server/ctdb_call.c

   1 /*
   2    ctdb_call protocol code
   3
   4    Copyright (C) Andrew Tridgell  2006
   5
   6    This program is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3 of the License, or
   9    (at your option) any later version.
  10
  11    This program is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    You should have received a copy of the GNU General Public License
  17    along with this program; if not, see <http://www.gnu.org/licenses/>.
  18 */
  19 /*
  20   see http://wiki.samba.org/index.php/Samba_%26_Clustering for
  21   protocol design and packet details
  22 */
  23 #include "includes.h"
  24 #include "lib/events/events.h"
  25 #include "lib/tdb/include/tdb.h"
  26 #include "lib/util/dlinklist.h"
  27 #include "system/network.h"
  28 #include "system/filesys.h"
  29 #include "../include/ctdb_private.h"
  30
  31 /*
  32   find the ctdb_db from a db index
  33  */
  34  struct ctdb_db_context *find_ctdb_db(struct ctdb_context *ctdb, uint32_t id)
  35 {
  36         struct ctdb_db_context *ctdb_db;
  37
  38         for (ctdb_db=ctdb->db_list; ctdb_db; ctdb_db=ctdb_db->next) {
  39                 if (ctdb_db->db_id == id) {
  40                         break;
  41                 }
  42         }
  43         return ctdb_db;
  44 }
  45
  46
  47 /*
  48   a varient of input packet that can be used in lock requeue
  49 */
  50 static void ctdb_call_input_pkt(void *p, struct ctdb_req_header *hdr)
  51 {
  52         struct ctdb_context *ctdb = talloc_get_type(p, struct ctdb_context);
  53         ctdb_input_pkt(ctdb, hdr);
  54 }
  55
  56
  57 /*
  58   send an error reply
  59 */
  60 static void ctdb_send_error(struct ctdb_context *ctdb,
  61                             struct ctdb_req_header *hdr, uint32_t status,
  62                             const char *fmt, ...) PRINTF_ATTRIBUTE(4,5);
  63 static void ctdb_send_error(struct ctdb_context *ctdb,
  64                             struct ctdb_req_header *hdr, uint32_t status,
  65                             const char *fmt, ...)
  66 {
  67         va_list ap;
  68         struct ctdb_reply_error *r;
  69         char *msg;
  70         int msglen, len;
  71
  72         if (ctdb->methods == NULL) {
  73                 DEBUG(DEBUG_ERR,(__location__ " Failed to send error. Transport is DOWN\n"));
  74                 return;
  75         }
  76
  77         va_start(ap, fmt);
  78         msg = talloc_vasprintf(ctdb, fmt, ap);
  79         if (msg == NULL) {
  80                 ctdb_fatal(ctdb, "Unable to allocate error in ctdb_send_error\n");
  81         }
  82         va_end(ap);
  83
  84         msglen = strlen(msg)+1;
  85         len = offsetof(struct ctdb_reply_error, msg);
  86         r = ctdb_transport_allocate(ctdb, msg, CTDB_REPLY_ERROR, len + msglen,
  87                                     struct ctdb_reply_error);
  88         CTDB_NO_MEMORY_FATAL(ctdb, r);
  89
  90         r->hdr.destnode  = hdr->srcnode;
  91         r->hdr.reqid     = hdr->reqid;
  92         r->status        = status;
  93         r->msglen        = msglen;
  94         memcpy(&r->msg[0], msg, msglen);
  95
  96         ctdb_queue_packet(ctdb, &r->hdr);
  97
  98         talloc_free(msg);
  99 }
 100
 101
 102 /*
 103   send a redirect reply
 104 */
 105 static void ctdb_call_send_redirect(struct ctdb_context *ctdb,
 106                                     TDB_DATA key,
 107                                     struct ctdb_req_call *c,
 108                                     struct ctdb_ltdb_header *header)
 109 {
 110
 111         uint32_t lmaster = ctdb_lmaster(ctdb, &key);
 112         if (ctdb->pnn == lmaster) {
 113                 c->hdr.destnode = header->dmaster;
 114         } else if ((c->hopcount % ctdb->tunable.max_redirect_count) == 0) {
 115                 c->hdr.destnode = lmaster;
 116         } else {
 117                 c->hdr.destnode = header->dmaster;
 118         }
 119         c->hopcount++;
 120         ctdb_queue_packet(ctdb, &c->hdr);
 121 }
 122
 123
 124 /*
 125   send a dmaster reply
 126
 127   caller must have the chainlock before calling this routine. Caller must be
 128   the lmaster
 129 */
 130 static void ctdb_send_dmaster_reply(struct ctdb_db_context *ctdb_db,
 131                                     struct ctdb_ltdb_header *header,
 132                                     TDB_DATA key, TDB_DATA data,
 133                                     uint32_t new_dmaster,
 134                                     uint32_t reqid)
 135 {
 136         struct ctdb_context *ctdb = ctdb_db->ctdb;
 137         struct ctdb_reply_dmaster *r;
 138         int ret, len;
 139         TALLOC_CTX *tmp_ctx;
 140
 141         if (ctdb->pnn != ctdb_lmaster(ctdb, &key)) {
 142                 DEBUG(DEBUG_ALERT,(__location__ " Caller is not lmaster!\n"));
 143                 return;
 144         }
 145
 146         header->dmaster = new_dmaster;
 147         ret = ctdb_ltdb_store(ctdb_db, key, header, data);
 148         if (ret != 0) {
 149                 ctdb_fatal(ctdb, "ctdb_send_dmaster_reply unable to update dmaster");
 150                 return;
 151         }
 152
 153         if (ctdb->methods == NULL) {
 154                 ctdb_fatal(ctdb, "ctdb_send_dmaster_reply cant update dmaster sicne transport is down");
 155                 return;
 156         }
 157
 158         /* put the packet on a temporary context, allowing us to safely free
 159            it below even if ctdb_reply_dmaster() has freed it already */
 160         tmp_ctx = talloc_new(ctdb);
 161
 162         /* send the CTDB_REPLY_DMASTER */
 163         len = offsetof(struct ctdb_reply_dmaster, data) + key.dsize + data.dsize;
 164         r = ctdb_transport_allocate(ctdb, tmp_ctx, CTDB_REPLY_DMASTER, len,
 165                                     struct ctdb_reply_dmaster);
 166         CTDB_NO_MEMORY_FATAL(ctdb, r);
 167
 168         r->hdr.destnode  = new_dmaster;
 169         r->hdr.reqid     = reqid;
 170         r->rsn           = header->rsn;
 171         r->keylen        = key.dsize;
 172         r->datalen       = data.dsize;
 173         r->db_id         = ctdb_db->db_id;
 174         memcpy(&r->data[0], key.dptr, key.dsize);
 175         memcpy(&r->data[key.dsize], data.dptr, data.dsize);
 176
 177         ctdb_queue_packet(ctdb, &r->hdr);
 178
 179         talloc_free(tmp_ctx);
 180 }
 181
 182 /*
 183   send a dmaster request (give another node the dmaster for a record)
 184
 185   This is always sent to the lmaster, which ensures that the lmaster
 186   always knows who the dmaster is. The lmaster will then send a
 187   CTDB_REPLY_DMASTER to the new dmaster
 188 */
 189 static void ctdb_call_send_dmaster(struct ctdb_db_context *ctdb_db,
 190                                    struct ctdb_req_call *c,
 191                                    struct ctdb_ltdb_header *header,
 192                                    TDB_DATA *key, TDB_DATA *data)
 193 {
 194         struct ctdb_req_dmaster *r;
 195         struct ctdb_context *ctdb = ctdb_db->ctdb;
 196         int len;
 197         uint32_t lmaster = ctdb_lmaster(ctdb, key);
 198
 199         if (ctdb->methods == NULL) {
 200                 ctdb_fatal(ctdb, "Failed ctdb_call_send_dmaster since transport is down");
 201                 return;
 202         }
 203
 204         if (lmaster == ctdb->pnn) {
 205                 ctdb_send_dmaster_reply(ctdb_db, header, *key, *data,
 206                                         c->hdr.srcnode, c->hdr.reqid);
 207                 return;
 208         }
 209
 210         len = offsetof(struct ctdb_req_dmaster, data) + key->dsize + data->dsize;
 211         r = ctdb_transport_allocate(ctdb, ctdb, CTDB_REQ_DMASTER, len,
 212                                     struct ctdb_req_dmaster);
 213         CTDB_NO_MEMORY_FATAL(ctdb, r);
 214         r->hdr.destnode  = lmaster;
 215         r->hdr.reqid     = c->hdr.reqid;
 216         r->db_id         = c->db_id;
 217         r->rsn           = header->rsn;
 218         r->dmaster       = c->hdr.srcnode;
 219         r->keylen        = key->dsize;
 220         r->datalen       = data->dsize;
 221         memcpy(&r->data[0], key->dptr, key->dsize);
 222         memcpy(&r->data[key->dsize], data->dptr, data->dsize);
 223
 224         header->dmaster = c->hdr.srcnode;
 225         if (ctdb_ltdb_store(ctdb_db, *key, header, *data) != 0) {
 226                 ctdb_fatal(ctdb, "Failed to store record in ctdb_call_send_dmaster");
 227         }
 228
 229         ctdb_queue_packet(ctdb, &r->hdr);
 230
 231         talloc_free(r);
 232 }
 233
 234 static void ctdb_hold_back_key(struct ctdb_db_context *db, TDB_DATA key)
 235 {
 236         size_t num_keys;
 237         uint8_t **tmp;
 238
 239         DEBUG(DEBUG_INFO, ("Holding back key %08x\n", ctdb_hash(&key)));
 240
 241         num_keys = talloc_array_length(db->holdback_keys);
 242         tmp = talloc_realloc(db, db->holdback_keys, uint8_t *, num_keys+1);
 243         if (tmp == NULL) {
 244                 DEBUG(DEBUG_ERR, ("talloc_realloc failed\n"));
 245                 return;
 246         }
 247         db->holdback_keys = tmp;
 248
 249         db->holdback_keys[num_keys] = (uint8_t *)talloc_memdup(
 250                 db->holdback_keys, key.dptr, key.dsize);
 251         if (db->holdback_keys[num_keys] == NULL) {
 252                 DEBUG(DEBUG_ERR, ("talloc_memdup failed\n"));
 253                 db->holdback_keys = talloc_realloc(db, db->holdback_keys,
 254                                                    uint8_t *, num_keys);
 255         }
 256 }
 257
 258 /*
 259   called when a CTDB_REPLY_DMASTER packet comes in, or when the lmaster
 260   gets a CTDB_REQUEST_DMASTER for itself. We become the dmaster.
 261
 262   must be called with the chainlock held. This function releases the chainlock
 263 */
 264 static void ctdb_become_dmaster(struct ctdb_db_context *ctdb_db,
 265                                 struct ctdb_req_header *hdr,
 266                                 TDB_DATA key, TDB_DATA data,
 267                                 uint64_t rsn)
 268 {
 269         struct ctdb_call_state *state;
 270         struct ctdb_context *ctdb = ctdb_db->ctdb;
 271         struct ctdb_ltdb_header header;
 272
 273         DEBUG(DEBUG_INFO,("pnn %u dmaster response %08x\n", ctdb->pnn, ctdb_hash(&key)));
 274
 275         ZERO_STRUCT(header);
 276         header.rsn = rsn + 1;
 277         header.dmaster = ctdb->pnn;
 278
 279         if (ctdb_ltdb_store(ctdb_db, key, &header, data) != 0) {
 280                 ctdb_fatal(ctdb, "ctdb_reply_dmaster store failed\n");
 281                 ctdb_ltdb_unlock(ctdb_db, key);
 282                 return;
 283         }
 284
 285         state = ctdb_reqid_find(ctdb, hdr->reqid, struct ctdb_call_state);
 286
 287         if (state == NULL) {
 288                 DEBUG(DEBUG_ERR,("pnn %u Invalid reqid %u in ctdb_become_dmaster from node %u\n",
 289                          ctdb->pnn, hdr->reqid, hdr->srcnode));
 290                 ctdb_ltdb_unlock(ctdb_db, key);
 291                 return;
 292         }
 293
 294         if (hdr->reqid != state->reqid) {
 295                 /* we found a record  but it was the wrong one */
 296                 DEBUG(DEBUG_ERR, ("Dropped orphan in ctdb_become_dmaster with reqid:%u\n from node %u", hdr->reqid, hdr->srcnode));
 297                 ctdb_ltdb_unlock(ctdb_db, key);
 298                 return;
 299         }
 300
 301         ctdb_hold_back_key(ctdb_db, key);
 302
 303         ctdb_call_local(ctdb_db, state->call, &header, state, &data, ctdb->pnn);
 304
 305         ctdb_ltdb_unlock(ctdb_db, state->call->key);
 306
 307         state->state = CTDB_CALL_DONE;
 308         if (state->async.fn) {
 309                 state->async.fn(state);
 310         }
 311 }
 312
 313
 314
 315 /*
 316   called when a CTDB_REQ_DMASTER packet comes in
 317
 318   this comes into the lmaster for a record when the current dmaster
 319   wants to give up the dmaster role and give it to someone else
 320 */
 321 void ctdb_request_dmaster(struct ctdb_context *ctdb, struct ctdb_req_header *hdr)
 322 {
 323         struct ctdb_req_dmaster *c = (struct ctdb_req_dmaster *)hdr;
 324         TDB_DATA key, data, data2;
 325         struct ctdb_ltdb_header header;
 326         struct ctdb_db_context *ctdb_db;
 327         int ret;
 328
 329         key.dptr = c->data;
 330         key.dsize = c->keylen;
 331         data.dptr = c->data + c->keylen;
 332         data.dsize = c->datalen;
 333
 334         ctdb_db = find_ctdb_db(ctdb, c->db_id);
 335         if (!ctdb_db) {
 336                 ctdb_send_error(ctdb, hdr, -1,
 337                                 "Unknown database in request. db_id==0x%08x",
 338                                 c->db_id);
 339                 return;
 340         }
 341
 342         /* fetch the current record */
 343         ret = ctdb_ltdb_lock_fetch_requeue(ctdb_db, key, &header, hdr, &data2,
 344                                            ctdb_call_input_pkt, ctdb, False);
 345         if (ret == -1) {
 346                 ctdb_fatal(ctdb, "ctdb_req_dmaster failed to fetch record");
 347                 return;
 348         }
 349         if (ret == -2) {
 350                 DEBUG(DEBUG_INFO,(__location__ " deferring ctdb_request_dmaster\n"));
 351                 return;
 352         }
 353
 354         if (ctdb_lmaster(ctdb, &key) != ctdb->pnn) {
 355                 DEBUG(DEBUG_ALERT,("pnn %u dmaster request to non-lmaster lmaster=%u gen=%u curgen=%u\n",
 356                          ctdb->pnn, ctdb_lmaster(ctdb, &key),
 357                          hdr->generation, ctdb->vnn_map->generation));
 358                 ctdb_fatal(ctdb, "ctdb_req_dmaster to non-lmaster");
 359         }
 360
 361         DEBUG(DEBUG_INFO,("pnn %u dmaster request on %08x for %u from %u\n",
 362                  ctdb->pnn, ctdb_hash(&key), c->dmaster, c->hdr.srcnode));
 363
 364         /* its a protocol error if the sending node is not the current dmaster */
 365         if (header.dmaster != hdr->srcnode) {
 366                 DEBUG(DEBUG_ALERT,("pnn %u dmaster request for new-dmaster %u from non-master %u real-dmaster=%u key %08x dbid 0x%08x gen=%u curgen=%u c->rsn=%llu header.rsn=%llu reqid=%u keyval=0x%08x\n",
 367                          ctdb->pnn, c->dmaster, hdr->srcnode, header.dmaster, ctdb_hash(&key),
 368                          ctdb_db->db_id, hdr->generation, ctdb->vnn_map->generation,
 369                          (unsigned long long)c->rsn, (unsigned long long)header.rsn, c->hdr.reqid,
 370                          (key.dsize >= 4)?(*(uint32_t *)key.dptr):0));
 371                 if (header.rsn != 0 || header.dmaster != ctdb->pnn) {
 372                         ctdb_fatal(ctdb, "ctdb_req_dmaster from non-master");
 373                         return;
 374                 }
 375         }
 376
 377         if (header.rsn > c->rsn) {
 378                 DEBUG(DEBUG_ALERT,("pnn %u dmaster request with older RSN new-dmaster %u from %u real-dmaster=%u key %08x dbid 0x%08x gen=%u curgen=%u c->rsn=%llu header.rsn=%llu reqid=%u\n",
 379                          ctdb->pnn, c->dmaster, hdr->srcnode, header.dmaster, ctdb_hash(&key),
 380                          ctdb_db->db_id, hdr->generation, ctdb->vnn_map->generation,
 381                          (unsigned long long)c->rsn, (unsigned long long)header.rsn, c->hdr.reqid));
 382         }
 383
 384         /* use the rsn from the sending node */
 385         header.rsn = c->rsn;
 386
 387         /* check if the new dmaster is the lmaster, in which case we
 388            skip the dmaster reply */
 389         if (c->dmaster == ctdb->pnn) {
 390                 ctdb_become_dmaster(ctdb_db, hdr, key, data, c->rsn);
 391         } else {
 392                 ctdb_send_dmaster_reply(ctdb_db, &header, key, data, c->dmaster, hdr->reqid);
 393                 ctdb_ltdb_unlock(ctdb_db, key);
 394         }
 395 }
 396
 397 /*
 398  * Did we just pull the dmaster of the record for a client fetch_lock,
 399  * so should we hold it back a while and thus give our client the
 400  * chance to do its own tdb_lock?
 401  */
 402
 403 static bool ctdb_held_back(struct ctdb_db_context *db, TDB_DATA key,
 404                            struct ctdb_req_header *hdr)
 405 {
 406         int i;
 407         size_t num_keys = talloc_array_length(db->holdback_keys);
 408         size_t num_hdrs;
 409         struct ctdb_req_header **tmp;
 410
 411         for (i=0; i<num_keys; i++) {
 412                 uint8_t *hold_key = db->holdback_keys[i];
 413                 size_t keylength = talloc_array_length(hold_key);
 414
 415                 if ((keylength == key.dsize)
 416                     && (memcmp(hold_key, key.dptr, keylength) == 0)) {
 417                         break;
 418                 }
 419         }
 420         if (i == num_keys) {
 421                 return false;
 422         }
 423         DEBUG(DEBUG_DEBUG, ("holding back record %08x after migration\n",
 424                             ctdb_hash(&key)));
 425
 426         num_hdrs = talloc_array_length(db->held_back);
 427
 428         tmp = talloc_realloc(db, db->held_back, struct ctdb_req_header *,
 429                              num_hdrs + 1);
 430         if (tmp == NULL) {
 431                 DEBUG(DEBUG_ERR, (__location__ "talloc_realloc failed\n"));
 432                 return false;
 433         }
 434         db->held_back = tmp;
 435         db->held_back[num_hdrs] = talloc_move(db->held_back, &hdr);
 436         return true;
 437 }
 438
 439 /*
 440   called when a CTDB_REQ_CALL packet comes in
 441 */
 442 void ctdb_request_call(struct ctdb_context *ctdb, struct ctdb_req_header *hdr)
 443 {
 444         struct ctdb_req_call *c = (struct ctdb_req_call *)hdr;
 445         TDB_DATA data;
 446         struct ctdb_reply_call *r;
 447         int ret, len;
 448         struct ctdb_ltdb_header header;
 449         struct ctdb_call *call;
 450         struct ctdb_db_context *ctdb_db;
 451
 452         if (ctdb->methods == NULL) {
 453                 DEBUG(DEBUG_ERR,(__location__ " Failed ctdb_request_call. Transport is DOWN\n"));
 454                 return;
 455         }
 456
 457
 458         ctdb_db = find_ctdb_db(ctdb, c->db_id);
 459         if (!ctdb_db) {
 460                 ctdb_send_error(ctdb, hdr, -1,
 461                                 "Unknown database in request. db_id==0x%08x",
 462                                 c->db_id);
 463                 return;
 464         }
 465
 466         call = talloc(hdr, struct ctdb_call);
 467         CTDB_NO_MEMORY_FATAL(ctdb, call);
 468
 469         call->call_id  = c->callid;
 470         call->key.dptr = c->data;
 471         call->key.dsize = c->keylen;
 472         call->call_data.dptr = c->data + c->keylen;
 473         call->call_data.dsize = c->calldatalen;
 474
 475         /* determine if we are the dmaster for this key. This also
 476            fetches the record data (if any), thus avoiding a 2nd fetch of the data
 477            if the call will be answered locally */
 478
 479         ret = ctdb_ltdb_lock_fetch_requeue(ctdb_db, call->key, &header, hdr, &data,
 480                                            ctdb_call_input_pkt, ctdb, False);
 481         if (ret == -1) {
 482                 ctdb_send_error(ctdb, hdr, ret, "ltdb fetch failed in ctdb_request_call");
 483                 return;
 484         }
 485         if (ret == -2) {
 486                 DEBUG(DEBUG_INFO,(__location__ " deferred ctdb_request_call\n"));
 487                 return;
 488         }
 489
 490         /* if we are not the dmaster, then send a redirect to the
 491            requesting node */
 492         if (header.dmaster != ctdb->pnn) {
 493                 talloc_free(data.dptr);
 494                 ctdb_call_send_redirect(ctdb, call->key, c, &header);
 495                 ctdb_ltdb_unlock(ctdb_db, call->key);
 496                 return;
 497         }
 498
 499         if (c->hopcount > ctdb->statistics.max_hop_count) {
 500                 ctdb->statistics.max_hop_count = c->hopcount;
 501         }
 502
 503         if ((c->flags & CTDB_IMMEDIATE_MIGRATION)
 504             && (ctdb_held_back(ctdb_db, call->key, hdr))) {
 505                 talloc_free(data.dptr);
 506                 ctdb_ltdb_unlock(ctdb_db, call->key);
 507                 return;
 508         }
 509
 510         /* if this nodes has done enough consecutive calls on the same record
 511            then give them the record
 512            or if the node requested an immediate migration
 513         */
 514         if ( c->hdr.srcnode != ctdb->pnn &&
 515              ((header.laccessor == c->hdr.srcnode
 516                && header.lacount >= ctdb->tunable.max_lacount)
 517               || (c->flags & CTDB_IMMEDIATE_MIGRATION)) ) {
 518                 if (ctdb_db->transaction_active) {
 519                         DEBUG(DEBUG_INFO, (__location__ " refusing migration"
 520                               " of key %s while transaction is active\n",
 521                               (char *)call->key.dptr));
 522                 } else {
 523                         DEBUG(DEBUG_INFO,("pnn %u starting migration of %08x to %u\n",
 524                                  ctdb->pnn, ctdb_hash(&(call->key)), c->hdr.srcnode));
 525                         ctdb_call_send_dmaster(ctdb_db, c, &header, &(call->key), &data);
 526                         talloc_free(data.dptr);
 527                         ctdb_ltdb_unlock(ctdb_db, call->key);
 528                         return;
 529                 }
 530         }
 531
 532         ctdb_call_local(ctdb_db, call, &header, hdr, &data, c->hdr.srcnode);
 533
 534         ctdb_ltdb_unlock(ctdb_db, call->key);
 535
 536         len = offsetof(struct ctdb_reply_call, data) + call->reply_data.dsize;
 537         r = ctdb_transport_allocate(ctdb, ctdb, CTDB_REPLY_CALL, len,
 538                                     struct ctdb_reply_call);
 539         CTDB_NO_MEMORY_FATAL(ctdb, r);
 540         r->hdr.destnode  = hdr->srcnode;
 541         r->hdr.reqid     = hdr->reqid;
 542         r->status        = call->status;
 543         r->datalen       = call->reply_data.dsize;
 544         if (call->reply_data.dsize) {
 545                 memcpy(&r->data[0], call->reply_data.dptr, call->reply_data.dsize);
 546         }
 547
 548         ctdb_queue_packet(ctdb, &r->hdr);
 549
 550         talloc_free(r);
 551 }
 552
 553 /*
 554   called when a CTDB_REPLY_CALL packet comes in
 555
 556   This packet comes in response to a CTDB_REQ_CALL request packet. It
 557   contains any reply data from the call
 558 */
 559 void ctdb_reply_call(struct ctdb_context *ctdb, struct ctdb_req_header *hdr)
 560 {
 561         struct ctdb_reply_call *c = (struct ctdb_reply_call *)hdr;
 562         struct ctdb_call_state *state;
 563
 564         state = ctdb_reqid_find(ctdb, hdr->reqid, struct ctdb_call_state);
 565         if (state == NULL) {
 566                 DEBUG(DEBUG_ERR, (__location__ " reqid %u not found\n", hdr->reqid));
 567                 return;
 568         }
 569
 570         if (hdr->reqid != state->reqid) {
 571                 /* we found a record  but it was the wrong one */
 572                 DEBUG(DEBUG_ERR, ("Dropped orphaned call reply with reqid:%u\n",hdr->reqid));
 573                 return;
 574         }
 575
 576         state->call->reply_data.dptr = c->data;
 577         state->call->reply_data.dsize = c->datalen;
 578         state->call->status = c->status;
 579
 580         talloc_steal(state, c);
 581
 582         state->state = CTDB_CALL_DONE;
 583         if (state->async.fn) {
 584                 state->async.fn(state);
 585         }
 586 }
 587
 588
 589 /*
 590   called when a CTDB_REPLY_DMASTER packet comes in
 591
 592   This packet comes in from the lmaster response to a CTDB_REQ_CALL
 593   request packet. It means that the current dmaster wants to give us
 594   the dmaster role
 595 */
 596 void ctdb_reply_dmaster(struct ctdb_context *ctdb, struct ctdb_req_header *hdr)
 597 {
 598         struct ctdb_reply_dmaster *c = (struct ctdb_reply_dmaster *)hdr;
 599         struct ctdb_db_context *ctdb_db;
 600         TDB_DATA key, data;
 601         int ret;
 602
 603         ctdb_db = find_ctdb_db(ctdb, c->db_id);
 604         if (ctdb_db == NULL) {
 605                 DEBUG(DEBUG_ERR,("Unknown db_id 0x%x in ctdb_reply_dmaster\n", c->db_id));
 606                 return;
 607         }
 608
 609         key.dptr = c->data;
 610         key.dsize = c->keylen;
 611         data.dptr = &c->data[key.dsize];
 612         data.dsize = c->datalen;
 613
 614         ret = ctdb_ltdb_lock_requeue(ctdb_db, key, hdr,
 615                                      ctdb_call_input_pkt, ctdb, False);
 616         if (ret == -2) {
 617                 return;
 618         }
 619         if (ret != 0) {
 620                 DEBUG(DEBUG_ERR,(__location__ " Failed to get lock in ctdb_reply_dmaster\n"));
 621                 return;
 622         }
 623
 624         ctdb_become_dmaster(ctdb_db, hdr, key, data, c->rsn);
 625 }
 626
 627
 628 /*
 629   called when a CTDB_REPLY_ERROR packet comes in
 630 */
 631 void ctdb_reply_error(struct ctdb_context *ctdb, struct ctdb_req_header *hdr)
 632 {
 633         struct ctdb_reply_error *c = (struct ctdb_reply_error *)hdr;
 634         struct ctdb_call_state *state;
 635
 636         state = ctdb_reqid_find(ctdb, hdr->reqid, struct ctdb_call_state);
 637         if (state == NULL) {
 638                 DEBUG(DEBUG_ERR,("pnn %u Invalid reqid %u in ctdb_reply_error\n",
 639                          ctdb->pnn, hdr->reqid));
 640                 return;
 641         }
 642
 643         if (hdr->reqid != state->reqid) {
 644                 /* we found a record  but it was the wrong one */
 645                 DEBUG(DEBUG_ERR, ("Dropped orphaned error reply with reqid:%u\n",hdr->reqid));
 646                 return;
 647         }
 648
 649         talloc_steal(state, c);
 650
 651         state->state  = CTDB_CALL_ERROR;
 652         state->errmsg = (char *)c->msg;
 653         if (state->async.fn) {
 654                 state->async.fn(state);
 655         }
 656 }
 657
 658
 659 /*
 660   destroy a ctdb_call
 661 */
 662 static int ctdb_call_destructor(struct ctdb_call_state *state)
 663 {
 664         DLIST_REMOVE(state->ctdb_db->ctdb->pending_calls, state);
 665         ctdb_reqid_remove(state->ctdb_db->ctdb, state->reqid);
 666         return 0;
 667 }
 668
 669
 670 /*
 671   called when a ctdb_call needs to be resent after a reconfigure event
 672 */
 673 static void ctdb_call_resend(struct ctdb_call_state *state)
 674 {
 675         struct ctdb_context *ctdb = state->ctdb_db->ctdb;
 676
 677         state->generation = ctdb->vnn_map->generation;
 678
 679         /* use a new reqid, in case the old reply does eventually come in */
 680         ctdb_reqid_remove(ctdb, state->reqid);
 681         state->reqid = ctdb_reqid_new(ctdb, state);
 682         state->c->hdr.reqid = state->reqid;
 683
 684         /* update the generation count for this request, so its valid with the new vnn_map */
 685         state->c->hdr.generation = state->generation;
 686
 687         /* send the packet to ourselves, it will be redirected appropriately */
 688         state->c->hdr.destnode = ctdb->pnn;
 689
 690         ctdb_queue_packet(ctdb, &state->c->hdr);
 691         DEBUG(DEBUG_NOTICE,("resent ctdb_call\n"));
 692 }
 693
 694 /*
 695   resend all pending calls on recovery
 696  */
 697 void ctdb_call_resend_all(struct ctdb_context *ctdb)
 698 {
 699         struct ctdb_call_state *state, *next;
 700         for (state=ctdb->pending_calls;state;state=next) {
 701                 next = state->next;
 702                 ctdb_call_resend(state);
 703         }
 704 }
 705
 706 /*
 707   this allows the caller to setup a async.fn
 708 */
 709 static void call_local_trigger(struct event_context *ev, struct timed_event *te,
 710                        struct timeval t, void *private_data)
 711 {
 712         struct ctdb_call_state *state = talloc_get_type(private_data, struct ctdb_call_state);
 713         if (state->async.fn) {
 714                 state->async.fn(state);
 715         }
 716 }
 717
 718
 719 /*
 720   construct an event driven local ctdb_call
 721
 722   this is used so that locally processed ctdb_call requests are processed
 723   in an event driven manner
 724 */
 725 struct ctdb_call_state *ctdb_call_local_send(struct ctdb_db_context *ctdb_db,
 726                                              struct ctdb_call *call,
 727                                              struct ctdb_ltdb_header *header,
 728                                              TDB_DATA *data)
 729 {
 730         struct ctdb_call_state *state;
 731         struct ctdb_context *ctdb = ctdb_db->ctdb;
 732         int ret;
 733
 734         state = talloc_zero(ctdb_db, struct ctdb_call_state);
 735         CTDB_NO_MEMORY_NULL(ctdb, state);
 736
 737         talloc_steal(state, data->dptr);
 738
 739         state->state = CTDB_CALL_DONE;
 740         state->call  = talloc(state, struct ctdb_call);
 741         CTDB_NO_MEMORY_NULL(ctdb, state->call);
 742         *(state->call) = *call;
 743         state->ctdb_db = ctdb_db;
 744
 745         ret = ctdb_call_local(ctdb_db, state->call, header, state, data, ctdb->pnn);
 746
 747         event_add_timed(ctdb->ev, state, timeval_zero(), call_local_trigger, state);
 748
 749         return state;
 750 }
 751
 752
 753 /*
 754   make a remote ctdb call - async send. Called in daemon context.
 755
 756   This constructs a ctdb_call request and queues it for processing.
 757   This call never blocks.
 758 */
 759 struct ctdb_call_state *ctdb_daemon_call_send_remote(struct ctdb_db_context *ctdb_db,
 760                                                      struct ctdb_call *call,
 761                                                      struct ctdb_ltdb_header *header)
 762 {
 763         uint32_t len;
 764         struct ctdb_call_state *state;
 765         struct ctdb_context *ctdb = ctdb_db->ctdb;
 766
 767         if (ctdb->methods == NULL) {
 768                 DEBUG(DEBUG_ERR,(__location__ " Failed send packet. Transport is down\n"));
 769                 return NULL;
 770         }
 771
 772         state = talloc_zero(ctdb_db, struct ctdb_call_state);
 773         CTDB_NO_MEMORY_NULL(ctdb, state);
 774         state->call = talloc(state, struct ctdb_call);
 775         CTDB_NO_MEMORY_NULL(ctdb, state->call);
 776
 777         state->reqid = ctdb_reqid_new(ctdb, state);
 778         state->ctdb_db = ctdb_db;
 779         talloc_set_destructor(state, ctdb_call_destructor);
 780
 781         len = offsetof(struct ctdb_req_call, data) + call->key.dsize + call->call_data.dsize;
 782         state->c = ctdb_transport_allocate(ctdb, state, CTDB_REQ_CALL, len,
 783                                            struct ctdb_req_call);
 784         CTDB_NO_MEMORY_NULL(ctdb, state->c);
 785         state->c->hdr.destnode  = header->dmaster;
 786
 787         /* this limits us to 16k outstanding messages - not unreasonable */
 788         state->c->hdr.reqid     = state->reqid;
 789         state->c->flags         = call->flags;
 790         state->c->db_id         = ctdb_db->db_id;
 791         state->c->callid        = call->call_id;
 792         state->c->hopcount      = 0;
 793         state->c->keylen        = call->key.dsize;
 794         state->c->calldatalen   = call->call_data.dsize;
 795         memcpy(&state->c->data[0], call->key.dptr, call->key.dsize);
 796         memcpy(&state->c->data[call->key.dsize],
 797                call->call_data.dptr, call->call_data.dsize);
 798         *(state->call)              = *call;
 799         state->call->call_data.dptr = &state->c->data[call->key.dsize];
 800         state->call->key.dptr       = &state->c->data[0];
 801
 802         state->state  = CTDB_CALL_WAIT;
 803         state->generation = ctdb->vnn_map->generation;
 804
 805         DLIST_ADD(ctdb->pending_calls, state);
 806
 807         ctdb_queue_packet(ctdb, &state->c->hdr);
 808
 809         return state;
 810 }
 811
 812 /*
 813   make a remote ctdb call - async recv - called in daemon context
 814
 815   This is called when the program wants to wait for a ctdb_call to complete and get the
 816   results. This call will block unless the call has already completed.
 817 */
 818 int ctdb_daemon_call_recv(struct ctdb_call_state *state, struct ctdb_call *call)
 819 {
 820         while (state->state < CTDB_CALL_DONE) {
 821                 event_loop_once(state->ctdb_db->ctdb->ev);
 822         }
 823         if (state->state != CTDB_CALL_DONE) {
 824                 ctdb_set_error(state->ctdb_db->ctdb, "%s", state->errmsg);
 825                 talloc_free(state);
 826                 return -1;
 827         }
 828
 829         if (state->call->reply_data.dsize) {
 830                 call->reply_data.dptr = talloc_memdup(call,
 831                                                       state->call->reply_data.dptr,
 832                                                       state->call->reply_data.dsize);
 833                 call->reply_data.dsize = state->call->reply_data.dsize;
 834         } else {
 835                 call->reply_data.dptr = NULL;
 836                 call->reply_data.dsize = 0;
 837         }
 838         call->status = state->call->status;
 839         talloc_free(state);
 840         return 0;
 841 }
 842
 843
 844 /*
 845    send a keepalive packet to the other node
 846 */
 847 void ctdb_send_keepalive(struct ctdb_context *ctdb, uint32_t destnode)
 848 {
 849         struct ctdb_req_keepalive *r;
 850
 851         if (ctdb->methods == NULL) {
 852                 DEBUG(DEBUG_ERR,(__location__ " Failed to send keepalive. Transport is DOWN\n"));
 853                 return;
 854         }
 855
 856         r = ctdb_transport_allocate(ctdb, ctdb, CTDB_REQ_KEEPALIVE,
 857                                     sizeof(struct ctdb_req_keepalive),
 858                                     struct ctdb_req_keepalive);
 859         CTDB_NO_MEMORY_FATAL(ctdb, r);
 860         r->hdr.destnode  = destnode;
 861         r->hdr.reqid     = 0;
 862
 863         ctdb->statistics.keepalive_packets_sent++;
 864
 865         ctdb_queue_packet(ctdb, &r->hdr);
 866
 867         talloc_free(r);
 868 }
 869
 870 static void ctdb_holdback_cleanup(struct event_context *ev,
 871                                   struct timed_event *te,
 872                                   struct timeval t, void *private_data)
 873 {
 874         struct ctdb_context *ctdb = talloc_get_type(private_data,
 875                                                     struct ctdb_context);
 876         struct ctdb_db_context *ctdb_db;
 877
 878         DEBUG(DEBUG_INFO, ("running ctdb_holdback_cleanup\n"));
 879
 880         if (te != ctdb->holdback_cleanup_te) {
 881                 ctdb_fatal(ctdb, "te != ctdb->holdback_cleanup_te");
 882         }
 883
 884         for (ctdb_db=ctdb->db_list; ctdb_db; ctdb_db=ctdb_db->next) {
 885                 size_t i, num_heldback;
 886
 887                 talloc_free(ctdb_db->holdback_keys);
 888                 ctdb_db->holdback_keys = NULL;
 889
 890                 num_heldback = talloc_array_length(ctdb_db->held_back);
 891                 for (i=0; i<num_heldback; i++) {
 892                         ctdb_queue_packet(ctdb, ctdb_db->held_back[i]);
 893                 }
 894                 talloc_free(ctdb_db->held_back);
 895                 ctdb_db->held_back = NULL;
 896         }
 897
 898         ctdb->holdback_cleanup_te = event_add_timed(
 899                 ctdb->ev, ctdb, timeval_current_ofs(
 900                         0, ctdb->tunable.holdback_cleanup_interval * 1000),
 901                 ctdb_holdback_cleanup, ctdb);
 902 }
 903
 904 void ctdb_start_holdback_cleanup(struct ctdb_context *ctdb)
 905 {
 906         ctdb->holdback_cleanup_te = event_add_timed(
 907                 ctdb->ev, ctdb, timeval_current_ofs(
 908                         0, ctdb->tunable.holdback_cleanup_interval * 1000),
 909                 ctdb_holdback_cleanup, ctdb);
 910
 911         CTDB_NO_MEMORY_FATAL(ctdb, ctdb->holdback_cleanup_te);
 912
 913         DEBUG(DEBUG_NOTICE,("Holdback cleanup has been started\n"));
 914 }
 915
 916 void ctdb_stop_holdback_cleanup(struct ctdb_context *ctdb)
 917 {
 918         talloc_free(ctdb->holdback_cleanup_te);
 919         ctdb->holdback_cleanup_te = NULL;
 920 }
 921
 922 int32_t ctdb_control_gotit(struct ctdb_context *ctdb, TDB_DATA indata)
 923 {
 924         struct ctdb_control_gotit *c =
 925                 (struct ctdb_control_gotit *)indata.dptr;
 926         struct ctdb_db_context *db;
 927         size_t i, num_keys, num_heldback;
 928         TDB_DATA key;
 929
 930         if (indata.dsize < sizeof(struct ctdb_control_gotit)) {
 931                 DEBUG(DEBUG_ERR, (__location__ "Invalid data size %d\n",
 932                                   (int)indata.dsize));
 933                 return -1;
 934         }
 935         db = find_ctdb_db(ctdb, c->db_id);
 936         if (db == NULL) {
 937                 DEBUG(DEBUG_ERR, ("Unknown db_id 0x%x in ctdb_reply_dmaster\n",
 938                                   (int)c->db_id));
 939                 return -1;
 940         }
 941
 942         key.dptr = c->key;
 943         key.dsize = indata.dsize - offsetof(struct ctdb_control_gotit, key);
 944
 945         num_keys = talloc_array_length(db->holdback_keys);
 946         for (i=0; i<num_keys; i++) {
 947                 uint8_t *hold_key = db->holdback_keys[i];
 948                 size_t keylength = talloc_array_length(hold_key);
 949
 950                 if ((keylength == key.dsize)
 951                     && (memcmp(hold_key, key.dptr, keylength) == 0)) {
 952                         break;
 953                 }
 954         }
 955         if (i == num_keys) {
 956                 /*
 957                  * ctdb_holdback_cleanup has kicked in. This is okay,
 958                  * we will just potentially have to retry.
 959                  */
 960                 return 0;
 961         }
 962
 963         talloc_free(db->holdback_keys[i]);
 964         if (i < num_keys-1) {
 965                 db->holdback_keys[i] = db->holdback_keys[num_keys-1];
 966         }
 967         db->holdback_keys = talloc_realloc(db, db->holdback_keys, uint8_t *,
 968                                            num_keys-1);
 969
 970         num_heldback = talloc_array_length(db->held_back);
 971         for (i=0; i<num_heldback; i++) {
 972                 ctdb_queue_packet(ctdb, db->held_back[i]);
 973         }
 974         talloc_free(db->held_back);
 975         db->held_back = NULL;
 976         return 0;
 977 }