fs/ocfs2/dlm/dlmdomain.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmdomain.c
   5  *
   6  * defines domain join / leave apis
   7  *
   8  * Copyright (C) 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  *
  25  */
  26
  27 #include <linux/module.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/utsname.h>
  32 #include <linux/init.h>
  33 #include <linux/spinlock.h>
  34 #include <linux/delay.h>
  35 #include <linux/err.h>
  36 #include <linux/debugfs.h>
  37
  38 #include "cluster/heartbeat.h"
  39 #include "cluster/nodemanager.h"
  40 #include "cluster/tcp.h"
  41
  42 #include "dlmapi.h"
  43 #include "dlmcommon.h"
  44 #include "dlmdomain.h"
  45 #include "dlmdebug.h"
  46
  47 #include "dlmver.h"
  48
  49 #define MLOG_MASK_PREFIX (ML_DLM|ML_DLM_DOMAIN)
  50 #include "cluster/masklog.h"
  51
  52 /*
  53  * ocfs2 node maps are array of long int, which limits to send them freely
  54  * across the wire due to endianness issues. To workaround this, we convert
  55  * long ints to byte arrays. Following 3 routines are helper functions to
  56  * set/test/copy bits within those array of bytes
  57  */
  58 static inline void byte_set_bit(u8 nr, u8 map[])
  59 {
  60         map[nr >> 3] |= (1UL << (nr & 7));
  61 }
  62
  63 static inline int byte_test_bit(u8 nr, u8 map[])
  64 {
  65         return ((1UL << (nr & 7)) & (map[nr >> 3])) != 0;
  66 }
  67
  68 static inline void byte_copymap(u8 dmap[], unsigned long smap[],
  69                         unsigned int sz)
  70 {
  71         unsigned int nn;
  72
  73         if (!sz)
  74                 return;
  75
  76         memset(dmap, 0, ((sz + 7) >> 3));
  77         for (nn = 0 ; nn < sz; nn++)
  78                 if (test_bit(nn, smap))
  79                         byte_set_bit(nn, dmap);
  80 }
  81
  82 static void dlm_free_pagevec(void **vec, int pages)
  83 {
  84         while (pages--)
  85                 free_page((unsigned long)vec[pages]);
  86         kfree(vec);
  87 }
  88
  89 static void **dlm_alloc_pagevec(int pages)
  90 {
  91         void **vec = kmalloc(pages * sizeof(void *), GFP_KERNEL);
  92         int i;
  93
  94         if (!vec)
  95                 return NULL;
  96
  97         for (i = 0; i < pages; i++)
  98                 if (!(vec[i] = (void *)__get_free_page(GFP_KERNEL)))
  99                         goto out_free;
 100
 101         mlog(0, "Allocated DLM hash pagevec; %d pages (%lu expected), %lu buckets per page\n",
 102              pages, (unsigned long)DLM_HASH_PAGES,
 103              (unsigned long)DLM_BUCKETS_PER_PAGE);
 104         return vec;
 105 out_free:
 106         dlm_free_pagevec(vec, i);
 107         return NULL;
 108 }
 109
 110 /*
 111  *
 112  * spinlock lock ordering: if multiple locks are needed, obey this ordering:
 113  *    dlm_domain_lock
 114  *    struct dlm_ctxt->spinlock
 115  *    struct dlm_lock_resource->spinlock
 116  *    struct dlm_ctxt->master_lock
 117  *    struct dlm_ctxt->ast_lock
 118  *    dlm_master_list_entry->spinlock
 119  *    dlm_lock->spinlock
 120  *
 121  */
 122
 123 DEFINE_SPINLOCK(dlm_domain_lock);
 124 LIST_HEAD(dlm_domains);
 125 static DECLARE_WAIT_QUEUE_HEAD(dlm_domain_events);
 126
 127 /*
 128  * The supported protocol version for DLM communication.  Running domains
 129  * will have a negotiated version with the same major number and a minor
 130  * number equal or smaller.  The dlm_ctxt->dlm_locking_proto field should
 131  * be used to determine what a running domain is actually using.
 132  */
 133 static const struct dlm_protocol_version dlm_protocol = {
 134         .pv_major = 1,
 135         .pv_minor = 0,
 136 };
 137
 138 #define DLM_DOMAIN_BACKOFF_MS 200
 139
 140 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 141                                   void **ret_data);
 142 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 143                                      void **ret_data);
 144 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 145                                    void **ret_data);
 146 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 147                                    void **ret_data);
 148 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
 149                                 struct dlm_protocol_version *request);
 150
 151 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm);
 152
 153 void __dlm_unhash_lockres(struct dlm_lock_resource *lockres)
 154 {
 155         if (!hlist_unhashed(&lockres->hash_node)) {
 156                 hlist_del_init(&lockres->hash_node);
 157                 dlm_lockres_put(lockres);
 158         }
 159 }
 160
 161 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 162                        struct dlm_lock_resource *res)
 163 {
 164         struct hlist_head *bucket;
 165         struct qstr *q;
 166
 167         assert_spin_locked(&dlm->spinlock);
 168
 169         q = &res->lockname;
 170         bucket = dlm_lockres_hash(dlm, q->hash);
 171
 172         /* get a reference for our hashtable */
 173         dlm_lockres_get(res);
 174
 175         hlist_add_head(&res->hash_node, bucket);
 176 }
 177
 178 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 179                                                      const char *name,
 180                                                      unsigned int len,
 181                                                      unsigned int hash)
 182 {
 183         struct hlist_head *bucket;
 184         struct hlist_node *list;
 185
 186         mlog_entry("%.*s\n", len, name);
 187
 188         assert_spin_locked(&dlm->spinlock);
 189
 190         bucket = dlm_lockres_hash(dlm, hash);
 191
 192         hlist_for_each(list, bucket) {
 193                 struct dlm_lock_resource *res = hlist_entry(list,
 194                         struct dlm_lock_resource, hash_node);
 195                 if (res->lockname.name[0] != name[0])
 196                         continue;
 197                 if (unlikely(res->lockname.len != len))
 198                         continue;
 199                 if (memcmp(res->lockname.name + 1, name + 1, len - 1))
 200                         continue;
 201                 dlm_lockres_get(res);
 202                 return res;
 203         }
 204         return NULL;
 205 }
 206
 207 /* intended to be called by functions which do not care about lock
 208  * resources which are being purged (most net _handler functions).
 209  * this will return NULL for any lock resource which is found but
 210  * currently in the process of dropping its mastery reference.
 211  * use __dlm_lookup_lockres_full when you need the lock resource
 212  * regardless (e.g. dlm_get_lock_resource) */
 213 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 214                                                 const char *name,
 215                                                 unsigned int len,
 216                                                 unsigned int hash)
 217 {
 218         struct dlm_lock_resource *res = NULL;
 219
 220         mlog_entry("%.*s\n", len, name);
 221
 222         assert_spin_locked(&dlm->spinlock);
 223
 224         res = __dlm_lookup_lockres_full(dlm, name, len, hash);
 225         if (res) {
 226                 spin_lock(&res->spinlock);
 227                 if (res->state & DLM_LOCK_RES_DROPPING_REF) {
 228                         spin_unlock(&res->spinlock);
 229                         dlm_lockres_put(res);
 230                         return NULL;
 231                 }
 232                 spin_unlock(&res->spinlock);
 233         }
 234
 235         return res;
 236 }
 237
 238 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 239                                     const char *name,
 240                                     unsigned int len)
 241 {
 242         struct dlm_lock_resource *res;
 243         unsigned int hash = dlm_lockid_hash(name, len);
 244
 245         spin_lock(&dlm->spinlock);
 246         res = __dlm_lookup_lockres(dlm, name, len, hash);
 247         spin_unlock(&dlm->spinlock);
 248         return res;
 249 }
 250
 251 static struct dlm_ctxt * __dlm_lookup_domain_full(const char *domain, int len)
 252 {
 253         struct dlm_ctxt *tmp = NULL;
 254         struct list_head *iter;
 255
 256         assert_spin_locked(&dlm_domain_lock);
 257
 258         /* tmp->name here is always NULL terminated,
 259          * but domain may not be! */
 260         list_for_each(iter, &dlm_domains) {
 261                 tmp = list_entry (iter, struct dlm_ctxt, list);
 262                 if (strlen(tmp->name) == len &&
 263                     memcmp(tmp->name, domain, len)==0)
 264                         break;
 265                 tmp = NULL;
 266         }
 267
 268         return tmp;
 269 }
 270
 271 /* For null terminated domain strings ONLY */
 272 static struct dlm_ctxt * __dlm_lookup_domain(const char *domain)
 273 {
 274         assert_spin_locked(&dlm_domain_lock);
 275
 276         return __dlm_lookup_domain_full(domain, strlen(domain));
 277 }
 278
 279
 280 /* returns true on one of two conditions:
 281  * 1) the domain does not exist
 282  * 2) the domain exists and it's state is "joined" */
 283 static int dlm_wait_on_domain_helper(const char *domain)
 284 {
 285         int ret = 0;
 286         struct dlm_ctxt *tmp = NULL;
 287
 288         spin_lock(&dlm_domain_lock);
 289
 290         tmp = __dlm_lookup_domain(domain);
 291         if (!tmp)
 292                 ret = 1;
 293         else if (tmp->dlm_state == DLM_CTXT_JOINED)
 294                 ret = 1;
 295
 296         spin_unlock(&dlm_domain_lock);
 297         return ret;
 298 }
 299
 300 static void dlm_free_ctxt_mem(struct dlm_ctxt *dlm)
 301 {
 302         dlm_destroy_debugfs_subroot(dlm);
 303
 304         if (dlm->lockres_hash)
 305                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
 306
 307         if (dlm->name)
 308                 kfree(dlm->name);
 309
 310         kfree(dlm);
 311 }
 312
 313 /* A little strange - this function will be called while holding
 314  * dlm_domain_lock and is expected to be holding it on the way out. We
 315  * will however drop and reacquire it multiple times */
 316 static void dlm_ctxt_release(struct kref *kref)
 317 {
 318         struct dlm_ctxt *dlm;
 319
 320         dlm = container_of(kref, struct dlm_ctxt, dlm_refs);
 321
 322         BUG_ON(dlm->num_joins);
 323         BUG_ON(dlm->dlm_state == DLM_CTXT_JOINED);
 324
 325         /* we may still be in the list if we hit an error during join. */
 326         list_del_init(&dlm->list);
 327
 328         spin_unlock(&dlm_domain_lock);
 329
 330         mlog(0, "freeing memory from domain %s\n", dlm->name);
 331
 332         wake_up(&dlm_domain_events);
 333
 334         dlm_free_ctxt_mem(dlm);
 335
 336         spin_lock(&dlm_domain_lock);
 337 }
 338
 339 void dlm_put(struct dlm_ctxt *dlm)
 340 {
 341         spin_lock(&dlm_domain_lock);
 342         kref_put(&dlm->dlm_refs, dlm_ctxt_release);
 343         spin_unlock(&dlm_domain_lock);
 344 }
 345
 346 static void __dlm_get(struct dlm_ctxt *dlm)
 347 {
 348         kref_get(&dlm->dlm_refs);
 349 }
 350
 351 /* given a questionable reference to a dlm object, gets a reference if
 352  * it can find it in the list, otherwise returns NULL in which case
 353  * you shouldn't trust your pointer. */
 354 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm)
 355 {
 356         struct list_head *iter;
 357         struct dlm_ctxt *target = NULL;
 358
 359         spin_lock(&dlm_domain_lock);
 360
 361         list_for_each(iter, &dlm_domains) {
 362                 target = list_entry (iter, struct dlm_ctxt, list);
 363
 364                 if (target == dlm) {
 365                         __dlm_get(target);
 366                         break;
 367                 }
 368
 369                 target = NULL;
 370         }
 371
 372         spin_unlock(&dlm_domain_lock);
 373
 374         return target;
 375 }
 376
 377 int dlm_domain_fully_joined(struct dlm_ctxt *dlm)
 378 {
 379         int ret;
 380
 381         spin_lock(&dlm_domain_lock);
 382         ret = (dlm->dlm_state == DLM_CTXT_JOINED) ||
 383                 (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN);
 384         spin_unlock(&dlm_domain_lock);
 385
 386         return ret;
 387 }
 388
 389 static void dlm_destroy_dlm_worker(struct dlm_ctxt *dlm)
 390 {
 391         if (dlm->dlm_worker) {
 392                 flush_workqueue(dlm->dlm_worker);
 393                 destroy_workqueue(dlm->dlm_worker);
 394                 dlm->dlm_worker = NULL;
 395         }
 396 }
 397
 398 static void dlm_complete_dlm_shutdown(struct dlm_ctxt *dlm)
 399 {
 400         dlm_unregister_domain_handlers(dlm);
 401         dlm_debug_shutdown(dlm);
 402         dlm_complete_thread(dlm);
 403         dlm_complete_recovery_thread(dlm);
 404         dlm_destroy_dlm_worker(dlm);
 405
 406         /* We've left the domain. Now we can take ourselves out of the
 407          * list and allow the kref stuff to help us free the
 408          * memory. */
 409         spin_lock(&dlm_domain_lock);
 410         list_del_init(&dlm->list);
 411         spin_unlock(&dlm_domain_lock);
 412
 413         /* Wake up anyone waiting for us to remove this domain */
 414         wake_up(&dlm_domain_events);
 415 }
 416
 417 static int dlm_migrate_all_locks(struct dlm_ctxt *dlm)
 418 {
 419         int i, num, n, ret = 0;
 420         struct dlm_lock_resource *res;
 421         struct hlist_node *iter;
 422         struct hlist_head *bucket;
 423         int dropped;
 424
 425         mlog(0, "Migrating locks from domain %s\n", dlm->name);
 426
 427         num = 0;
 428         spin_lock(&dlm->spinlock);
 429         for (i = 0; i < DLM_HASH_BUCKETS; i++) {
 430 redo_bucket:
 431                 n = 0;
 432                 bucket = dlm_lockres_hash(dlm, i);
 433                 iter = bucket->first;
 434                 while (iter) {
 435                         n++;
 436                         res = hlist_entry(iter, struct dlm_lock_resource,
 437                                           hash_node);
 438                         dlm_lockres_get(res);
 439                         /* migrate, if necessary.  this will drop the dlm
 440                          * spinlock and retake it if it does migration. */
 441                         dropped = dlm_empty_lockres(dlm, res);
 442
 443                         spin_lock(&res->spinlock);
 444                         __dlm_lockres_calc_usage(dlm, res);
 445                         iter = res->hash_node.next;
 446                         spin_unlock(&res->spinlock);
 447
 448                         dlm_lockres_put(res);
 449
 450                         if (dropped)
 451                                 goto redo_bucket;
 452                 }
 453                 cond_resched_lock(&dlm->spinlock);
 454                 num += n;
 455                 mlog(0, "%s: touched %d lockreses in bucket %d "
 456                      "(tot=%d)\n", dlm->name, n, i, num);
 457         }
 458         spin_unlock(&dlm->spinlock);
 459         wake_up(&dlm->dlm_thread_wq);
 460
 461         /* let the dlm thread take care of purging, keep scanning until
 462          * nothing remains in the hash */
 463         if (num) {
 464                 mlog(0, "%s: %d lock resources in hash last pass\n",
 465                      dlm->name, num);
 466                 ret = -EAGAIN;
 467         }
 468         mlog(0, "DONE Migrating locks from domain %s\n", dlm->name);
 469         return ret;
 470 }
 471
 472 static int dlm_no_joining_node(struct dlm_ctxt *dlm)
 473 {
 474         int ret;
 475
 476         spin_lock(&dlm->spinlock);
 477         ret = dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN;
 478         spin_unlock(&dlm->spinlock);
 479
 480         return ret;
 481 }
 482
 483 static void dlm_mark_domain_leaving(struct dlm_ctxt *dlm)
 484 {
 485         /* Yikes, a double spinlock! I need domain_lock for the dlm
 486          * state and the dlm spinlock for join state... Sorry! */
 487 again:
 488         spin_lock(&dlm_domain_lock);
 489         spin_lock(&dlm->spinlock);
 490
 491         if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 492                 mlog(0, "Node %d is joining, we wait on it.\n",
 493                           dlm->joining_node);
 494                 spin_unlock(&dlm->spinlock);
 495                 spin_unlock(&dlm_domain_lock);
 496
 497                 wait_event(dlm->dlm_join_events, dlm_no_joining_node(dlm));
 498                 goto again;
 499         }
 500
 501         dlm->dlm_state = DLM_CTXT_LEAVING;
 502         spin_unlock(&dlm->spinlock);
 503         spin_unlock(&dlm_domain_lock);
 504 }
 505
 506 static void __dlm_print_nodes(struct dlm_ctxt *dlm)
 507 {
 508         int node = -1;
 509
 510         assert_spin_locked(&dlm->spinlock);
 511
 512         printk(KERN_INFO "ocfs2_dlm: Nodes in domain (\"%s\"): ", dlm->name);
 513
 514         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 515                                      node + 1)) < O2NM_MAX_NODES) {
 516                 printk("%d ", node);
 517         }
 518         printk("\n");
 519 }
 520
 521 static int dlm_exit_domain_handler(struct o2net_msg *msg, u32 len, void *data,
 522                                    void **ret_data)
 523 {
 524         struct dlm_ctxt *dlm = data;
 525         unsigned int node;
 526         struct dlm_exit_domain *exit_msg = (struct dlm_exit_domain *) msg->buf;
 527
 528         mlog_entry("%p %u %p", msg, len, data);
 529
 530         if (!dlm_grab(dlm))
 531                 return 0;
 532
 533         node = exit_msg->node_idx;
 534
 535         printk(KERN_INFO "ocfs2_dlm: Node %u leaves domain %s\n", node, dlm->name);
 536
 537         spin_lock(&dlm->spinlock);
 538         clear_bit(node, dlm->domain_map);
 539         __dlm_print_nodes(dlm);
 540
 541         /* notify anything attached to the heartbeat events */
 542         dlm_hb_event_notify_attached(dlm, node, 0);
 543
 544         spin_unlock(&dlm->spinlock);
 545
 546         dlm_put(dlm);
 547
 548         return 0;
 549 }
 550
 551 static int dlm_send_one_domain_exit(struct dlm_ctxt *dlm,
 552                                     unsigned int node)
 553 {
 554         int status;
 555         struct dlm_exit_domain leave_msg;
 556
 557         mlog(0, "Asking node %u if we can leave the domain %s me = %u\n",
 558                   node, dlm->name, dlm->node_num);
 559
 560         memset(&leave_msg, 0, sizeof(leave_msg));
 561         leave_msg.node_idx = dlm->node_num;
 562
 563         status = o2net_send_message(DLM_EXIT_DOMAIN_MSG, dlm->key,
 564                                     &leave_msg, sizeof(leave_msg), node,
 565                                     NULL);
 566
 567         mlog(0, "status return %d from o2net_send_message\n", status);
 568
 569         return status;
 570 }
 571
 572
 573 static void dlm_leave_domain(struct dlm_ctxt *dlm)
 574 {
 575         int node, clear_node, status;
 576
 577         /* At this point we've migrated away all our locks and won't
 578          * accept mastership of new ones. The dlm is responsible for
 579          * almost nothing now. We make sure not to confuse any joining
 580          * nodes and then commence shutdown procedure. */
 581
 582         spin_lock(&dlm->spinlock);
 583         /* Clear ourselves from the domain map */
 584         clear_bit(dlm->node_num, dlm->domain_map);
 585         while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
 586                                      0)) < O2NM_MAX_NODES) {
 587                 /* Drop the dlm spinlock. This is safe wrt the domain_map.
 588                  * -nodes cannot be added now as the
 589                  *   query_join_handlers knows to respond with OK_NO_MAP
 590                  * -we catch the right network errors if a node is
 591                  *   removed from the map while we're sending him the
 592                  *   exit message. */
 593                 spin_unlock(&dlm->spinlock);
 594
 595                 clear_node = 1;
 596
 597                 status = dlm_send_one_domain_exit(dlm, node);
 598                 if (status < 0 &&
 599                     status != -ENOPROTOOPT &&
 600                     status != -ENOTCONN) {
 601                         mlog(ML_NOTICE, "Error %d sending domain exit message "
 602                              "to node %d\n", status, node);
 603
 604                         /* Not sure what to do here but lets sleep for
 605                          * a bit in case this was a transient
 606                          * error... */
 607                         msleep(DLM_DOMAIN_BACKOFF_MS);
 608                         clear_node = 0;
 609                 }
 610
 611                 spin_lock(&dlm->spinlock);
 612                 /* If we're not clearing the node bit then we intend
 613                  * to loop back around to try again. */
 614                 if (clear_node)
 615                         clear_bit(node, dlm->domain_map);
 616         }
 617         spin_unlock(&dlm->spinlock);
 618 }
 619
 620 int dlm_joined(struct dlm_ctxt *dlm)
 621 {
 622         int ret = 0;
 623
 624         spin_lock(&dlm_domain_lock);
 625
 626         if (dlm->dlm_state == DLM_CTXT_JOINED)
 627                 ret = 1;
 628
 629         spin_unlock(&dlm_domain_lock);
 630
 631         return ret;
 632 }
 633
 634 int dlm_shutting_down(struct dlm_ctxt *dlm)
 635 {
 636         int ret = 0;
 637
 638         spin_lock(&dlm_domain_lock);
 639
 640         if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
 641                 ret = 1;
 642
 643         spin_unlock(&dlm_domain_lock);
 644
 645         return ret;
 646 }
 647
 648 void dlm_unregister_domain(struct dlm_ctxt *dlm)
 649 {
 650         int leave = 0;
 651         struct dlm_lock_resource *res;
 652
 653         spin_lock(&dlm_domain_lock);
 654         BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
 655         BUG_ON(!dlm->num_joins);
 656
 657         dlm->num_joins--;
 658         if (!dlm->num_joins) {
 659                 /* We mark it "in shutdown" now so new register
 660                  * requests wait until we've completely left the
 661                  * domain. Don't use DLM_CTXT_LEAVING yet as we still
 662                  * want new domain joins to communicate with us at
 663                  * least until we've completed migration of our
 664                  * resources. */
 665                 dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
 666                 leave = 1;
 667         }
 668         spin_unlock(&dlm_domain_lock);
 669
 670         if (leave) {
 671                 mlog(0, "shutting down domain %s\n", dlm->name);
 672
 673                 /* We changed dlm state, notify the thread */
 674                 dlm_kick_thread(dlm, NULL);
 675
 676                 while (dlm_migrate_all_locks(dlm)) {
 677                         /* Give dlm_thread time to purge the lockres' */
 678                         msleep(500);
 679                         mlog(0, "%s: more migration to do\n", dlm->name);
 680                 }
 681
 682                 /* This list should be empty. If not, print remaining lockres */
 683                 if (!list_empty(&dlm->tracking_list)) {
 684                         mlog(ML_ERROR, "Following lockres' are still on the "
 685                              "tracking list:\n");
 686                         list_for_each_entry(res, &dlm->tracking_list, tracking)
 687                                 dlm_print_one_lock_resource(res);
 688                 }
 689
 690                 dlm_mark_domain_leaving(dlm);
 691                 dlm_leave_domain(dlm);
 692                 dlm_complete_dlm_shutdown(dlm);
 693         }
 694         dlm_put(dlm);
 695 }
 696 EXPORT_SYMBOL_GPL(dlm_unregister_domain);
 697
 698 static int dlm_query_join_proto_check(char *proto_type, int node,
 699                                       struct dlm_protocol_version *ours,
 700                                       struct dlm_protocol_version *request)
 701 {
 702         int rc;
 703         struct dlm_protocol_version proto = *request;
 704
 705         if (!dlm_protocol_compare(ours, &proto)) {
 706                 mlog(0,
 707                      "node %u wanted to join with %s locking protocol "
 708                      "%u.%u, we respond with %u.%u\n",
 709                      node, proto_type,
 710                      request->pv_major,
 711                      request->pv_minor,
 712                      proto.pv_major, proto.pv_minor);
 713                 request->pv_minor = proto.pv_minor;
 714                 rc = 0;
 715         } else {
 716                 mlog(ML_NOTICE,
 717                      "Node %u wanted to join with %s locking "
 718                      "protocol %u.%u, but we have %u.%u, disallowing\n",
 719                      node, proto_type,
 720                      request->pv_major,
 721                      request->pv_minor,
 722                      ours->pv_major,
 723                      ours->pv_minor);
 724                 rc = 1;
 725         }
 726
 727         return rc;
 728 }
 729
 730 /*
 731  * struct dlm_query_join_packet is made up of four one-byte fields.  They
 732  * are effectively in big-endian order already.  However, little-endian
 733  * machines swap them before putting the packet on the wire (because
 734  * query_join's response is a status, and that status is treated as a u32
 735  * on the wire).  Thus, a big-endian and little-endian machines will treat
 736  * this structure differently.
 737  *
 738  * The solution is to have little-endian machines swap the structure when
 739  * converting from the structure to the u32 representation.  This will
 740  * result in the structure having the correct format on the wire no matter
 741  * the host endian format.
 742  */
 743 static void dlm_query_join_packet_to_wire(struct dlm_query_join_packet *packet,
 744                                           u32 *wire)
 745 {
 746         union dlm_query_join_response response;
 747
 748         response.packet = *packet;
 749         *wire = cpu_to_be32(response.intval);
 750 }
 751
 752 static void dlm_query_join_wire_to_packet(u32 wire,
 753                                           struct dlm_query_join_packet *packet)
 754 {
 755         union dlm_query_join_response response;
 756
 757         response.intval = cpu_to_be32(wire);
 758         *packet = response.packet;
 759 }
 760
 761 static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
 762                                   void **ret_data)
 763 {
 764         struct dlm_query_join_request *query;
 765         struct dlm_query_join_packet packet = {
 766                 .code = JOIN_DISALLOW,
 767         };
 768         struct dlm_ctxt *dlm = NULL;
 769         u32 response;
 770         u8 nodenum;
 771
 772         query = (struct dlm_query_join_request *) msg->buf;
 773
 774         mlog(0, "node %u wants to join domain %s\n", query->node_idx,
 775                   query->domain);
 776
 777         /*
 778          * If heartbeat doesn't consider the node live, tell it
 779          * to back off and try again.  This gives heartbeat a chance
 780          * to catch up.
 781          */
 782         if (!o2hb_check_node_heartbeating(query->node_idx)) {
 783                 mlog(0, "node %u is not in our live map yet\n",
 784                      query->node_idx);
 785
 786                 packet.code = JOIN_DISALLOW;
 787                 goto respond;
 788         }
 789
 790         packet.code = JOIN_OK_NO_MAP;
 791
 792         spin_lock(&dlm_domain_lock);
 793         dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
 794         if (!dlm)
 795                 goto unlock_respond;
 796
 797         /*
 798          * There is a small window where the joining node may not see the
 799          * node(s) that just left but still part of the cluster. DISALLOW
 800          * join request if joining node has different node map.
 801          */
 802         nodenum=0;
 803         while (nodenum < O2NM_MAX_NODES) {
 804                 if (test_bit(nodenum, dlm->domain_map)) {
 805                         if (!byte_test_bit(nodenum, query->node_map)) {
 806                                 mlog(0, "disallow join as node %u does not "
 807                                      "have node %u in its nodemap\n",
 808                                      query->node_idx, nodenum);
 809                                 packet.code = JOIN_DISALLOW;
 810                                 goto unlock_respond;
 811                         }
 812                 }
 813                 nodenum++;
 814         }
 815
 816         /* Once the dlm ctxt is marked as leaving then we don't want
 817          * to be put in someone's domain map.
 818          * Also, explicitly disallow joining at certain troublesome
 819          * times (ie. during recovery). */
 820         if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
 821                 int bit = query->node_idx;
 822                 spin_lock(&dlm->spinlock);
 823
 824                 if (dlm->dlm_state == DLM_CTXT_NEW &&
 825                     dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
 826                         /*If this is a brand new context and we
 827                          * haven't started our join process yet, then
 828                          * the other node won the race. */
 829                         packet.code = JOIN_OK_NO_MAP;
 830                 } else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
 831                         /* Disallow parallel joins. */
 832                         packet.code = JOIN_DISALLOW;
 833                 } else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
 834                         mlog(0, "node %u trying to join, but recovery "
 835                              "is ongoing.\n", bit);
 836                         packet.code = JOIN_DISALLOW;
 837                 } else if (test_bit(bit, dlm->recovery_map)) {
 838                         mlog(0, "node %u trying to join, but it "
 839                              "still needs recovery.\n", bit);
 840                         packet.code = JOIN_DISALLOW;
 841                 } else if (test_bit(bit, dlm->domain_map)) {
 842                         mlog(0, "node %u trying to join, but it "
 843                              "is still in the domain! needs recovery?\n",
 844                              bit);
 845                         packet.code = JOIN_DISALLOW;
 846                 } else {
 847                         /* Alright we're fully a part of this domain
 848                          * so we keep some state as to who's joining
 849                          * and indicate to him that needs to be fixed
 850                          * up. */
 851
 852                         /* Make sure we speak compatible locking protocols.  */
 853                         if (dlm_query_join_proto_check("DLM", bit,
 854                                                        &dlm->dlm_locking_proto,
 855                                                        &query->dlm_proto)) {
 856                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 857                         } else if (dlm_query_join_proto_check("fs", bit,
 858                                                               &dlm->fs_locking_proto,
 859                                                               &query->fs_proto)) {
 860                                 packet.code = JOIN_PROTOCOL_MISMATCH;
 861                         } else {
 862                                 packet.dlm_minor = query->dlm_proto.pv_minor;
 863                                 packet.fs_minor = query->fs_proto.pv_minor;
 864                                 packet.code = JOIN_OK;
 865                                 __dlm_set_joining_node(dlm, query->node_idx);
 866                         }
 867                 }
 868
 869                 spin_unlock(&dlm->spinlock);
 870         }
 871 unlock_respond:
 872         spin_unlock(&dlm_domain_lock);
 873
 874 respond:
 875         mlog(0, "We respond with %u\n", packet.code);
 876
 877         dlm_query_join_packet_to_wire(&packet, &response);
 878         return response;
 879 }
 880
 881 static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
 882                                      void **ret_data)
 883 {
 884         struct dlm_assert_joined *assert;
 885         struct dlm_ctxt *dlm = NULL;
 886
 887         assert = (struct dlm_assert_joined *) msg->buf;
 888
 889         mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
 890                   assert->domain);
 891
 892         spin_lock(&dlm_domain_lock);
 893         dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
 894         /* XXX should we consider no dlm ctxt an error? */
 895         if (dlm) {
 896                 spin_lock(&dlm->spinlock);
 897
 898                 /* Alright, this node has officially joined our
 899                  * domain. Set him in the map and clean up our
 900                  * leftover join state. */
 901                 BUG_ON(dlm->joining_node != assert->node_idx);
 902                 set_bit(assert->node_idx, dlm->domain_map);
 903                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 904
 905                 printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
 906                        assert->node_idx, dlm->name);
 907                 __dlm_print_nodes(dlm);
 908
 909                 /* notify anything attached to the heartbeat events */
 910                 dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);
 911
 912                 spin_unlock(&dlm->spinlock);
 913         }
 914         spin_unlock(&dlm_domain_lock);
 915
 916         return 0;
 917 }
 918
 919 static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
 920                                    void **ret_data)
 921 {
 922         struct dlm_cancel_join *cancel;
 923         struct dlm_ctxt *dlm = NULL;
 924
 925         cancel = (struct dlm_cancel_join *) msg->buf;
 926
 927         mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
 928                   cancel->domain);
 929
 930         spin_lock(&dlm_domain_lock);
 931         dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);
 932
 933         if (dlm) {
 934                 spin_lock(&dlm->spinlock);
 935
 936                 /* Yikes, this guy wants to cancel his join. No
 937                  * problem, we simply cleanup our join state. */
 938                 BUG_ON(dlm->joining_node != cancel->node_idx);
 939                 __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
 940
 941                 spin_unlock(&dlm->spinlock);
 942         }
 943         spin_unlock(&dlm_domain_lock);
 944
 945         return 0;
 946 }
 947
 948 static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
 949                                     unsigned int node)
 950 {
 951         int status;
 952         struct dlm_cancel_join cancel_msg;
 953
 954         memset(&cancel_msg, 0, sizeof(cancel_msg));
 955         cancel_msg.node_idx = dlm->node_num;
 956         cancel_msg.name_len = strlen(dlm->name);
 957         memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);
 958
 959         status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
 960                                     &cancel_msg, sizeof(cancel_msg), node,
 961                                     NULL);
 962         if (status < 0) {
 963                 mlog_errno(status);
 964                 goto bail;
 965         }
 966
 967 bail:
 968         return status;
 969 }
 970
 971 /* map_size should be in bytes. */
 972 static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
 973                                  unsigned long *node_map,
 974                                  unsigned int map_size)
 975 {
 976         int status, tmpstat;
 977         unsigned int node;
 978
 979         if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
 980                          sizeof(unsigned long))) {
 981                 mlog(ML_ERROR,
 982                      "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
 983                      map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
 984                 return -EINVAL;
 985         }
 986
 987         status = 0;
 988         node = -1;
 989         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
 990                                      node + 1)) < O2NM_MAX_NODES) {
 991                 if (node == dlm->node_num)
 992                         continue;
 993
 994                 tmpstat = dlm_send_one_join_cancel(dlm, node);
 995                 if (tmpstat) {
 996                         mlog(ML_ERROR, "Error return %d cancelling join on "
 997                              "node %d\n", tmpstat, node);
 998                         if (!status)
 999                                 status = tmpstat;
1000                 }
1001         }
1002
1003         if (status)
1004                 mlog_errno(status);
1005         return status;
1006 }
1007
1008 static int dlm_request_join(struct dlm_ctxt *dlm,
1009                             int node,
1010                             enum dlm_query_join_response_code *response)
1011 {
1012         int status;
1013         struct dlm_query_join_request join_msg;
1014         struct dlm_query_join_packet packet;
1015         u32 join_resp;
1016
1017         mlog(0, "querying node %d\n", node);
1018
1019         memset(&join_msg, 0, sizeof(join_msg));
1020         join_msg.node_idx = dlm->node_num;
1021         join_msg.name_len = strlen(dlm->name);
1022         memcpy(join_msg.domain, dlm->name, join_msg.name_len);
1023         join_msg.dlm_proto = dlm->dlm_locking_proto;
1024         join_msg.fs_proto = dlm->fs_locking_proto;
1025
1026         /* copy live node map to join message */
1027         byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);
1028
1029         status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
1030                                     sizeof(join_msg), node,
1031                                     &join_resp);
1032         if (status < 0 && status != -ENOPROTOOPT) {
1033                 mlog_errno(status);
1034                 goto bail;
1035         }
1036         dlm_query_join_wire_to_packet(join_resp, &packet);
1037
1038         /* -ENOPROTOOPT from the net code means the other side isn't
1039             listening for our message type -- that's fine, it means
1040             his dlm isn't up, so we can consider him a 'yes' but not
1041             joined into the domain.  */
1042         if (status == -ENOPROTOOPT) {
1043                 status = 0;
1044                 *response = JOIN_OK_NO_MAP;
1045         } else if (packet.code == JOIN_DISALLOW ||
1046                    packet.code == JOIN_OK_NO_MAP) {
1047                 *response = packet.code;
1048         } else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
1049                 mlog(ML_NOTICE,
1050                      "This node requested DLM locking protocol %u.%u and "
1051                      "filesystem locking protocol %u.%u.  At least one of "
1052                      "the protocol versions on node %d is not compatible, "
1053                      "disconnecting\n",
1054                      dlm->dlm_locking_proto.pv_major,
1055                      dlm->dlm_locking_proto.pv_minor,
1056                      dlm->fs_locking_proto.pv_major,
1057                      dlm->fs_locking_proto.pv_minor,
1058                      node);
1059                 status = -EPROTO;
1060                 *response = packet.code;
1061         } else if (packet.code == JOIN_OK) {
1062                 *response = packet.code;
1063                 /* Use the same locking protocol as the remote node */
1064                 dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
1065                 dlm->fs_locking_proto.pv_minor = packet.fs_minor;
1066                 mlog(0,
1067                      "Node %d responds JOIN_OK with DLM locking protocol "
1068                      "%u.%u and fs locking protocol %u.%u\n",
1069                      node,
1070                      dlm->dlm_locking_proto.pv_major,
1071                      dlm->dlm_locking_proto.pv_minor,
1072                      dlm->fs_locking_proto.pv_major,
1073                      dlm->fs_locking_proto.pv_minor);
1074         } else {
1075                 status = -EINVAL;
1076                 mlog(ML_ERROR, "invalid response %d from node %u\n",
1077                      packet.code, node);
1078         }
1079
1080         mlog(0, "status %d, node %d response is %d\n", status, node,
1081              *response);
1082
1083 bail:
1084         return status;
1085 }
1086
1087 static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
1088                                     unsigned int node)
1089 {
1090         int status;
1091         struct dlm_assert_joined assert_msg;
1092
1093         mlog(0, "Sending join assert to node %u\n", node);
1094
1095         memset(&assert_msg, 0, sizeof(assert_msg));
1096         assert_msg.node_idx = dlm->node_num;
1097         assert_msg.name_len = strlen(dlm->name);
1098         memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);
1099
1100         status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1101                                     &assert_msg, sizeof(assert_msg), node,
1102                                     NULL);
1103         if (status < 0)
1104                 mlog_errno(status);
1105
1106         return status;
1107 }
1108
1109 static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
1110                                   unsigned long *node_map)
1111 {
1112         int status, node, live;
1113
1114         status = 0;
1115         node = -1;
1116         while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
1117                                      node + 1)) < O2NM_MAX_NODES) {
1118                 if (node == dlm->node_num)
1119                         continue;
1120
1121                 do {
1122                         /* It is very important that this message be
1123                          * received so we spin until either the node
1124                          * has died or it gets the message. */
1125                         status = dlm_send_one_join_assert(dlm, node);
1126
1127                         spin_lock(&dlm->spinlock);
1128                         live = test_bit(node, dlm->live_nodes_map);
1129                         spin_unlock(&dlm->spinlock);
1130
1131                         if (status) {
1132                                 mlog(ML_ERROR, "Error return %d asserting "
1133                                      "join on node %d\n", status, node);
1134
1135                                 /* give us some time between errors... */
1136                                 if (live)
1137                                         msleep(DLM_DOMAIN_BACKOFF_MS);
1138                         }
1139                 } while (status && live);
1140         }
1141 }
1142
1143 struct domain_join_ctxt {
1144         unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1145         unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
1146 };
1147
1148 static int dlm_should_restart_join(struct dlm_ctxt *dlm,
1149                                    struct domain_join_ctxt *ctxt,
1150                                    enum dlm_query_join_response_code response)
1151 {
1152         int ret;
1153
1154         if (response == JOIN_DISALLOW) {
1155                 mlog(0, "Latest response of disallow -- should restart\n");
1156                 return 1;
1157         }
1158
1159         spin_lock(&dlm->spinlock);
1160         /* For now, we restart the process if the node maps have
1161          * changed at all */
1162         ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
1163                      sizeof(dlm->live_nodes_map));
1164         spin_unlock(&dlm->spinlock);
1165
1166         if (ret)
1167                 mlog(0, "Node maps changed -- should restart\n");
1168
1169         return ret;
1170 }
1171
1172 static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
1173 {
1174         int status = 0, tmpstat, node;
1175         struct domain_join_ctxt *ctxt;
1176         enum dlm_query_join_response_code response = JOIN_DISALLOW;
1177
1178         mlog_entry("%p", dlm);
1179
1180         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
1181         if (!ctxt) {
1182                 status = -ENOMEM;
1183                 mlog_errno(status);
1184                 goto bail;
1185         }
1186
1187         /* group sem locking should work for us here -- we're already
1188          * registered for heartbeat events so filling this should be
1189          * atomic wrt getting those handlers called. */
1190         o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));
1191
1192         spin_lock(&dlm->spinlock);
1193         memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));
1194
1195         __dlm_set_joining_node(dlm, dlm->node_num);
1196
1197         spin_unlock(&dlm->spinlock);
1198
1199         node = -1;
1200         while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
1201                                      node + 1)) < O2NM_MAX_NODES) {
1202                 if (node == dlm->node_num)
1203                         continue;
1204
1205                 status = dlm_request_join(dlm, node, &response);
1206                 if (status < 0) {
1207                         mlog_errno(status);
1208                         goto bail;
1209                 }
1210
1211                 /* Ok, either we got a response or the node doesn't have a
1212                  * dlm up. */
1213                 if (response == JOIN_OK)
1214                         set_bit(node, ctxt->yes_resp_map);
1215
1216                 if (dlm_should_restart_join(dlm, ctxt, response)) {
1217                         status = -EAGAIN;
1218                         goto bail;
1219                 }
1220         }
1221
1222         mlog(0, "Yay, done querying nodes!\n");
1223
1224         /* Yay, everyone agree's we can join the domain. My domain is
1225          * comprised of all nodes who were put in the
1226          * yes_resp_map. Copy that into our domain map and send a join
1227          * assert message to clean up everyone elses state. */
1228         spin_lock(&dlm->spinlock);
1229         memcpy(dlm->domain_map, ctxt->yes_resp_map,
1230                sizeof(ctxt->yes_resp_map));
1231         set_bit(dlm->node_num, dlm->domain_map);
1232         spin_unlock(&dlm->spinlock);
1233
1234         dlm_send_join_asserts(dlm, ctxt->yes_resp_map);
1235
1236         /* Joined state *must* be set before the joining node
1237          * information, otherwise the query_join handler may read no
1238          * current joiner but a state of NEW and tell joining nodes
1239          * we're not in the domain. */
1240         spin_lock(&dlm_domain_lock);
1241         dlm->dlm_state = DLM_CTXT_JOINED;
1242         dlm->num_joins++;
1243         spin_unlock(&dlm_domain_lock);
1244
1245 bail:
1246         spin_lock(&dlm->spinlock);
1247         __dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
1248         if (!status)
1249                 __dlm_print_nodes(dlm);
1250         spin_unlock(&dlm->spinlock);
1251
1252         if (ctxt) {
1253                 /* Do we need to send a cancel message to any nodes? */
1254                 if (status < 0) {
1255                         tmpstat = dlm_send_join_cancels(dlm,
1256                                                         ctxt->yes_resp_map,
1257                                                         sizeof(ctxt->yes_resp_map));
1258                         if (tmpstat < 0)
1259                                 mlog_errno(tmpstat);
1260                 }
1261                 kfree(ctxt);
1262         }
1263
1264         mlog(0, "returning %d\n", status);
1265         return status;
1266 }
1267
1268 static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
1269 {
1270         o2hb_unregister_callback(NULL, &dlm->dlm_hb_up);
1271         o2hb_unregister_callback(NULL, &dlm->dlm_hb_down);
1272         o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
1273 }
1274
1275 static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
1276 {
1277         int status;
1278
1279         mlog(0, "registering handlers.\n");
1280
1281         o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
1282                             dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
1283         status = o2hb_register_callback(NULL, &dlm->dlm_hb_down);
1284         if (status)
1285                 goto bail;
1286
1287         o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
1288                             dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
1289         status = o2hb_register_callback(NULL, &dlm->dlm_hb_up);
1290         if (status)
1291                 goto bail;
1292
1293         status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
1294                                         sizeof(struct dlm_master_request),
1295                                         dlm_master_request_handler,
1296                                         dlm, NULL, &dlm->dlm_domain_handlers);
1297         if (status)
1298                 goto bail;
1299
1300         status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
1301                                         sizeof(struct dlm_assert_master),
1302                                         dlm_assert_master_handler,
1303                                         dlm, dlm_assert_master_post_handler,
1304                                         &dlm->dlm_domain_handlers);
1305         if (status)
1306                 goto bail;
1307
1308         status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
1309                                         sizeof(struct dlm_create_lock),
1310                                         dlm_create_lock_handler,
1311                                         dlm, NULL, &dlm->dlm_domain_handlers);
1312         if (status)
1313                 goto bail;
1314
1315         status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
1316                                         DLM_CONVERT_LOCK_MAX_LEN,
1317                                         dlm_convert_lock_handler,
1318                                         dlm, NULL, &dlm->dlm_domain_handlers);
1319         if (status)
1320                 goto bail;
1321
1322         status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
1323                                         DLM_UNLOCK_LOCK_MAX_LEN,
1324                                         dlm_unlock_lock_handler,
1325                                         dlm, NULL, &dlm->dlm_domain_handlers);
1326         if (status)
1327                 goto bail;
1328
1329         status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
1330                                         DLM_PROXY_AST_MAX_LEN,
1331                                         dlm_proxy_ast_handler,
1332                                         dlm, NULL, &dlm->dlm_domain_handlers);
1333         if (status)
1334                 goto bail;
1335
1336         status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
1337                                         sizeof(struct dlm_exit_domain),
1338                                         dlm_exit_domain_handler,
1339                                         dlm, NULL, &dlm->dlm_domain_handlers);
1340         if (status)
1341                 goto bail;
1342
1343         status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
1344                                         sizeof(struct dlm_deref_lockres),
1345                                         dlm_deref_lockres_handler,
1346                                         dlm, NULL, &dlm->dlm_domain_handlers);
1347         if (status)
1348                 goto bail;
1349
1350         status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
1351                                         sizeof(struct dlm_migrate_request),
1352                                         dlm_migrate_request_handler,
1353                                         dlm, NULL, &dlm->dlm_domain_handlers);
1354         if (status)
1355                 goto bail;
1356
1357         status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
1358                                         DLM_MIG_LOCKRES_MAX_LEN,
1359                                         dlm_mig_lockres_handler,
1360                                         dlm, NULL, &dlm->dlm_domain_handlers);
1361         if (status)
1362                 goto bail;
1363
1364         status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
1365                                         sizeof(struct dlm_master_requery),
1366                                         dlm_master_requery_handler,
1367                                         dlm, NULL, &dlm->dlm_domain_handlers);
1368         if (status)
1369                 goto bail;
1370
1371         status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
1372                                         sizeof(struct dlm_lock_request),
1373                                         dlm_request_all_locks_handler,
1374                                         dlm, NULL, &dlm->dlm_domain_handlers);
1375         if (status)
1376                 goto bail;
1377
1378         status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
1379                                         sizeof(struct dlm_reco_data_done),
1380                                         dlm_reco_data_done_handler,
1381                                         dlm, NULL, &dlm->dlm_domain_handlers);
1382         if (status)
1383                 goto bail;
1384
1385         status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
1386                                         sizeof(struct dlm_begin_reco),
1387                                         dlm_begin_reco_handler,
1388                                         dlm, NULL, &dlm->dlm_domain_handlers);
1389         if (status)
1390                 goto bail;
1391
1392         status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
1393                                         sizeof(struct dlm_finalize_reco),
1394                                         dlm_finalize_reco_handler,
1395                                         dlm, NULL, &dlm->dlm_domain_handlers);
1396         if (status)
1397                 goto bail;
1398
1399 bail:
1400         if (status)
1401                 dlm_unregister_domain_handlers(dlm);
1402
1403         return status;
1404 }
1405
1406 static int dlm_join_domain(struct dlm_ctxt *dlm)
1407 {
1408         int status;
1409         unsigned int backoff;
1410         unsigned int total_backoff = 0;
1411
1412         BUG_ON(!dlm);
1413
1414         mlog(0, "Join domain %s\n", dlm->name);
1415
1416         status = dlm_register_domain_handlers(dlm);
1417         if (status) {
1418                 mlog_errno(status);
1419                 goto bail;
1420         }
1421
1422         status = dlm_debug_init(dlm);
1423         if (status < 0) {
1424                 mlog_errno(status);
1425                 goto bail;
1426         }
1427
1428         status = dlm_launch_thread(dlm);
1429         if (status < 0) {
1430                 mlog_errno(status);
1431                 goto bail;
1432         }
1433
1434         status = dlm_launch_recovery_thread(dlm);
1435         if (status < 0) {
1436                 mlog_errno(status);
1437                 goto bail;
1438         }
1439
1440         dlm->dlm_worker = create_singlethread_workqueue("dlm_wq");
1441         if (!dlm->dlm_worker) {
1442                 status = -ENOMEM;
1443                 mlog_errno(status);
1444                 goto bail;
1445         }
1446
1447         do {
1448                 status = dlm_try_to_join_domain(dlm);
1449
1450                 /* If we're racing another node to the join, then we
1451                  * need to back off temporarily and let them
1452                  * complete. */
1453 #define DLM_JOIN_TIMEOUT_MSECS  90000
1454                 if (status == -EAGAIN) {
1455                         if (signal_pending(current)) {
1456                                 status = -ERESTARTSYS;
1457                                 goto bail;
1458                         }
1459
1460                         if (total_backoff >
1461                             msecs_to_jiffies(DLM_JOIN_TIMEOUT_MSECS)) {
1462                                 status = -ERESTARTSYS;
1463                                 mlog(ML_NOTICE, "Timed out joining dlm domain "
1464                                      "%s after %u msecs\n", dlm->name,
1465                                      jiffies_to_msecs(total_backoff));
1466                                 goto bail;
1467                         }
1468
1469                         /*
1470                          * <chip> After you!
1471                          * <dale> No, after you!
1472                          * <chip> I insist!
1473                          * <dale> But you first!
1474                          * ...
1475                          */
1476                         backoff = (unsigned int)(jiffies & 0x3);
1477                         backoff *= DLM_DOMAIN_BACKOFF_MS;
1478                         total_backoff += backoff;
1479                         mlog(0, "backoff %d\n", backoff);
1480                         msleep(backoff);
1481                 }
1482         } while (status == -EAGAIN);
1483
1484         if (status < 0) {
1485                 mlog_errno(status);
1486                 goto bail;
1487         }
1488
1489         status = 0;
1490 bail:
1491         wake_up(&dlm_domain_events);
1492
1493         if (status) {
1494                 dlm_unregister_domain_handlers(dlm);
1495                 dlm_debug_shutdown(dlm);
1496                 dlm_complete_thread(dlm);
1497                 dlm_complete_recovery_thread(dlm);
1498                 dlm_destroy_dlm_worker(dlm);
1499         }
1500
1501         return status;
1502 }
1503
1504 static struct dlm_ctxt *dlm_alloc_ctxt(const char *domain,
1505                                 u32 key)
1506 {
1507         int i;
1508         int ret;
1509         struct dlm_ctxt *dlm = NULL;
1510
1511         dlm = kzalloc(sizeof(*dlm), GFP_KERNEL);
1512         if (!dlm) {
1513                 mlog_errno(-ENOMEM);
1514                 goto leave;
1515         }
1516
1517         dlm->name = kmalloc(strlen(domain) + 1, GFP_KERNEL);
1518         if (dlm->name == NULL) {
1519                 mlog_errno(-ENOMEM);
1520                 kfree(dlm);
1521                 dlm = NULL;
1522                 goto leave;
1523         }
1524
1525         dlm->lockres_hash = (struct hlist_head **)dlm_alloc_pagevec(DLM_HASH_PAGES);
1526         if (!dlm->lockres_hash) {
1527                 mlog_errno(-ENOMEM);
1528                 kfree(dlm->name);
1529                 kfree(dlm);
1530                 dlm = NULL;
1531                 goto leave;
1532         }
1533
1534         for (i = 0; i < DLM_HASH_BUCKETS; i++)
1535                 INIT_HLIST_HEAD(dlm_lockres_hash(dlm, i));
1536
1537         strcpy(dlm->name, domain);
1538         dlm->key = key;
1539         dlm->node_num = o2nm_this_node();
1540
1541         ret = dlm_create_debugfs_subroot(dlm);
1542         if (ret < 0) {
1543                 dlm_free_pagevec((void **)dlm->lockres_hash, DLM_HASH_PAGES);
1544                 kfree(dlm->name);
1545                 kfree(dlm);
1546                 dlm = NULL;
1547                 goto leave;
1548         }
1549
1550         spin_lock_init(&dlm->spinlock);
1551         spin_lock_init(&dlm->master_lock);
1552         spin_lock_init(&dlm->ast_lock);
1553         spin_lock_init(&dlm->track_lock);
1554         INIT_LIST_HEAD(&dlm->list);
1555         INIT_LIST_HEAD(&dlm->dirty_list);
1556         INIT_LIST_HEAD(&dlm->reco.resources);
1557         INIT_LIST_HEAD(&dlm->reco.received);
1558         INIT_LIST_HEAD(&dlm->reco.node_data);
1559         INIT_LIST_HEAD(&dlm->purge_list);
1560         INIT_LIST_HEAD(&dlm->dlm_domain_handlers);
1561         INIT_LIST_HEAD(&dlm->tracking_list);
1562         dlm->reco.state = 0;
1563
1564         INIT_LIST_HEAD(&dlm->pending_asts);
1565         INIT_LIST_HEAD(&dlm->pending_basts);
1566
1567         mlog(0, "dlm->recovery_map=%p, &(dlm->recovery_map[0])=%p\n",
1568                   dlm->recovery_map, &(dlm->recovery_map[0]));
1569
1570         memset(dlm->recovery_map, 0, sizeof(dlm->recovery_map));
1571         memset(dlm->live_nodes_map, 0, sizeof(dlm->live_nodes_map));
1572         memset(dlm->domain_map, 0, sizeof(dlm->domain_map));
1573
1574         dlm->dlm_thread_task = NULL;
1575         dlm->dlm_reco_thread_task = NULL;
1576         dlm->dlm_worker = NULL;
1577         init_waitqueue_head(&dlm->dlm_thread_wq);
1578         init_waitqueue_head(&dlm->dlm_reco_thread_wq);
1579         init_waitqueue_head(&dlm->reco.event);
1580         init_waitqueue_head(&dlm->ast_wq);
1581         init_waitqueue_head(&dlm->migration_wq);
1582         INIT_LIST_HEAD(&dlm->master_list);
1583         INIT_LIST_HEAD(&dlm->mle_hb_events);
1584
1585         dlm->joining_node = DLM_LOCK_RES_OWNER_UNKNOWN;
1586         init_waitqueue_head(&dlm->dlm_join_events);
1587
1588         dlm->reco.new_master = O2NM_INVALID_NODE_NUM;
1589         dlm->reco.dead_node = O2NM_INVALID_NODE_NUM;
1590         atomic_set(&dlm->local_resources, 0);
1591         atomic_set(&dlm->remote_resources, 0);
1592         atomic_set(&dlm->unknown_resources, 0);
1593
1594         spin_lock_init(&dlm->work_lock);
1595         INIT_LIST_HEAD(&dlm->work_list);
1596         INIT_WORK(&dlm->dispatched_work, dlm_dispatch_work);
1597
1598         kref_init(&dlm->dlm_refs);
1599         dlm->dlm_state = DLM_CTXT_NEW;
1600
1601         INIT_LIST_HEAD(&dlm->dlm_eviction_callbacks);
1602
1603         mlog(0, "context init: refcount %u\n",
1604                   atomic_read(&dlm->dlm_refs.refcount));
1605
1606 leave:
1607         return dlm;
1608 }
1609
1610 /*
1611  * Compare a requested locking protocol version against the current one.
1612  *
1613  * If the major numbers are different, they are incompatible.
1614  * If the current minor is greater than the request, they are incompatible.
1615  * If the current minor is less than or equal to the request, they are
1616  * compatible, and the requester should run at the current minor version.
1617  */
1618 static int dlm_protocol_compare(struct dlm_protocol_version *existing,
1619                                 struct dlm_protocol_version *request)
1620 {
1621         if (existing->pv_major != request->pv_major)
1622                 return 1;
1623
1624         if (existing->pv_minor > request->pv_minor)
1625                 return 1;
1626
1627         if (existing->pv_minor < request->pv_minor)
1628                 request->pv_minor = existing->pv_minor;
1629
1630         return 0;
1631 }
1632
1633 /*
1634  * dlm_register_domain: one-time setup per "domain".
1635  *
1636  * The filesystem passes in the requested locking version via proto.
1637  * If registration was successful, proto will contain the negotiated
1638  * locking protocol.
1639  */
1640 struct dlm_ctxt * dlm_register_domain(const char *domain,
1641                                u32 key,
1642                                struct dlm_protocol_version *fs_proto)
1643 {
1644         int ret;
1645         struct dlm_ctxt *dlm = NULL;
1646         struct dlm_ctxt *new_ctxt = NULL;
1647
1648         if (strlen(domain) > O2NM_MAX_NAME_LEN) {
1649                 ret = -ENAMETOOLONG;
1650                 mlog(ML_ERROR, "domain name length too long\n");
1651                 goto leave;
1652         }
1653
1654         if (!o2hb_check_local_node_heartbeating()) {
1655                 mlog(ML_ERROR, "the local node has not been configured, or is "
1656                      "not heartbeating\n");
1657                 ret = -EPROTO;
1658                 goto leave;
1659         }
1660
1661         mlog(0, "register called for domain \"%s\"\n", domain);
1662
1663 retry:
1664         dlm = NULL;
1665         if (signal_pending(current)) {
1666                 ret = -ERESTARTSYS;
1667                 mlog_errno(ret);
1668                 goto leave;
1669         }
1670
1671         spin_lock(&dlm_domain_lock);
1672
1673         dlm = __dlm_lookup_domain(domain);
1674         if (dlm) {
1675                 if (dlm->dlm_state != DLM_CTXT_JOINED) {
1676                         spin_unlock(&dlm_domain_lock);
1677
1678                         mlog(0, "This ctxt is not joined yet!\n");
1679                         wait_event_interruptible(dlm_domain_events,
1680                                                  dlm_wait_on_domain_helper(
1681                                                          domain));
1682                         goto retry;
1683                 }
1684
1685                 if (dlm_protocol_compare(&dlm->fs_locking_proto, fs_proto)) {
1686                         mlog(ML_ERROR,
1687                              "Requested locking protocol version is not "
1688                              "compatible with already registered domain "
1689                              "\"%s\"\n", domain);
1690                         ret = -EPROTO;
1691                         goto leave;
1692                 }
1693
1694                 __dlm_get(dlm);
1695                 dlm->num_joins++;
1696
1697                 spin_unlock(&dlm_domain_lock);
1698
1699                 ret = 0;
1700                 goto leave;
1701         }
1702
1703         /* doesn't exist */
1704         if (!new_ctxt) {
1705                 spin_unlock(&dlm_domain_lock);
1706
1707                 new_ctxt = dlm_alloc_ctxt(domain, key);
1708                 if (new_ctxt)
1709                         goto retry;
1710
1711                 ret = -ENOMEM;
1712                 mlog_errno(ret);
1713                 goto leave;
1714         }
1715
1716         /* a little variable switch-a-roo here... */
1717         dlm = new_ctxt;
1718         new_ctxt = NULL;
1719
1720         /* add the new domain */
1721         list_add_tail(&dlm->list, &dlm_domains);
1722         spin_unlock(&dlm_domain_lock);
1723
1724         /*
1725          * Pass the locking protocol version into the join.  If the join
1726          * succeeds, it will have the negotiated protocol set.
1727          */
1728         dlm->dlm_locking_proto = dlm_protocol;
1729         dlm->fs_locking_proto = *fs_proto;
1730
1731         ret = dlm_join_domain(dlm);
1732         if (ret) {
1733                 mlog_errno(ret);
1734                 dlm_put(dlm);
1735                 goto leave;
1736         }
1737
1738         /* Tell the caller what locking protocol we negotiated */
1739         *fs_proto = dlm->fs_locking_proto;
1740
1741         ret = 0;
1742 leave:
1743         if (new_ctxt)
1744                 dlm_free_ctxt_mem(new_ctxt);
1745
1746         if (ret < 0)
1747                 dlm = ERR_PTR(ret);
1748
1749         return dlm;
1750 }
1751 EXPORT_SYMBOL_GPL(dlm_register_domain);
1752
1753 static LIST_HEAD(dlm_join_handlers);
1754
1755 static void dlm_unregister_net_handlers(void)
1756 {
1757         o2net_unregister_handler_list(&dlm_join_handlers);
1758 }
1759
1760 static int dlm_register_net_handlers(void)
1761 {
1762         int status = 0;
1763
1764         status = o2net_register_handler(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY,
1765                                         sizeof(struct dlm_query_join_request),
1766                                         dlm_query_join_handler,
1767                                         NULL, NULL, &dlm_join_handlers);
1768         if (status)
1769                 goto bail;
1770
1771         status = o2net_register_handler(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
1772                                         sizeof(struct dlm_assert_joined),
1773                                         dlm_assert_joined_handler,
1774                                         NULL, NULL, &dlm_join_handlers);
1775         if (status)
1776                 goto bail;
1777
1778         status = o2net_register_handler(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
1779                                         sizeof(struct dlm_cancel_join),
1780                                         dlm_cancel_join_handler,
1781                                         NULL, NULL, &dlm_join_handlers);
1782
1783 bail:
1784         if (status < 0)
1785                 dlm_unregister_net_handlers();
1786
1787         return status;
1788 }
1789
1790 /* Domain eviction callback handling.
1791  *
1792  * The file system requires notification of node death *before* the
1793  * dlm completes it's recovery work, otherwise it may be able to
1794  * acquire locks on resources requiring recovery. Since the dlm can
1795  * evict a node from it's domain *before* heartbeat fires, a similar
1796  * mechanism is required. */
1797
1798 /* Eviction is not expected to happen often, so a per-domain lock is
1799  * not necessary. Eviction callbacks are allowed to sleep for short
1800  * periods of time. */
1801 static DECLARE_RWSEM(dlm_callback_sem);
1802
1803 void dlm_fire_domain_eviction_callbacks(struct dlm_ctxt *dlm,
1804                                         int node_num)
1805 {
1806         struct list_head *iter;
1807         struct dlm_eviction_cb *cb;
1808
1809         down_read(&dlm_callback_sem);
1810         list_for_each(iter, &dlm->dlm_eviction_callbacks) {
1811                 cb = list_entry(iter, struct dlm_eviction_cb, ec_item);
1812
1813                 cb->ec_func(node_num, cb->ec_data);
1814         }
1815         up_read(&dlm_callback_sem);
1816 }
1817
1818 void dlm_setup_eviction_cb(struct dlm_eviction_cb *cb,
1819                            dlm_eviction_func *f,
1820                            void *data)
1821 {
1822         INIT_LIST_HEAD(&cb->ec_item);
1823         cb->ec_func = f;
1824         cb->ec_data = data;
1825 }
1826 EXPORT_SYMBOL_GPL(dlm_setup_eviction_cb);
1827
1828 void dlm_register_eviction_cb(struct dlm_ctxt *dlm,
1829                               struct dlm_eviction_cb *cb)
1830 {
1831         down_write(&dlm_callback_sem);
1832         list_add_tail(&cb->ec_item, &dlm->dlm_eviction_callbacks);
1833         up_write(&dlm_callback_sem);
1834 }
1835 EXPORT_SYMBOL_GPL(dlm_register_eviction_cb);
1836
1837 void dlm_unregister_eviction_cb(struct dlm_eviction_cb *cb)
1838 {
1839         down_write(&dlm_callback_sem);
1840         list_del_init(&cb->ec_item);
1841         up_write(&dlm_callback_sem);
1842 }
1843 EXPORT_SYMBOL_GPL(dlm_unregister_eviction_cb);
1844
1845 static int __init dlm_init(void)
1846 {
1847         int status;
1848
1849         dlm_print_version();
1850
1851         status = dlm_init_mle_cache();
1852         if (status) {
1853                 mlog(ML_ERROR, "Could not create o2dlm_mle slabcache\n");
1854                 goto error;
1855         }
1856
1857         status = dlm_init_master_caches();
1858         if (status) {
1859                 mlog(ML_ERROR, "Could not create o2dlm_lockres and "
1860                      "o2dlm_lockname slabcaches\n");
1861                 goto error;
1862         }
1863
1864         status = dlm_init_lock_cache();
1865         if (status) {
1866                 mlog(ML_ERROR, "Count not create o2dlm_lock slabcache\n");
1867                 goto error;
1868         }
1869
1870         status = dlm_register_net_handlers();
1871         if (status) {
1872                 mlog(ML_ERROR, "Unable to register network handlers\n");
1873                 goto error;
1874         }
1875
1876         status = dlm_create_debugfs_root();
1877         if (status)
1878                 goto error;
1879
1880         return 0;
1881 error:
1882         dlm_unregister_net_handlers();
1883         dlm_destroy_lock_cache();
1884         dlm_destroy_master_caches();
1885         dlm_destroy_mle_cache();
1886         return -1;
1887 }
1888
1889 static void __exit dlm_exit (void)
1890 {
1891         dlm_destroy_debugfs_root();
1892         dlm_unregister_net_handlers();
1893         dlm_destroy_lock_cache();
1894         dlm_destroy_master_caches();
1895         dlm_destroy_mle_cache();
1896 }
1897
1898 MODULE_AUTHOR("Oracle");
1899 MODULE_LICENSE("GPL");
1900
1901 module_init(dlm_init);
1902 module_exit(dlm_exit);