kernel/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/init.h>
  32 #include <linux/spinlock.h>
  33 #include <linux/delay.h>
  34 #include <linux/err.h>
  35 #include <linux/debugfs.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 #include "dlmdomain.h"
  44 #include "dlmdebug.h"
  45
  46 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
  47 #include "cluster/masklog.h"
  48
  49 /*
  50  * ocfs2 node maps are array of long int, which limits to send them freely
  51  * across the wire due to endianness issues. To workaround this, we convert
  52  * long ints to byte arrays. Following 3 routines are helper functions to
  53  * set/test/copy bits within those array of bytes
  54  */
  55 static inline void byte_set_bit(u8 nr, u8 map[])
  56 {
  57         map[nr >> 3] |= (1UL << (nr & 7));
  58 }
  59
  60 static inline int byte_test_bit(u8 nr, u8 map[])
  61 {
  62         return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
  63 }
  64
  65 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
  66                         unsigned int sz)
  67 {
  68         unsigned int nn;
  69
  70         if (!sz)
  71                 return;
  72
  73         memset(dmap, 0, ((sz + 7) >> 3));
  74         for (nn = 0 ; nn < sz; nn++)
  75                 if (test_bit(nn, smap))
  76                         byte_set_bit(nn, dmap);
  77 }
  78
  79 static void dlm_free_pagevec(void **vec, int pages)
  80 {
  81         while (pages--)
  82                 free_page((unsigned long)vec[pages]);
  83         kfree(vec);
  84 }
  85
  86 static void **dlm_alloc_pagevec(int pages)
  87 {
  88         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
  89         int i;
  90
  91         if (!vec)
  92                 return NULL;
  93
  94         for (i = 0; i < pages; i++)
  95                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
  96                         goto out_free;
  97
  98         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
  99              pages, (unsigned long)DLM_HASH_PAGES,
 100              (unsigned long)DLM_BUCKETS_PER_PAGE);
 101         return vec;
 102 out_free:
 103         dlm_free_pagevec(vec, i);
 104         return NULL;
 105 }
 106
 107 /*
 108  *
 109  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
 110  *    dlm_domain_lock
 111  *    struct dlm_ctxt->spinlock
 112  *    struct dlm_lock_resource->spinlock
 113  *    struct dlm_ctxt->master_lock
 114  *    struct dlm_ctxt->ast_lock
 115  *    dlm_master_list_entry->spinlock
 116  *    dlm_lock->spinlock
 117  *
 118  */
 119
 120 DEFINE_SPINLOCK(dlm_domain_lock);
 121 LIST_HEAD(dlm_domains);
 122 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
 123
 124 /*
 125  * The supported protocol version for DLM communication.  Running domains
 126  * will have a negotiated version with the same major number and a minor
 127  * number equal or smaller.  The dlm_ctxt->dlm_locking_proto field should
 128  * be used to determine what a running domain is actually using.
 129  *
 130  * New in version 1.1:
 131  *      - Message DLM_QUERY_REGION added to support global heartbeat
 132  *      - Message DLM_QUERY_NODEINFO added to allow online node removes
 133  * New in version 1.2:
 134  *      - Message DLM_BEGIN_EXIT_DOMAIN_MSG added to mark start of exit domain
 135  */
 136 static const struct dlm_protocol_version dlm_protocol = {
 137         .pv_major = 1,
 138         .pv_minor = 2,
 139 };
 140
 141 #define DLM_DOMAIN_BACKOFF_MS 200
 142
 143 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 144                                   void **ret_data);
 145 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 146                                      void **ret_data);
 147 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 148                                    void **ret_data);
 149 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
 150                                     void *data, void **ret_data);
 151 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 152                                    void **ret_data);
 153 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
 154                                 struct dlm_protocol_version *request);
 155
 156 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 157
 158 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 159 {
 160         if (hlist_unhashed(&res->hash_node))
 161                 return;
 162
 163         mlog(0, "%s: Unhash res %.*s\n", dlm->name, res->lockname.len,
 164              res->lockname.name);
 165         hlist_del_init(&res->hash_node);
 166         dlm_lockres_put(res);
 167 }
 168
 169 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
 170 {
 171         struct hlist_head *bucket;
 172         struct qstr *q;
 173
 174         assert_spin_locked(&dlm->spinlock);
 175
 176         q = &res->lockname;
 177         bucket = dlm_lockres_hash(dlm, q->hash);
 178
 179         /* get a reference for our hashtable */
 180         dlm_lockres_get(res);
 181
 182         hlist_add_head(&res->hash_node, bucket);
 183
 184         mlog(0, "%s: Hash res %.*s\n", dlm->name, res->lockname.len,
 185              res->lockname.name);
 186 }
 187
 188 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 189                                                      const char *name,
 190                                                      unsigned int len,
 191                                                      unsigned int hash)
 192 {
 193         struct hlist_head *bucket;
 194         struct dlm_lock_resource *res;
 195
 196         mlog(0, "%.*s\n", len, name);
 197
 198         assert_spin_locked(&dlm->spinlock);
 199
 200         bucket = dlm_lockres_hash(dlm, hash);
 201
 202         hlist_for_each_entry(res, bucket, hash_node) {
 203                 if (res->lockname.name[0] != name[0])
 204                         continue;
 205                 if (unlikely(res->lockname.len != len))
 206                         continue;
 207                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 208                         continue;
 209                 dlm_lockres_get(res);
 210                 return res;
 211         }
 212         return NULL;
 213 }
 214
 215 /* intended to be called by functions which do not care about lock
 216  * resources which are being purged (most net _handler functions).
 217  * this will return NULL for any lock resource which is found but
 218  * currently in the process of dropping its mastery reference.
 219  * use __dlm_lookup_lockres_full when you need the lock resource
 220  * regardless (e.g. dlm_get_lock_resource) */
 221 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 222                                                 const char *name,
 223                                                 unsigned int len,
 224                                                 unsigned int hash)
 225 {
 226         struct dlm_lock_resource *res = NULL;
 227
 228         mlog(0, "%.*s\n", len, name);
 229
 230         assert_spin_locked(&dlm->spinlock);
 231
 232         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 233         if (res) {
 234                 spin_lock(&res->spinlock);
 235                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 236                         spin_unlock(&res->spinlock);
 237                         dlm_lockres_put(res);
 238                         return NULL;
 239                 }
 240                 spin_unlock(&res->spinlock);
 241         }
 242
 243         return res;
 244 }
 245
 246 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 247                                     const char *name,
 248                                     unsigned int len)
 249 {
 250         struct dlm_lock_resource *res;
 251         unsigned int hash = dlm_lockid_hash(name, len);
 252
 253         spin_lock(&dlm->spinlock);
 254         res = __dlm_lookup_lockres(dlm, name, len, hash);
 255         spin_unlock(&dlm->spinlock);
 256         return res;
 257 }
 258
 259 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 260 {
 261         struct dlm_ctxt *tmp;
 262
 263         assert_spin_locked(&dlm_domain_lock);
 264
 265         /* tmp->name here is always NULL terminated,
 266          * but domain may not be! */
 267         list_for_each_entry(tmp, &dlm_domains, list) {
 268                 if (strlen(tmp->name) == len &&
 269                     memcmp(tmp->name, domain, len)==0)
 270                         return tmp;
 271         }
 272
 273         return NULL;
 274 }
 275
 276 /* For null terminated domain strings ONLY */
 277 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 278 {
 279         assert_spin_locked(&dlm_domain_lock);
 280
 281         return __dlm_lookup_domain_full(domain, strlen(domain));
 282 }
 283
 284
 285 /* returns true on one of two conditions:
 286  * 1) the domain does not exist
 287  * 2) the domain exists and it's state is "joined" */
 288 static int dlm_wait_on_domain_helper(const char *domain)
 289 {
 290         int ret = 0;
 291         struct dlm_ctxt *tmp = NULL;
 292
 293         spin_lock(&dlm_domain_lock);
 294
 295         tmp = __dlm_lookup_domain(domain);
 296         if (!tmp)
 297                 ret = 1;
 298         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 299                 ret = 1;
 300
 301         spin_unlock(&dlm_domain_lock);
 302         return ret;
 303 }
 304
 305 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 306 {
 307         dlm_destroy_debugfs_subroot(dlm);
 308
 309         if (dlm->lockres_hash)
 310                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 311
 312         if (dlm->master_hash)
 313                 dlm_free_pagevec((void **)dlm->master_hash, DLM_HASH_PAGES);
 314
 315         kfree(dlm->name);
 316         kfree(dlm);
 317 }
 318
 319 /* A little strange - this function will be called while holding
 320  * dlm_domain_lock and is expected to be holding it on the way out. We
 321  * will however drop and reacquire it multiple times */
 322 static void dlm_ctxt_release(struct kref *kref)
 323 {
 324         struct dlm_ctxt *dlm;
 325
 326         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 327
 328         BUG_ON(dlm->num_joins);
 329         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 330
 331         /* we may still be in the list if we hit an error during join. */
 332         list_del_init(&dlm->list);
 333
 334         spin_unlock(&dlm_domain_lock);
 335
 336         mlog(0, "freeing memory from domain %s\n", dlm->name);
 337
 338         wake_up(&dlm_domain_events);
 339
 340         dlm_free_ctxt_mem(dlm);
 341
 342         spin_lock(&dlm_domain_lock);
 343 }
 344
 345 void dlm_put(struct dlm_ctxt *dlm)
 346 {
 347         spin_lock(&dlm_domain_lock);
 348         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 349         spin_unlock(&dlm_domain_lock);
 350 }
 351
 352 static void __dlm_get(struct dlm_ctxt *dlm)
 353 {
 354         kref_get(&dlm->dlm_refs);
 355 }
 356
 357 /* given a questionable reference to a dlm object, gets a reference if
 358  * it can find it in the list, otherwise returns NULL in which case
 359  * you shouldn't trust your pointer. */
 360 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 361 {
 362         struct dlm_ctxt *target;
 363         struct dlm_ctxt *ret = NULL;
 364
 365         spin_lock(&dlm_domain_lock);
 366
 367         list_for_each_entry(target, &dlm_domains, list) {
 368                 if (target == dlm) {
 369                         __dlm_get(target);
 370                         ret = target;
 371                         break;
 372                 }
 373         }
 374
 375         spin_unlock(&dlm_domain_lock);
 376
 377         return ret;
 378 }
 379
 380 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 381 {
 382         int ret;
 383
 384         spin_lock(&dlm_domain_lock);
 385         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 386                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 387         spin_unlock(&dlm_domain_lock);
 388
 389         return ret;
 390 }
 391
 392 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 393 {
 394         if (dlm->dlm_worker) {
 395                 flush_workqueue(dlm->dlm_worker);
 396                 destroy_workqueue(dlm->dlm_worker);
 397                 dlm->dlm_worker = NULL;
 398         }
 399 }
 400
 401 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 402 {
 403         dlm_unregister_domain_handlers(dlm);
 404         dlm_debug_shutdown(dlm);
 405         dlm_complete_thread(dlm);
 406         dlm_complete_recovery_thread(dlm);
 407         dlm_destroy_dlm_worker(dlm);
 408
 409         /* We've left the domain. Now we can take ourselves out of the
 410          * list and allow the kref stuff to help us free the
 411          * memory. */
 412         spin_lock(&dlm_domain_lock);
 413         list_del_init(&dlm->list);
 414         spin_unlock(&dlm_domain_lock);
 415
 416         /* Wake up anyone waiting for us to remove this domain */
 417         wake_up(&dlm_domain_events);
 418 }
 419
 420 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 421 {
 422         int i, num, n, ret = 0;
 423         struct dlm_lock_resource *res;
 424         struct hlist_node *iter;
 425         struct hlist_head *bucket;
 426         int dropped;
 427
 428         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 429
 430         num = 0;
 431         spin_lock(&dlm->spinlock);
 432         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 433 redo_bucket:
 434                 n = 0;
 435                 bucket = dlm_lockres_hash(dlm, i);
 436                 iter = bucket->first;
 437                 while (iter) {
 438                         n++;
 439                         res = hlist_entry(iter, struct dlm_lock_resource,
 440                                           hash_node);
 441                         dlm_lockres_get(res);
 442                         /* migrate, if necessary.  this will drop the dlm
 443                          * spinlock and retake it if it does migration. */
 444                         dropped = dlm_empty_lockres(dlm, res);
 445
 446                         spin_lock(&res->spinlock);
 447                         if (dropped)
 448                                 __dlm_lockres_calc_usage(dlm, res);
 449                         else
 450                                 iter = res->hash_node.next;
 451                         spin_unlock(&res->spinlock);
 452
 453                         dlm_lockres_put(res);
 454
 455                         if (dropped) {
 456                                 cond_resched_lock(&dlm->spinlock);
 457                                 goto redo_bucket;
 458                         }
 459                 }
 460                 cond_resched_lock(&dlm->spinlock);
 461                 num += n;
 462         }
 463         spin_unlock(&dlm->spinlock);
 464         wake_up(&dlm->dlm_thread_wq);
 465
 466         /* let the dlm thread take care of purging, keep scanning until
 467          * nothing remains in the hash */
 468         if (num) {
 469                 mlog(0, "%s: %d lock resources in hash last pass\n",
 470                      dlm->name, num);
 471                 ret = -EAGAIN;
 472         }
 473         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 474         return ret;
 475 }
 476
 477 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 478 {
 479         int ret;
 480
 481         spin_lock(&dlm->spinlock);
 482         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 483         spin_unlock(&dlm->spinlock);
 484
 485         return ret;
 486 }
 487
 488 static int dlm_begin_exit_domain_handler(struct o2net_msg *msg, u32 len,
 489                                          void *data, void **ret_data)
 490 {
 491         struct dlm_ctxt *dlm = data;
 492         unsigned int node;
 493         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 494
 495         if (!dlm_grab(dlm))
 496                 return 0;
 497
 498         node = exit_msg->node_idx;
 499         mlog(0, "%s: Node %u sent a begin exit domain message\n", dlm->name, node);
 500
 501         spin_lock(&dlm->spinlock);
 502         set_bit(node, dlm->exit_domain_map);
 503         spin_unlock(&dlm->spinlock);
 504
 505         dlm_put(dlm);
 506
 507         return 0;
 508 }
 509
 510 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 511 {
 512         /* Yikes, a double spinlock! I need domain_lock for the dlm
 513          * state and the dlm spinlock for join state... Sorry! */
 514 again:
 515         spin_lock(&dlm_domain_lock);
 516         spin_lock(&dlm->spinlock);
 517
 518         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 519                 mlog(0, "Node %d is joining, we wait on it.\n",
 520                           dlm->joining_node);
 521                 spin_unlock(&dlm->spinlock);
 522                 spin_unlock(&dlm_domain_lock);
 523
 524                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 525                 goto again;
 526         }
 527
 528         dlm->dlm_state = DLM_CTXT_LEAVING;
 529         spin_unlock(&dlm->spinlock);
 530         spin_unlock(&dlm_domain_lock);
 531 }
 532
 533 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 534 {
 535         int node = -1, num = 0;
 536
 537         assert_spin_locked(&dlm->spinlock);
 538
 539         printk("( ");
 540         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 541                                      node + 1)) < O2NM_MAX_NODES) {
 542                 printk("%d ", node);
 543                 ++num;
 544         }
 545         printk(") %u nodes\n", num);
 546 }
 547
 548 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 549                                    void **ret_data)
 550 {
 551         struct dlm_ctxt *dlm = data;
 552         unsigned int node;
 553         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 554
 555         mlog(0, "%p %u %p", msg, len, data);
 556
 557         if (!dlm_grab(dlm))
 558                 return 0;
 559
 560         node = exit_msg->node_idx;
 561
 562         spin_lock(&dlm->spinlock);
 563         clear_bit(node, dlm->domain_map);
 564         clear_bit(node, dlm->exit_domain_map);
 565         printk(KERN_NOTICE "o2dlm: Node %u leaves domain %s ", node, dlm->name);
 566         __dlm_print_nodes(dlm);
 567
 568         /* notify anything attached to the heartbeat events */
 569         dlm_hb_event_notify_attached(dlm, node, 0);
 570
 571         spin_unlock(&dlm->spinlock);
 572
 573         dlm_put(dlm);
 574
 575         return 0;
 576 }
 577
 578 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm, u32 msg_type,
 579                                     unsigned int node)
 580 {
 581         int status;
 582         struct dlm_exit_domain leave_msg;
 583
 584         mlog(0, "%s: Sending domain exit message %u to node %u\n", dlm->name,
 585              msg_type, node);
 586
 587         memset(&leave_msg, 0, sizeof(leave_msg));
 588         leave_msg.node_idx = dlm->node_num;
 589
 590         status = o2net_send_message(msg_type, dlm->key, &leave_msg,
 591                                     sizeof(leave_msg), node, NULL);
 592         if (status < 0)
 593                 mlog(ML_ERROR, "Error %d sending domain exit message %u "
 594                      "to node %u on domain %s\n", status, msg_type, node,
 595                      dlm->name);
 596
 597         return status;
 598 }
 599
 600 static void dlm_begin_exit_domain(struct dlm_ctxt *dlm)
 601 {
 602         int node = -1;
 603
 604         /* Support for begin exit domain was added in 1.2 */
 605         if (dlm->dlm_locking_proto.pv_major == 1 &&
 606             dlm->dlm_locking_proto.pv_minor < 2)
 607                 return;
 608
 609         /*
 610          * Unlike DLM_EXIT_DOMAIN_MSG, DLM_BEGIN_EXIT_DOMAIN_MSG is purely
 611          * informational. Meaning if a node does not receive the message,
 612          * so be it.
 613          */
 614         spin_lock(&dlm->spinlock);
 615         while (1) {
 616                 node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES, node + 1);
 617                 if (node >= O2NM_MAX_NODES)
 618                         break;
 619                 if (node == dlm->node_num)
 620                         continue;
 621
 622                 spin_unlock(&dlm->spinlock);
 623                 dlm_send_one_domain_exit(dlm, DLM_BEGIN_EXIT_DOMAIN_MSG, node);
 624                 spin_lock(&dlm->spinlock);
 625         }
 626         spin_unlock(&dlm->spinlock);
 627 }
 628
 629 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 630 {
 631         int node, clear_node, status;
 632
 633         /* At this point we've migrated away all our locks and won't
 634          * accept mastership of new ones. The dlm is responsible for
 635          * almost nothing now. We make sure not to confuse any joining
 636          * nodes and then commence shutdown procedure. */
 637
 638         spin_lock(&dlm->spinlock);
 639         /* Clear ourselves from the domain map */
 640         clear_bit(dlm->node_num, dlm->domain_map);
 641         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 642                                      0)) < O2NM_MAX_NODES) {
 643                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 644                  * -nodes cannot be added now as the
 645                  *   query_join_handlers knows to respond with OK_NO_MAP
 646                  * -we catch the right network errors if a node is
 647                  *   removed from the map while we're sending him the
 648                  *   exit message. */
 649                 spin_unlock(&dlm->spinlock);
 650
 651                 clear_node = 1;
 652
 653                 status = dlm_send_one_domain_exit(dlm, DLM_EXIT_DOMAIN_MSG,
 654                                                   node);
 655                 if (status < 0 &&
 656                     status != -ENOPROTOOPT &&
 657                     status != -ENOTCONN) {
 658                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 659                              "to node %d\n", status, node);
 660
 661                         /* Not sure what to do here but lets sleep for
 662                          * a bit in case this was a transient
 663                          * error... */
 664                         msleep(DLM_DOMAIN_BACKOFF_MS);
 665                         clear_node = 0;
 666                 }
 667
 668                 spin_lock(&dlm->spinlock);
 669                 /* If we're not clearing the node bit then we intend
 670                  * to loop back around to try again. */
 671                 if (clear_node)
 672                         clear_bit(node, dlm->domain_map);
 673         }
 674         spin_unlock(&dlm->spinlock);
 675 }
 676
 677 int dlm_shutting_down(struct dlm_ctxt *dlm)
 678 {
 679         int ret = 0;
 680
 681         spin_lock(&dlm_domain_lock);
 682
 683         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 684                 ret = 1;
 685
 686         spin_unlock(&dlm_domain_lock);
 687
 688         return ret;
 689 }
 690
 691 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 692 {
 693         int leave = 0;
 694         struct dlm_lock_resource *res;
 695
 696         spin_lock(&dlm_domain_lock);
 697         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 698         BUG_ON(!dlm->num_joins);
 699
 700         dlm->num_joins--;
 701         if (!dlm->num_joins) {
 702                 /* We mark it "in shutdown" now so new register
 703                  * requests wait until we've completely left the
 704                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 705                  * want new domain joins to communicate with us at
 706                  * least until we've completed migration of our
 707                  * resources. */
 708                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 709                 leave = 1;
 710         }
 711         spin_unlock(&dlm_domain_lock);
 712
 713         if (leave) {
 714                 mlog(0, "shutting down domain %s\n", dlm->name);
 715                 dlm_begin_exit_domain(dlm);
 716
 717                 /* We changed dlm state, notify the thread */
 718                 dlm_kick_thread(dlm, NULL);
 719
 720                 while (dlm_migrate_all_locks(dlm)) {
 721                         /* Give dlm_thread time to purge the lockres' */
 722                         msleep(500);
 723                         mlog(0, "%s: more migration to do\n", dlm->name);
 724                 }
 725
 726                 /* This list should be empty. If not, print remaining lockres */
 727                 if (!list_empty(&dlm->tracking_list)) {
 728                         mlog(ML_ERROR, "Following lockres' are still on the "
 729                              "tracking list:\n");
 730                         list_for_each_entry(res, &dlm->tracking_list, tracking)
 731                                 dlm_print_one_lock_resource(res);
 732                 }
 733
 734                 dlm_mark_domain_leaving(dlm);
 735                 dlm_leave_domain(dlm);
 736                 printk(KERN_NOTICE "o2dlm: Leaving domain %s\n", dlm->name);
 737                 dlm_force_free_mles(dlm);
 738                 dlm_complete_dlm_shutdown(dlm);
 739         }
 740         dlm_put(dlm);
 741 }
 742 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 743
 744 static int dlm_query_join_proto_check(char *proto_type, int node,
 745                                       struct dlm_protocol_version *ours,
 746                                       struct dlm_protocol_version *request)
 747 {
 748         int rc;
 749         struct dlm_protocol_version proto = *request;
 750
 751         if (!dlm_protocol_compare(ours, &proto)) {
 752                 mlog(0,
 753                      "node %u wanted to join with %s locking protocol "
 754                      "%u.%u, we respond with %u.%u\n",
 755                      node, proto_type,
 756                      request->pv_major,
 757                      request->pv_minor,
 758                      proto.pv_major, proto.pv_minor);
 759                 request->pv_minor = proto.pv_minor;
 760                 rc = 0;
 761         } else {
 762                 mlog(ML_NOTICE,
 763                      "Node %u wanted to join with %s locking "
 764                      "protocol %u.%u, but we have %u.%u, disallowing\n",
 765                      node, proto_type,
 766                      request->pv_major,
 767                      request->pv_minor,
 768                      ours->pv_major,
 769                      ours->pv_minor);
 770                 rc = 1;
 771         }
 772
 773         return rc;
 774 }
 775
 776 /*
 777  * struct dlm_query_join_packet is made up of four one-byte fields.  They
 778  * are effectively in big-endian order already.  However, little-endian
 779  * machines swap them before putting the packet on the wire (because
 780  * query_join's response is a status, and that status is treated as a u32
 781  * on the wire).  Thus, a big-endian and little-endian machines will treat
 782  * this structure differently.
 783  *
 784  * The solution is to have little-endian machines swap the structure when
 785  * converting from the structure to the u32 representation.  This will
 786  * result in the structure having the correct format on the wire no matter
 787  * the host endian format.
 788  */
 789 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
 790                                           u32 *wire)
 791 {
 792         union dlm_query_join_response response;
 793
 794         response.packet = *packet;
 795         *wire = be32_to_cpu(response.intval);
 796 }
 797
 798 static void dlm_query_join_wire_to_packet(u32 wire,
 799                                           struct dlm_query_join_packet *packet)
 800 {
 801         union dlm_query_join_response response;
 802
 803         response.intval = cpu_to_be32(wire);
 804         *packet = response.packet;
 805 }
 806
 807 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 808                                   void **ret_data)
 809 {
 810         struct dlm_query_join_request *query;
 811         struct dlm_query_join_packet packet = {
 812                 .code = JOIN_DISALLOW,
 813         };
 814         struct dlm_ctxt *dlm = NULL;
 815         u32 response;
 816         u8 nodenum;
 817
 818         query = (struct dlm_query_join_request *) msg->buf;
 819
 820         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 821                   query->domain);
 822
 823         /*
 824          * If heartbeat doesn't consider the node live, tell it
 825          * to back off and try again.  This gives heartbeat a chance
 826          * to catch up.
 827          */
 828         if (!o2hb_check_node_heartbeating_no_sem(query->node_idx)) {
 829                 mlog(0, "node %u is not in our live map yet\n",
 830                      query->node_idx);
 831
 832                 packet.code = JOIN_DISALLOW;
 833                 goto respond;
 834         }
 835
 836         packet.code = JOIN_OK_NO_MAP;
 837
 838         spin_lock(&dlm_domain_lock);
 839         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 840         if (!dlm)
 841                 goto unlock_respond;
 842
 843         /*
 844          * There is a small window where the joining node may not see the
 845          * node(s) that just left but still part of the cluster. DISALLOW
 846          * join request if joining node has different node map.
 847          */
 848         nodenum=0;
 849         while (nodenum < O2NM_MAX_NODES) {
 850                 if (test_bit(nodenum, dlm->domain_map)) {
 851                         if (!byte_test_bit(nodenum, query->node_map)) {
 852                                 mlog(0, "disallow join as node %u does not "
 853                                      "have node %u in its nodemap\n",
 854                                      query->node_idx, nodenum);
 855                                 packet.code = JOIN_DISALLOW;
 856                                 goto unlock_respond;
 857                         }
 858                 }
 859                 nodenum++;
 860         }
 861
 862         /* Once the dlm ctxt is marked as leaving then we don't want
 863          * to be put in someone's domain map.
 864          * Also, explicitly disallow joining at certain troublesome
 865          * times (ie. during recovery). */
 866         if (dlm->dlm_state != DLM_CTXT_LEAVING) {
 867                 int bit = query->node_idx;
 868                 spin_lock(&dlm->spinlock);
 869
 870                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 871                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 872                         /*If this is a brand new context and we
 873                          * haven't started our join process yet, then
 874                          * the other node won the race. */
 875                         packet.code = JOIN_OK_NO_MAP;
 876                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 877                         /* Disallow parallel joins. */
 878                         packet.code = JOIN_DISALLOW;
 879                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 880                         mlog(0, "node %u trying to join, but recovery "
 881                              "is ongoing.\n", bit);
 882                         packet.code = JOIN_DISALLOW;
 883                 } else if (test_bit(bit, dlm->recovery_map)) {
 884                         mlog(0, "node %u trying to join, but it "
 885                              "still needs recovery.\n", bit);
 886                         packet.code = JOIN_DISALLOW;
 887                 } else if (test_bit(bit, dlm->domain_map)) {
 888                         mlog(0, "node %u trying to join, but it "
 889                              "is still in the domain! needs recovery?\n",
 890                              bit);
 891                         packet.code = JOIN_DISALLOW;
 892                 } else {
 893                         /* Alright we're fully a part of this domain
 894                          * so we keep some state as to who's joining
 895                          * and indicate to him that needs to be fixed
 896                          * up. */
 897
 898                         /* Make sure we speak compatible locking protocols.  */
 899                         if (dlm_query_join_proto_check("DLM", bit,
 900                                                        &dlm->dlm_locking_proto,
 901                                                        &query->dlm_proto)) {
 902                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 903                         } else if (dlm_query_join_proto_check("fs", bit,
 904                                                               &dlm->fs_locking_proto,
 905                                                               &query->fs_proto)) {
 906                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 907                         } else {
 908                                 packet.dlm_minor = query->dlm_proto.pv_minor;
 909                                 packet.fs_minor = query->fs_proto.pv_minor;
 910                                 packet.code = JOIN_OK;
 911                                 __dlm_set_joining_node(dlm, query->node_idx);
 912                         }
 913                 }
 914
 915                 spin_unlock(&dlm->spinlock);
 916         }
 917 unlock_respond:
 918         spin_unlock(&dlm_domain_lock);
 919
 920 respond:
 921         mlog(0, "We respond with %u\n", packet.code);
 922
 923         dlm_query_join_packet_to_wire(&packet, &response);
 924         return response;
 925 }
 926
 927 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 928                                      void **ret_data)
 929 {
 930         struct dlm_assert_joined *assert;
 931         struct dlm_ctxt *dlm = NULL;
 932
 933         assert = (struct dlm_assert_joined *) msg->buf;
 934
 935         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 936                   assert->domain);
 937
 938         spin_lock(&dlm_domain_lock);
 939         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 940         /* XXX should we consider no dlm ctxt an error? */
 941         if (dlm) {
 942                 spin_lock(&dlm->spinlock);
 943
 944                 /* Alright, this node has officially joined our
 945                  * domain. Set him in the map and clean up our
 946                  * leftover join state. */
 947                 BUG_ON(dlm->joining_node != assert->node_idx);
 948
 949                 if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 950                         mlog(0, "dlm recovery is ongoing, disallow join\n");
 951                         spin_unlock(&dlm->spinlock);
 952                         spin_unlock(&dlm_domain_lock);
 953                         return -EAGAIN;
 954                 }
 955
 956                 set_bit(assert->node_idx, dlm->domain_map);
 957                 clear_bit(assert->node_idx, dlm->exit_domain_map);
 958                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 959
 960                 printk(KERN_NOTICE "o2dlm: Node %u joins domain %s ",
 961                        assert->node_idx, dlm->name);
 962                 __dlm_print_nodes(dlm);
 963
 964                 /* notify anything attached to the heartbeat events */
 965                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 966
 967                 spin_unlock(&dlm->spinlock);
 968         }
 969         spin_unlock(&dlm_domain_lock);
 970
 971         return 0;
 972 }
 973
 974 static int dlm_match_regions(struct dlm_ctxt *dlm,
 975                              struct dlm_query_region *qr,
 976                              char *local, int locallen)
 977 {
 978         char *remote = qr->qr_regions;
 979         char *l, *r;
 980         int localnr, i, j, foundit;
 981         int status = 0;
 982
 983         if (!o2hb_global_heartbeat_active()) {
 984                 if (qr->qr_numregions) {
 985                         mlog(ML_ERROR, "Domain %s: Joining node %d has global "
 986                              "heartbeat enabled but local node %d does not\n",
 987                              qr->qr_domain, qr->qr_node, dlm->node_num);
 988                         status = -EINVAL;
 989                 }
 990                 goto bail;
 991         }
 992
 993         if (o2hb_global_heartbeat_active() && !qr->qr_numregions) {
 994                 mlog(ML_ERROR, "Domain %s: Local node %d has global "
 995                      "heartbeat enabled but joining node %d does not\n",
 996                      qr->qr_domain, dlm->node_num, qr->qr_node);
 997                 status = -EINVAL;
 998                 goto bail;
 999         }
1000
1001         r = remote;
1002         for (i = 0; i < qr->qr_numregions; ++i) {
1003                 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, r);
1004                 r += O2HB_MAX_REGION_NAME_LEN;
1005         }
1006
1007         localnr = min(O2NM_MAX_REGIONS, locallen/O2HB_MAX_REGION_NAME_LEN);
1008         localnr = o2hb_get_all_regions(local, (u8)localnr);
1009
1010         /* compare local regions with remote */
1011         l = local;
1012         for (i = 0; i < localnr; ++i) {
1013                 foundit = 0;
1014                 r = remote;
1015                 for (j = 0; j <= qr->qr_numregions; ++j) {
1016                         if (!memcmp(l, r, O2HB_MAX_REGION_NAME_LEN)) {
1017                                 foundit = 1;
1018                                 break;
1019                         }
1020                         r += O2HB_MAX_REGION_NAME_LEN;
1021                 }
1022                 if (!foundit) {
1023                         status = -EINVAL;
1024                         mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1025                              "in local node %d but not in joining node %d\n",
1026                              qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, l,
1027                              dlm->node_num, qr->qr_node);
1028                         goto bail;
1029                 }
1030                 l += O2HB_MAX_REGION_NAME_LEN;
1031         }
1032
1033         /* compare remote with local regions */
1034         r = remote;
1035         for (i = 0; i < qr->qr_numregions; ++i) {
1036                 foundit = 0;
1037                 l = local;
1038                 for (j = 0; j < localnr; ++j) {
1039                         if (!memcmp(r, l, O2HB_MAX_REGION_NAME_LEN)) {
1040                                 foundit = 1;
1041                                 break;
1042                         }
1043                         l += O2HB_MAX_REGION_NAME_LEN;
1044                 }
1045                 if (!foundit) {
1046                         status = -EINVAL;
1047                         mlog(ML_ERROR, "Domain %s: Region '%.*s' registered "
1048                              "in joining node %d but not in local node %d\n",
1049                              qr->qr_domain, O2HB_MAX_REGION_NAME_LEN, r,
1050                              qr->qr_node, dlm->node_num);
1051                         goto bail;
1052                 }
1053                 r += O2HB_MAX_REGION_NAME_LEN;
1054         }
1055
1056 bail:
1057         return status;
1058 }
1059
1060 static int dlm_send_regions(struct dlm_ctxt *dlm, unsigned long *node_map)
1061 {
1062         struct dlm_query_region *qr = NULL;
1063         int status, ret = 0, i;
1064         char *p;
1065
1066         if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1067                 goto bail;
1068
1069         qr = kzalloc(sizeof(struct dlm_query_region), GFP_KERNEL);
1070         if (!qr) {
1071                 ret = -ENOMEM;
1072                 mlog_errno(ret);
1073                 goto bail;
1074         }
1075
1076         qr->qr_node = dlm->node_num;
1077         qr->qr_namelen = strlen(dlm->name);
1078         memcpy(qr->qr_domain, dlm->name, qr->qr_namelen);
1079         /* if local hb, the numregions will be zero */
1080         if (o2hb_global_heartbeat_active())
1081                 qr->qr_numregions = o2hb_get_all_regions(qr->qr_regions,
1082                                                          O2NM_MAX_REGIONS);
1083
1084         p = qr->qr_regions;
1085         for (i = 0; i < qr->qr_numregions; ++i, p += O2HB_MAX_REGION_NAME_LEN)
1086                 mlog(0, "Region %.*s\n", O2HB_MAX_REGION_NAME_LEN, p);
1087
1088         i = -1;
1089         while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1090                                   i + 1)) < O2NM_MAX_NODES) {
1091                 if (i == dlm->node_num)
1092                         continue;
1093
1094                 mlog(0, "Sending regions to node %d\n", i);
1095
1096                 ret = o2net_send_message(DLM_QUERY_REGION, DLM_MOD_KEY, qr,
1097                                          sizeof(struct dlm_query_region),
1098                                          i, &status);
1099                 if (ret >= 0)
1100                         ret = status;
1101                 if (ret) {
1102                         mlog(ML_ERROR, "Region mismatch %d, node %d\n",
1103                              ret, i);
1104                         break;
1105                 }
1106         }
1107
1108 bail:
1109         kfree(qr);
1110         return ret;
1111 }
1112
1113 static int dlm_query_region_handler(struct o2net_msg *msg, u32 len,
1114                                     void *data, void **ret_data)
1115 {
1116         struct dlm_query_region *qr;
1117         struct dlm_ctxt *dlm = NULL;
1118         char *local = NULL;
1119         int status = 0;
1120
1121         qr = (struct dlm_query_region *) msg->buf;
1122
1123         mlog(0, "Node %u queries hb regions on domain %s\n", qr->qr_node,
1124              qr->qr_domain);
1125
1126         /* buffer used in dlm_mast_regions() */
1127         local = kmalloc(sizeof(qr->qr_regions), GFP_KERNEL);
1128         if (!local)
1129                 return -ENOMEM;
1130
1131         status = -EINVAL;
1132
1133         spin_lock(&dlm_domain_lock);
1134         dlm = __dlm_lookup_domain_full(qr->qr_domain, qr->qr_namelen);
1135         if (!dlm) {
1136                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1137                      "before join domain\n", qr->qr_node, qr->qr_domain);
1138                 goto out_domain_lock;
1139         }
1140
1141         spin_lock(&dlm->spinlock);
1142         if (dlm->joining_node != qr->qr_node) {
1143                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1144                      "but joining node is %d\n", qr->qr_node, qr->qr_domain,
1145                      dlm->joining_node);
1146                 goto out_dlm_lock;
1147         }
1148
1149         /* Support for global heartbeat was added in 1.1 */
1150         if (dlm->dlm_locking_proto.pv_major == 1 &&
1151             dlm->dlm_locking_proto.pv_minor == 0) {
1152                 mlog(ML_ERROR, "Node %d queried hb regions on domain %s "
1153                      "but active dlm protocol is %d.%d\n", qr->qr_node,
1154                      qr->qr_domain, dlm->dlm_locking_proto.pv_major,
1155                      dlm->dlm_locking_proto.pv_minor);
1156                 goto out_dlm_lock;
1157         }
1158
1159         status = dlm_match_regions(dlm, qr, local, sizeof(qr->qr_regions));
1160
1161 out_dlm_lock:
1162         spin_unlock(&dlm->spinlock);
1163
1164 out_domain_lock:
1165         spin_unlock(&dlm_domain_lock);
1166
1167         kfree(local);
1168
1169         return status;
1170 }
1171
1172 static int dlm_match_nodes(struct dlm_ctxt *dlm, struct dlm_query_nodeinfo *qn)
1173 {
1174         struct o2nm_node *local;
1175         struct dlm_node_info *remote;
1176         int i, j;
1177         int status = 0;
1178
1179         for (j = 0; j < qn->qn_numnodes; ++j)
1180                 mlog(0, "Node %3d, %pI4:%u\n", qn->qn_nodes[j].ni_nodenum,
1181                      &(qn->qn_nodes[j].ni_ipv4_address),
1182                      ntohs(qn->qn_nodes[j].ni_ipv4_port));
1183
1184         for (i = 0; i < O2NM_MAX_NODES && !status; ++i) {
1185                 local = o2nm_get_node_by_num(i);
1186                 remote = NULL;
1187                 for (j = 0; j < qn->qn_numnodes; ++j) {
1188                         if (qn->qn_nodes[j].ni_nodenum == i) {
1189                                 remote = &(qn->qn_nodes[j]);
1190                                 break;
1191                         }
1192                 }
1193
1194                 if (!local && !remote)
1195                         continue;
1196
1197                 if ((local && !remote) || (!local && remote))
1198                         status = -EINVAL;
1199
1200                 if (!status &&
1201                     ((remote->ni_nodenum != local->nd_num) ||
1202                      (remote->ni_ipv4_port != local->nd_ipv4_port) ||
1203                      (remote->ni_ipv4_address != local->nd_ipv4_address)))
1204                         status = -EINVAL;
1205
1206                 if (status) {
1207                         if (remote && !local)
1208                                 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1209                                      "registered in joining node %d but not in "
1210                                      "local node %d\n", qn->qn_domain,
1211                                      remote->ni_nodenum,
1212                                      &(remote->ni_ipv4_address),
1213                                      ntohs(remote->ni_ipv4_port),
1214                                      qn->qn_nodenum, dlm->node_num);
1215                         if (local && !remote)
1216                                 mlog(ML_ERROR, "Domain %s: Node %d (%pI4:%u) "
1217                                      "registered in local node %d but not in "
1218                                      "joining node %d\n", qn->qn_domain,
1219                                      local->nd_num, &(local->nd_ipv4_address),
1220                                      ntohs(local->nd_ipv4_port),
1221                                      dlm->node_num, qn->qn_nodenum);
1222                         BUG_ON((!local && !remote));
1223                 }
1224
1225                 if (local)
1226                         o2nm_node_put(local);
1227         }
1228
1229         return status;
1230 }
1231
1232 static int dlm_send_nodeinfo(struct dlm_ctxt *dlm, unsigned long *node_map)
1233 {
1234         struct dlm_query_nodeinfo *qn = NULL;
1235         struct o2nm_node *node;
1236         int ret = 0, status, count, i;
1237
1238         if (find_next_bit(node_map, O2NM_MAX_NODES, 0) >= O2NM_MAX_NODES)
1239                 goto bail;
1240
1241         qn = kzalloc(sizeof(struct dlm_query_nodeinfo), GFP_KERNEL);
1242         if (!qn) {
1243                 ret = -ENOMEM;
1244                 mlog_errno(ret);
1245                 goto bail;
1246         }
1247
1248         for (i = 0, count = 0; i < O2NM_MAX_NODES; ++i) {
1249                 node = o2nm_get_node_by_num(i);
1250                 if (!node)
1251                         continue;
1252                 qn->qn_nodes[count].ni_nodenum = node->nd_num;
1253                 qn->qn_nodes[count].ni_ipv4_port = node->nd_ipv4_port;
1254                 qn->qn_nodes[count].ni_ipv4_address = node->nd_ipv4_address;
1255                 mlog(0, "Node %3d, %pI4:%u\n", node->nd_num,
1256                      &(node->nd_ipv4_address), ntohs(node->nd_ipv4_port));
1257                 ++count;
1258                 o2nm_node_put(node);
1259         }
1260
1261         qn->qn_nodenum = dlm->node_num;
1262         qn->qn_numnodes = count;
1263         qn->qn_namelen = strlen(dlm->name);
1264         memcpy(qn->qn_domain, dlm->name, qn->qn_namelen);
1265
1266         i = -1;
1267         while ((i = find_next_bit(node_map, O2NM_MAX_NODES,
1268                                   i + 1)) < O2NM_MAX_NODES) {
1269                 if (i == dlm->node_num)
1270                         continue;
1271
1272                 mlog(0, "Sending nodeinfo to node %d\n", i);
1273
1274                 ret = o2net_send_message(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
1275                                          qn, sizeof(struct dlm_query_nodeinfo),
1276                                          i, &status);
1277                 if (ret >= 0)
1278                         ret = status;
1279                 if (ret) {
1280                         mlog(ML_ERROR, "node mismatch %d, node %d\n", ret, i);
1281                         break;
1282                 }
1283         }
1284
1285 bail:
1286         kfree(qn);
1287         return ret;
1288 }
1289
1290 static int dlm_query_nodeinfo_handler(struct o2net_msg *msg, u32 len,
1291                                       void *data, void **ret_data)
1292 {
1293         struct dlm_query_nodeinfo *qn;
1294         struct dlm_ctxt *dlm = NULL;
1295         int locked = 0, status = -EINVAL;
1296
1297         qn = (struct dlm_query_nodeinfo *) msg->buf;
1298
1299         mlog(0, "Node %u queries nodes on domain %s\n", qn->qn_nodenum,
1300              qn->qn_domain);
1301
1302         spin_lock(&dlm_domain_lock);
1303         dlm = __dlm_lookup_domain_full(qn->qn_domain, qn->qn_namelen);
1304         if (!dlm) {
1305                 mlog(ML_ERROR, "Node %d queried nodes on domain %s before "
1306                      "join domain\n", qn->qn_nodenum, qn->qn_domain);
1307                 goto bail;
1308         }
1309
1310         spin_lock(&dlm->spinlock);
1311         locked = 1;
1312         if (dlm->joining_node != qn->qn_nodenum) {
1313                 mlog(ML_ERROR, "Node %d queried nodes on domain %s but "
1314                      "joining node is %d\n", qn->qn_nodenum, qn->qn_domain,
1315                      dlm->joining_node);
1316                 goto bail;
1317         }
1318
1319         /* Support for node query was added in 1.1 */
1320         if (dlm->dlm_locking_proto.pv_major == 1 &&
1321             dlm->dlm_locking_proto.pv_minor == 0) {
1322                 mlog(ML_ERROR, "Node %d queried nodes on domain %s "
1323                      "but active dlm protocol is %d.%d\n", qn->qn_nodenum,
1324                      qn->qn_domain, dlm->dlm_locking_proto.pv_major,
1325                      dlm->dlm_locking_proto.pv_minor);
1326                 goto bail;
1327         }
1328
1329         status = dlm_match_nodes(dlm, qn);
1330
1331 bail:
1332         if (locked)
1333                 spin_unlock(&dlm->spinlock);
1334         spin_unlock(&dlm_domain_lock);
1335
1336         return status;
1337 }
1338
1339 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
1340                                    void **ret_data)
1341 {
1342         struct dlm_cancel_join *cancel;
1343         struct dlm_ctxt *dlm = NULL;
1344
1345         cancel = (struct dlm_cancel_join *) msg->buf;
1346
1347         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
1348                   cancel->domain);
1349
1350         spin_lock(&dlm_domain_lock);
1351         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
1352
1353         if (dlm) {
1354                 spin_lock(&dlm->spinlock);
1355
1356                 /* Yikes, this guy wants to cancel his join. No
1357                  * problem, we simply cleanup our join state. */
1358                 BUG_ON(dlm->joining_node != cancel->node_idx);
1359                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1360
1361                 spin_unlock(&dlm->spinlock);
1362         }
1363         spin_unlock(&dlm_domain_lock);
1364
1365         return 0;
1366 }
1367
1368 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
1369                                     unsigned int node)
1370 {
1371         int status;
1372         struct dlm_cancel_join cancel_msg;
1373
1374         memset(&cancel_msg, 0, sizeof(cancel_msg));
1375         cancel_msg.node_idx = dlm->node_num;
1376         cancel_msg.name_len = strlen(dlm->name);
1377         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
1378
1379         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1380                                     &cancel_msg, sizeof(cancel_msg), node,
1381                                     NULL);
1382         if (status < 0) {
1383                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1384                      "node %u\n", status, DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1385                      node);
1386                 goto bail;
1387         }
1388
1389 bail:
1390         return status;
1391 }
1392
1393 /* map_size should be in bytes. */
1394 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
1395                                  unsigned long *node_map,
1396                                  unsigned int map_size)
1397 {
1398         int status, tmpstat;
1399         unsigned int node;
1400
1401         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
1402                          sizeof(unsigned long))) {
1403                 mlog(ML_ERROR,
1404                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
1405                      map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
1406                 return -EINVAL;
1407         }
1408
1409         status = 0;
1410         node = -1;
1411         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1412                                      node + 1)) < O2NM_MAX_NODES) {
1413                 if (node == dlm->node_num)
1414                         continue;
1415
1416                 tmpstat = dlm_send_one_join_cancel(dlm, node);
1417                 if (tmpstat) {
1418                         mlog(ML_ERROR, "Error return %d cancelling join on "
1419                              "node %d\n", tmpstat, node);
1420                         if (!status)
1421                                 status = tmpstat;
1422                 }
1423         }
1424
1425         if (status)
1426                 mlog_errno(status);
1427         return status;
1428 }
1429
1430 static int dlm_request_join(struct dlm_ctxt *dlm,
1431                             int node,
1432                             enum dlm_query_join_response_code *response)
1433 {
1434         int status;
1435         struct dlm_query_join_request join_msg;
1436         struct dlm_query_join_packet packet;
1437         u32 join_resp;
1438
1439         mlog(0, "querying node %d\n", node);
1440
1441         memset(&join_msg, 0, sizeof(join_msg));
1442         join_msg.node_idx = dlm->node_num;
1443         join_msg.name_len = strlen(dlm->name);
1444         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1445         join_msg.dlm_proto = dlm->dlm_locking_proto;
1446         join_msg.fs_proto = dlm->fs_locking_proto;
1447
1448         /* copy live node map to join message */
1449         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1450
1451         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1452                                     sizeof(join_msg), node, &join_resp);
1453         if (status < 0 && status != -ENOPROTOOPT) {
1454                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1455                      "node %u\n", status, DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1456                      node);
1457                 goto bail;
1458         }
1459         dlm_query_join_wire_to_packet(join_resp, &packet);
1460
1461         /* -ENOPROTOOPT from the net code means the other side isn't
1462             listening for our message type -- that's fine, it means
1463             his dlm isn't up, so we can consider him a 'yes' but not
1464             joined into the domain.  */
1465         if (status == -ENOPROTOOPT) {
1466                 status = 0;
1467                 *response = JOIN_OK_NO_MAP;
1468         } else {
1469                 *response = packet.code;
1470                 switch (packet.code) {
1471                 case JOIN_DISALLOW:
1472                 case JOIN_OK_NO_MAP:
1473                         break;
1474                 case JOIN_PROTOCOL_MISMATCH:
1475                         mlog(ML_NOTICE,
1476                              "This node requested DLM locking protocol %u.%u and "
1477                              "filesystem locking protocol %u.%u.  At least one of "
1478                              "the protocol versions on node %d is not compatible, "
1479                              "disconnecting\n",
1480                              dlm->dlm_locking_proto.pv_major,
1481                              dlm->dlm_locking_proto.pv_minor,
1482                              dlm->fs_locking_proto.pv_major,
1483                              dlm->fs_locking_proto.pv_minor,
1484                              node);
1485                         status = -EPROTO;
1486                         break;
1487                 case JOIN_OK:
1488                         /* Use the same locking protocol as the remote node */
1489                         dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1490                         dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1491                         mlog(0,
1492                              "Node %d responds JOIN_OK with DLM locking protocol "
1493                              "%u.%u and fs locking protocol %u.%u\n",
1494                              node,
1495                              dlm->dlm_locking_proto.pv_major,
1496                              dlm->dlm_locking_proto.pv_minor,
1497                              dlm->fs_locking_proto.pv_major,
1498                              dlm->fs_locking_proto.pv_minor);
1499                         break;
1500                 default:
1501                         status = -EINVAL;
1502                         mlog(ML_ERROR, "invalid response %d from node %u\n",
1503                              packet.code, node);
1504                         /* Reset response to JOIN_DISALLOW */
1505                         *response = JOIN_DISALLOW;
1506                         break;
1507                 }
1508         }
1509
1510         mlog(0, "status %d, node %d response is %d\n", status, node,
1511              *response);
1512
1513 bail:
1514         return status;
1515 }
1516
1517 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1518                                     unsigned int node)
1519 {
1520         int status;
1521         int ret;
1522         struct dlm_assert_joined assert_msg;
1523
1524         mlog(0, "Sending join assert to node %u\n", node);
1525
1526         memset(&assert_msg, 0, sizeof(assert_msg));
1527         assert_msg.node_idx = dlm->node_num;
1528         assert_msg.name_len = strlen(dlm->name);
1529         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1530
1531         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1532                                     &assert_msg, sizeof(assert_msg), node,
1533                                     &ret);
1534         if (status < 0)
1535                 mlog(ML_ERROR, "Error %d when sending message %u (key 0x%x) to "
1536                      "node %u\n", status, DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1537                      node);
1538         else
1539                 status = ret;
1540
1541         return status;
1542 }
1543
1544 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1545                                   unsigned long *node_map)
1546 {
1547         int status, node, live;
1548
1549         status = 0;
1550         node = -1;
1551         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1552                                      node + 1)) < O2NM_MAX_NODES) {
1553                 if (node == dlm->node_num)
1554                         continue;
1555
1556                 do {
1557                         /* It is very important that this message be
1558                          * received so we spin until either the node
1559                          * has died or it gets the message. */
1560                         status = dlm_send_one_join_assert(dlm, node);
1561
1562                         spin_lock(&dlm->spinlock);
1563                         live = test_bit(node, dlm->live_nodes_map);
1564                         spin_unlock(&dlm->spinlock);
1565
1566                         if (status) {
1567                                 mlog(ML_ERROR, "Error return %d asserting "
1568                                      "join on node %d\n", status, node);
1569
1570                                 /* give us some time between errors... */
1571                                 if (live)
1572                                         msleep(DLM_DOMAIN_BACKOFF_MS);
1573                         }
1574                 } while (status && live);
1575         }
1576 }
1577
1578 struct domain_join_ctxt {
1579         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1580         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1581 };
1582
1583 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1584                                    struct domain_join_ctxt *ctxt,
1585                                    enum dlm_query_join_response_code response)
1586 {
1587         int ret;
1588
1589         if (response == JOIN_DISALLOW) {
1590                 mlog(0, "Latest response of disallow -- should restart\n");
1591                 return 1;
1592         }
1593
1594         spin_lock(&dlm->spinlock);
1595         /* For now, we restart the process if the node maps have
1596          * changed at all */
1597         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1598                      sizeof(dlm->live_nodes_map));
1599         spin_unlock(&dlm->spinlock);
1600
1601         if (ret)
1602                 mlog(0, "Node maps changed -- should restart\n");
1603
1604         return ret;
1605 }
1606
1607 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1608 {
1609         int status = 0, tmpstat, node;
1610         struct domain_join_ctxt *ctxt;
1611         enum dlm_query_join_response_code response = JOIN_DISALLOW;
1612
1613         mlog(0, "%p", dlm);
1614
1615         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1616         if (!ctxt) {
1617                 status = -ENOMEM;
1618                 mlog_errno(status);
1619                 goto bail;
1620         }
1621
1622         /* group sem locking should work for us here -- we're already
1623          * registered for heartbeat events so filling this should be
1624          * atomic wrt getting those handlers called. */
1625         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1626
1627         spin_lock(&dlm->spinlock);
1628         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1629
1630         __dlm_set_joining_node(dlm, dlm->node_num);
1631
1632         spin_unlock(&dlm->spinlock);
1633
1634         node = -1;
1635         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1636                                      node + 1)) < O2NM_MAX_NODES) {
1637                 if (node == dlm->node_num)
1638                         continue;
1639
1640                 status = dlm_request_join(dlm, node, &response);
1641                 if (status < 0) {
1642                         mlog_errno(status);
1643                         goto bail;
1644                 }
1645
1646                 /* Ok, either we got a response or the node doesn't have a
1647                  * dlm up. */
1648                 if (response == JOIN_OK)
1649                         set_bit(node, ctxt->yes_resp_map);
1650
1651                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1652                         status = -EAGAIN;
1653                         goto bail;
1654                 }
1655         }
1656
1657         mlog(0, "Yay, done querying nodes!\n");
1658
1659         /* Yay, everyone agree's we can join the domain. My domain is
1660          * comprised of all nodes who were put in the
1661          * yes_resp_map. Copy that into our domain map and send a join
1662          * assert message to clean up everyone elses state. */
1663         spin_lock(&dlm->spinlock);
1664         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1665                sizeof(ctxt->yes_resp_map));
1666         set_bit(dlm->node_num, dlm->domain_map);
1667         spin_unlock(&dlm->spinlock);
1668
1669         /* Support for global heartbeat and node info was added in 1.1 */
1670         if (dlm->dlm_locking_proto.pv_major > 1 ||
1671             dlm->dlm_locking_proto.pv_minor > 0) {
1672                 status = dlm_send_nodeinfo(dlm, ctxt->yes_resp_map);
1673                 if (status) {
1674                         mlog_errno(status);
1675                         goto bail;
1676                 }
1677                 status = dlm_send_regions(dlm, ctxt->yes_resp_map);
1678                 if (status) {
1679                         mlog_errno(status);
1680                         goto bail;
1681                 }
1682         }
1683
1684         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1685
1686         /* Joined state *must* be set before the joining node
1687          * information, otherwise the query_join handler may read no
1688          * current joiner but a state of NEW and tell joining nodes
1689          * we're not in the domain. */
1690         spin_lock(&dlm_domain_lock);
1691         dlm->dlm_state = DLM_CTXT_JOINED;
1692         dlm->num_joins++;
1693         spin_unlock(&dlm_domain_lock);
1694
1695 bail:
1696         spin_lock(&dlm->spinlock);
1697         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1698         if (!status) {
1699                 printk(KERN_NOTICE "o2dlm: Joining domain %s ", dlm->name);
1700                 __dlm_print_nodes(dlm);
1701         }
1702         spin_unlock(&dlm->spinlock);
1703
1704         if (ctxt) {
1705                 /* Do we need to send a cancel message to any nodes? */
1706                 if (status < 0) {
1707                         tmpstat = dlm_send_join_cancels(dlm,
1708                                                         ctxt->yes_resp_map,
1709                                                         sizeof(ctxt->yes_resp_map));
1710                         if (tmpstat < 0)
1711                                 mlog_errno(tmpstat);
1712                 }
1713                 kfree(ctxt);
1714         }
1715
1716         mlog(0, "returning %d\n", status);
1717         return status;
1718 }
1719
1720 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1721 {
1722         o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_up);
1723         o2hb_unregister_callback(dlm->name, &dlm->dlm_hb_down);
1724         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1725 }
1726
1727 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1728 {
1729         int status;
1730
1731         mlog(0, "registering handlers.\n");
1732
1733         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1734                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1735         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1736                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1737
1738         status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_down);
1739         if (status)
1740                 goto bail;
1741
1742         status = o2hb_register_callback(dlm->name, &dlm->dlm_hb_up);
1743         if (status)
1744                 goto bail;
1745
1746         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1747                                         sizeof(struct dlm_master_request),
1748                                         dlm_master_request_handler,
1749                                         dlm, NULL, &dlm->dlm_domain_handlers);
1750         if (status)
1751                 goto bail;
1752
1753         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1754                                         sizeof(struct dlm_assert_master),
1755                                         dlm_assert_master_handler,
1756                                         dlm, dlm_assert_master_post_handler,
1757                                         &dlm->dlm_domain_handlers);
1758         if (status)
1759                 goto bail;
1760
1761         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1762                                         sizeof(struct dlm_create_lock),
1763                                         dlm_create_lock_handler,
1764                                         dlm, NULL, &dlm->dlm_domain_handlers);
1765         if (status)
1766                 goto bail;
1767
1768         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1769                                         DLM_CONVERT_LOCK_MAX_LEN,
1770                                         dlm_convert_lock_handler,
1771                                         dlm, NULL, &dlm->dlm_domain_handlers);
1772         if (status)
1773                 goto bail;
1774
1775         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1776                                         DLM_UNLOCK_LOCK_MAX_LEN,
1777                                         dlm_unlock_lock_handler,
1778                                         dlm, NULL, &dlm->dlm_domain_handlers);
1779         if (status)
1780                 goto bail;
1781
1782         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1783                                         DLM_PROXY_AST_MAX_LEN,
1784                                         dlm_proxy_ast_handler,
1785                                         dlm, NULL, &dlm->dlm_domain_handlers);
1786         if (status)
1787                 goto bail;
1788
1789         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1790                                         sizeof(struct dlm_exit_domain),
1791                                         dlm_exit_domain_handler,
1792                                         dlm, NULL, &dlm->dlm_domain_handlers);
1793         if (status)
1794                 goto bail;
1795
1796         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1797                                         sizeof(struct dlm_deref_lockres),
1798                                         dlm_deref_lockres_handler,
1799                                         dlm, NULL, &dlm->dlm_domain_handlers);
1800         if (status)
1801                 goto bail;
1802
1803         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1804                                         sizeof(struct dlm_migrate_request),
1805                                         dlm_migrate_request_handler,
1806                                         dlm, NULL, &dlm->dlm_domain_handlers);
1807         if (status)
1808                 goto bail;
1809
1810         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1811                                         DLM_MIG_LOCKRES_MAX_LEN,
1812                                         dlm_mig_lockres_handler,
1813                                         dlm, NULL, &dlm->dlm_domain_handlers);
1814         if (status)
1815                 goto bail;
1816
1817         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1818                                         sizeof(struct dlm_master_requery),
1819                                         dlm_master_requery_handler,
1820                                         dlm, NULL, &dlm->dlm_domain_handlers);
1821         if (status)
1822                 goto bail;
1823
1824         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1825                                         sizeof(struct dlm_lock_request),
1826                                         dlm_request_all_locks_handler,
1827                                         dlm, NULL, &dlm->dlm_domain_handlers);
1828         if (status)
1829                 goto bail;
1830
1831         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1832                                         sizeof(struct dlm_reco_data_done),
1833                                         dlm_reco_data_done_handler,
1834                                         dlm, NULL, &dlm->dlm_domain_handlers);
1835         if (status)
1836                 goto bail;
1837
1838         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1839                                         sizeof(struct dlm_begin_reco),
1840                                         dlm_begin_reco_handler,
1841                                         dlm, NULL, &dlm->dlm_domain_handlers);
1842         if (status)
1843                 goto bail;
1844
1845         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1846                                         sizeof(struct dlm_finalize_reco),
1847                                         dlm_finalize_reco_handler,
1848                                         dlm, NULL, &dlm->dlm_domain_handlers);
1849         if (status)
1850                 goto bail;
1851
1852         status = o2net_register_handler(DLM_BEGIN_EXIT_DOMAIN_MSG, dlm->key,
1853                                         sizeof(struct dlm_exit_domain),
1854                                         dlm_begin_exit_domain_handler,
1855                                         dlm, NULL, &dlm->dlm_domain_handlers);
1856
1857 bail:
1858         if (status)
1859                 dlm_unregister_domain_handlers(dlm);
1860
1861         return status;
1862 }
1863
1864 static int dlm_join_domain(struct dlm_ctxt *dlm)
1865 {
1866         int status;
1867         unsigned int backoff;
1868         unsigned int total_backoff = 0;
1869         char wq_name[O2NM_MAX_NAME_LEN];
1870
1871         BUG_ON(!dlm);
1872
1873         mlog(0, "Join domain %s\n", dlm->name);
1874
1875         status = dlm_register_domain_handlers(dlm);
1876         if (status) {
1877                 mlog_errno(status);
1878                 goto bail;
1879         }
1880
1881         status = dlm_launch_thread(dlm);
1882         if (status < 0) {
1883                 mlog_errno(status);
1884                 goto bail;
1885         }
1886
1887         status = dlm_launch_recovery_thread(dlm);
1888         if (status < 0) {
1889                 mlog_errno(status);
1890                 goto bail;
1891         }
1892
1893         status = dlm_debug_init(dlm);
1894         if (status < 0) {
1895                 mlog_errno(status);
1896                 goto bail;
1897         }
1898
1899         snprintf(wq_name, O2NM_MAX_NAME_LEN, "dlm_wq-%s", dlm->name);
1900         dlm->dlm_worker = create_singlethread_workqueue(wq_name);
1901         if (!dlm->dlm_worker) {
1902                 status = -ENOMEM;
1903                 mlog_errno(status);
1904                 goto bail;
1905         }
1906
1907         do {
1908                 status = dlm_try_to_join_domain(dlm);
1909
1910                 /* If we're racing another node to the join, then we
1911                  * need to back off temporarily and let them
1912                  * complete. */
1913 #define DLM_JOIN_TIMEOUT_MSECS  90000
1914                 if (status == -EAGAIN) {
1915                         if (signal_pending(current)) {
1916                                 status = -ERESTARTSYS;
1917                                 goto bail;
1918                         }
1919
1920                         if (total_backoff > DLM_JOIN_TIMEOUT_MSECS) {
1921                                 status = -ERESTARTSYS;
1922                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1923                                      "%s after %u msecs\n", dlm->name,
1924                                      total_backoff);
1925                                 goto bail;
1926                         }
1927
1928                         /*
1929                          * <chip> After you!
1930                          * <dale> No, after you!
1931                          * <chip> I insist!
1932                          * <dale> But you first!
1933                          * ...
1934                          */
1935                         backoff = (unsigned int)(jiffies & 0x3);
1936                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1937                         total_backoff += backoff;
1938                         mlog(0, "backoff %d\n", backoff);
1939                         msleep(backoff);
1940                 }
1941         } while (status == -EAGAIN);
1942
1943         if (status < 0) {
1944                 mlog_errno(status);
1945                 goto bail;
1946         }
1947
1948         status = 0;
1949 bail:
1950         wake_up(&dlm_domain_events);
1951
1952         if (status) {
1953                 dlm_unregister_domain_handlers(dlm);
1954                 dlm_debug_shutdown(dlm);
1955                 dlm_complete_thread(dlm);
1956                 dlm_complete_recovery_thread(dlm);
1957                 dlm_destroy_dlm_worker(dlm);
1958         }
1959
1960         return status;
1961 }
1962
1963 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1964                                 u32 key)
1965 {
1966         int i;
1967         int ret;
1968         struct dlm_ctxt *dlm = NULL;
1969
1970         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1971         if (!dlm) {
1972                 ret = -ENOMEM;
1973                 mlog_errno(ret);
1974                 goto leave;
1975         }
1976
1977         dlm->name = kstrdup(domain, GFP_KERNEL);
1978         if (dlm->name == NULL) {
1979                 ret = -ENOMEM;
1980                 mlog_errno(ret);
1981                 goto leave;
1982         }
1983
1984         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1985         if (!dlm->lockres_hash) {
1986                 ret = -ENOMEM;
1987                 mlog_errno(ret);
1988                 goto leave;
1989         }
1990
1991         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1992                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1993
1994         dlm->master_hash = (struct hlist_head **)
1995                                 dlm_alloc_pagevec(DLM_HASH_PAGES);
1996         if (!dlm->master_hash) {
1997                 ret = -ENOMEM;
1998                 mlog_errno(ret);
1999                 goto leave;
2000         }
2001
2002         for (i = 0; i < DLM_HASH_BUCKETS; i++)
2003                 INIT_HLIST_HEAD(dlm_master_hash(dlm, i));
2004
2005         dlm->key = key;
2006         dlm->node_num = o2nm_this_node();
2007
2008         ret = dlm_create_debugfs_subroot(dlm);
2009         if (ret < 0)
2010                 goto leave;
2011
2012         spin_lock_init(&dlm->spinlock);
2013         spin_lock_init(&dlm->master_lock);
2014         spin_lock_init(&dlm->ast_lock);
2015         spin_lock_init(&dlm->track_lock);
2016         INIT_LIST_HEAD(&dlm->list);
2017         INIT_LIST_HEAD(&dlm->dirty_list);
2018         INIT_LIST_HEAD(&dlm->reco.resources);
2019         INIT_LIST_HEAD(&dlm->reco.node_data);
2020         INIT_LIST_HEAD(&dlm->purge_list);
2021         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
2022         INIT_LIST_HEAD(&dlm->tracking_list);
2023         dlm->reco.state = 0;
2024
2025         INIT_LIST_HEAD(&dlm->pending_asts);
2026         INIT_LIST_HEAD(&dlm->pending_basts);
2027
2028         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
2029                   dlm->recovery_map, &(dlm->recovery_map[0]));
2030
2031         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
2032         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
2033         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
2034
2035         dlm->dlm_thread_task = NULL;
2036         dlm->dlm_reco_thread_task = NULL;
2037         dlm->dlm_worker = NULL;
2038         init_waitqueue_head(&dlm->dlm_thread_wq);
2039         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
2040         init_waitqueue_head(&dlm->reco.event);
2041         init_waitqueue_head(&dlm->ast_wq);
2042         init_waitqueue_head(&dlm->migration_wq);
2043         INIT_LIST_HEAD(&dlm->mle_hb_events);
2044
2045         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
2046         init_waitqueue_head(&dlm->dlm_join_events);
2047
2048         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
2049         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
2050
2051         atomic_set(&dlm->res_tot_count, 0);
2052         atomic_set(&dlm->res_cur_count, 0);
2053         for (i = 0; i < DLM_MLE_NUM_TYPES; ++i) {
2054                 atomic_set(&dlm->mle_tot_count[i], 0);
2055                 atomic_set(&dlm->mle_cur_count[i], 0);
2056         }
2057
2058         spin_lock_init(&dlm->work_lock);
2059         INIT_LIST_HEAD(&dlm->work_list);
2060         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
2061
2062         kref_init(&dlm->dlm_refs);
2063         dlm->dlm_state = DLM_CTXT_NEW;
2064
2065         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
2066
2067         mlog(0, "context init: refcount %u\n",
2068                   atomic_read(&dlm->dlm_refs.refcount));
2069
2070 leave:
2071         if (ret < 0 && dlm) {
2072                 if (dlm->master_hash)
2073                         dlm_free_pagevec((void **)dlm->master_hash,
2074                                         DLM_HASH_PAGES);
2075
2076                 if (dlm->lockres_hash)
2077                         dlm_free_pagevec((void **)dlm->lockres_hash,
2078                                         DLM_HASH_PAGES);
2079
2080                 kfree(dlm->name);
2081                 kfree(dlm);
2082                 dlm = NULL;
2083         }
2084         return dlm;
2085 }
2086
2087 /*
2088  * Compare a requested locking protocol version against the current one.
2089  *
2090  * If the major numbers are different, they are incompatible.
2091  * If the current minor is greater than the request, they are incompatible.
2092  * If the current minor is less than or equal to the request, they are
2093  * compatible, and the requester should run at the current minor version.
2094  */
2095 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
2096                                 struct dlm_protocol_version *request)
2097 {
2098         if (existing->pv_major != request->pv_major)
2099                 return 1;
2100
2101         if (existing->pv_minor > request->pv_minor)
2102                 return 1;
2103
2104         if (existing->pv_minor < request->pv_minor)
2105                 request->pv_minor = existing->pv_minor;
2106
2107         return 0;
2108 }
2109
2110 /*
2111  * dlm_register_domain: one-time setup per "domain".
2112  *
2113  * The filesystem passes in the requested locking version via proto.
2114  * If registration was successful, proto will contain the negotiated
2115  * locking protocol.
2116  */
2117 struct dlm_ctxt * dlm_register_domain(const char *domain,
2118                                u32 key,
2119                                struct dlm_protocol_version *fs_proto)
2120 {
2121         int ret;
2122         struct dlm_ctxt *dlm = NULL;
2123         struct dlm_ctxt *new_ctxt = NULL;
2124
2125         if (strlen(domain) >= O2NM_MAX_NAME_LEN) {
2126                 ret = -ENAMETOOLONG;
2127                 mlog(ML_ERROR, "domain name length too long\n");
2128                 goto leave;
2129         }
2130
2131         mlog(0, "register called for domain \"%s\"\n", domain);
2132
2133 retry:
2134         dlm = NULL;
2135         if (signal_pending(current)) {
2136                 ret = -ERESTARTSYS;
2137                 mlog_errno(ret);
2138                 goto leave;
2139         }
2140
2141         spin_lock(&dlm_domain_lock);
2142
2143         dlm = __dlm_lookup_domain(domain);
2144         if (dlm) {
2145                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
2146                         spin_unlock(&dlm_domain_lock);
2147
2148                         mlog(0, "This ctxt is not joined yet!\n");
2149                         wait_event_interruptible(dlm_domain_events,
2150                                                  dlm_wait_on_domain_helper(
2151                                                          domain));
2152                         goto retry;
2153                 }
2154
2155                 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
2156                         spin_unlock(&dlm_domain_lock);
2157                         mlog(ML_ERROR,
2158                              "Requested locking protocol version is not "
2159                              "compatible with already registered domain "
2160                              "\"%s\"\n", domain);
2161                         ret = -EPROTO;
2162                         goto leave;
2163                 }
2164
2165                 __dlm_get(dlm);
2166                 dlm->num_joins++;
2167
2168                 spin_unlock(&dlm_domain_lock);
2169
2170                 ret = 0;
2171                 goto leave;
2172         }
2173
2174         /* doesn't exist */
2175         if (!new_ctxt) {
2176                 spin_unlock(&dlm_domain_lock);
2177
2178                 new_ctxt = dlm_alloc_ctxt(domain, key);
2179                 if (new_ctxt)
2180                         goto retry;
2181
2182                 ret = -ENOMEM;
2183                 mlog_errno(ret);
2184                 goto leave;
2185         }
2186
2187         /* a little variable switch-a-roo here... */
2188         dlm = new_ctxt;
2189         new_ctxt = NULL;
2190
2191         /* add the new domain */
2192         list_add_tail(&dlm->list, &dlm_domains);
2193         spin_unlock(&dlm_domain_lock);
2194
2195         /*
2196          * Pass the locking protocol version into the join.  If the join
2197          * succeeds, it will have the negotiated protocol set.
2198          */
2199         dlm->dlm_locking_proto = dlm_protocol;
2200         dlm->fs_locking_proto = *fs_proto;
2201
2202         ret = dlm_join_domain(dlm);
2203         if (ret) {
2204                 mlog_errno(ret);
2205                 dlm_put(dlm);
2206                 goto leave;
2207         }
2208
2209         /* Tell the caller what locking protocol we negotiated */
2210         *fs_proto = dlm->fs_locking_proto;
2211
2212         ret = 0;
2213 leave:
2214         if (new_ctxt)
2215                 dlm_free_ctxt_mem(new_ctxt);
2216
2217         if (ret < 0)
2218                 dlm = ERR_PTR(ret);
2219
2220         return dlm;
2221 }
2222 EXPORT_SYMBOL_GPL(dlm_register_domain);
2223
2224 static LIST_HEAD(dlm_join_handlers);
2225
2226 static void dlm_unregister_net_handlers(void)
2227 {
2228         o2net_unregister_handler_list(&dlm_join_handlers);
2229 }
2230
2231 static int dlm_register_net_handlers(void)
2232 {
2233         int status = 0;
2234
2235         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
2236                                         sizeof(struct dlm_query_join_request),
2237                                         dlm_query_join_handler,
2238                                         NULL, NULL, &dlm_join_handlers);
2239         if (status)
2240                 goto bail;
2241
2242         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
2243                                         sizeof(struct dlm_assert_joined),
2244                                         dlm_assert_joined_handler,
2245                                         NULL, NULL, &dlm_join_handlers);
2246         if (status)
2247                 goto bail;
2248
2249         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
2250                                         sizeof(struct dlm_cancel_join),
2251                                         dlm_cancel_join_handler,
2252                                         NULL, NULL, &dlm_join_handlers);
2253         if (status)
2254                 goto bail;
2255
2256         status = o2net_register_handler(DLM_QUERY_REGION, DLM_MOD_KEY,
2257                                         sizeof(struct dlm_query_region),
2258                                         dlm_query_region_handler,
2259                                         NULL, NULL, &dlm_join_handlers);
2260
2261         if (status)
2262                 goto bail;
2263
2264         status = o2net_register_handler(DLM_QUERY_NODEINFO, DLM_MOD_KEY,
2265                                         sizeof(struct dlm_query_nodeinfo),
2266                                         dlm_query_nodeinfo_handler,
2267                                         NULL, NULL, &dlm_join_handlers);
2268 bail:
2269         if (status < 0)
2270                 dlm_unregister_net_handlers();
2271
2272         return status;
2273 }
2274
2275 /* Domain eviction callback handling.
2276  *
2277  * The file system requires notification of node death *before* the
2278  * dlm completes it's recovery work, otherwise it may be able to
2279  * acquire locks on resources requiring recovery. Since the dlm can
2280  * evict a node from it's domain *before* heartbeat fires, a similar
2281  * mechanism is required. */
2282
2283 /* Eviction is not expected to happen often, so a per-domain lock is
2284  * not necessary. Eviction callbacks are allowed to sleep for short
2285  * periods of time. */
2286 static DECLARE_RWSEM(dlm_callback_sem);
2287
2288 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
2289                                         int node_num)
2290 {
2291         struct dlm_eviction_cb *cb;
2292
2293         down_read(&dlm_callback_sem);
2294         list_for_each_entry(cb, &dlm->dlm_eviction_callbacks, ec_item) {
2295                 cb->ec_func(node_num, cb->ec_data);
2296         }
2297         up_read(&dlm_callback_sem);
2298 }
2299
2300 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
2301                            dlm_eviction_func *f,
2302                            void *data)
2303 {
2304         INIT_LIST_HEAD(&cb->ec_item);
2305         cb->ec_func = f;
2306         cb->ec_data = data;
2307 }
2308 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
2309
2310 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
2311                               struct dlm_eviction_cb *cb)
2312 {
2313         down_write(&dlm_callback_sem);
2314         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
2315         up_write(&dlm_callback_sem);
2316 }
2317 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
2318
2319 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
2320 {
2321         down_write(&dlm_callback_sem);
2322         list_del_init(&cb->ec_item);
2323         up_write(&dlm_callback_sem);
2324 }
2325 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
2326
2327 static int __init dlm_init(void)
2328 {
2329         int status;
2330
2331         status = dlm_init_mle_cache();
2332         if (status) {
2333                 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
2334                 goto error;
2335         }
2336
2337         status = dlm_init_master_caches();
2338         if (status) {
2339                 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
2340                      "o2dlm_lockname slabcaches\n");
2341                 goto error;
2342         }
2343
2344         status = dlm_init_lock_cache();
2345         if (status) {
2346                 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
2347                 goto error;
2348         }
2349
2350         status = dlm_register_net_handlers();
2351         if (status) {
2352                 mlog(ML_ERROR, "Unable to register network handlers\n");
2353                 goto error;
2354         }
2355
2356         status = dlm_create_debugfs_root();
2357         if (status)
2358                 goto error;
2359
2360         return 0;
2361 error:
2362         dlm_unregister_net_handlers();
2363         dlm_destroy_lock_cache();
2364         dlm_destroy_master_caches();
2365         dlm_destroy_mle_cache();
2366         return -1;
2367 }
2368
2369 static void __exit dlm_exit (void)
2370 {
2371         dlm_destroy_debugfs_root();
2372         dlm_unregister_net_handlers();
2373         dlm_destroy_lock_cache();
2374         dlm_destroy_master_caches();
2375         dlm_destroy_mle_cache();
2376 }
2377
2378 MODULE_AUTHOR("Oracle");
2379 MODULE_LICENSE("GPL");
2380 MODULE_DESCRIPTION("OCFS2 Distributed Lock Management");
2381
2382 module_init(dlm_init);
2383 module_exit(dlm_exit);