fs/ocfs2/dlm/dlmdomain.c

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