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 NTSTATUS register_with_ctdbd(struct ctdbd_connection *conn, uint64_t srvid)
 111 {
 112
 113         int cstatus;
 114         return ctdbd_control(conn, CTDB_CURRENT_NODE,
 115                              CTDB_CONTROL_REGISTER_SRVID, srvid, 0,
 116                              tdb_null, NULL, NULL, &cstatus);
 117 }
 118
 119 /*
 120  * get our vnn from the cluster
 121  */
 122 static NTSTATUS get_cluster_vnn(struct ctdbd_connection *conn, uint32 *vnn)
 123 {
 124         int32_t cstatus=-1;
 125         NTSTATUS status;
 126         status = ctdbd_control(conn,
 127                                CTDB_CURRENT_NODE, CTDB_CONTROL_GET_PNN, 0, 0,
 128                                tdb_null, NULL, NULL, &cstatus);
 129         if (!NT_STATUS_IS_OK(status)) {
 130                 cluster_fatal("ctdbd_control failed\n");
 131         }
 132         *vnn = (uint32_t)cstatus;
 133         return status;
 134 }
 135
 136 /*
 137  * Are we active (i.e. not banned or stopped?)
 138  */
 139 static bool ctdbd_working(struct ctdbd_connection *conn, uint32_t vnn)
 140 {
 141         int32_t cstatus=-1;
 142         NTSTATUS status;
 143         TDB_DATA outdata;
 144         struct ctdb_node_map *m;
 145         uint32_t failure_flags;
 146         bool ret = false;
 147         int i;
 148
 149         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
 150                                CTDB_CONTROL_GET_NODEMAP, 0, 0,
 151                                tdb_null, talloc_tos(), &outdata, &cstatus);
 152         if (!NT_STATUS_IS_OK(status)) {
 153                 cluster_fatal("ctdbd_control failed\n");
 154         }
 155         if ((cstatus != 0) || (outdata.dptr == NULL)) {
 156                 DEBUG(2, ("Received invalid ctdb data\n"));
 157                 return false;
 158         }
 159
 160         m = (struct ctdb_node_map *)outdata.dptr;
 161
 162         for (i=0; i<m->num; i++) {
 163                 if (vnn == m->nodes[i].pnn) {
 164                         break;
 165                 }
 166         }
 167
 168         if (i == m->num) {
 169                 DEBUG(2, ("Did not find ourselves (node %d) in nodemap\n",
 170                           (int)vnn));
 171                 goto fail;
 172         }
 173
 174         failure_flags = NODE_FLAGS_BANNED | NODE_FLAGS_DISCONNECTED
 175                 | NODE_FLAGS_PERMANENTLY_DISABLED | NODE_FLAGS_STOPPED;
 176
 177         if ((m->nodes[i].flags & failure_flags) != 0) {
 178                 DEBUG(2, ("Node has status %x, not active\n",
 179                           (int)m->nodes[i].flags));
 180                 goto fail;
 181         }
 182
 183         ret = true;
 184 fail:
 185         TALLOC_FREE(outdata.dptr);
 186         return ret;
 187 }
 188
 189 uint32 ctdbd_vnn(const struct ctdbd_connection *conn)
 190 {
 191         return conn->our_vnn;
 192 }
 193
 194 /*
 195  * Get us a ctdb connection
 196  */
 197
 198 static NTSTATUS ctdbd_connect(TALLOC_CTX *mem_ctx,
 199                               struct ctdb_packet_context **presult)
 200 {
 201         struct ctdb_packet_context *result;
 202         const char *sockname = lp_ctdbd_socket();
 203         struct sockaddr_un addr;
 204         int fd;
 205
 206         if (!sockname || !*sockname) {
 207                 sockname = CTDB_PATH;
 208         }
 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         ZERO_STRUCT(addr);
 217         addr.sun_family = AF_UNIX;
 218         strncpy(addr.sun_path, sockname, sizeof(addr.sun_path));
 219
 220         if (sys_connect(fd, (struct sockaddr *)(void *)&addr) == -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 msglen;
 245
 246         if (available < sizeof(msglen)) {
 247                 return False;
 248         }
 249
 250         msglen = *((uint32 *)buf);
 251
 252         DEBUG(10, ("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 event_context *event_ctx,
 284                                       struct timed_event *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 >= 10) {
 353                 DEBUG(10, ("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 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(10, ("Received ctdb packet\n"));
 415         ctdb_packet_dump(hdr);
 416
 417         if (hdr->operation == CTDB_REQ_MESSAGE) {
 418                 struct timed_event *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 = event_add_timed(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)sys_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 != sys_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 event_context *event_ctx,
 690                                  struct fd_event *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 = event_add_fd(msg_ctx->event_ctx, conn,
 726                                        ctdb_packet_get_fd(conn->pkt),
 727                                        EVENT_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 dst_vnn, uint64 dst_srvid,
 745                               struct messaging_rec *msg)
 746 {
 747         struct ctdb_req_message r;
 748         TALLOC_CTX *mem_ctx;
 749         DATA_BLOB blob;
 750         NTSTATUS status;
 751         enum ndr_err_code ndr_err;
 752
 753         if (!(mem_ctx = talloc_init("ctdbd_messaging_send"))) {
 754                 DEBUG(0, ("talloc failed\n"));
 755                 return NT_STATUS_NO_MEMORY;
 756         }
 757
 758         ndr_err = ndr_push_struct_blob(
 759                 &blob, mem_ctx, msg,
 760                 (ndr_push_flags_fn_t)ndr_push_messaging_rec);
 761
 762         if (!NDR_ERR_CODE_IS_SUCCESS(ndr_err)) {
 763                 DEBUG(0, ("ndr_push_struct_blob failed: %s\n",
 764                           ndr_errstr(ndr_err)));
 765                 status = ndr_map_error2ntstatus(ndr_err);
 766                 goto fail;
 767         }
 768
 769         r.hdr.length = offsetof(struct ctdb_req_message, data) + blob.length;
 770         r.hdr.ctdb_magic = CTDB_MAGIC;
 771         r.hdr.ctdb_version = CTDB_VERSION;
 772         r.hdr.generation = 1;
 773         r.hdr.operation  = CTDB_REQ_MESSAGE;
 774         r.hdr.destnode   = dst_vnn;
 775         r.hdr.srcnode    = conn->our_vnn;
 776         r.hdr.reqid      = 0;
 777         r.srvid          = dst_srvid;
 778         r.datalen        = blob.length;
 779
 780         DEBUG(10, ("ctdbd_messaging_send: Sending ctdb packet\n"));
 781         ctdb_packet_dump(&r.hdr);
 782
 783         status = ctdb_packet_send(
 784                 conn->pkt, 2,
 785                 data_blob_const(&r, offsetof(struct ctdb_req_message, data)),
 786                 blob);
 787
 788         if (!NT_STATUS_IS_OK(status)) {
 789                 DEBUG(0, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
 790                 goto fail;
 791         }
 792
 793         status = ctdb_packet_flush(conn->pkt);
 794
 795         if (!NT_STATUS_IS_OK(status)) {
 796                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
 797                 cluster_fatal("cluster dispatch daemon msg write error\n");
 798         }
 799
 800         status = NT_STATUS_OK;
 801  fail:
 802         TALLOC_FREE(mem_ctx);
 803         return status;
 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 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 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                 goto fail;
1060         }
1061
1062         for (i=0; i<num_pids; i++) {
1063                 unsigned j;
1064                 uint32_t vnn = pids[i].vnn;
1065
1066                 for (j=0; j<num_vnns; j++) {
1067                         if (vnn == vnns[j].vnn) {
1068                                 break;
1069                         }
1070                 }
1071                 vnn_indexes[i] = j;
1072
1073                 if (j < num_vnns) {
1074                         /*
1075                          * Already in the array
1076                          */
1077                         vnns[j].num_srvids += 1;
1078                         continue;
1079                 }
1080                 vnns = talloc_realloc(mem_ctx, vnns, struct ctdb_vnn_list,
1081                                       num_vnns+1);
1082                 if (vnns == NULL) {
1083                         goto fail;
1084                 }
1085                 vnns[num_vnns].vnn = vnn;
1086                 vnns[num_vnns].num_srvids = 1;
1087                 vnns[num_vnns].num_filled = 0;
1088                 num_vnns += 1;
1089         }
1090         for (i=0; i<num_vnns; i++) {
1091                 struct ctdb_vnn_list *vnn = &vnns[i];
1092
1093                 vnn->srvids = talloc_array(vnns, uint64_t, vnn->num_srvids);
1094                 if (vnn->srvids == NULL) {
1095                         goto fail;
1096                 }
1097                 vnn->pid_indexes = talloc_array(vnns, unsigned,
1098                                                 vnn->num_srvids);
1099                 if (vnn->pid_indexes == NULL) {
1100                         goto fail;
1101                 }
1102         }
1103         for (i=0; i<num_pids; i++) {
1104                 struct ctdb_vnn_list *vnn = &vnns[vnn_indexes[i]];
1105                 vnn->srvids[vnn->num_filled] = pids[i].unique_id;
1106                 vnn->pid_indexes[vnn->num_filled] = i;
1107                 vnn->num_filled += 1;
1108         }
1109
1110         TALLOC_FREE(vnn_indexes);
1111         *pvnns = vnns;
1112         *pnum_vnns = num_vnns;
1113         return true;
1114 fail:
1115         TALLOC_FREE(vnns);
1116         TALLOC_FREE(vnn_indexes);
1117         return false;
1118 }
1119
1120 bool ctdb_serverids_exist(struct ctdbd_connection *conn,
1121                           const struct server_id *pids, unsigned num_pids,
1122                           bool *results)
1123 {
1124         unsigned i, num_received;
1125         NTSTATUS status;
1126         struct ctdb_vnn_list *vnns = NULL;
1127         unsigned num_vnns;
1128         bool result = false;
1129
1130         if (!ctdb_collect_vnns(talloc_tos(), pids, num_pids,
1131                                &vnns, &num_vnns)) {
1132                 goto fail;
1133         }
1134
1135         for (i=0; i<num_vnns; i++) {
1136                 struct ctdb_vnn_list *vnn = &vnns[i];
1137                 struct ctdb_req_control req;
1138
1139                 vnn->reqid = ctdbd_next_reqid(conn);
1140
1141                 ZERO_STRUCT(req);
1142
1143                 DEBUG(10, ("Requesting VNN %d, reqid=%d, num_srvids=%u\n",
1144                            (int)vnn->vnn, (int)vnn->reqid, vnn->num_srvids));
1145
1146                 req.hdr.length = offsetof(struct ctdb_req_control, data);
1147                 req.hdr.ctdb_magic   = CTDB_MAGIC;
1148                 req.hdr.ctdb_version = CTDB_VERSION;
1149                 req.hdr.operation    = CTDB_REQ_CONTROL;
1150                 req.hdr.reqid        = vnn->reqid;
1151                 req.hdr.destnode     = vnn->vnn;
1152                 req.opcode           = CTDB_CONTROL_CHECK_SRVIDS;
1153                 req.srvid            = 0;
1154                 req.datalen          = sizeof(uint64_t) * vnn->num_srvids;
1155                 req.hdr.length      += req.datalen;
1156                 req.flags            = 0;
1157
1158                 DEBUG(10, ("ctdbd_control: Sending ctdb packet\n"));
1159                 ctdb_packet_dump(&req.hdr);
1160
1161                 status = ctdb_packet_send(
1162                         conn->pkt, 2,
1163                         data_blob_const(
1164                                 &req, offsetof(struct ctdb_req_control,
1165                                                data)),
1166                         data_blob_const(vnn->srvids, req.datalen));
1167                 if (!NT_STATUS_IS_OK(status)) {
1168                         DEBUG(10, ("ctdb_packet_send failed: %s\n",
1169                                    nt_errstr(status)));
1170                         goto fail;
1171                 }
1172         }
1173
1174         status = ctdb_packet_flush(conn->pkt);
1175         if (!NT_STATUS_IS_OK(status)) {
1176                 DEBUG(10, ("ctdb_packet_flush failed: %s\n",
1177                            nt_errstr(status)));
1178                 goto fail;
1179         }
1180
1181         num_received = 0;
1182
1183         while (num_received < num_vnns) {
1184                 struct ctdb_reply_control *reply = NULL;
1185                 struct ctdb_vnn_list *vnn;
1186                 uint32_t reqid;
1187
1188                 status = ctdb_read_req(conn, 0, talloc_tos(), (void *)&reply);
1189                 if (!NT_STATUS_IS_OK(status)) {
1190                         DEBUG(10, ("ctdb_read_req failed: %s\n",
1191                                    nt_errstr(status)));
1192                         goto fail;
1193                 }
1194
1195                 if (reply->hdr.operation != CTDB_REPLY_CONTROL) {
1196                         DEBUG(10, ("Received invalid reply\n"));
1197                         goto fail;
1198                 }
1199
1200                 reqid = reply->hdr.reqid;
1201
1202                 DEBUG(10, ("Received reqid %d\n", (int)reqid));
1203
1204                 for (i=0; i<num_vnns; i++) {
1205                         if (reqid == vnns[i].reqid) {
1206                                 break;
1207                         }
1208                 }
1209                 if (i == num_vnns) {
1210                         DEBUG(10, ("Received unknown reqid number %u\n",
1211                                    (unsigned)reqid));
1212                         goto fail;
1213                 }
1214
1215                 DEBUG(10, ("Found index %u\n", i));
1216
1217                 vnn = &vnns[i];
1218
1219                 DEBUG(10, ("Received vnn %u, vnn->num_srvids %u, datalen %u\n",
1220                            (unsigned)vnn->vnn, vnn->num_srvids,
1221                            (unsigned)reply->datalen));
1222
1223                 if (reply->datalen < ((vnn->num_srvids+7)/8)) {
1224                         DEBUG(10, ("Received short reply\n"));
1225                         goto fail;
1226                 }
1227
1228                 for (i=0; i<vnn->num_srvids; i++) {
1229                         results[vnn->pid_indexes[i]] =
1230                                 ((reply->data[i/8] & (1<<(i%8))) != 0);
1231                 }
1232
1233                 TALLOC_FREE(reply);
1234                 num_received += 1;
1235         }
1236
1237         result = true;
1238 fail:
1239         TALLOC_FREE(vnns);
1240         return result;
1241 }
1242
1243 /*
1244  * Get a db path
1245  */
1246 char *ctdbd_dbpath(struct ctdbd_connection *conn,
1247                    TALLOC_CTX *mem_ctx, uint32_t db_id)
1248 {
1249         NTSTATUS status;
1250         TDB_DATA data;
1251         int32_t cstatus;
1252
1253         data.dptr = (uint8_t*)&db_id;
1254         data.dsize = sizeof(db_id);
1255
1256         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1257                                CTDB_CONTROL_GETDBPATH, 0, 0, data,
1258                                mem_ctx, &data, &cstatus);
1259         if (!NT_STATUS_IS_OK(status) || cstatus != 0) {
1260                 DEBUG(0,(__location__ " ctdb_control for getdbpath failed\n"));
1261                 return NULL;
1262         }
1263
1264         return (char *)data.dptr;
1265 }
1266
1267 /*
1268  * attach to a ctdb database
1269  */
1270 NTSTATUS ctdbd_db_attach(struct ctdbd_connection *conn,
1271                          const char *name, uint32_t *db_id, int tdb_flags)
1272 {
1273         NTSTATUS status;
1274         TDB_DATA data;
1275         int32_t cstatus;
1276         bool persistent = (tdb_flags & TDB_CLEAR_IF_FIRST) == 0;
1277
1278         data.dptr = (uint8_t*)name;
1279         data.dsize = strlen(name)+1;
1280
1281         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1282                                persistent
1283                                ? CTDB_CONTROL_DB_ATTACH_PERSISTENT
1284                                : CTDB_CONTROL_DB_ATTACH,
1285                                tdb_flags, 0, data, NULL, &data, &cstatus);
1286         if (!NT_STATUS_IS_OK(status)) {
1287                 DEBUG(0, (__location__ " ctdb_control for db_attach "
1288                           "failed: %s\n", nt_errstr(status)));
1289                 return status;
1290         }
1291
1292         if (cstatus != 0 || data.dsize != sizeof(uint32_t)) {
1293                 DEBUG(0,(__location__ " ctdb_control for db_attach failed\n"));
1294                 return NT_STATUS_INTERNAL_ERROR;
1295         }
1296
1297         *db_id = *(uint32_t *)data.dptr;
1298         talloc_free(data.dptr);
1299
1300         if (!(tdb_flags & TDB_SEQNUM)) {
1301                 return NT_STATUS_OK;
1302         }
1303
1304         data.dptr = (uint8_t *)db_id;
1305         data.dsize = sizeof(*db_id);
1306
1307         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1308                                CTDB_CONTROL_ENABLE_SEQNUM, 0, 0, data,
1309                                NULL, NULL, &cstatus);
1310         if (!NT_STATUS_IS_OK(status) || cstatus != 0) {
1311                 DEBUG(0,(__location__ " ctdb_control for enable seqnum "
1312                          "failed\n"));
1313                 return NT_STATUS_IS_OK(status) ? NT_STATUS_INTERNAL_ERROR :
1314                         status;
1315         }
1316
1317         return NT_STATUS_OK;
1318 }
1319
1320 /*
1321  * force the migration of a record to this node
1322  */
1323 NTSTATUS ctdbd_migrate(struct ctdbd_connection *conn, uint32 db_id,
1324                        TDB_DATA key)
1325 {
1326         struct ctdb_req_call req;
1327         struct ctdb_reply_call *reply;
1328         NTSTATUS status;
1329
1330         ZERO_STRUCT(req);
1331
1332         req.hdr.length = offsetof(struct ctdb_req_call, data) + key.dsize;
1333         req.hdr.ctdb_magic   = CTDB_MAGIC;
1334         req.hdr.ctdb_version = CTDB_VERSION;
1335         req.hdr.operation    = CTDB_REQ_CALL;
1336         req.hdr.reqid        = ctdbd_next_reqid(conn);
1337         req.flags            = CTDB_IMMEDIATE_MIGRATION;
1338         req.callid           = CTDB_NULL_FUNC;
1339         req.db_id            = db_id;
1340         req.keylen           = key.dsize;
1341
1342         DEBUG(10, ("ctdbd_migrate: Sending ctdb packet\n"));
1343         ctdb_packet_dump(&req.hdr);
1344
1345         status = ctdb_packet_send(
1346                 conn->pkt, 2,
1347                 data_blob_const(&req, offsetof(struct ctdb_req_call, data)),
1348                 data_blob_const(key.dptr, key.dsize));
1349
1350         if (!NT_STATUS_IS_OK(status)) {
1351                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
1352                 return status;
1353         }
1354
1355         status = ctdb_packet_flush(conn->pkt);
1356
1357         if (!NT_STATUS_IS_OK(status)) {
1358                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
1359                 cluster_fatal("cluster dispatch daemon control write error\n");
1360         }
1361
1362         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
1363
1364         if (!NT_STATUS_IS_OK(status)) {
1365                 DEBUG(0, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
1366                 goto fail;
1367         }
1368
1369         if (reply->hdr.operation != CTDB_REPLY_CALL) {
1370                 DEBUG(0, ("received invalid reply\n"));
1371                 status = NT_STATUS_INTERNAL_ERROR;
1372                 goto fail;
1373         }
1374
1375         status = NT_STATUS_OK;
1376  fail:
1377
1378         TALLOC_FREE(reply);
1379         return status;
1380 }
1381
1382 /*
1383  * remotely fetch a record without locking it or forcing a migration
1384  */
1385 NTSTATUS ctdbd_fetch(struct ctdbd_connection *conn, uint32 db_id,
1386                      TDB_DATA key, TALLOC_CTX *mem_ctx, TDB_DATA *data)
1387 {
1388         struct ctdb_req_call req;
1389         struct ctdb_reply_call *reply;
1390         NTSTATUS status;
1391
1392         ZERO_STRUCT(req);
1393
1394         req.hdr.length = offsetof(struct ctdb_req_call, data) + key.dsize;
1395         req.hdr.ctdb_magic   = CTDB_MAGIC;
1396         req.hdr.ctdb_version = CTDB_VERSION;
1397         req.hdr.operation    = CTDB_REQ_CALL;
1398         req.hdr.reqid        = ctdbd_next_reqid(conn);
1399         req.flags            = 0;
1400         req.callid           = CTDB_FETCH_FUNC;
1401         req.db_id            = db_id;
1402         req.keylen           = key.dsize;
1403
1404         status = ctdb_packet_send(
1405                 conn->pkt, 2,
1406                 data_blob_const(&req, offsetof(struct ctdb_req_call, data)),
1407                 data_blob_const(key.dptr, key.dsize));
1408
1409         if (!NT_STATUS_IS_OK(status)) {
1410                 DEBUG(3, ("ctdb_packet_send failed: %s\n", nt_errstr(status)));
1411                 return status;
1412         }
1413
1414         status = ctdb_packet_flush(conn->pkt);
1415
1416         if (!NT_STATUS_IS_OK(status)) {
1417                 DEBUG(3, ("write to ctdbd failed: %s\n", nt_errstr(status)));
1418                 cluster_fatal("cluster dispatch daemon control write error\n");
1419         }
1420
1421         status = ctdb_read_req(conn, req.hdr.reqid, NULL, (void *)&reply);
1422
1423         if (!NT_STATUS_IS_OK(status)) {
1424                 DEBUG(0, ("ctdb_read_req failed: %s\n", nt_errstr(status)));
1425                 goto fail;
1426         }
1427
1428         if (reply->hdr.operation != CTDB_REPLY_CALL) {
1429                 DEBUG(0, ("received invalid reply\n"));
1430                 status = NT_STATUS_INTERNAL_ERROR;
1431                 goto fail;
1432         }
1433
1434         data->dsize = reply->datalen;
1435         if (data->dsize == 0) {
1436                 data->dptr = NULL;
1437                 goto done;
1438         }
1439
1440         data->dptr = (uint8 *)talloc_memdup(mem_ctx, &reply->data[0],
1441                                             reply->datalen);
1442         if (data->dptr == NULL) {
1443                 DEBUG(0, ("talloc failed\n"));
1444                 status = NT_STATUS_NO_MEMORY;
1445                 goto fail;
1446         }
1447
1448  done:
1449         status = NT_STATUS_OK;
1450  fail:
1451         TALLOC_FREE(reply);
1452         return status;
1453 }
1454
1455 struct ctdbd_traverse_state {
1456         void (*fn)(TDB_DATA key, TDB_DATA data, void *private_data);
1457         void *private_data;
1458 };
1459
1460 /*
1461  * Handle a traverse record coming in on the ctdbd connection
1462  */
1463
1464 static NTSTATUS ctdb_traverse_handler(uint8_t *buf, size_t length,
1465                                       void *private_data)
1466 {
1467         struct ctdbd_traverse_state *state =
1468                 (struct ctdbd_traverse_state *)private_data;
1469
1470         struct ctdb_req_message *m;
1471         struct ctdb_rec_data *d;
1472         TDB_DATA key, data;
1473
1474         m = (struct ctdb_req_message *)buf;
1475
1476         if (length < sizeof(*m) || m->hdr.length != length) {
1477                 DEBUG(0, ("Got invalid message of length %d\n", (int)length));
1478                 TALLOC_FREE(buf);
1479                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1480         }
1481
1482         d = (struct ctdb_rec_data *)&m->data[0];
1483         if (m->datalen < sizeof(uint32_t) || m->datalen != d->length) {
1484                 DEBUG(0, ("Got invalid traverse data of length %d\n",
1485                           (int)m->datalen));
1486                 TALLOC_FREE(buf);
1487                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1488         }
1489
1490         key.dsize = d->keylen;
1491         key.dptr  = &d->data[0];
1492         data.dsize = d->datalen;
1493         data.dptr = &d->data[d->keylen];
1494
1495         if (key.dsize == 0 && data.dsize == 0) {
1496                 /* end of traverse */
1497                 return NT_STATUS_END_OF_FILE;
1498         }
1499
1500         if (data.dsize < sizeof(struct ctdb_ltdb_header)) {
1501                 DEBUG(0, ("Got invalid ltdb header length %d\n",
1502                           (int)data.dsize));
1503                 TALLOC_FREE(buf);
1504                 return NT_STATUS_UNEXPECTED_IO_ERROR;
1505         }
1506         data.dsize -= sizeof(struct ctdb_ltdb_header);
1507         data.dptr += sizeof(struct ctdb_ltdb_header);
1508
1509         if (state->fn) {
1510                 state->fn(key, data, state->private_data);
1511         }
1512
1513         TALLOC_FREE(buf);
1514         return NT_STATUS_OK;
1515 }
1516
1517 /*
1518   Traverse a ctdb database. This uses a kind-of hackish way to open a second
1519   connection to ctdbd to avoid the hairy recursive and async problems with
1520   everything in-line.
1521 */
1522
1523 NTSTATUS ctdbd_traverse(uint32 db_id,
1524                         void (*fn)(TDB_DATA key, TDB_DATA data,
1525                                    void *private_data),
1526                         void *private_data)
1527 {
1528         struct ctdbd_connection *conn;
1529         NTSTATUS status;
1530
1531         TDB_DATA data;
1532         struct ctdb_traverse_start t;
1533         int cstatus;
1534         struct ctdbd_traverse_state state;
1535
1536         status = ctdbd_init_connection(NULL, &conn);
1537         if (!NT_STATUS_IS_OK(status)) {
1538                 DEBUG(0, ("ctdbd_init_connection failed: %s\n",
1539                           nt_errstr(status)));
1540                 return status;
1541         }
1542
1543         t.db_id = db_id;
1544         t.srvid = conn->rand_srvid;
1545         t.reqid = ctdbd_next_reqid(conn);
1546
1547         data.dptr = (uint8_t *)&t;
1548         data.dsize = sizeof(t);
1549
1550         status = ctdbd_control(conn, CTDB_CURRENT_NODE,
1551                                CTDB_CONTROL_TRAVERSE_START, conn->rand_srvid, 0,
1552                                data, NULL, NULL, &cstatus);
1553
1554         if (!NT_STATUS_IS_OK(status) || (cstatus != 0)) {
1555
1556                 DEBUG(0,("ctdbd_control failed: %s, %d\n", nt_errstr(status),
1557                          cstatus));
1558
1559                 if (NT_STATUS_IS_OK(status)) {
1560                         /*
1561                          * We need a mapping here
1562                          */
1563                         status = NT_STATUS_UNSUCCESSFUL;
1564                 }
1565                 goto done;
1566         }
1567
1568         state.fn = fn;
1569         state.private_data = private_data;
1570
1571         while (True) {
1572
1573                 status = NT_STATUS_OK;
1574
1575                 if (ctdb_packet_handler(conn->pkt, ctdb_req_complete,
1576                                    ctdb_traverse_handler, &state, &status)) {
1577
1578                         if (NT_STATUS_EQUAL(status, NT_STATUS_END_OF_FILE)) {
1579                                 status = NT_STATUS_OK;
1580                                 break;
1581                         }
1582
1583                         /*
1584                          * There might be more in the queue
1585                          */
1586                         continue;
1587                 }
1588
1589                 if (!NT_STATUS_IS_OK(status)) {
1590                         break;
1591                 }
1592
1593                 status = ctdb_packet_fd_read_sync(conn->pkt);
1594
1595                 if (NT_STATUS_EQUAL(status, NT_STATUS_RETRY)) {
1596                         /*
1597                          * There might be more in the queue
1598                          */
1599                         continue;
1600                 }
1601
1602                 if (NT_STATUS_EQUAL(status, NT_STATUS_END_OF_FILE)) {
1603                         status = NT_STATUS_OK;
1604                         break;
1605                 }
1606
1607                 if (!NT_STATUS_IS_OK(status)) {
1608                         DEBUG(0, ("ctdb_packet_fd_read_sync failed: %s\n", nt_errstr(status)));
1609                         cluster_fatal("ctdbd died\n");
1610                 }
1611         }
1612
1613  done:
1614         TALLOC_FREE(conn);
1615         return status;
1616 }
1617
1618 /*
1619    This is used to canonicalize a ctdb_sock_addr structure.
1620 */
1621 static void smbd_ctdb_canonicalize_ip(const struct sockaddr_storage *in,
1622                                       struct sockaddr_storage *out)
1623 {
1624         memcpy(out, in, sizeof (*out));
1625
1626 #ifdef HAVE_IPV6
1627         if (in->ss_family == AF_INET6) {
1628                 const char prefix[12] = { 0,0,0,0,0,0,0,0,0,0,0xff,0xff };
1629                 const struct sockaddr_in6 *in6 = (struct sockaddr_in6 *)in;
1630                 struct sockaddr_in *out4 = (struct sockaddr_in *)out;
1631                 if (memcmp(&in6->sin6_addr, prefix, 12) == 0) {
1632                         memset(out, 0, sizeof(*out));
1633 #ifdef HAVE_SOCK_SIN_LEN
1634                         out4->sin_len = sizeof(*out);
1635 #endif
1636                         out4->sin_family = AF_INET;
1637                         out4->sin_port   = in6->sin6_port;
1638                         memcpy(&out4->sin_addr, &in6->sin6_addr.s6_addr32[3], 4);
1639                 }
1640         }
1641 #endif
1642 }
1643
1644 /*
1645  * Register us as a server for a particular tcp connection
1646  */
1647
1648 NTSTATUS ctdbd_register_ips(struct ctdbd_connection *conn,
1649                             const struct sockaddr_storage *_server,
1650                             const struct sockaddr_storage *_client,
1651                             void (*release_ip_handler)(const char *ip_addr,
1652                                                        void *private_data),
1653                             void *private_data)
1654 {
1655         /*
1656          * we still use ctdb_control_tcp for ipv4
1657          * because we want to work against older ctdb
1658          * versions at runtime
1659          */
1660         struct ctdb_control_tcp p4;
1661 #ifdef HAVE_STRUCT_CTDB_CONTROL_TCP_ADDR
1662         struct ctdb_control_tcp_addr p;
1663 #endif
1664         TDB_DATA data;
1665         NTSTATUS status;
1666         struct sockaddr_storage client;
1667         struct sockaddr_storage server;
1668
1669         /*
1670          * Only one connection so far
1671          */
1672         SMB_ASSERT(conn->release_ip_handler == NULL);
1673
1674         smbd_ctdb_canonicalize_ip(_client, &client);
1675         smbd_ctdb_canonicalize_ip(_server, &server);
1676
1677         switch (client.ss_family) {
1678         case AF_INET:
1679                 memcpy(&p4.dest, &server, sizeof(p4.dest));
1680                 memcpy(&p4.src, &client, sizeof(p4.src));
1681                 data.dptr = (uint8_t *)&p4;
1682                 data.dsize = sizeof(p4);
1683                 break;
1684 #ifdef HAVE_STRUCT_CTDB_CONTROL_TCP_ADDR
1685         case AF_INET6:
1686                 memcpy(&p.dest.ip6, &server, sizeof(p.dest.ip6));
1687                 memcpy(&p.src.ip6, &client, sizeof(p.src.ip6));
1688                 data.dptr = (uint8_t *)&p;
1689                 data.dsize = sizeof(p);
1690                 break;
1691 #endif
1692         default:
1693                 return NT_STATUS_INTERNAL_ERROR;
1694         }
1695
1696         conn->release_ip_handler = release_ip_handler;
1697         conn->release_ip_priv = private_data;
1698
1699         /*
1700          * We want to be told about IP releases
1701          */
1702
1703         status = register_with_ctdbd(conn, CTDB_SRVID_RELEASE_IP);
1704         if (!NT_STATUS_IS_OK(status)) {
1705                 return status;
1706         }
1707
1708         /*
1709          * inform ctdb of our tcp connection, so if IP takeover happens ctdb
1710          * can send an extra ack to trigger a reset for our client, so it
1711          * immediately reconnects
1712          */
1713         return ctdbd_control(conn, CTDB_CURRENT_NODE,
1714                              CTDB_CONTROL_TCP_CLIENT, 0,
1715                              CTDB_CTRL_FLAG_NOREPLY, data, NULL, NULL, NULL);
1716 }
1717
1718 /*
1719  * We want to handle reconfigure events
1720  */
1721 NTSTATUS ctdbd_register_reconfigure(struct ctdbd_connection *conn)
1722 {
1723         return register_with_ctdbd(conn, CTDB_SRVID_RECONFIGURE);
1724 }
1725
1726 /*
1727   call a control on the local node
1728  */
1729 NTSTATUS ctdbd_control_local(struct ctdbd_connection *conn, uint32 opcode,
1730                              uint64_t srvid, uint32_t flags, TDB_DATA data,
1731                              TALLOC_CTX *mem_ctx, TDB_DATA *outdata,
1732                              int *cstatus)
1733 {
1734         return ctdbd_control(conn, CTDB_CURRENT_NODE, opcode, srvid, flags, data, mem_ctx, outdata, cstatus);
1735 }
1736
1737 NTSTATUS ctdb_watch_us(struct ctdbd_connection *conn)
1738 {
1739         struct ctdb_client_notify_register reg_data;
1740         size_t struct_len;
1741         NTSTATUS status;
1742         int cstatus;
1743
1744         reg_data.srvid = CTDB_SRVID_SAMBA_NOTIFY;
1745         reg_data.len = 1;
1746         reg_data.notify_data[0] = 0;
1747
1748         struct_len = offsetof(struct ctdb_client_notify_register,
1749                               notify_data) + reg_data.len;
1750
1751         status = ctdbd_control_local(
1752                 conn, CTDB_CONTROL_REGISTER_NOTIFY, conn->rand_srvid, 0,
1753                 make_tdb_data((uint8_t *)&reg_data, struct_len),
1754                 NULL, NULL, &cstatus);
1755         if (!NT_STATUS_IS_OK(status)) {
1756                 DEBUG(1, ("ctdbd_control_local failed: %s\n",
1757                           nt_errstr(status)));
1758         }
1759         return status;
1760 }
1761
1762 NTSTATUS ctdb_unwatch(struct ctdbd_connection *conn)
1763 {
1764         struct ctdb_client_notify_deregister dereg_data;
1765         NTSTATUS status;
1766         int cstatus;
1767
1768         dereg_data.srvid = CTDB_SRVID_SAMBA_NOTIFY;
1769
1770         status = ctdbd_control_local(
1771                 conn, CTDB_CONTROL_DEREGISTER_NOTIFY, conn->rand_srvid, 0,
1772                 make_tdb_data((uint8_t *)&dereg_data, sizeof(dereg_data)),
1773                 NULL, NULL, &cstatus);
1774         if (!NT_STATUS_IS_OK(status)) {
1775                 DEBUG(1, ("ctdbd_control_local failed: %s\n",
1776                           nt_errstr(status)));
1777         }
1778         return status;
1779 }
1780
1781 #endif