", nt_errstr(status)));
 403                         cluster_fatal("ctdbd died\n");
 404                 }
 405         }
 406
 407         if (!NT_STATUS_IS_OK(status)) {
 408                 DEBUG(0, ("Could not read ctdb_packet: %s\n", nt_errstr(status)));
 409                 cluster_fatal("ctdbd died\n");
 410         }
 411
 412         hdr = (struct ctdb_req_header *)state.req.data;
 413
 414         DEBUG(11, ("Received ctdb packet\n"));
 415         ctdb_packet_dump(hdr);
 416
 417         if (hdr->operation == CTDB_REQ_MESSAGE) {
 418                 struct tevent_timer *evt;
 419                 struct deferred_msg_state *msg_state;
 420                 struct ctdb_req_message *msg = (struct ctdb_req_message *)hdr;
 421
 422                 if (conn->msg_ctx == NULL) {
 423                         DEBUG(1, ("Got a message without having a msg ctx, "
 424                                   "dropping msg %llu\n",
 425                                   (long long unsigned)msg->srvid));
 426                         goto next_pkt;
 427                 }
 428
 429                 if ((conn->release_ip_handler != NULL)
 430                     && (msg->srvid == CTDB_SRVID_RELEASE_IP)) {
 431                         /* must be dispatched immediately */
 432                         DEBUG(10, ("received CTDB_SRVID_RELEASE_IP\n"));
 433                         conn->release_ip_handler((const char *)msg->data,
 434                                                  conn->release_ip_priv);
 435                         TALLOC_FREE(hdr);
 436                         goto next_pkt;
 437                 }
 438
 439                 if ((msg->srvid == CTDB_SRVID_RECONFIGURE)
 440                     || (msg->srvid == CTDB_SRVID_SAMBA_NOTIFY)) {
 441
 442                         DEBUG(1, ("ctdb_read_req: Got %s message\n",
 443                                   (msg->srvid == CTDB_SRVID_RECONFIGURE)
 444                                   ? "cluster reconfigure" : "SAMBA_NOTIFY"));
 445
 446                         messaging_send(conn->msg_ctx,
 447                                        messaging_server_id(conn->msg_ctx),
 448                                        MSG_SMB_BRL_VALIDATE, &data_blob_null);
 449                         messaging_send(conn->msg_ctx,
 450                                        messaging_server_id(conn->msg_ctx),
 451                                        MSG_DBWRAP_G_LOCK_RETRY,
 452                                        &data_blob_null);
 453                         TALLOC_FREE(hdr);
 454                         goto next_pkt;
 455                 }
 456
 457                 msg_state = talloc(NULL, struct deferred_msg_state);
 458                 if (msg_state == NULL) {
 459                         DEBUG(0, ("talloc failed\n"));
 460                         TALLOC_FREE(hdr);
 461                         goto next_pkt;
 462                 }
 463
 464                 if (!(msg_state->rec = ctdb_pull_messaging_rec(
 465                               msg_state, state.req.length, msg))) {
 466                         DEBUG(0, ("ctdbd_pull_messaging_rec failed\n"));
 467                         TALLOC_FREE(msg_state);
 468                         TALLOC_FREE(hdr);
 469                         goto next_pkt;
 470                 }
 471
 472                 TALLOC_FREE(hdr);
 473
 474                 msg_state->msg_ctx = conn->msg_ctx;
 475
 476                 /*
 477                  * We're waiting for a call reply, but an async message has
 478                  * crossed. Defer dispatching to the toplevel event loop.
 479                  */
 480                 evt = tevent_add_timer(conn->msg_ctx->event_ctx,
 481                                       conn->msg_ctx->event_ctx,
 482                                       timeval_zero(),
 483                                       deferred_message_dispatch,
 484                                       msg_state);
 485                 if (evt == NULL) {
 486                         DEBUG(0, ("event_add_timed failed\n"));
 487                         TALLOC_FREE(msg_state);
 488                         TALLOC_FREE(hdr);
 489                         goto next_pkt;
 490                 }
 491
 492                 goto next_pkt;
 493         }
 494
 495         if ((reqid != 0) && (hdr->reqid != reqid)) {
 496                 /* we got the wrong reply */
 497                 DEBUG(0,("Discarding mismatched ctdb reqid %u should have "
 498                          "been %u\n", hdr->reqid, reqid));
 499                 TALLOC_FREE(hdr);
 500                 goto next_pkt;
 501         }
 502
 503         *((void **)result) = talloc_move(mem_ctx, &hdr);
 504
 505         return NT_STATUS_OK;
 506 }
 507
 508 /*
 509  * Get us a ctdbd connection
 510  */
 511
 512 static NTSTATUS ctdbd_init_connection(TALLOC_CTX *mem_ctx,
 513                                       struct ctdbd_connection **pconn)
 514 {
 515         struct ctdbd_connection *conn;
 516         NTSTATUS status;
 517
 518         if (!(conn = talloc_zero(mem_ctx, struct ctdbd_connection))) {
 519                 DEBUG(0, ("talloc failed\n"));
 520                 return NT_STATUS_NO_MEMORY;
 521         }
 522
 523         status = ctdbd_connect(conn, &conn->pkt);
 524
 525         if (!NT_STATUS_IS_OK(status)) {
 526                 DEBUG(10, ("ctdbd_connect failed: %s\n", nt_errstr(status)));
 527                 goto fail;
 528         }
 529
 530         status = get_cluster_vnn(conn, &conn->our_vnn);
 531
 532         if (!NT_STATUS_IS_OK(status)) {
 533                 DEBUG(10, ("get_cluster_vnn failed: %s\n", nt_errstr(status)));
 534                 goto fail;
 535         }
 536
 537         if (!ctdbd_working(conn, conn->our_vnn)) {
 538                 DEBUG(2, ("Node is not working, can not connect\n"));
 539                 status = NT_STATUS_INTERNAL_DB_ERROR;
 540                 goto fail;
 541         }
 542
 543         generate_random_buffer((unsigned char *)&conn->rand_srvid,
 544                                sizeof(conn->rand_srvid));
 545
 546         status = register_with_ctdbd(conn, conn->rand_srvid);
 547
 548         if (!NT_STATUS_IS_OK(status)) {
 549                 DEBUG(5, ("Could not register random srvid: %s\n",
 550                           nt_errstr(status)));
 551                 goto fail;
 552         }
 553
 554         *pconn = conn;
 555         return NT_STATUS_OK;
 556
 557  fail:
 558         TALLOC_FREE(conn);
 559         return status;
 560 }
 561
 562 /*
 563  * Get us a ctdbd connection and register us as a process
 564  */
 565
 566 NTSTATUS ctdbd_messaging_connection(TALLOC_CTX *mem_ctx,
 567                                     struct ctdbd_connection **pconn)
 568 {
 569         struct ctdbd_connection *conn;
 570         NTSTATUS status;
 571
 572         status = ctdbd_init_connection(mem_ctx, &conn);
 573
 574         if (!NT_STATUS_IS_OK(status)) {
 575                 return status;
 576         }
 577
 578         status = register_with_ctdbd(conn, (uint64_t)getpid());
 579         if (!NT_STATUS_IS_OK(status)) {
 580                 goto fail;
 581         }
 582
 583         status = register_with_ctdbd(conn, MSG_SRVID_SAMBA);
 584         if (!NT_STATUS_IS_OK(status)) {
 585                 goto fail;
 586         }
 587
 588         status = register_with_ctdbd(conn, CTDB_SRVID_SAMBA_NOTIFY);
 589         if (!NT_STATUS_IS_OK(status)) {
 590                 goto fail;
 591         }
 592
 593         *pconn = conn;
 594         return NT_STATUS_OK;
 595
 596  fail:
 597         TALLOC_FREE(conn);
 598         return status;
 599 }
 600
 601 struct messaging_context *ctdb_conn_msg_ctx(struct ctdbd_connection *conn)
 602 {
 603         return conn->msg_ctx;
 604 }
 605
 606 int ctdbd_conn_get_fd(struct ctdbd_connection *conn)
 607 {
 608         return ctdb_packet_get_fd(conn->pkt);
 609 }
 610
 611 /*
 612  * Packet handler to receive and handle a ctdb message
 613  */
 614 static NTSTATUS ctdb_handle_message(uint8_t *buf, size_t length,
 615                                     void *private_data)
 616 {
 617         struct ctdbd_connection *conn = talloc_get_type_abort(
 618                 private_data, struct ctdbd_connection);
 619         struct ctdb_req_message *msg;
 620         struct messaging_rec *msg_rec;
 621
 622         msg = (struct ctdb_req_message *)buf;
 623
 624         if (msg->hdr.operation != CTDB_REQ_MESSAGE) {
 625                 DEBUG(0, ("Received async msg of type %u, discarding\n",
 626                           msg->hdr.operation));
 627                 TALLOC_FREE(buf);
 628                 return NT_STATUS_INVALID_PARAMETER;
 629         }
 630
 631         if ((conn->release_ip_handler != NULL)
 632             && (msg->srvid == CTDB_SRVID_RELEASE_IP)) {
 633                 /* must be dispatched immediately */
 634                 DEBUG(10, ("received CTDB_SRVID_RELEASE_IP\n"));
 635                 conn->release_ip_handler((const char *)msg->data,
 636                                          conn->release_ip_priv);
 637                 TALLOC_FREE(buf);
 638                 return NT_STATUS_OK;
 639         }
 640
 641         SMB_ASSERT(conn->msg_ctx != NULL);
 642
 643         if ((msg->srvid == CTDB_SRVID_RECONFIGURE)
 644             || (msg->srvid == CTDB_SRVID_SAMBA_NOTIFY)){
 645                 DEBUG(0,("Got cluster reconfigure message\n"));
 646                 /*
 647                  * when the cluster is reconfigured or someone of the
 648                  * family has passed away (SAMBA_NOTIFY), we need to
 649                  * clean the brl database
 650                  */
 651                 messaging_send(conn->msg_ctx,
 652                                messaging_server_id(conn->msg_ctx),
 653                                MSG_SMB_BRL_VALIDATE, &data_blob_null);
 654
 655                 messaging_send(conn->msg_ctx,
 656                                messaging_server_id(conn->msg_ctx),
 657                                MSG_DBWRAP_G_LOCK_RETRY,
 658                                &data_blob_null);
 659
 660                 TALLOC_FREE(buf);
 661                 return NT_STATUS_OK;
 662         }
 663
 664         /* only messages to our pid or the broadcast are valid here */
 665         if (msg->srvid != getpid() && msg->srvid != MSG_SRVID_SAMBA) {
 666                 DEBUG(0,("Got unexpected message with srvid=%llu\n",
 667                          (unsigned long long)msg->srvid));
 668                 TALLOC_FREE(buf);
 669                 return NT_STATUS_OK;
 670         }
 671
 672         if (!(msg_rec = ctdb_pull_messaging_rec(NULL, length, msg))) {
 673                 DEBUG(10, ("ctdb_pull_messaging_rec failed\n"));
 674                 TALLOC_FREE(buf);
 675                 return NT_STATUS_NO_MEMORY;
 676         }
 677
 678         messaging_dispatch_rec(conn->msg_ctx, msg_rec);
 679
 680         TALLOC_FREE(msg_rec);
 681         TALLOC_FREE(buf);
 682         return NT_STATUS_OK;
 683 }
 684
 685 /*
 686  * The ctdbd socket is readable asynchronuously
 687  */
 688
 689 static void ctdbd_socket_handler(struct tevent_context *event_ctx,
 690                                  struct tevent_fd *event,
 691                                  uint16 flags,
 692                                  void *private_data)
 693 {
 694         struct ctdbd_connection *conn = talloc_get_type_abort(
 695                 private_data, struct ctdbd_connection);
 696
 697         NTSTATUS status;
 698
 699         status = ctdb_packet_fd_read(conn->pkt);
 700
 701         if (!NT_STATUS_IS_OK(status)) {
 702                 DEBUG(0, ("packet_fd_read failed: %s\n", nt_errstr(status)));
 703                 cluster_fatal("ctdbd died\n");
 704         }
 705
 706         while (ctdb_packet_handler(conn->pkt, ctdb_req_complete,
 707                               ctdb_handle_message, conn, &status)) {
 708                 if (!NT_STATUS_IS_OK(status)) {
 709                         DEBUG(10, ("could not handle incoming message: %s\n",
 710                                    nt_errstr(status)));
 711                 }
 712         }
 713 }
 714
 715 /*
 716  * Prepare a ctdbd connection to receive messages
 717  */
 718
 719 NTSTATUS ctdbd_register_msg_ctx(struct ctdbd_connection *conn,
 720                                 struct messaging_context *msg_ctx)
 721 {
 722         SMB_ASSERT(conn->msg_ctx == NULL);
 723         SMB_ASSERT(conn->fde == NULL);
 724
 725         if (!(conn->fde = tevent_add_fd(msg_ctx->event_ctx, conn,
 726                                        ctdb_packet_get_fd(conn->pkt),
 727                                        TEVENT_FD_READ,
 728                                        ctdbd_socket_handler,
 729                                        conn))) {
 730                 DEBUG(0, ("event_add_fd failed\n"));
 731                 return NT_STATUS_NO_MEMORY;
 732         }
 733
 734         conn->msg_ctx = msg_ctx;
 735
 736         return NT_STATUS_OK;
 737 }
 738
 739 /*
 740  * Send a messaging message across a ctdbd
 741  */
 742
 743 NTSTATUS ctdbd_messaging_send(struct ctdbd_connection *conn,
 744                               uint32_t dst_vnn, uint64_t dst_srvid,
 745                               struct messaging_rec *msg)
 746 {
 747         DATA_BLOB blob;
 748         NTSTATUS status;
 749         enum ndr_err_code ndr_err;
 750
 751         ndr_err = ndr_push_struct_blob(
 752                 &blob, talloc_tos(), msg,
 753                 (ndr_push_flags_fn_t)ndr_push_messaging_rec);
 754
 755         if (!NDR_ERR_CODE_IS_SUCCESS(ndr_err)) {
 756                 DEBUG(0, ("ndr_push_struct_blob failed: %s\n",
 757                           ndr_errstr(ndr_err)));
 758                 return ndr_map_error2ntstatus(ndr_err);
 759         }
 760
 761         status = ctdbd_messaging_send_blob(conn, dst_vnn, dst_srvid,
 762                                            blob.data, blob.length);
 763         TALLOC_FREE(blob.data);
 764         return status;
 765 }
 766
 767 NTSTATUS ctdbd_messaging_send_blob(struct ctdbd_connection *conn,
 768                                    uint32_t dst_vnn, uint64_t dst_srvid,
 769                                    const uint8_t *buf, size_t buflen)
 770 {
 771         struct ctdb_req_message r;
 772         NTSTATUS status;
 773
 774         r.hdr.length = offsetof(struct ctdb_req_message, data) + buflen;
 775         r.hdr.ctdb_magic = CTDB_MAGIC;
 776         r.hdr.ctdb_version = CTDB_VERSION;
 777         r.hdr.generation = 1;
 778         r.hdr.operation  = CTDB_REQ_MESSAGE;
 779         r.hdr.destnode   = dst_vnn;
 780         r.hdr.srcnode    = conn->our_vnn;
 781         r.hdr.reqid      = 0;
 782         r.srvid          = dst_srvid;
 783         r.datalen        = buflen;
 784
 785         DEBUG(10, ("ctdbd_messaging_send: Sending ctdb packet\n"));
 786         ctdb_packet_dump(&r.hdr);
 787
 788         status = ctdb_packet_send(
 789                 conn->pkt, 2,
 790                 data_blob_const(&r, offsetof(struct ctdb_req_message, data)),
 791                 data_blob_const(buf, buflen));
 792
 793         if (!NT_STATUS_IS_OK(status)) {
 794                 DEBUG(0, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
 795                 return status;
 796         }
 797
 798         status = ctdb_packet_flush(conn->pkt);
 799         if (!NT_STATUS_IS_OK(status)) {
 800                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
 801                 cluster_fatal("cluster dispatch daemon msg write error\n");
 802         }
 803         return NT_STATUS_OK;
 804 }
 805
 806 /*
 807  * send/recv a generic ctdb control message
 808  */
 809 static NTSTATUS ctdbd_control(struct ctdbd_connection *conn,
 810                               uint32_t vnn, uint32_t opcode,
 811                               uint64_t srvid, uint32_t flags,
 812                               TDB_DATA data,
 813                               TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
 814                               int *cstatus)
 815 {
 816         struct ctdb_req_control req;
 817         struct ctdb_reply_control *reply = NULL;
 818         struct ctdbd_connection *new_conn = NULL;
 819         NTSTATUS status;
 820
 821         if (conn == NULL) {
 822                 status = ctdbd_init_connection(NULL, &new_conn);
 823
 824                 if (!NT_STATUS_IS_OK(status)) {
 825                         DEBUG(10, ("Could not init temp connection: %s\n",
 826                                    nt_errstr(status)));
 827                         goto fail;
 828                 }
 829
 830                 conn = new_conn;
 831         }
 832
 833         ZERO_STRUCT(req);
 834         req.hdr.length = offsetof(struct ctdb_req_control, data) + data.dsize;
 835         req.hdr.ctdb_magic   = CTDB_MAGIC;
 836         req.hdr.ctdb_version = CTDB_VERSION;
 837         req.hdr.operation    = CTDB_REQ_CONTROL;
 838         req.hdr.reqid        = ctdbd_next_reqid(conn);
 839         req.hdr.destnode     = vnn;
 840         req.opcode           = opcode;
 841         req.srvid            = srvid;
 842         req.datalen          = data.dsize;
 843         req.flags            = flags;
 844
 845         DEBUG(10, ("ctdbd_control: Sending ctdb packet\n"));
 846         ctdb_packet_dump(&req.hdr);
 847
 848         status = ctdb_packet_send(
 849                 conn->pkt, 2,
 850                 data_blob_const(&req, offsetof(struct ctdb_req_control, data)),
 851                 data_blob_const(data.dptr, data.dsize));
 852
 853         if (!NT_STATUS_IS_OK(status)) {
 854                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
 855                 goto fail;
 856         }
 857
 858         status = ctdb_packet_flush(conn->pkt);
 859
 860         if (!NT_STATUS_IS_OK(status)) {
 861                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
 862                 cluster_fatal("cluster dispatch daemon control write error\n");
 863         }
 864
 865         if (flags & CTDB_CTRL_FLAG_NOREPLY) {
 866                 TALLOC_FREE(new_conn);
 867                 if (cstatus) {
 868                         *cstatus = 0;
 869                 }
 870                 return NT_STATUS_OK;
 871         }
 872
 873         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
 874
 875         if (!NT_STATUS_IS_OK(status)) {
 876                 DEBUG(10, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
 877                 goto fail;
 878         }
 879
 880         if (reply->hdr.operation != CTDB_REPLY_CONTROL) {
 881                 DEBUG(0, ("received invalid reply\n"));
 882                 goto fail;
 883         }
 884
 885         if (outdata) {
 886                 if (!(outdata->dptr = (uint8 *)talloc_memdup(
 887                               mem_ctx, reply->data, reply->datalen))) {
 888                         TALLOC_FREE(reply);
 889                         return NT_STATUS_NO_MEMORY;
 890                 }
 891                 outdata->dsize = reply->datalen;
 892         }
 893         if (cstatus) {
 894                 (*cstatus) = reply->status;
 895         }
 896
 897         status = NT_STATUS_OK;
 898
 899  fail:
 900         TALLOC_FREE(new_conn);
 901         TALLOC_FREE(reply);
 902         return status;
 903 }
 904
 905 /*
 906  * see if a remote process exists
 907  */
 908 bool ctdbd_process_exists(struct ctdbd_connection *conn, uint32_t vnn, pid_t pid)
 909 {
 910         struct server_id id;
 911         bool result;
 912
 913         id.pid = pid;
 914         id.vnn = vnn;
 915
 916         if (!ctdb_processes_exist(conn, &id, 1, &result)) {
 917                 DEBUG(10, ("ctdb_processes_exist failed\n"));
 918                 return false;
 919         }
 920         return result;
 921 }
 922
 923 bool ctdb_processes_exist(struct ctdbd_connection *conn,
 924                           const struct server_id *pids, int num_pids,
 925                           bool *results)
 926 {
 927         TALLOC_CTX *frame = talloc_stackframe();
 928         int i, num_received;
 929         NTSTATUS status;
 930         uint32_t *reqids;
 931         bool result = false;
 932
 933         reqids = talloc_array(talloc_tos(), uint32_t, num_pids);
 934         if (reqids == NULL) {
 935                 goto fail;
 936         }
 937
 938         for (i=0; i<num_pids; i++) {
 939                 struct ctdb_req_control req;
 940                 pid_t pid;
 941
 942                 results[i] = false;
 943                 reqids[i] = ctdbd_next_reqid(conn);
 944
 945                 ZERO_STRUCT(req);
 946
 947                 /*
 948                  * pids[i].pid is uint64_t, scale down to pid_t which
 949                  * is the wire protocol towards ctdb.
 950                  */
 951                 pid = pids[i].pid;
 952
 953                 DEBUG(10, ("Requesting PID %d/%d, reqid=%d\n",
 954                            (int)pids[i].vnn, (int)pid,
 955                            (int)reqids[i]));
 956
 957                 req.hdr.length = offsetof(struct ctdb_req_control, data);
 958                 req.hdr.length += sizeof(pid);
 959                 req.hdr.ctdb_magic   = CTDB_MAGIC;
 960                 req.hdr.ctdb_version = CTDB_VERSION;
 961                 req.hdr.operation    = CTDB_REQ_CONTROL;
 962                 req.hdr.reqid        = reqids[i];
 963                 req.hdr.destnode     = pids[i].vnn;
 964                 req.opcode           = CTDB_CONTROL_PROCESS_EXISTS;
 965                 req.srvid            = 0;
 966                 req.datalen          = sizeof(pid);
 967                 req.flags            = 0;
 968
 969                 DEBUG(10, ("ctdbd_control: Sending ctdb packet\n"));
 970                 ctdb_packet_dump(&req.hdr);
 971
 972                 status = ctdb_packet_send(
 973                         conn->pkt, 2,
 974                         data_blob_const(
 975                                 &req, offsetof(struct ctdb_req_control, data)),
 976                         data_blob_const(&pid, sizeof(pid)));
 977                 if (!NT_STATUS_IS_OK(status)) {
 978                         DEBUG(10, ("ctdb_packet_send failed: %s\n",
 979                                    nt_errstr(status)));
 980                         goto fail;
 981                 }
 982         }
 983
 984         status = ctdb_packet_flush(conn->pkt);
 985         if (!NT_STATUS_IS_OK(status)) {
 986                 DEBUG(10, ("ctdb_packet_flush failed: %s\n",
 987                            nt_errstr(status)));
 988                 goto fail;
 989         }
 990
 991         num_received = 0;
 992
 993         while (num_received < num_pids) {
 994                 struct ctdb_reply_control *reply = NULL;
 995                 uint32_t reqid;
 996
 997                 status = ctdb_read_req(conn, 0, talloc_tos(), (void *)&reply);
 998                 if (!NT_STATUS_IS_OK(status)) {
 999                         DEBUG(10, ("ctdb_read_req failed: %s\n",
1000                                    nt_errstr(status)));
1001                         goto fail;
1002                 }
1003
1004                 if (reply->hdr.operation != CTDB_REPLY_CONTROL) {
1005                         DEBUG(10, ("Received invalid reply\n"));
1006                         goto fail;
1007                 }
1008
1009                 reqid = reply->hdr.reqid;
1010
1011                 DEBUG(10, ("Received reqid %d\n", (int)reqid));
1012
1013                 for (i=0; i<num_pids; i++) {
1014                         if (reqid == reqids[i]) {
1015                                 break;
1016                         }
1017                 }
1018                 if (i == num_pids) {
1019                         DEBUG(10, ("Received unknown record number %u\n",
1020                                    (unsigned)reqid));
1021                         goto fail;
1022                 }
1023                 results[i] = ((reply->status) == 0);
1024                 TALLOC_FREE(reply);
1025                 num_received += 1;
1026         }
1027
1028         result = true;
1029 fail:
1030         TALLOC_FREE(frame);
1031         return result;
1032 }
1033
1034 struct ctdb_vnn_list {
1035         uint32_t vnn;
1036         uint32_t reqid;
1037         unsigned num_srvids;
1038         unsigned num_filled;
1039         uint64_t *srvids;
1040         unsigned *pid_indexes;
1041 };
1042
1043 /*
1044  * Get a list of all vnns mentioned in a list of
1045  * server_ids. vnn_indexes tells where in the vnns array we have to
1046  * place the pids.
1047  */
1048 static bool ctdb_collect_vnns(TALLOC_CTX *mem_ctx,
1049                               const struct server_id *pids, unsigned num_pids,
1050                               struct ctdb_vnn_list **pvnns,
1051                               unsigned *pnum_vnns)
1052 {
1053         struct ctdb_vnn_list *vnns = NULL;
1054         unsigned *vnn_indexes = NULL;
1055         unsigned i, num_vnns = 0;
1056
1057         vnn_indexes = talloc_array(mem_ctx, unsigned, num_pids);
1058         if (vnn_indexes == NULL) {
1059                 DEBUG(1, ("talloc_array failed\n"));
1060                 goto fail;
1061         }
1062
1063         for (i=0; i<num_pids; i++) {
1064                 unsigned j;
1065                 uint32_t vnn = pids[i].vnn;
1066
1067                 for (j=0; j<num_vnns; j++) {
1068                         if (vnn == vnns[j].vnn) {
1069                                 break;
1070                         }
1071                 }
1072                 vnn_indexes[i] = j;
1073
1074                 if (j < num_vnns) {
1075                         /*
1076                          * Already in the array
1077                          */
1078                         vnns[j].num_srvids += 1;
1079                         continue;
1080                 }
1081                 vnns = talloc_realloc(mem_ctx, vnns, struct ctdb_vnn_list,
1082                                       num_vnns+1);
1083                 if (vnns == NULL) {
1084                         DEBUG(1, ("talloc_realloc failed\n"));
1085                         goto fail;
1086                 }
1087                 vnns[num_vnns].vnn = vnn;
1088                 vnns[num_vnns].num_srvids = 1;
1089                 vnns[num_vnns].num_filled = 0;
1090                 num_vnns += 1;
1091         }
1092         for (i=0; i<num_vnns; i++) {
1093                 struct ctdb_vnn_list *vnn = &vnns[i];
1094
1095                 vnn->srvids = talloc_array(vnns, uint64_t, vnn->num_srvids);
1096                 if (vnn->srvids == NULL) {
1097                         DEBUG(1, ("talloc_array failed\n"));
1098                         goto fail;
1099                 }
1100                 vnn->pid_indexes = talloc_array(vnns, unsigned,
1101                                                 vnn->num_srvids);
1102                 if (vnn->pid_indexes == NULL) {
1103                         DEBUG(1, ("talloc_array failed\n"));
1104                         goto fail;
1105                 }
1106         }
1107         for (i=0; i<num_pids; i++) {
1108                 struct ctdb_vnn_list *vnn = &vnns[vnn_indexes[i]];
1109                 vnn->srvids[vnn->num_filled] = pids[i].unique_id;
1110                 vnn->pid_indexes[vnn->num_filled] = i;
1111                 vnn->num_filled += 1;
1112         }
1113
1114         TALLOC_FREE(vnn_indexes);
1115         *pvnns = vnns;
1116         *pnum_vnns = num_vnns;
1117         return true;
1118 fail:
1119         TALLOC_FREE(vnns);
1120         TALLOC_FREE(vnn_indexes);
1121         return false;
1122 }
1123
1124 #ifdef HAVE_CTDB_CONTROL_CHECK_SRVIDS_DECL
1125
1126 bool ctdb_serverids_exist(struct ctdbd_connection *conn,
1127                           const struct server_id *pids, unsigned num_pids,
1128                           bool *results)
1129 {
1130         unsigned i, num_received;
1131         NTSTATUS status;
1132         struct ctdb_vnn_list *vnns = NULL;
1133         unsigned num_vnns;
1134         bool result = false;
1135
1136         if (!ctdb_collect_vnns(talloc_tos(), pids, num_pids,
1137                                &vnns, &num_vnns)) {
1138                 DEBUG(1, ("ctdb_collect_vnns failed\n"));
1139                 goto fail;
1140         }
1141
1142         for (i=0; i<num_vnns; i++) {
1143                 struct ctdb_vnn_list *vnn = &vnns[i];
1144                 struct ctdb_req_control req;
1145
1146                 vnn->reqid = ctdbd_next_reqid(conn);
1147
1148                 ZERO_STRUCT(req);
1149
1150                 DEBUG(10, ("Requesting VNN %d, reqid=%d, num_srvids=%u\n",
1151                            (int)vnn->vnn, (int)vnn->reqid, vnn->num_srvids));
1152
1153                 req.hdr.length = offsetof(struct ctdb_req_control, data);
1154                 req.hdr.ctdb_magic   = CTDB_MAGIC;
1155                 req.hdr.ctdb_version = CTDB_VERSION;
1156                 req.hdr.operation    = CTDB_REQ_CONTROL;
1157                 req.hdr.reqid        = vnn->reqid;
1158                 req.hdr.destnode     = vnn->vnn;
1159                 req.opcode           = CTDB_CONTROL_CHECK_SRVIDS;
1160                 req.srvid            = 0;
1161                 req.datalen          = sizeof(uint64_t) * vnn->num_srvids;
1162                 req.hdr.length      += req.datalen;
1163                 req.flags            = 0;
1164
1165                 DEBUG(10, ("ctdbd_control: Sending ctdb packet\n"));
1166                 ctdb_packet_dump(&req.hdr);
1167
1168                 status = ctdb_packet_send(
1169                         conn->pkt, 2,
1170                         data_blob_const(
1171                                 &req, offsetof(struct ctdb_req_control,
1172                                                data)),
1173                         data_blob_const(vnn->srvids, req.datalen));
1174                 if (!NT_STATUS_IS_OK(status)) {
1175                         DEBUG(1, ("ctdb_packet_send failed: %s\n",
1176                                   nt_errstr(status)));
1177                         goto fail;
1178                 }
1179         }
1180
1181         status = ctdb_packet_flush(conn->pkt);
1182         if (!NT_STATUS_IS_OK(status)) {
1183                 DEBUG(1, ("ctdb_packet_flush failed: %s\n",
1184                           nt_errstr(status)));
1185                 goto fail;
1186         }
1187
1188         num_received = 0;
1189
1190         while (num_received < num_vnns) {
1191                 struct ctdb_reply_control *reply = NULL;
1192                 struct ctdb_vnn_list *vnn;
1193                 uint32_t reqid;
1194                 uint8_t *reply_data;
1195
1196                 status = ctdb_read_req(conn, 0, talloc_tos(), (void *)&reply);
1197                 if (!NT_STATUS_IS_OK(status)) {
1198                         DEBUG(1, ("ctdb_read_req failed: %s\n",
1199                                   nt_errstr(status)));
1200                         goto fail;
1201                 }
1202
1203                 if (reply->hdr.operation != CTDB_REPLY_CONTROL) {
1204                         DEBUG(1, ("Received invalid reply %u\n",
1205                                   (unsigned)reply->hdr.operation));
1206                         goto fail;
1207                 }
1208
1209                 reqid = reply->hdr.reqid;
1210
1211                 DEBUG(10, ("Received reqid %d\n", (int)reqid));
1212
1213                 for (i=0; i<num_vnns; i++) {
1214                         if (reqid == vnns[i].reqid) {
1215                                 break;
1216                         }
1217                 }
1218                 if (i == num_vnns) {
1219                         DEBUG(1, ("Received unknown reqid number %u\n",
1220                                   (unsigned)reqid));
1221                         goto fail;
1222                 }
1223
1224                 DEBUG(10, ("Found index %u\n", i));
1225
1226                 vnn = &vnns[i];
1227
1228                 DEBUG(10, ("Received vnn %u, vnn->num_srvids %u, datalen %u\n",
1229                            (unsigned)vnn->vnn, vnn->num_srvids,
1230                            (unsigned)reply->datalen));
1231
1232                 if (reply->datalen >= ((vnn->num_srvids+7)/8)) {
1233                         /*
1234                          * Got a real reply
1235                          */
1236                         reply_data = reply->data;
1237                 } else {
1238                         /*
1239                          * Got an error reply
1240                          */
1241                         DEBUG(5, ("Received short reply len %d, status %u, "
1242                                   "errorlen %u\n",
1243                                   (unsigned)reply->datalen,
1244                                   (unsigned)reply->status,
1245                                   (unsigned)reply->errorlen));
1246                         dump_data(5, reply->data, reply->errorlen);
1247
1248                         /*
1249                          * This will trigger everything set to false
1250                          */
1251                         reply_data = NULL;
1252                 }
1253
1254                 for (i=0; i<vnn->num_srvids; i++) {
1255                         int idx = vnn->pid_indexes[i];
1256
1257                         if (pids[i].unique_id ==
1258                             SERVERID_UNIQUE_ID_NOT_TO_VERIFY) {
1259                                 results[idx] = true;
1260                                 continue;
1261                         }
1262                         results[idx] =
1263                                 (reply_data != NULL) &&
1264                                 ((reply_data[i/8] & (1<<(i%8))) != 0);
1265                 }
1266
1267                 TALLOC_FREE(reply);
1268                 num_received += 1;
1269         }
1270
1271         result = true;
1272 fail:
1273         TALLOC_FREE(vnns);
1274         return result;
1275 }
1276
1277 #endif /* HAVE_CTDB_CONTROL_CHECK_SRVIDS_DECL */
1278
1279 /*
1280  * Get a db path
1281  */
1282 char *ctdbd_dbpath(struct ctdbd_connection *conn,
1283                    TALLOC_CTX *mem_ctx, uint32_t db_id)
1284 {
1285         NTSTATUS status;
1286         TDB_DATA data;
1287         int32_t cstatus;
1288
1289         data.dptr = (uint8_t*)&db_id;
1290         data.dsize = sizeof(db_id);
1291
1292         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1293                                CTDB_CONTROL_GETDBPATH, 0, 0, data,
1294                                mem_ctx, &data, &cstatus);
1295         if (!NT_STATUS_IS_OK(status) || cstatus != 0) {
1296                 DEBUG(0,(__location__ " ctdb_control for getdbpath failed\n"));
1297                 return NULL;
1298         }
1299
1300         return (char *)data.dptr;
1301 }
1302
1303 /*
1304  * attach to a ctdb database
1305  */
1306 NTSTATUS ctdbd_db_attach(struct ctdbd_connection *conn,
1307                          const char *name, uint32_t *db_id, int tdb_flags)
1308 {
1309         NTSTATUS status;
1310         TDB_DATA data;
1311         int32_t cstatus;
1312         bool persistent = (tdb_flags & TDB_CLEAR_IF_FIRST) == 0;
1313
1314         data = string_term_tdb_data(name);
1315
1316         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1317                                persistent
1318                                ? CTDB_CONTROL_DB_ATTACH_PERSISTENT
1319                                : CTDB_CONTROL_DB_ATTACH,
1320                                tdb_flags, 0, data, NULL, &data, &cstatus);
1321         if (!NT_STATUS_IS_OK(status)) {
1322                 DEBUG(0, (__location__ " ctdb_control for db_attach "
1323                           "failed: %s\n", nt_errstr(status)));
1324                 return status;
1325         }
1326
1327         if (cstatus != 0 || data.dsize != sizeof(uint32_t)) {
1328                 DEBUG(0,(__location__ " ctdb_control for db_attach failed\n"));
1329                 return NT_STATUS_INTERNAL_ERROR;
1330         }
1331
1332         *db_id = *(uint32_t *)data.dptr;
1333         talloc_free(data.dptr);
1334
1335         if (!(tdb_flags & TDB_SEQNUM)) {
1336                 return NT_STATUS_OK;
1337         }
1338
1339         data.dptr = (uint8_t *)db_id;
1340         data.dsize = sizeof(*db_id);
1341
1342         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1343                                CTDB_CONTROL_ENABLE_SEQNUM, 0, 0, data,
1344                                NULL, NULL, &cstatus);
1345         if (!NT_STATUS_IS_OK(status) || cstatus != 0) {
1346                 DEBUG(0,(__location__ " ctdb_control for enable seqnum "
1347                          "failed\n"));
1348                 return NT_STATUS_IS_OK(status) ? NT_STATUS_INTERNAL_ERROR :
1349                         status;
1350         }
1351
1352         return NT_STATUS_OK;
1353 }
1354
1355 /*
1356  * force the migration of a record to this node
1357  */
1358 NTSTATUS ctdbd_migrate(struct ctdbd_connection *conn, uint32_t db_id,
1359                        TDB_DATA key)
1360 {
1361         struct ctdb_req_call req;
1362         struct ctdb_reply_call *reply;
1363         NTSTATUS status;
1364
1365         ZERO_STRUCT(req);
1366
1367         req.hdr.length = offsetof(struct ctdb_req_call, data) + key.dsize;
1368         req.hdr.ctdb_magic   = CTDB_MAGIC;
1369         req.hdr.ctdb_version = CTDB_VERSION;
1370         req.hdr.operation    = CTDB_REQ_CALL;
1371         req.hdr.reqid        = ctdbd_next_reqid(conn);
1372         req.flags            = CTDB_IMMEDIATE_MIGRATION;
1373         req.callid           = CTDB_NULL_FUNC;
1374         req.db_id            = db_id;
1375         req.keylen           = key.dsize;
1376
1377         DEBUG(10, ("ctdbd_migrate: Sending ctdb packet\n"));
1378         ctdb_packet_dump(&req.hdr);
1379
1380         status = ctdb_packet_send(
1381                 conn->pkt, 2,
1382                 data_blob_const(&req, offsetof(struct ctdb_req_call, data)),
1383                 data_blob_const(key.dptr, key.dsize));
1384
1385         if (!NT_STATUS_IS_OK(status)) {
1386                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
1387                 return status;
1388         }
1389
1390         status = ctdb_packet_flush(conn->pkt);
1391
1392         if (!NT_STATUS_IS_OK(status)) {
1393                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
1394                 cluster_fatal("cluster dispatch daemon control write error\n");
1395         }
1396
1397         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
1398
1399         if (!NT_STATUS_IS_OK(status)) {
1400                 DEBUG(0, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
1401                 goto fail;
1402         }
1403
1404         if (reply->hdr.operation != CTDB_REPLY_CALL) {
1405                 DEBUG(0, ("received invalid reply\n"));
1406                 status = NT_STATUS_INTERNAL_ERROR;
1407                 goto fail;
1408         }
1409
1410         status = NT_STATUS_OK;
1411  fail:
1412
1413         TALLOC_FREE(reply);
1414         return status;
1415 }
1416
1417 /*
1418  * Fetch a record and parse it
1419  */
1420 NTSTATUS ctdbd_parse(struct ctdbd_connection *conn, uint32_t db_id,
1421                      TDB_DATA key, bool local_copy,
1422                      void (*parser)(TDB_DATA key, TDB_DATA data,
1423                                     void *private_data),
1424                      void *private_data)
1425 {
1426         struct ctdb_req_call req;
1427         struct ctdb_reply_call *reply;
1428         NTSTATUS status;
1429         uint32_t flags;
1430
1431 #ifdef HAVE_CTDB_WANT_READONLY_DECL
1432         flags = local_copy ? CTDB_WANT_READONLY : 0;
1433 #else
1434         flags = 0;
1435 #endif
1436
1437         ZERO_STRUCT(req);
1438
1439         req.hdr.length = offsetof(struct ctdb_req_call, data) + key.dsize;
1440         req.hdr.ctdb_magic   = CTDB_MAGIC;
1441         req.hdr.ctdb_version = CTDB_VERSION;
1442         req.hdr.operation    = CTDB_REQ_CALL;
1443         req.hdr.reqid        = ctdbd_next_reqid(conn);
1444         req.flags            = flags;
1445         req.callid           = CTDB_FETCH_FUNC;
1446         req.db_id            = db_id;
1447         req.keylen           = key.dsize;
1448
1449         status = ctdb_packet_send(
1450                 conn->pkt, 2,
1451                 data_blob_const(&req, offsetof(struct ctdb_req_call, data)),
1452                 data_blob_const(key.dptr, key.dsize));
1453
1454         if (!NT_STATUS_IS_OK(status)) {
1455                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
1456                 return status;
1457         }
1458
1459         status = ctdb_packet_flush(conn->pkt);
1460
1461         if (!NT_STATUS_IS_OK(status)) {
1462                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
1463                 cluster_fatal("cluster dispatch daemon control write error\n");
1464         }
1465
1466         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
1467
1468         if (!NT_STATUS_IS_OK(status)) {
1469                 DEBUG(0, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
1470                 goto fail;
1471         }
1472
1473         if (reply->hdr.operation != CTDB_REPLY_CALL) {
1474                 DEBUG(0, ("received invalid reply\n"));
1475                 status = NT_STATUS_INTERNAL_ERROR;
1476                 goto fail;
1477         }
1478
1479         if (reply->datalen == 0) {
1480                 /*
1481                  * Treat an empty record as non-existing
1482                  */
1483                 status = NT_STATUS_NOT_FOUND;
1484                 goto fail;
1485         }
1486
1487         parser(key, make_tdb_data(&reply->data[0], reply->datalen),
1488                private_data);
1489
1490         status = NT_STATUS_OK;
1491  fail:
1492         TALLOC_FREE(reply);
1493         return status;
1494 }
1495
1496 struct ctdbd_traverse_state {
1497         void (*fn)(TDB_DATA key, TDB_DATA data, void *private_data);
1498         void *private_data;
1499 };
1500
1501 /*
1502  * Handle a traverse record coming in on the ctdbd connection
1503  */
1504
1505 static NTSTATUS ctdb_traverse_handler(uint8_t *buf, size_t length,
1506                                       void *private_data)
1507 {
1508         struct ctdbd_traverse_state *state =
1509                 (struct ctdbd_traverse_state *)private_data;
1510
1511         struct ctdb_req_message *m;
1512         struct ctdb_rec_data *d;
1513         TDB_DATA key, data;
1514
1515         m = (struct ctdb_req_message *)buf;
1516
1517         if (length < sizeof(*m) || m->hdr.length != length) {
1518                 DEBUG(0, ("Got invalid message of length %d\n", (int)length));
1519                 TALLOC_FREE(buf);
1520                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1521         }
1522
1523         d = (struct ctdb_rec_data *)&m->data[0];
1524         if (m->datalen < sizeof(uint32_t) || m->datalen != d->length) {
1525                 DEBUG(0, ("Got invalid traverse data of length %d\n",
1526                           (int)m->datalen));
1527                 TALLOC_FREE(buf);
1528                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1529         }
1530
1531         key.dsize = d->keylen;
1532         key.dptr  = &d->data[0];
1533         data.dsize = d->datalen;
1534         data.dptr = &d->data[d->keylen];
1535
1536         if (key.dsize == 0 && data.dsize == 0) {
1537                 /* end of traverse */
1538                 return NT_STATUS_END_OF_FILE;
1539         }
1540
1541         if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
1542                 DEBUG(0, ("Got invalid ltdb header length %d\n",
1543                           (int)data.dsize));
1544                 TALLOC_FREE(buf);
1545                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1546         }
1547         data.dsize -= sizeof(struct ctdb_ltdb_header);
1548         data.dptr += sizeof(struct ctdb_ltdb_header);
1549
1550         if (state->fn) {
1551                 state->fn(key, data, state->private_data);
1552         }
1553
1554         TALLOC_FREE(buf);
1555         return NT_STATUS_OK;
1556 }
1557
1558 /*
1559   Traverse a ctdb database. This uses a kind-of hackish way to open a second
1560   connection to ctdbd to avoid the hairy recursive and async problems with
1561   everything in-line.
1562 */
1563
1564 NTSTATUS ctdbd_traverse(uint32_t db_id,
1565                         void (*fn)(TDB_DATA key, TDB_DATA data,
1566                                    void *private_data),
1567                         void *private_data)
1568 {
1569         struct ctdbd_connection *conn;
1570         NTSTATUS status;
1571
1572         TDB_DATA data;
1573         struct ctdb_traverse_start t;
1574         int cstatus;
1575         struct ctdbd_traverse_state state;
1576
1577         become_root();
1578         status = ctdbd_init_connection(NULL, &conn);
1579         unbecome_root();
1580         if (!NT_STATUS_IS_OK(status)) {
1581                 DEBUG(0, ("ctdbd_init_connection failed: %s\n",
1582                           nt_errstr(status)));
1583                 return status;
1584         }
1585
1586         t.db_id = db_id;
1587         t.srvid = conn->rand_srvid;
1588         t.reqid = ctdbd_next_reqid(conn);
1589
1590         data.dptr = (uint8_t *)&t;
1591         data.dsize = sizeof(t);
1592
1593         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1594                                CTDB_CONTROL_TRAVERSE_START, conn->rand_srvid, 0,
1595                                data, NULL, NULL, &cstatus);
1596
1597         if (!NT_STATUS_IS_OK(status) || (cstatus != 0)) {
1598
1599                 DEBUG(0,("ctdbd_control failed: %s, %d\n", nt_errstr(status),
1600                          cstatus));
1601
1602                 if (NT_STATUS_IS_OK(status)) {
1603                         /*
1604                          * We need a mapping here
1605                          */
1606                         status = NT_STATUS_UNSUCCESSFUL;
1607                 }
1608                 goto done;
1609         }
1610
1611         state.fn = fn;
1612         state.private_data = private_data;
1613
1614         while (True) {
1615
1616                 status = NT_STATUS_OK;
1617
1618                 if (ctdb_packet_handler(conn->pkt, ctdb_req_complete,
1619                                    ctdb_traverse_handler, &state, &status)) {
1620
1621                         if (NT_STATUS_EQUAL(status, NT_STATUS_END_OF_FILE)) {
1622                                 status = NT_STATUS_OK;
1623                                 break;
1624                         }
1625
1626                         /*
1627                          * There might be more in the queue
1628                          */
1629                         continue;
1630                 }
1631
1632                 if (!NT_STATUS_IS_OK(status)) {
1633                         break;
1634                 }
1635
1636                 status = ctdb_packet_fd_read_sync(conn->pkt);
1637
1638                 if (NT_STATUS_EQUAL(status, NT_STATUS_RETRY)) {
1639                         /*
1640                          * There might be more in the queue
1641                          */
1642                         continue;
1643                 }
1644
1645                 if (NT_STATUS_EQUAL(status, NT_STATUS_END_OF_FILE)) {
1646                         status = NT_STATUS_OK;
1647                         break;
1648                 }
1649
1650                 if (!NT_STATUS_IS_OK(status)) {
1651                         DEBUG(0, ("ctdb_packet_fd_read_sync failed: %s\n", nt_errstr(status)));
1652                         cluster_fatal("ctdbd died\n");
1653                 }
1654         }
1655
1656  done:
1657         TALLOC_FREE(conn);
1658         return status;
1659 }
1660
1661 /*
1662    This is used to canonicalize a ctdb_sock_addr structure.
1663 */
1664 static void smbd_ctdb_canonicalize_ip(const struct sockaddr_storage *in,
1665                                       struct sockaddr_storage *out)
1666 {
1667         memcpy(out, in, sizeof (*out));
1668
1669 #ifdef HAVE_IPV6
1670         if (in->ss_family == AF_INET6) {
1671                 const char prefix[12] = { 0,0,0,0,0,0,0,0,0,0,0xff,0xff };
1672                 const struct sockaddr_in6 *in6 =
1673                         (const struct sockaddr_in6 *)in;
1674                 struct sockaddr_in *out4 = (struct sockaddr_in *)out;
1675                 if (memcmp(&in6->sin6_addr, prefix, 12) == 0) {
1676                         memset(out, 0, sizeof(*out));
1677 #ifdef HAVE_SOCK_SIN_LEN
1678                         out4->sin_len = sizeof(*out);
1679 #endif
1680                         out4->sin_family = AF_INET;
1681                         out4->sin_port   = in6->sin6_port;
1682                         memcpy(&out4->sin_addr, &in6->sin6_addr.s6_addr[12], 4);
1683                 }
1684         }
1685 #endif
1686 }
1687
1688 /*
1689  * Register us as a server for a particular tcp connection
1690  */
1691
1692 NTSTATUS ctdbd_register_ips(struct ctdbd_connection *conn,
1693                             const struct sockaddr_storage *_server,
1694                             const struct sockaddr_storage *_client,
1695                             void (*release_ip_handler)(const char *ip_addr,
1696                                                        void *private_data),
1697                             void *private_data)
1698 {
1699         /*
1700          * we still use ctdb_control_tcp for ipv4
1701          * because we want to work against older ctdb
1702          * versions at runtime
1703          */
1704         struct ctdb_control_tcp p4;
1705 #ifdef HAVE_STRUCT_CTDB_CONTROL_TCP_ADDR
1706         struct ctdb_control_tcp_addr p;
1707 #endif
1708         TDB_DATA data;
1709         NTSTATUS status;
1710         struct sockaddr_storage client;
1711         struct sockaddr_storage server;
1712
1713         /*
1714          * Only one connection so far
1715          */
1716         SMB_ASSERT(conn->release_ip_handler == NULL);
1717
1718         smbd_ctdb_canonicalize_ip(_client, &client);
1719         smbd_ctdb_canonicalize_ip(_server, &server);
1720
1721         switch (client.ss_family) {
1722         case AF_INET:
1723                 memcpy(&p4.dest, &server, sizeof(p4.dest));
1724                 memcpy(&p4.src, &client, sizeof(p4.src));
1725                 data.dptr = (uint8_t *)&p4;
1726                 data.dsize = sizeof(p4);
1727                 break;
1728 #ifdef HAVE_STRUCT_CTDB_CONTROL_TCP_ADDR
1729         case AF_INET6:
1730                 memcpy(&p.dest.ip6, &server, sizeof(p.dest.ip6));
1731                 memcpy(&p.src.ip6, &client, sizeof(p.src.ip6));
1732                 data.dptr = (uint8_t *)&p;
1733                 data.dsize = sizeof(p);
1734                 break;
1735 #endif
1736         default:
1737                 return NT_STATUS_INTERNAL_ERROR;
1738         }
1739
1740         conn->release_ip_handler = release_ip_handler;
1741         conn->release_ip_priv = private_data;
1742
1743         /*
1744          * We want to be told about IP releases
1745          */
1746
1747         status = register_with_ctdbd(conn, CTDB_SRVID_RELEASE_IP);
1748         if (!NT_STATUS_IS_OK(status)) {
1749                 return status;
1750         }
1751
1752         /*
1753          * inform ctdb of our tcp connection, so if IP takeover happens ctdb
1754          * can send an extra ack to trigger a reset for our client, so it
1755          * immediately reconnects
1756          */
1757         return ctdbd_control(conn, CTDB_CURRENT_NODE,
1758                              CTDB_CONTROL_TCP_CLIENT, 0,
1759                              CTDB_CTRL_FLAG_NOREPLY, data, NULL, NULL, NULL);
1760 }
1761
1762 /*
1763  * We want to handle reconfigure events
1764  */
1765 NTSTATUS ctdbd_register_reconfigure(struct ctdbd_connection *conn)
1766 {
1767         return register_with_ctdbd(conn, CTDB_SRVID_RECONFIGURE);
1768 }
1769
1770 /*
1771   call a control on the local node
1772  */
1773 NTSTATUS ctdbd_control_local(struct ctdbd_connection *conn, uint32_t opcode,
1774                              uint64_t srvid, uint32_t flags, TDB_DATA data,
1775                              TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
1776                              int *cstatus)
1777 {
1778         return ctdbd_control(conn, CTDB_CURRENT_NODE, opcode, srvid, flags, data, mem_ctx, outdata, cstatus);
1779 }
1780
1781 NTSTATUS ctdb_watch_us(struct ctdbd_connection *conn)
1782 {
1783         struct ctdb_client_notify_register reg_data;
1784         size_t struct_len;
1785         NTSTATUS status;
1786         int cstatus;
1787
1788         reg_data.srvid = CTDB_SRVID_SAMBA_NOTIFY;
1789         reg_data.len = 1;
1790         reg_data.notify_data[0] = 0;
1791
1792         struct_len = offsetof(struct ctdb_client_notify_register,
1793                               notify_data) + reg_data.len;
1794
1795         status = ctdbd_control_local(
1796                 conn, CTDB_CONTROL_REGISTER_NOTIFY, conn->rand_srvid, 0,
1797                 make_tdb_data((uint8_t *)&reg_data, struct_len),
1798                 NULL, NULL, &cstatus);
1799         if (!NT_STATUS_IS_OK(status)) {
1800                 DEBUG(1, ("ctdbd_control_local failed: %s\n",
1801                           nt_errstr(status)));
1802         }
1803         return status;
1804 }
1805
1806 NTSTATUS ctdb_unwatch(struct ctdbd_connection *conn)
1807 {
1808         struct ctdb_client_notify_deregister dereg_data;
1809         NTSTATUS status;
1810         int cstatus;
1811
1812         dereg_data.srvid = CTDB_SRVID_SAMBA_NOTIFY;
1813
1814         status = ctdbd_control_local(
1815                 conn, CTDB_CONTROL_DEREGISTER_NOTIFY, conn->rand_srvid, 0,
1816                 make_tdb_data((uint8_t *)&dereg_data, sizeof(dereg_data)),
1817                 NULL, NULL, &cstatus);
1818         if (!NT_STATUS_IS_OK(status)) {
1819                 DEBUG(1, ("ctdbd_control_local failed: %s\n",
1820                           nt_errstr(status)));
1821         }
1822         return status;
1823 }
1824
1825 #else
1826
1827 NTSTATUS ctdbd_messaging_send_blob(struct ctdbd_connection *conn,
1828                                    uint32_t dst_vnn, uint64_t dst_srvid,
1829                                    const uint8_t *buf, size_t buflen)
1830 {
1831         return NT_STATUS_NOT_IMPLEMENTED;
1832 }
1833
1834 #endif