fs/ocfs2/dlm/dlmdomain.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmdomain.c
   5  *
   6  * defines domain join / leave apis
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27 #include <linux/module.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/utsname.h>
  32 #include <linux/init.h>
  33 #include <linux/spinlock.h>
  34 #include <linux/delay.h>
  35 #include <linux/err.h>
  36
  37 #include "cluster/heartbeat.h"
  38 #include "cluster/nodemanager.h"
  39 #include "cluster/tcp.h"
  40
  41 #include "dlmapi.h"
  42 #include "dlmcommon.h"
  43
  44 #include "dlmdomain.h"
  45
  46 #include "dlmver.h"
  47
  48 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
  49 #include "cluster/masklog.h"
  50
  51 static void dlm_free_pagevec(void **vec, int pages)
  52 {
  53         while (pages--)
  54                 free_page((unsigned long)vec[pages]);
  55         kfree(vec);
  56 }
  57
  58 static void **dlm_alloc_pagevec(int pages)
  59 {
  60         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
  61         int i;
  62
  63         if (!vec)
  64                 return NULL;
  65
  66         for (i = 0; i < pages; i++)
  67                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
  68                         goto out_free;
  69
  70         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
  71              pages, (unsigned long)DLM_HASH_PAGES,
  72              (unsigned long)DLM_BUCKETS_PER_PAGE);
  73         return vec;
  74 out_free:
  75         dlm_free_pagevec(vec, i);
  76         return NULL;
  77 }
  78
  79 /*
  80  *
  81  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
  82  *    dlm_domain_lock
  83  *    struct dlm_ctxt->spinlock
  84  *    struct dlm_lock_resource->spinlock
  85  *    struct dlm_ctxt->master_lock
  86  *    struct dlm_ctxt->ast_lock
  87  *    dlm_master_list_entry->spinlock
  88  *    dlm_lock->spinlock
  89  *
  90  */
  91
  92 DEFINE_SPINLOCK(dlm_domain_lock);
  93 LIST_HEAD(dlm_domains);
  94 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
  95
  96 #define DLM_DOMAIN_BACKOFF_MS 200
  97
  98 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
  99                                   void **ret_data);
 100 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 101                                      void **ret_data);
 102 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 103                                    void **ret_data);
 104 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 105                                    void **ret_data);
 106
 107 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 108
 109 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 110 {
 111         hlist_del_init(&lockres->hash_node);
 112         dlm_lockres_put(lockres);
 113 }
 114
 115 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 116                        struct dlm_lock_resource *res)
 117 {
 118         struct hlist_head *bucket;
 119         struct qstr *q;
 120
 121         assert_spin_locked(&dlm->spinlock);
 122
 123         q = &res->lockname;
 124         bucket = dlm_lockres_hash(dlm, q->hash);
 125
 126         /* get a reference for our hashtable */
 127         dlm_lockres_get(res);
 128
 129         hlist_add_head(&res->hash_node, bucket);
 130 }
 131
 132 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 133                                                      const char *name,
 134                                                      unsigned int len,
 135                                                      unsigned int hash)
 136 {
 137         struct hlist_head *bucket;
 138         struct hlist_node *list;
 139
 140         mlog_entry("%.*s\n", len, name);
 141
 142         assert_spin_locked(&dlm->spinlock);
 143
 144         bucket = dlm_lockres_hash(dlm, hash);
 145
 146         hlist_for_each(list, bucket) {
 147                 struct dlm_lock_resource *res = hlist_entry(list,
 148                         struct dlm_lock_resource, hash_node);
 149                 if (res->lockname.name[0] != name[0])
 150                         continue;
 151                 if (unlikely(res->lockname.len != len))
 152                         continue;
 153                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 154                         continue;
 155                 dlm_lockres_get(res);
 156                 return res;
 157         }
 158         return NULL;
 159 }
 160
 161 /* intended to be called by functions which do not care about lock
 162  * resources which are being purged (most net _handler functions).
 163  * this will return NULL for any lock resource which is found but
 164  * currently in the process of dropping its mastery reference.
 165  * use __dlm_lookup_lockres_full when you need the lock resource
 166  * regardless (e.g. dlm_get_lock_resource) */
 167 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 168                                                 const char *name,
 169                                                 unsigned int len,
 170                                                 unsigned int hash)
 171 {
 172         struct dlm_lock_resource *res = NULL;
 173
 174         mlog_entry("%.*s\n", len, name);
 175
 176         assert_spin_locked(&dlm->spinlock);
 177
 178         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 179         if (res) {
 180                 spin_lock(&res->spinlock);
 181                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 182                         spin_unlock(&res->spinlock);
 183                         dlm_lockres_put(res);
 184                         return NULL;
 185                 }
 186                 spin_unlock(&res->spinlock);
 187         }
 188
 189         return res;
 190 }
 191
 192 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 193                                     const char *name,
 194                                     unsigned int len)
 195 {
 196         struct dlm_lock_resource *res;
 197         unsigned int hash = dlm_lockid_hash(name, len);
 198
 199         spin_lock(&dlm->spinlock);
 200         res = __dlm_lookup_lockres(dlm, name, len, hash);
 201         spin_unlock(&dlm->spinlock);
 202         return res;
 203 }
 204
 205 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 206 {
 207         struct dlm_ctxt *tmp = NULL;
 208         struct list_head *iter;
 209
 210         assert_spin_locked(&dlm_domain_lock);
 211
 212         /* tmp->name here is always NULL terminated,
 213          * but domain may not be! */
 214         list_for_each(iter, &dlm_domains) {
 215                 tmp = list_entry (iter, struct dlm_ctxt, list);
 216                 if (strlen(tmp->name) == len &&
 217                     memcmp(tmp->name, domain, len)==0)
 218                         break;
 219                 tmp = NULL;
 220         }
 221
 222         return tmp;
 223 }
 224
 225 /* For null terminated domain strings ONLY */
 226 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 227 {
 228         assert_spin_locked(&dlm_domain_lock);
 229
 230         return __dlm_lookup_domain_full(domain, strlen(domain));
 231 }
 232
 233
 234 /* returns true on one of two conditions:
 235  * 1) the domain does not exist
 236  * 2) the domain exists and it's state is "joined" */
 237 static int dlm_wait_on_domain_helper(const char *domain)
 238 {
 239         int ret = 0;
 240         struct dlm_ctxt *tmp = NULL;
 241
 242         spin_lock(&dlm_domain_lock);
 243
 244         tmp = __dlm_lookup_domain(domain);
 245         if (!tmp)
 246                 ret = 1;
 247         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 248                 ret = 1;
 249
 250         spin_unlock(&dlm_domain_lock);
 251         return ret;
 252 }
 253
 254 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 255 {
 256         if (dlm->lockres_hash)
 257                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 258
 259         if (dlm->name)
 260                 kfree(dlm->name);
 261
 262         kfree(dlm);
 263 }
 264
 265 /* A little strange - this function will be called while holding
 266  * dlm_domain_lock and is expected to be holding it on the way out. We
 267  * will however drop and reacquire it multiple times */
 268 static void dlm_ctxt_release(struct kref *kref)
 269 {
 270         struct dlm_ctxt *dlm;
 271
 272         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 273
 274         BUG_ON(dlm->num_joins);
 275         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 276
 277         /* we may still be in the list if we hit an error during join. */
 278         list_del_init(&dlm->list);
 279
 280         spin_unlock(&dlm_domain_lock);
 281
 282         mlog(0, "freeing memory from domain %s\n", dlm->name);
 283
 284         wake_up(&dlm_domain_events);
 285
 286         dlm_free_ctxt_mem(dlm);
 287
 288         spin_lock(&dlm_domain_lock);
 289 }
 290
 291 void dlm_put(struct dlm_ctxt *dlm)
 292 {
 293         spin_lock(&dlm_domain_lock);
 294         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 295         spin_unlock(&dlm_domain_lock);
 296 }
 297
 298 static void __dlm_get(struct dlm_ctxt *dlm)
 299 {
 300         kref_get(&dlm->dlm_refs);
 301 }
 302
 303 /* given a questionable reference to a dlm object, gets a reference if
 304  * it can find it in the list, otherwise returns NULL in which case
 305  * you shouldn't trust your pointer. */
 306 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 307 {
 308         struct list_head *iter;
 309         struct dlm_ctxt *target = NULL;
 310
 311         spin_lock(&dlm_domain_lock);
 312
 313         list_for_each(iter, &dlm_domains) {
 314                 target = list_entry (iter, struct dlm_ctxt, list);
 315
 316                 if (target == dlm) {
 317                         __dlm_get(target);
 318                         break;
 319                 }
 320
 321                 target = NULL;
 322         }
 323
 324         spin_unlock(&dlm_domain_lock);
 325
 326         return target;
 327 }
 328
 329 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 330 {
 331         int ret;
 332
 333         spin_lock(&dlm_domain_lock);
 334         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 335                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 336         spin_unlock(&dlm_domain_lock);
 337
 338         return ret;
 339 }
 340
 341 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 342 {
 343         if (dlm->dlm_worker) {
 344                 flush_workqueue(dlm->dlm_worker);
 345                 destroy_workqueue(dlm->dlm_worker);
 346                 dlm->dlm_worker = NULL;
 347         }
 348 }
 349
 350 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 351 {
 352         dlm_unregister_domain_handlers(dlm);
 353         dlm_complete_thread(dlm);
 354         dlm_complete_recovery_thread(dlm);
 355         dlm_destroy_dlm_worker(dlm);
 356
 357         /* We've left the domain. Now we can take ourselves out of the
 358          * list and allow the kref stuff to help us free the
 359          * memory. */
 360         spin_lock(&dlm_domain_lock);
 361         list_del_init(&dlm->list);
 362         spin_unlock(&dlm_domain_lock);
 363
 364         /* Wake up anyone waiting for us to remove this domain */
 365         wake_up(&dlm_domain_events);
 366 }
 367
 368 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 369 {
 370         int i, num, n, ret = 0;
 371         struct dlm_lock_resource *res;
 372         struct hlist_node *iter;
 373         struct hlist_head *bucket;
 374         int dropped;
 375
 376         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 377
 378         num = 0;
 379         spin_lock(&dlm->spinlock);
 380         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 381 redo_bucket:
 382                 n = 0;
 383                 bucket = dlm_lockres_hash(dlm, i);
 384                 iter = bucket->first;
 385                 while (iter) {
 386                         n++;
 387                         res = hlist_entry(iter, struct dlm_lock_resource,
 388                                           hash_node);
 389                         dlm_lockres_get(res);
 390                         /* migrate, if necessary.  this will drop the dlm
 391                          * spinlock and retake it if it does migration. */
 392                         dropped = dlm_empty_lockres(dlm, res);
 393
 394                         spin_lock(&res->spinlock);
 395                         __dlm_lockres_calc_usage(dlm, res);
 396                         iter = res->hash_node.next;
 397                         spin_unlock(&res->spinlock);
 398
 399                         dlm_lockres_put(res);
 400
 401                         cond_resched_lock(&dlm->spinlock);
 402
 403                         if (dropped)
 404                                 goto redo_bucket;
 405                 }
 406                 num += n;
 407                 mlog(0, "%s: touched %d lockreses in bucket %d "
 408                      "(tot=%d)\n", dlm->name, n, i, num);
 409         }
 410         spin_unlock(&dlm->spinlock);
 411         wake_up(&dlm->dlm_thread_wq);
 412
 413         /* let the dlm thread take care of purging, keep scanning until
 414          * nothing remains in the hash */
 415         if (num) {
 416                 mlog(0, "%s: %d lock resources in hash last pass\n",
 417                      dlm->name, num);
 418                 ret = -EAGAIN;
 419         }
 420         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 421         return ret;
 422 }
 423
 424 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 425 {
 426         int ret;
 427
 428         spin_lock(&dlm->spinlock);
 429         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 430         spin_unlock(&dlm->spinlock);
 431
 432         return ret;
 433 }
 434
 435 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 436 {
 437         /* Yikes, a double spinlock! I need domain_lock for the dlm
 438          * state and the dlm spinlock for join state... Sorry! */
 439 again:
 440         spin_lock(&dlm_domain_lock);
 441         spin_lock(&dlm->spinlock);
 442
 443         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 444                 mlog(0, "Node %d is joining, we wait on it.\n",
 445                           dlm->joining_node);
 446                 spin_unlock(&dlm->spinlock);
 447                 spin_unlock(&dlm_domain_lock);
 448
 449                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 450                 goto again;
 451         }
 452
 453         dlm->dlm_state = DLM_CTXT_LEAVING;
 454         spin_unlock(&dlm->spinlock);
 455         spin_unlock(&dlm_domain_lock);
 456 }
 457
 458 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 459 {
 460         int node = -1;
 461
 462         assert_spin_locked(&dlm->spinlock);
 463
 464         printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name);
 465
 466         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 467                                      node + 1)) < O2NM_MAX_NODES) {
 468                 printk("%d ", node);
 469         }
 470         printk("\n");
 471 }
 472
 473 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 474                                    void **ret_data)
 475 {
 476         struct dlm_ctxt *dlm = data;
 477         unsigned int node;
 478         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 479
 480         mlog_entry("%p %u %p", msg, len, data);
 481
 482         if (!dlm_grab(dlm))
 483                 return 0;
 484
 485         node = exit_msg->node_idx;
 486
 487         printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name);
 488
 489         spin_lock(&dlm->spinlock);
 490         clear_bit(node, dlm->domain_map);
 491         __dlm_print_nodes(dlm);
 492
 493         /* notify anything attached to the heartbeat events */
 494         dlm_hb_event_notify_attached(dlm, node, 0);
 495
 496         spin_unlock(&dlm->spinlock);
 497
 498         dlm_put(dlm);
 499
 500         return 0;
 501 }
 502
 503 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
 504                                     unsigned int node)
 505 {
 506         int status;
 507         struct dlm_exit_domain leave_msg;
 508
 509         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
 510                   node, dlm->name, dlm->node_num);
 511
 512         memset(&leave_msg, 0, sizeof(leave_msg));
 513         leave_msg.node_idx = dlm->node_num;
 514
 515         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
 516                                     &leave_msg, sizeof(leave_msg), node,
 517                                     NULL);
 518
 519         mlog(0, "status return %d from o2net_send_message\n", status);
 520
 521         return status;
 522 }
 523
 524
 525 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 526 {
 527         int node, clear_node, status;
 528
 529         /* At this point we've migrated away all our locks and won't
 530          * accept mastership of new ones. The dlm is responsible for
 531          * almost nothing now. We make sure not to confuse any joining
 532          * nodes and then commence shutdown procedure. */
 533
 534         spin_lock(&dlm->spinlock);
 535         /* Clear ourselves from the domain map */
 536         clear_bit(dlm->node_num, dlm->domain_map);
 537         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 538                                      0)) < O2NM_MAX_NODES) {
 539                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 540                  * -nodes cannot be added now as the
 541                  *   query_join_handlers knows to respond with OK_NO_MAP
 542                  * -we catch the right network errors if a node is
 543                  *   removed from the map while we're sending him the
 544                  *   exit message. */
 545                 spin_unlock(&dlm->spinlock);
 546
 547                 clear_node = 1;
 548
 549                 status = dlm_send_one_domain_exit(dlm, node);
 550                 if (status < 0 &&
 551                     status != -ENOPROTOOPT &&
 552                     status != -ENOTCONN) {
 553                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 554                              "to node %d\n", status, node);
 555
 556                         /* Not sure what to do here but lets sleep for
 557                          * a bit in case this was a transient
 558                          * error... */
 559                         msleep(DLM_DOMAIN_BACKOFF_MS);
 560                         clear_node = 0;
 561                 }
 562
 563                 spin_lock(&dlm->spinlock);
 564                 /* If we're not clearing the node bit then we intend
 565                  * to loop back around to try again. */
 566                 if (clear_node)
 567                         clear_bit(node, dlm->domain_map);
 568         }
 569         spin_unlock(&dlm->spinlock);
 570 }
 571
 572 int dlm_joined(struct dlm_ctxt *dlm)
 573 {
 574         int ret = 0;
 575
 576         spin_lock(&dlm_domain_lock);
 577
 578         if (dlm->dlm_state == DLM_CTXT_JOINED)
 579                 ret = 1;
 580
 581         spin_unlock(&dlm_domain_lock);
 582
 583         return ret;
 584 }
 585
 586 int dlm_shutting_down(struct dlm_ctxt *dlm)
 587 {
 588         int ret = 0;
 589
 590         spin_lock(&dlm_domain_lock);
 591
 592         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 593                 ret = 1;
 594
 595         spin_unlock(&dlm_domain_lock);
 596
 597         return ret;
 598 }
 599
 600 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 601 {
 602         int leave = 0;
 603
 604         spin_lock(&dlm_domain_lock);
 605         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 606         BUG_ON(!dlm->num_joins);
 607
 608         dlm->num_joins--;
 609         if (!dlm->num_joins) {
 610                 /* We mark it "in shutdown" now so new register
 611                  * requests wait until we've completely left the
 612                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 613                  * want new domain joins to communicate with us at
 614                  * least until we've completed migration of our
 615                  * resources. */
 616                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 617                 leave = 1;
 618         }
 619         spin_unlock(&dlm_domain_lock);
 620
 621         if (leave) {
 622                 mlog(0, "shutting down domain %s\n", dlm->name);
 623
 624                 /* We changed dlm state, notify the thread */
 625                 dlm_kick_thread(dlm, NULL);
 626
 627                 while (dlm_migrate_all_locks(dlm)) {
 628                         mlog(0, "%s: more migration to do\n", dlm->name);
 629                 }
 630                 dlm_mark_domain_leaving(dlm);
 631                 dlm_leave_domain(dlm);
 632                 dlm_complete_dlm_shutdown(dlm);
 633         }
 634         dlm_put(dlm);
 635 }
 636 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 637
 638 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 639                                   void **ret_data)
 640 {
 641         struct dlm_query_join_request *query;
 642         enum dlm_query_join_response response;
 643         struct dlm_ctxt *dlm = NULL;
 644
 645         query = (struct dlm_query_join_request *) msg->buf;
 646
 647         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 648                   query->domain);
 649
 650         /*
 651          * If heartbeat doesn't consider the node live, tell it
 652          * to back off and try again.  This gives heartbeat a chance
 653          * to catch up.
 654          */
 655         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 656                 mlog(0, "node %u is not in our live map yet\n",
 657                      query->node_idx);
 658
 659                 response = JOIN_DISALLOW;
 660                 goto respond;
 661         }
 662
 663         response = JOIN_OK_NO_MAP;
 664
 665         spin_lock(&dlm_domain_lock);
 666         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 667         /* Once the dlm ctxt is marked as leaving then we don't want
 668          * to be put in someone's domain map.
 669          * Also, explicitly disallow joining at certain troublesome
 670          * times (ie. during recovery). */
 671         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 672                 int bit = query->node_idx;
 673                 spin_lock(&dlm->spinlock);
 674
 675                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 676                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 677                         /*If this is a brand new context and we
 678                          * haven't started our join process yet, then
 679                          * the other node won the race. */
 680                         response = JOIN_OK_NO_MAP;
 681                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 682                         /* Disallow parallel joins. */
 683                         response = JOIN_DISALLOW;
 684                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 685                         mlog(ML_NOTICE, "node %u trying to join, but recovery "
 686                              "is ongoing.\n", bit);
 687                         response = JOIN_DISALLOW;
 688                 } else if (test_bit(bit, dlm->recovery_map)) {
 689                         mlog(ML_NOTICE, "node %u trying to join, but it "
 690                              "still needs recovery.\n", bit);
 691                         response = JOIN_DISALLOW;
 692                 } else if (test_bit(bit, dlm->domain_map)) {
 693                         mlog(ML_NOTICE, "node %u trying to join, but it "
 694                              "is still in the domain! needs recovery?\n",
 695                              bit);
 696                         response = JOIN_DISALLOW;
 697                 } else {
 698                         /* Alright we're fully a part of this domain
 699                          * so we keep some state as to who's joining
 700                          * and indicate to him that needs to be fixed
 701                          * up. */
 702                         response = JOIN_OK;
 703                         __dlm_set_joining_node(dlm, query->node_idx);
 704                 }
 705
 706                 spin_unlock(&dlm->spinlock);
 707         }
 708         spin_unlock(&dlm_domain_lock);
 709
 710 respond:
 711         mlog(0, "We respond with %u\n", response);
 712
 713         return response;
 714 }
 715
 716 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 717                                      void **ret_data)
 718 {
 719         struct dlm_assert_joined *assert;
 720         struct dlm_ctxt *dlm = NULL;
 721
 722         assert = (struct dlm_assert_joined *) msg->buf;
 723
 724         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 725                   assert->domain);
 726
 727         spin_lock(&dlm_domain_lock);
 728         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 729         /* XXX should we consider no dlm ctxt an error? */
 730         if (dlm) {
 731                 spin_lock(&dlm->spinlock);
 732
 733                 /* Alright, this node has officially joined our
 734                  * domain. Set him in the map and clean up our
 735                  * leftover join state. */
 736                 BUG_ON(dlm->joining_node != assert->node_idx);
 737                 set_bit(assert->node_idx, dlm->domain_map);
 738                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 739
 740                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
 741                        assert->node_idx, dlm->name);
 742                 __dlm_print_nodes(dlm);
 743
 744                 /* notify anything attached to the heartbeat events */
 745                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 746
 747                 spin_unlock(&dlm->spinlock);
 748         }
 749         spin_unlock(&dlm_domain_lock);
 750
 751         return 0;
 752 }
 753
 754 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 755                                    void **ret_data)
 756 {
 757         struct dlm_cancel_join *cancel;
 758         struct dlm_ctxt *dlm = NULL;
 759
 760         cancel = (struct dlm_cancel_join *) msg->buf;
 761
 762         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 763                   cancel->domain);
 764
 765         spin_lock(&dlm_domain_lock);
 766         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 767
 768         if (dlm) {
 769                 spin_lock(&dlm->spinlock);
 770
 771                 /* Yikes, this guy wants to cancel his join. No
 772                  * problem, we simply cleanup our join state. */
 773                 BUG_ON(dlm->joining_node != cancel->node_idx);
 774                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 775
 776                 spin_unlock(&dlm->spinlock);
 777         }
 778         spin_unlock(&dlm_domain_lock);
 779
 780         return 0;
 781 }
 782
 783 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 784                                     unsigned int node)
 785 {
 786         int status;
 787         struct dlm_cancel_join cancel_msg;
 788
 789         memset(&cancel_msg, 0, sizeof(cancel_msg));
 790         cancel_msg.node_idx = dlm->node_num;
 791         cancel_msg.name_len = strlen(dlm->name);
 792         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 793
 794         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 795                                     &cancel_msg, sizeof(cancel_msg), node,
 796                                     NULL);
 797         if (status < 0) {
 798                 mlog_errno(status);
 799                 goto bail;
 800         }
 801
 802 bail:
 803         return status;
 804 }
 805
 806 /* map_size should be in bytes. */
 807 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 808                                  unsigned long *node_map,
 809                                  unsigned int map_size)
 810 {
 811         int status, tmpstat;
 812         unsigned int node;
 813
 814         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 815                          sizeof(unsigned long))) {
 816                 mlog(ML_ERROR,
 817                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 818                      map_size, BITS_TO_LONGS(O2NM_MAX_NODES));
 819                 return -EINVAL;
 820         }
 821
 822         status = 0;
 823         node = -1;
 824         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 825                                      node + 1)) < O2NM_MAX_NODES) {
 826                 if (node == dlm->node_num)
 827                         continue;
 828
 829                 tmpstat = dlm_send_one_join_cancel(dlm, node);
 830                 if (tmpstat) {
 831                         mlog(ML_ERROR, "Error return %d cancelling join on "
 832                              "node %d\n", tmpstat, node);
 833                         if (!status)
 834                                 status = tmpstat;
 835                 }
 836         }
 837
 838         if (status)
 839                 mlog_errno(status);
 840         return status;
 841 }
 842
 843 static int dlm_request_join(struct dlm_ctxt *dlm,
 844                             int node,
 845                             enum dlm_query_join_response *response)
 846 {
 847         int status, retval;
 848         struct dlm_query_join_request join_msg;
 849
 850         mlog(0, "querying node %d\n", node);
 851
 852         memset(&join_msg, 0, sizeof(join_msg));
 853         join_msg.node_idx = dlm->node_num;
 854         join_msg.name_len = strlen(dlm->name);
 855         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
 856
 857         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
 858                                     sizeof(join_msg), node, &retval);
 859         if (status < 0 && status != -ENOPROTOOPT) {
 860                 mlog_errno(status);
 861                 goto bail;
 862         }
 863
 864         /* -ENOPROTOOPT from the net code means the other side isn't
 865             listening for our message type -- that's fine, it means
 866             his dlm isn't up, so we can consider him a 'yes' but not
 867             joined into the domain.  */
 868         if (status == -ENOPROTOOPT) {
 869                 status = 0;
 870                 *response = JOIN_OK_NO_MAP;
 871         } else if (retval == JOIN_DISALLOW ||
 872                    retval == JOIN_OK ||
 873                    retval == JOIN_OK_NO_MAP) {
 874                 *response = retval;
 875         } else {
 876                 status = -EINVAL;
 877                 mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
 878                      node);
 879         }
 880
 881         mlog(0, "status %d, node %d response is %d\n", status, node,
 882                   *response);
 883
 884 bail:
 885         return status;
 886 }
 887
 888 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
 889                                     unsigned int node)
 890 {
 891         int status;
 892         struct dlm_assert_joined assert_msg;
 893
 894         mlog(0, "Sending join assert to node %u\n", node);
 895
 896         memset(&assert_msg, 0, sizeof(assert_msg));
 897         assert_msg.node_idx = dlm->node_num;
 898         assert_msg.name_len = strlen(dlm->name);
 899         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
 900
 901         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
 902                                     &assert_msg, sizeof(assert_msg), node,
 903                                     NULL);
 904         if (status < 0)
 905                 mlog_errno(status);
 906
 907         return status;
 908 }
 909
 910 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
 911                                   unsigned long *node_map)
 912 {
 913         int status, node, live;
 914
 915         status = 0;
 916         node = -1;
 917         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 918                                      node + 1)) < O2NM_MAX_NODES) {
 919                 if (node == dlm->node_num)
 920                         continue;
 921
 922                 do {
 923                         /* It is very important that this message be
 924                          * received so we spin until either the node
 925                          * has died or it gets the message. */
 926                         status = dlm_send_one_join_assert(dlm, node);
 927
 928                         spin_lock(&dlm->spinlock);
 929                         live = test_bit(node, dlm->live_nodes_map);
 930                         spin_unlock(&dlm->spinlock);
 931
 932                         if (status) {
 933                                 mlog(ML_ERROR, "Error return %d asserting "
 934                                      "join on node %d\n", status, node);
 935
 936                                 /* give us some time between errors... */
 937                                 if (live)
 938                                         msleep(DLM_DOMAIN_BACKOFF_MS);
 939                         }
 940                 } while (status && live);
 941         }
 942 }
 943
 944 struct domain_join_ctxt {
 945         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 946         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 947 };
 948
 949 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
 950                                    struct domain_join_ctxt *ctxt,
 951                                    enum dlm_query_join_response response)
 952 {
 953         int ret;
 954
 955         if (response == JOIN_DISALLOW) {
 956                 mlog(0, "Latest response of disallow -- should restart\n");
 957                 return 1;
 958         }
 959
 960         spin_lock(&dlm->spinlock);
 961         /* For now, we restart the process if the node maps have
 962          * changed at all */
 963         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
 964                      sizeof(dlm->live_nodes_map));
 965         spin_unlock(&dlm->spinlock);
 966
 967         if (ret)
 968                 mlog(0, "Node maps changed -- should restart\n");
 969
 970         return ret;
 971 }
 972
 973 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
 974 {
 975         int status = 0, tmpstat, node;
 976         struct domain_join_ctxt *ctxt;
 977         enum dlm_query_join_response response;
 978
 979         mlog_entry("%p", dlm);
 980
 981         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 982         if (!ctxt) {
 983                 status = -ENOMEM;
 984                 mlog_errno(status);
 985                 goto bail;
 986         }
 987
 988         /* group sem locking should work for us here -- we're already
 989          * registered for heartbeat events so filling this should be
 990          * atomic wrt getting those handlers called. */
 991         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
 992
 993         spin_lock(&dlm->spinlock);
 994         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
 995
 996         __dlm_set_joining_node(dlm, dlm->node_num);
 997
 998         spin_unlock(&dlm->spinlock);
 999
1000         node = -1;
1001         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1002                                      node + 1)) < O2NM_MAX_NODES) {
1003                 if (node == dlm->node_num)
1004                         continue;
1005
1006                 status = dlm_request_join(dlm, node, &response);
1007                 if (status < 0) {
1008                         mlog_errno(status);
1009                         goto bail;
1010                 }
1011
1012                 /* Ok, either we got a response or the node doesn't have a
1013                  * dlm up. */
1014                 if (response == JOIN_OK)
1015                         set_bit(node, ctxt->yes_resp_map);
1016
1017                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1018                         status = -EAGAIN;
1019                         goto bail;
1020                 }
1021         }
1022
1023         mlog(0, "Yay, done querying nodes!\n");
1024
1025         /* Yay, everyone agree's we can join the domain. My domain is
1026          * comprised of all nodes who were put in the
1027          * yes_resp_map. Copy that into our domain map and send a join
1028          * assert message to clean up everyone elses state. */
1029         spin_lock(&dlm->spinlock);
1030         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1031                sizeof(ctxt->yes_resp_map));
1032         set_bit(dlm->node_num, dlm->domain_map);
1033         spin_unlock(&dlm->spinlock);
1034
1035         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1036
1037         /* Joined state *must* be set before the joining node
1038          * information, otherwise the query_join handler may read no
1039          * current joiner but a state of NEW and tell joining nodes
1040          * we're not in the domain. */
1041         spin_lock(&dlm_domain_lock);
1042         dlm->dlm_state = DLM_CTXT_JOINED;
1043         dlm->num_joins++;
1044         spin_unlock(&dlm_domain_lock);
1045
1046 bail:
1047         spin_lock(&dlm->spinlock);
1048         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1049         if (!status)
1050                 __dlm_print_nodes(dlm);
1051         spin_unlock(&dlm->spinlock);
1052
1053         if (ctxt) {
1054                 /* Do we need to send a cancel message to any nodes? */
1055                 if (status < 0) {
1056                         tmpstat = dlm_send_join_cancels(dlm,
1057                                                         ctxt->yes_resp_map,
1058                                                         sizeof(ctxt->yes_resp_map));
1059                         if (tmpstat < 0)
1060                                 mlog_errno(tmpstat);
1061                 }
1062                 kfree(ctxt);
1063         }
1064
1065         mlog(0, "returning %d\n", status);
1066         return status;
1067 }
1068
1069 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1070 {
1071         o2hb_unregister_callback(&dlm->dlm_hb_up);
1072         o2hb_unregister_callback(&dlm->dlm_hb_down);
1073         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1074 }
1075
1076 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1077 {
1078         int status;
1079
1080         mlog(0, "registering handlers.\n");
1081
1082         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1083                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1084         status = o2hb_register_callback(&dlm->dlm_hb_down);
1085         if (status)
1086                 goto bail;
1087
1088         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1089                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1090         status = o2hb_register_callback(&dlm->dlm_hb_up);
1091         if (status)
1092                 goto bail;
1093
1094         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1095                                         sizeof(struct dlm_master_request),
1096                                         dlm_master_request_handler,
1097                                         dlm, NULL, &dlm->dlm_domain_handlers);
1098         if (status)
1099                 goto bail;
1100
1101         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1102                                         sizeof(struct dlm_assert_master),
1103                                         dlm_assert_master_handler,
1104                                         dlm, dlm_assert_master_post_handler,
1105                                         &dlm->dlm_domain_handlers);
1106         if (status)
1107                 goto bail;
1108
1109         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1110                                         sizeof(struct dlm_create_lock),
1111                                         dlm_create_lock_handler,
1112                                         dlm, NULL, &dlm->dlm_domain_handlers);
1113         if (status)
1114                 goto bail;
1115
1116         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1117                                         DLM_CONVERT_LOCK_MAX_LEN,
1118                                         dlm_convert_lock_handler,
1119                                         dlm, NULL, &dlm->dlm_domain_handlers);
1120         if (status)
1121                 goto bail;
1122
1123         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1124                                         DLM_UNLOCK_LOCK_MAX_LEN,
1125                                         dlm_unlock_lock_handler,
1126                                         dlm, NULL, &dlm->dlm_domain_handlers);
1127         if (status)
1128                 goto bail;
1129
1130         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1131                                         DLM_PROXY_AST_MAX_LEN,
1132                                         dlm_proxy_ast_handler,
1133                                         dlm, NULL, &dlm->dlm_domain_handlers);
1134         if (status)
1135                 goto bail;
1136
1137         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1138                                         sizeof(struct dlm_exit_domain),
1139                                         dlm_exit_domain_handler,
1140                                         dlm, NULL, &dlm->dlm_domain_handlers);
1141         if (status)
1142                 goto bail;
1143
1144         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1145                                         sizeof(struct dlm_deref_lockres),
1146                                         dlm_deref_lockres_handler,
1147                                         dlm, NULL, &dlm->dlm_domain_handlers);
1148         if (status)
1149                 goto bail;
1150
1151         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1152                                         sizeof(struct dlm_migrate_request),
1153                                         dlm_migrate_request_handler,
1154                                         dlm, NULL, &dlm->dlm_domain_handlers);
1155         if (status)
1156                 goto bail;
1157
1158         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1159                                         DLM_MIG_LOCKRES_MAX_LEN,
1160                                         dlm_mig_lockres_handler,
1161                                         dlm, NULL, &dlm->dlm_domain_handlers);
1162         if (status)
1163                 goto bail;
1164
1165         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1166                                         sizeof(struct dlm_master_requery),
1167                                         dlm_master_requery_handler,
1168                                         dlm, NULL, &dlm->dlm_domain_handlers);
1169         if (status)
1170                 goto bail;
1171
1172         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1173                                         sizeof(struct dlm_lock_request),
1174                                         dlm_request_all_locks_handler,
1175                                         dlm, NULL, &dlm->dlm_domain_handlers);
1176         if (status)
1177                 goto bail;
1178
1179         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1180                                         sizeof(struct dlm_reco_data_done),
1181                                         dlm_reco_data_done_handler,
1182                                         dlm, NULL, &dlm->dlm_domain_handlers);
1183         if (status)
1184                 goto bail;
1185
1186         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1187                                         sizeof(struct dlm_begin_reco),
1188                                         dlm_begin_reco_handler,
1189                                         dlm, NULL, &dlm->dlm_domain_handlers);
1190         if (status)
1191                 goto bail;
1192
1193         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1194                                         sizeof(struct dlm_finalize_reco),
1195                                         dlm_finalize_reco_handler,
1196                                         dlm, NULL, &dlm->dlm_domain_handlers);
1197         if (status)
1198                 goto bail;
1199
1200 bail:
1201         if (status)
1202                 dlm_unregister_domain_handlers(dlm);
1203
1204         return status;
1205 }
1206
1207 static int dlm_join_domain(struct dlm_ctxt *dlm)
1208 {
1209         int status;
1210
1211         BUG_ON(!dlm);
1212
1213         mlog(0, "Join domain %s\n", dlm->name);
1214
1215         status = dlm_register_domain_handlers(dlm);
1216         if (status) {
1217                 mlog_errno(status);
1218                 goto bail;
1219         }
1220
1221         status = dlm_launch_thread(dlm);
1222         if (status < 0) {
1223                 mlog_errno(status);
1224                 goto bail;
1225         }
1226
1227         status = dlm_launch_recovery_thread(dlm);
1228         if (status < 0) {
1229                 mlog_errno(status);
1230                 goto bail;
1231         }
1232
1233         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1234         if (!dlm->dlm_worker) {
1235                 status = -ENOMEM;
1236                 mlog_errno(status);
1237                 goto bail;
1238         }
1239
1240         do {
1241                 unsigned int backoff;
1242                 status = dlm_try_to_join_domain(dlm);
1243
1244                 /* If we're racing another node to the join, then we
1245                  * need to back off temporarily and let them
1246                  * complete. */
1247                 if (status == -EAGAIN) {
1248                         if (signal_pending(current)) {
1249                                 status = -ERESTARTSYS;
1250                                 goto bail;
1251                         }
1252
1253                         /*
1254                          * <chip> After you!
1255                          * <dale> No, after you!
1256                          * <chip> I insist!
1257                          * <dale> But you first!
1258                          * ...
1259                          */
1260                         backoff = (unsigned int)(jiffies & 0x3);
1261                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1262                         mlog(0, "backoff %d\n", backoff);
1263                         msleep(backoff);
1264                 }
1265         } while (status == -EAGAIN);
1266
1267         if (status < 0) {
1268                 mlog_errno(status);
1269                 goto bail;
1270         }
1271
1272         status = 0;
1273 bail:
1274         wake_up(&dlm_domain_events);
1275
1276         if (status) {
1277                 dlm_unregister_domain_handlers(dlm);
1278                 dlm_complete_thread(dlm);
1279                 dlm_complete_recovery_thread(dlm);
1280                 dlm_destroy_dlm_worker(dlm);
1281         }
1282
1283         return status;
1284 }
1285
1286 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1287                                 u32 key)
1288 {
1289         int i;
1290         struct dlm_ctxt *dlm = NULL;
1291
1292         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1293         if (!dlm) {
1294                 mlog_errno(-ENOMEM);
1295                 goto leave;
1296         }
1297
1298         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1299         if (dlm->name == NULL) {
1300                 mlog_errno(-ENOMEM);
1301                 kfree(dlm);
1302                 dlm = NULL;
1303                 goto leave;
1304         }
1305
1306         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1307         if (!dlm->lockres_hash) {
1308                 mlog_errno(-ENOMEM);
1309                 kfree(dlm->name);
1310                 kfree(dlm);
1311                 dlm = NULL;
1312                 goto leave;
1313         }
1314
1315         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1316                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1317
1318         strcpy(dlm->name, domain);
1319         dlm->key = key;
1320         dlm->node_num = o2nm_this_node();
1321
1322         spin_lock_init(&dlm->spinlock);
1323         spin_lock_init(&dlm->master_lock);
1324         spin_lock_init(&dlm->ast_lock);
1325         INIT_LIST_HEAD(&dlm->list);
1326         INIT_LIST_HEAD(&dlm->dirty_list);
1327         INIT_LIST_HEAD(&dlm->reco.resources);
1328         INIT_LIST_HEAD(&dlm->reco.received);
1329         INIT_LIST_HEAD(&dlm->reco.node_data);
1330         INIT_LIST_HEAD(&dlm->purge_list);
1331         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1332         dlm->reco.state = 0;
1333
1334         INIT_LIST_HEAD(&dlm->pending_asts);
1335         INIT_LIST_HEAD(&dlm->pending_basts);
1336
1337         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1338                   dlm->recovery_map, &(dlm->recovery_map[0]));
1339
1340         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1341         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1342         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1343
1344         dlm->dlm_thread_task = NULL;
1345         dlm->dlm_reco_thread_task = NULL;
1346         dlm->dlm_worker = NULL;
1347         init_waitqueue_head(&dlm->dlm_thread_wq);
1348         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1349         init_waitqueue_head(&dlm->reco.event);
1350         init_waitqueue_head(&dlm->ast_wq);
1351         init_waitqueue_head(&dlm->migration_wq);
1352         INIT_LIST_HEAD(&dlm->master_list);
1353         INIT_LIST_HEAD(&dlm->mle_hb_events);
1354
1355         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1356         init_waitqueue_head(&dlm->dlm_join_events);
1357
1358         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1359         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1360         atomic_set(&dlm->local_resources, 0);
1361         atomic_set(&dlm->remote_resources, 0);
1362         atomic_set(&dlm->unknown_resources, 0);
1363
1364         spin_lock_init(&dlm->work_lock);
1365         INIT_LIST_HEAD(&dlm->work_list);
1366         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1367
1368         kref_init(&dlm->dlm_refs);
1369         dlm->dlm_state = DLM_CTXT_NEW;
1370
1371         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1372
1373         mlog(0, "context init: refcount %u\n",
1374                   atomic_read(&dlm->dlm_refs.refcount));
1375
1376 leave:
1377         return dlm;
1378 }
1379
1380 /*
1381  * dlm_register_domain: one-time setup per "domain"
1382  */
1383 struct dlm_ctxt * dlm_register_domain(const char *domain,
1384                                u32 key)
1385 {
1386         int ret;
1387         struct dlm_ctxt *dlm = NULL;
1388         struct dlm_ctxt *new_ctxt = NULL;
1389
1390         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1391                 ret = -ENAMETOOLONG;
1392                 mlog(ML_ERROR, "domain name length too long\n");
1393                 goto leave;
1394         }
1395
1396         if (!o2hb_check_local_node_heartbeating()) {
1397                 mlog(ML_ERROR, "the local node has not been configured, or is "
1398                      "not heartbeating\n");
1399                 ret = -EPROTO;
1400                 goto leave;
1401         }
1402
1403         mlog(0, "register called for domain \"%s\"\n", domain);
1404
1405 retry:
1406         dlm = NULL;
1407         if (signal_pending(current)) {
1408                 ret = -ERESTARTSYS;
1409                 mlog_errno(ret);
1410                 goto leave;
1411         }
1412
1413         spin_lock(&dlm_domain_lock);
1414
1415         dlm = __dlm_lookup_domain(domain);
1416         if (dlm) {
1417                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1418                         spin_unlock(&dlm_domain_lock);
1419
1420                         mlog(0, "This ctxt is not joined yet!\n");
1421                         wait_event_interruptible(dlm_domain_events,
1422                                                  dlm_wait_on_domain_helper(
1423                                                          domain));
1424                         goto retry;
1425                 }
1426
1427                 __dlm_get(dlm);
1428                 dlm->num_joins++;
1429
1430                 spin_unlock(&dlm_domain_lock);
1431
1432                 ret = 0;
1433                 goto leave;
1434         }
1435
1436         /* doesn't exist */
1437         if (!new_ctxt) {
1438                 spin_unlock(&dlm_domain_lock);
1439
1440                 new_ctxt = dlm_alloc_ctxt(domain, key);
1441                 if (new_ctxt)
1442                         goto retry;
1443
1444                 ret = -ENOMEM;
1445                 mlog_errno(ret);
1446                 goto leave;
1447         }
1448
1449         /* a little variable switch-a-roo here... */
1450         dlm = new_ctxt;
1451         new_ctxt = NULL;
1452
1453         /* add the new domain */
1454         list_add_tail(&dlm->list, &dlm_domains);
1455         spin_unlock(&dlm_domain_lock);
1456
1457         ret = dlm_join_domain(dlm);
1458         if (ret) {
1459                 mlog_errno(ret);
1460                 dlm_put(dlm);
1461                 goto leave;
1462         }
1463
1464         ret = 0;
1465 leave:
1466         if (new_ctxt)
1467                 dlm_free_ctxt_mem(new_ctxt);
1468
1469         if (ret < 0)
1470                 dlm = ERR_PTR(ret);
1471
1472         return dlm;
1473 }
1474 EXPORT_SYMBOL_GPL(dlm_register_domain);
1475
1476 static LIST_HEAD(dlm_join_handlers);
1477
1478 static void dlm_unregister_net_handlers(void)
1479 {
1480         o2net_unregister_handler_list(&dlm_join_handlers);
1481 }
1482
1483 static int dlm_register_net_handlers(void)
1484 {
1485         int status = 0;
1486
1487         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1488                                         sizeof(struct dlm_query_join_request),
1489                                         dlm_query_join_handler,
1490                                         NULL, NULL, &dlm_join_handlers);
1491         if (status)
1492                 goto bail;
1493
1494         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1495                                         sizeof(struct dlm_assert_joined),
1496                                         dlm_assert_joined_handler,
1497                                         NULL, NULL, &dlm_join_handlers);
1498         if (status)
1499                 goto bail;
1500
1501         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1502                                         sizeof(struct dlm_cancel_join),
1503                                         dlm_cancel_join_handler,
1504                                         NULL, NULL, &dlm_join_handlers);
1505
1506 bail:
1507         if (status < 0)
1508                 dlm_unregister_net_handlers();
1509
1510         return status;
1511 }
1512
1513 /* Domain eviction callback handling.
1514  *
1515  * The file system requires notification of node death *before* the
1516  * dlm completes it's recovery work, otherwise it may be able to
1517  * acquire locks on resources requiring recovery. Since the dlm can
1518  * evict a node from it's domain *before* heartbeat fires, a similar
1519  * mechanism is required. */
1520
1521 /* Eviction is not expected to happen often, so a per-domain lock is
1522  * not necessary. Eviction callbacks are allowed to sleep for short
1523  * periods of time. */
1524 static DECLARE_RWSEM(dlm_callback_sem);
1525
1526 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1527                                         int node_num)
1528 {
1529         struct list_head *iter;
1530         struct dlm_eviction_cb *cb;
1531
1532         down_read(&dlm_callback_sem);
1533         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1534                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1535
1536                 cb->ec_func(node_num, cb->ec_data);
1537         }
1538         up_read(&dlm_callback_sem);
1539 }
1540
1541 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1542                            dlm_eviction_func *f,
1543                            void *data)
1544 {
1545         INIT_LIST_HEAD(&cb->ec_item);
1546         cb->ec_func = f;
1547         cb->ec_data = data;
1548 }
1549 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1550
1551 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1552                               struct dlm_eviction_cb *cb)
1553 {
1554         down_write(&dlm_callback_sem);
1555         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1556         up_write(&dlm_callback_sem);
1557 }
1558 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1559
1560 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1561 {
1562         down_write(&dlm_callback_sem);
1563         list_del_init(&cb->ec_item);
1564         up_write(&dlm_callback_sem);
1565 }
1566 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1567
1568 static int __init dlm_init(void)
1569 {
1570         int status;
1571
1572         dlm_print_version();
1573
1574         status = dlm_init_mle_cache();
1575         if (status)
1576                 return -1;
1577
1578         status = dlm_register_net_handlers();
1579         if (status) {
1580                 dlm_destroy_mle_cache();
1581                 return -1;
1582         }
1583
1584         return 0;
1585 }
1586
1587 static void __exit dlm_exit (void)
1588 {
1589         dlm_unregister_net_handlers();
1590         dlm_destroy_mle_cache();
1591 }
1592
1593 MODULE_AUTHOR("Oracle");
1594 MODULE_LICENSE("GPL");
1595
1596 module_init(dlm_init);
1597 module_exit(dlm_exit);