fs/ocfs2/dlm/dlmcommon.h

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * dlmcommon.h
   5  *
   6  * Copyright (C) 2004 Oracle.  All rights reserved.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2 of the License, or (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public
  19  * License along with this program; if not, write to the
  20  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  21  * Boston, MA 021110-1307, USA.
  22  *
  23  */
  24
  25 #ifndef DLMCOMMON_H
  26 #define DLMCOMMON_H
  27
  28 #include <linux/kref.h>
  29
  30 #define DLM_HB_NODE_DOWN_PRI     (0xf000000)
  31 #define DLM_HB_NODE_UP_PRI       (0x8000000)
  32
  33 #define DLM_LOCKID_NAME_MAX    32
  34
  35 #define DLM_DOMAIN_NAME_MAX_LEN    255
  36 #define DLM_LOCK_RES_OWNER_UNKNOWN     O2NM_MAX_NODES
  37 #define DLM_THREAD_SHUFFLE_INTERVAL    5     // flush everything every 5 passes
  38 #define DLM_THREAD_MS                  200   // flush at least every 200 ms
  39
  40 #define DLM_HASH_SIZE_DEFAULT   (1 << 17)
  41 #if DLM_HASH_SIZE_DEFAULT < PAGE_SIZE
  42 # define DLM_HASH_PAGES         1
  43 #else
  44 # define DLM_HASH_PAGES         (DLM_HASH_SIZE_DEFAULT / PAGE_SIZE)
  45 #endif
  46 #define DLM_BUCKETS_PER_PAGE    (PAGE_SIZE / sizeof(struct hlist_head))
  47 #define DLM_HASH_BUCKETS        (DLM_HASH_PAGES * DLM_BUCKETS_PER_PAGE)
  48
  49 /* Intended to make it easier for us to switch out hash functions */
  50 #define dlm_lockid_hash(_n, _l) full_name_hash(_n, _l)
  51
  52 enum dlm_mle_type {
  53         DLM_MLE_BLOCK = 0,
  54         DLM_MLE_MASTER = 1,
  55         DLM_MLE_MIGRATION = 2,
  56         DLM_MLE_NUM_TYPES = 3,
  57 };
  58
  59 struct dlm_master_list_entry {
  60         struct hlist_node master_hash_node;
  61         struct list_head hb_events;
  62         struct dlm_ctxt *dlm;
  63         spinlock_t spinlock;
  64         wait_queue_head_t wq;
  65         atomic_t woken;
  66         struct kref mle_refs;
  67         int inuse;
  68         unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  69         unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  70         unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  71         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  72         u8 master;
  73         u8 new_master;
  74         enum dlm_mle_type type;
  75         struct o2hb_callback_func mle_hb_up;
  76         struct o2hb_callback_func mle_hb_down;
  77         struct dlm_lock_resource *mleres;
  78         unsigned char mname[DLM_LOCKID_NAME_MAX];
  79         unsigned int mnamelen;
  80         unsigned int mnamehash;
  81 };
  82
  83 enum dlm_ast_type {
  84         DLM_AST = 0,
  85         DLM_BAST = 1,
  86         DLM_ASTUNLOCK = 2,
  87 };
  88
  89
  90 #define LKM_VALID_FLAGS (LKM_VALBLK | LKM_CONVERT | LKM_UNLOCK | \
  91                          LKM_CANCEL | LKM_INVVALBLK | LKM_FORCE | \
  92                          LKM_RECOVERY | LKM_LOCAL | LKM_NOQUEUE)
  93
  94 #define DLM_RECOVERY_LOCK_NAME       "$RECOVERY"
  95 #define DLM_RECOVERY_LOCK_NAME_LEN   9
  96
  97 static inline int dlm_is_recovery_lock(const char *lock_name, int name_len)
  98 {
  99         if (name_len == DLM_RECOVERY_LOCK_NAME_LEN &&
 100             memcmp(lock_name, DLM_RECOVERY_LOCK_NAME, name_len)==0)
 101                 return 1;
 102         return 0;
 103 }
 104
 105 #define DLM_RECO_STATE_ACTIVE    0x0001
 106 #define DLM_RECO_STATE_FINALIZE  0x0002
 107
 108 struct dlm_recovery_ctxt
 109 {
 110         struct list_head resources;
 111         struct list_head received;
 112         struct list_head node_data;
 113         u8  new_master;
 114         u8  dead_node;
 115         u16 state;
 116         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 117         wait_queue_head_t event;
 118 };
 119
 120 enum dlm_ctxt_state {
 121         DLM_CTXT_NEW = 0,
 122         DLM_CTXT_JOINED = 1,
 123         DLM_CTXT_IN_SHUTDOWN = 2,
 124         DLM_CTXT_LEAVING = 3,
 125 };
 126
 127 struct dlm_ctxt
 128 {
 129         struct list_head list;
 130         struct hlist_head **lockres_hash;
 131         struct list_head dirty_list;
 132         struct list_head purge_list;
 133         struct list_head pending_asts;
 134         struct list_head pending_basts;
 135         struct list_head tracking_list;
 136         unsigned int purge_count;
 137         spinlock_t spinlock;
 138         spinlock_t ast_lock;
 139         spinlock_t track_lock;
 140         char *name;
 141         u8 node_num;
 142         u32 key;
 143         u8  joining_node;
 144         wait_queue_head_t dlm_join_events;
 145         unsigned long live_nodes_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 146         unsigned long domain_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 147         unsigned long recovery_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 148         struct dlm_recovery_ctxt reco;
 149         spinlock_t master_lock;
 150         struct hlist_head **master_hash;
 151         struct list_head mle_hb_events;
 152
 153         /* these give a really vague idea of the system load */
 154         atomic_t mle_tot_count[DLM_MLE_NUM_TYPES];
 155         atomic_t mle_cur_count[DLM_MLE_NUM_TYPES];
 156         atomic_t res_tot_count;
 157         atomic_t res_cur_count;
 158
 159         struct dlm_debug_ctxt *dlm_debug_ctxt;
 160         struct dentry *dlm_debugfs_subroot;
 161
 162         /* NOTE: Next three are protected by dlm_domain_lock */
 163         struct kref dlm_refs;
 164         enum dlm_ctxt_state dlm_state;
 165         unsigned int num_joins;
 166
 167         struct o2hb_callback_func dlm_hb_up;
 168         struct o2hb_callback_func dlm_hb_down;
 169         struct task_struct *dlm_thread_task;
 170         struct task_struct *dlm_reco_thread_task;
 171         struct workqueue_struct *dlm_worker;
 172         wait_queue_head_t dlm_thread_wq;
 173         wait_queue_head_t dlm_reco_thread_wq;
 174         wait_queue_head_t ast_wq;
 175         wait_queue_head_t migration_wq;
 176
 177         struct work_struct dispatched_work;
 178         struct list_head work_list;
 179         spinlock_t work_lock;
 180         struct list_head dlm_domain_handlers;
 181         struct list_head        dlm_eviction_callbacks;
 182
 183         /* The filesystem specifies this at domain registration.  We
 184          * cache it here to know what to tell other nodes. */
 185         struct dlm_protocol_version fs_locking_proto;
 186         /* This is the inter-dlm communication version */
 187         struct dlm_protocol_version dlm_locking_proto;
 188 };
 189
 190 static inline struct hlist_head *dlm_lockres_hash(struct dlm_ctxt *dlm, unsigned i)
 191 {
 192         return dlm->lockres_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES] + (i % DLM_BUCKETS_PER_PAGE);
 193 }
 194
 195 static inline struct hlist_head *dlm_master_hash(struct dlm_ctxt *dlm,
 196                                                  unsigned i)
 197 {
 198         return dlm->master_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES] +
 199                         (i % DLM_BUCKETS_PER_PAGE);
 200 }
 201
 202 /* these keventd work queue items are for less-frequently
 203  * called functions that cannot be directly called from the
 204  * net message handlers for some reason, usually because
 205  * they need to send net messages of their own. */
 206 void dlm_dispatch_work(struct work_struct *work);
 207
 208 struct dlm_lock_resource;
 209 struct dlm_work_item;
 210
 211 typedef void (dlm_workfunc_t)(struct dlm_work_item *, void *);
 212
 213 struct dlm_request_all_locks_priv
 214 {
 215         u8 reco_master;
 216         u8 dead_node;
 217 };
 218
 219 struct dlm_mig_lockres_priv
 220 {
 221         struct dlm_lock_resource *lockres;
 222         u8 real_master;
 223         u8 extra_ref;
 224 };
 225
 226 struct dlm_assert_master_priv
 227 {
 228         struct dlm_lock_resource *lockres;
 229         u8 request_from;
 230         u32 flags;
 231         unsigned ignore_higher:1;
 232 };
 233
 234 struct dlm_deref_lockres_priv
 235 {
 236         struct dlm_lock_resource *deref_res;
 237         u8 deref_node;
 238 };
 239
 240 struct dlm_work_item
 241 {
 242         struct list_head list;
 243         dlm_workfunc_t *func;
 244         struct dlm_ctxt *dlm;
 245         void *data;
 246         union {
 247                 struct dlm_request_all_locks_priv ral;
 248                 struct dlm_mig_lockres_priv ml;
 249                 struct dlm_assert_master_priv am;
 250                 struct dlm_deref_lockres_priv dl;
 251         } u;
 252 };
 253
 254 static inline void dlm_init_work_item(struct dlm_ctxt *dlm,
 255                                       struct dlm_work_item *i,
 256                                       dlm_workfunc_t *f, void *data)
 257 {
 258         memset(i, 0, sizeof(*i));
 259         i->func = f;
 260         INIT_LIST_HEAD(&i->list);
 261         i->data = data;
 262         i->dlm = dlm;  /* must have already done a dlm_grab on this! */
 263 }
 264
 265
 266
 267 static inline void __dlm_set_joining_node(struct dlm_ctxt *dlm,
 268                                           u8 node)
 269 {
 270         assert_spin_locked(&dlm->spinlock);
 271
 272         dlm->joining_node = node;
 273         wake_up(&dlm->dlm_join_events);
 274 }
 275
 276 #define DLM_LOCK_RES_UNINITED             0x00000001
 277 #define DLM_LOCK_RES_RECOVERING           0x00000002
 278 #define DLM_LOCK_RES_READY                0x00000004
 279 #define DLM_LOCK_RES_DIRTY                0x00000008
 280 #define DLM_LOCK_RES_IN_PROGRESS          0x00000010
 281 #define DLM_LOCK_RES_MIGRATING            0x00000020
 282 #define DLM_LOCK_RES_DROPPING_REF         0x00000040
 283 #define DLM_LOCK_RES_BLOCK_DIRTY          0x00001000
 284 #define DLM_LOCK_RES_SETREF_INPROG        0x00002000
 285
 286 /* max milliseconds to wait to sync up a network failure with a node death */
 287 #define DLM_NODE_DEATH_WAIT_MAX (5 * 1000)
 288
 289 #define DLM_PURGE_INTERVAL_MS   (8 * 1000)
 290
 291 struct dlm_lock_resource
 292 {
 293         /* WARNING: Please see the comment in dlm_init_lockres before
 294          * adding fields here. */
 295         struct hlist_node hash_node;
 296         struct qstr lockname;
 297         struct kref      refs;
 298
 299         /*
 300          * Please keep granted, converting, and blocked in this order,
 301          * as some funcs want to iterate over all lists.
 302          *
 303          * All four lists are protected by the hash's reference.
 304          */
 305         struct list_head granted;
 306         struct list_head converting;
 307         struct list_head blocked;
 308         struct list_head purge;
 309
 310         /*
 311          * These two lists require you to hold an additional reference
 312          * while they are on the list.
 313          */
 314         struct list_head dirty;
 315         struct list_head recovering; // dlm_recovery_ctxt.resources list
 316
 317         /* Added during init and removed during release */
 318         struct list_head tracking;      /* dlm->tracking_list */
 319
 320         /* unused lock resources have their last_used stamped and are
 321          * put on a list for the dlm thread to run. */
 322         unsigned long    last_used;
 323
 324         struct dlm_ctxt *dlm;
 325
 326         unsigned migration_pending:1;
 327         atomic_t asts_reserved;
 328         spinlock_t spinlock;
 329         wait_queue_head_t wq;
 330         u8  owner;              //node which owns the lock resource, or unknown
 331         u16 state;
 332         char lvb[DLM_LVB_LEN];
 333         unsigned int inflight_locks;
 334         unsigned long refmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
 335 };
 336
 337 struct dlm_migratable_lock
 338 {
 339         __be64 cookie;
 340
 341         /* these 3 are just padding for the in-memory structure, but
 342          * list and flags are actually used when sent over the wire */
 343         __be16 pad1;
 344         u8 list;  // 0=granted, 1=converting, 2=blocked
 345         u8 flags;
 346
 347         s8 type;
 348         s8 convert_type;
 349         s8 highest_blocked;
 350         u8 node;
 351 };  // 16 bytes
 352
 353 struct dlm_lock
 354 {
 355         struct dlm_migratable_lock ml;
 356
 357         struct list_head list;
 358         struct list_head ast_list;
 359         struct list_head bast_list;
 360         struct dlm_lock_resource *lockres;
 361         spinlock_t spinlock;
 362         struct kref lock_refs;
 363
 364         // ast and bast must be callable while holding a spinlock!
 365         dlm_astlockfunc_t *ast;
 366         dlm_bastlockfunc_t *bast;
 367         void *astdata;
 368         struct dlm_lockstatus *lksb;
 369         unsigned ast_pending:1,
 370                  bast_pending:1,
 371                  convert_pending:1,
 372                  lock_pending:1,
 373                  cancel_pending:1,
 374                  unlock_pending:1,
 375                  lksb_kernel_allocated:1;
 376 };
 377
 378
 379 #define DLM_LKSB_UNUSED1           0x01
 380 #define DLM_LKSB_PUT_LVB           0x02
 381 #define DLM_LKSB_GET_LVB           0x04
 382 #define DLM_LKSB_UNUSED2           0x08
 383 #define DLM_LKSB_UNUSED3           0x10
 384 #define DLM_LKSB_UNUSED4           0x20
 385 #define DLM_LKSB_UNUSED5           0x40
 386 #define DLM_LKSB_UNUSED6           0x80
 387
 388
 389 enum dlm_lockres_list {
 390         DLM_GRANTED_LIST = 0,
 391         DLM_CONVERTING_LIST = 1,
 392         DLM_BLOCKED_LIST = 2,
 393 };
 394
 395 static inline int dlm_lvb_is_empty(char *lvb)
 396 {
 397         int i;
 398         for (i=0; i<DLM_LVB_LEN; i++)
 399                 if (lvb[i])
 400                         return 0;
 401         return 1;
 402 }
 403
 404 static inline char *dlm_list_in_text(enum dlm_lockres_list idx)
 405 {
 406         if (idx == DLM_GRANTED_LIST)
 407                 return "granted";
 408         else if (idx == DLM_CONVERTING_LIST)
 409                 return "converting";
 410         else if (idx == DLM_BLOCKED_LIST)
 411                 return "blocked";
 412         else
 413                 return "unknown";
 414 }
 415
 416 static inline struct list_head *
 417 dlm_list_idx_to_ptr(struct dlm_lock_resource *res, enum dlm_lockres_list idx)
 418 {
 419         struct list_head *ret = NULL;
 420         if (idx == DLM_GRANTED_LIST)
 421                 ret = &res->granted;
 422         else if (idx == DLM_CONVERTING_LIST)
 423                 ret = &res->converting;
 424         else if (idx == DLM_BLOCKED_LIST)
 425                 ret = &res->blocked;
 426         else
 427                 BUG();
 428         return ret;
 429 }
 430
 431
 432
 433
 434 struct dlm_node_iter
 435 {
 436         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 437         int curnode;
 438 };
 439
 440
 441 enum {
 442         DLM_MASTER_REQUEST_MSG          = 500,
 443         DLM_UNUSED_MSG1                 = 501,
 444         DLM_ASSERT_MASTER_MSG           = 502,
 445         DLM_CREATE_LOCK_MSG             = 503,
 446         DLM_CONVERT_LOCK_MSG            = 504,
 447         DLM_PROXY_AST_MSG               = 505,
 448         DLM_UNLOCK_LOCK_MSG             = 506,
 449         DLM_DEREF_LOCKRES_MSG           = 507,
 450         DLM_MIGRATE_REQUEST_MSG         = 508,
 451         DLM_MIG_LOCKRES_MSG             = 509,
 452         DLM_QUERY_JOIN_MSG              = 510,
 453         DLM_ASSERT_JOINED_MSG           = 511,
 454         DLM_CANCEL_JOIN_MSG             = 512,
 455         DLM_EXIT_DOMAIN_MSG             = 513,
 456         DLM_MASTER_REQUERY_MSG          = 514,
 457         DLM_LOCK_REQUEST_MSG            = 515,
 458         DLM_RECO_DATA_DONE_MSG          = 516,
 459         DLM_BEGIN_RECO_MSG              = 517,
 460         DLM_FINALIZE_RECO_MSG           = 518,
 461         DLM_QUERY_REGION                = 519,
 462         DLM_QUERY_NODEINFO              = 520,
 463 };
 464
 465 struct dlm_reco_node_data
 466 {
 467         int state;
 468         u8 node_num;
 469         struct list_head list;
 470 };
 471
 472 enum {
 473         DLM_RECO_NODE_DATA_DEAD = -1,
 474         DLM_RECO_NODE_DATA_INIT = 0,
 475         DLM_RECO_NODE_DATA_REQUESTING = 1,
 476         DLM_RECO_NODE_DATA_REQUESTED = 2,
 477         DLM_RECO_NODE_DATA_RECEIVING = 3,
 478         DLM_RECO_NODE_DATA_DONE = 4,
 479         DLM_RECO_NODE_DATA_FINALIZE_SENT = 5,
 480 };
 481
 482
 483 enum {
 484         DLM_MASTER_RESP_NO = 0,
 485         DLM_MASTER_RESP_YES = 1,
 486         DLM_MASTER_RESP_MAYBE = 2,
 487         DLM_MASTER_RESP_ERROR = 3,
 488 };
 489
 490
 491 struct dlm_master_request
 492 {
 493         u8 node_idx;
 494         u8 namelen;
 495         __be16 pad1;
 496         __be32 flags;
 497
 498         u8 name[O2NM_MAX_NAME_LEN];
 499 };
 500
 501 #define DLM_ASSERT_RESPONSE_REASSERT       0x00000001
 502 #define DLM_ASSERT_RESPONSE_MASTERY_REF    0x00000002
 503
 504 #define DLM_ASSERT_MASTER_MLE_CLEANUP      0x00000001
 505 #define DLM_ASSERT_MASTER_REQUERY          0x00000002
 506 #define DLM_ASSERT_MASTER_FINISH_MIGRATION 0x00000004
 507 struct dlm_assert_master
 508 {
 509         u8 node_idx;
 510         u8 namelen;
 511         __be16 pad1;
 512         __be32 flags;
 513
 514         u8 name[O2NM_MAX_NAME_LEN];
 515 };
 516
 517 #define DLM_MIGRATE_RESPONSE_MASTERY_REF   0x00000001
 518
 519 struct dlm_migrate_request
 520 {
 521         u8 master;
 522         u8 new_master;
 523         u8 namelen;
 524         u8 pad1;
 525         __be32 pad2;
 526         u8 name[O2NM_MAX_NAME_LEN];
 527 };
 528
 529 struct dlm_master_requery
 530 {
 531         u8 pad1;
 532         u8 pad2;
 533         u8 node_idx;
 534         u8 namelen;
 535         __be32 pad3;
 536         u8 name[O2NM_MAX_NAME_LEN];
 537 };
 538
 539 #define DLM_MRES_RECOVERY   0x01
 540 #define DLM_MRES_MIGRATION  0x02
 541 #define DLM_MRES_ALL_DONE   0x04
 542
 543 /*
 544  * We would like to get one whole lockres into a single network
 545  * message whenever possible.  Generally speaking, there will be
 546  * at most one dlm_lock on a lockres for each node in the cluster,
 547  * plus (infrequently) any additional locks coming in from userdlm.
 548  *
 549  * struct _dlm_lockres_page
 550  * {
 551  *      dlm_migratable_lockres mres;
 552  *      dlm_migratable_lock ml[DLM_MAX_MIGRATABLE_LOCKS];
 553  *      u8 pad[DLM_MIG_LOCKRES_RESERVED];
 554  * };
 555  *
 556  * from ../cluster/tcp.h
 557  *    NET_MAX_PAYLOAD_BYTES  (4096 - sizeof(net_msg))
 558  *    (roughly 4080 bytes)
 559  * and sizeof(dlm_migratable_lockres) = 112 bytes
 560  * and sizeof(dlm_migratable_lock) = 16 bytes
 561  *
 562  * Choosing DLM_MAX_MIGRATABLE_LOCKS=240 and
 563  * DLM_MIG_LOCKRES_RESERVED=128 means we have this:
 564  *
 565  *  (DLM_MAX_MIGRATABLE_LOCKS * sizeof(dlm_migratable_lock)) +
 566  *     sizeof(dlm_migratable_lockres) + DLM_MIG_LOCKRES_RESERVED =
 567  *        NET_MAX_PAYLOAD_BYTES
 568  *  (240 * 16) + 112 + 128 = 4080
 569  *
 570  * So a lockres would need more than 240 locks before it would
 571  * use more than one network packet to recover.  Not too bad.
 572  */
 573 #define DLM_MAX_MIGRATABLE_LOCKS   240
 574
 575 struct dlm_migratable_lockres
 576 {
 577         u8 master;
 578         u8 lockname_len;
 579         u8 num_locks;    // locks sent in this structure
 580         u8 flags;
 581         __be32 total_locks; // locks to be sent for this migration cookie
 582         __be64 mig_cookie;  // cookie for this lockres migration
 583                          // or zero if not needed
 584         // 16 bytes
 585         u8 lockname[DLM_LOCKID_NAME_MAX];
 586         // 48 bytes
 587         u8 lvb[DLM_LVB_LEN];
 588         // 112 bytes
 589         struct dlm_migratable_lock ml[0];  // 16 bytes each, begins at byte 112
 590 };
 591 #define DLM_MIG_LOCKRES_MAX_LEN  \
 592         (sizeof(struct dlm_migratable_lockres) + \
 593          (sizeof(struct dlm_migratable_lock) * \
 594           DLM_MAX_MIGRATABLE_LOCKS) )
 595
 596 /* from above, 128 bytes
 597  * for some undetermined future use */
 598 #define DLM_MIG_LOCKRES_RESERVED   (NET_MAX_PAYLOAD_BYTES - \
 599                                     DLM_MIG_LOCKRES_MAX_LEN)
 600
 601 struct dlm_create_lock
 602 {
 603         __be64 cookie;
 604
 605         __be32 flags;
 606         u8 pad1;
 607         u8 node_idx;
 608         s8 requested_type;
 609         u8 namelen;
 610
 611         u8 name[O2NM_MAX_NAME_LEN];
 612 };
 613
 614 struct dlm_convert_lock
 615 {
 616         __be64 cookie;
 617
 618         __be32 flags;
 619         u8 pad1;
 620         u8 node_idx;
 621         s8 requested_type;
 622         u8 namelen;
 623
 624         u8 name[O2NM_MAX_NAME_LEN];
 625
 626         s8 lvb[0];
 627 };
 628 #define DLM_CONVERT_LOCK_MAX_LEN  (sizeof(struct dlm_convert_lock)+DLM_LVB_LEN)
 629
 630 struct dlm_unlock_lock
 631 {
 632         __be64 cookie;
 633
 634         __be32 flags;
 635         __be16 pad1;
 636         u8 node_idx;
 637         u8 namelen;
 638
 639         u8 name[O2NM_MAX_NAME_LEN];
 640
 641         s8 lvb[0];
 642 };
 643 #define DLM_UNLOCK_LOCK_MAX_LEN  (sizeof(struct dlm_unlock_lock)+DLM_LVB_LEN)
 644
 645 struct dlm_proxy_ast
 646 {
 647         __be64 cookie;
 648
 649         __be32 flags;
 650         u8 node_idx;
 651         u8 type;
 652         u8 blocked_type;
 653         u8 namelen;
 654
 655         u8 name[O2NM_MAX_NAME_LEN];
 656
 657         s8 lvb[0];
 658 };
 659 #define DLM_PROXY_AST_MAX_LEN  (sizeof(struct dlm_proxy_ast)+DLM_LVB_LEN)
 660
 661 #define DLM_MOD_KEY (0x666c6172)
 662 enum dlm_query_join_response_code {
 663         JOIN_DISALLOW = 0,
 664         JOIN_OK = 1,
 665         JOIN_OK_NO_MAP = 2,
 666         JOIN_PROTOCOL_MISMATCH = 3,
 667 };
 668
 669 struct dlm_query_join_packet {
 670         u8 code;        /* Response code.  dlm_minor and fs_minor
 671                            are only valid if this is JOIN_OK */
 672         u8 dlm_minor;   /* The minor version of the protocol the
 673                            dlm is speaking. */
 674         u8 fs_minor;    /* The minor version of the protocol the
 675                            filesystem is speaking. */
 676         u8 reserved;
 677 };
 678
 679 union dlm_query_join_response {
 680         u32 intval;
 681         struct dlm_query_join_packet packet;
 682 };
 683
 684 struct dlm_lock_request
 685 {
 686         u8 node_idx;
 687         u8 dead_node;
 688         __be16 pad1;
 689         __be32 pad2;
 690 };
 691
 692 struct dlm_reco_data_done
 693 {
 694         u8 node_idx;
 695         u8 dead_node;
 696         __be16 pad1;
 697         __be32 pad2;
 698
 699         /* unused for now */
 700         /* eventually we can use this to attempt
 701          * lvb recovery based on each node's info */
 702         u8 reco_lvb[DLM_LVB_LEN];
 703 };
 704
 705 struct dlm_begin_reco
 706 {
 707         u8 node_idx;
 708         u8 dead_node;
 709         __be16 pad1;
 710         __be32 pad2;
 711 };
 712
 713
 714 #define BITS_PER_BYTE 8
 715 #define BITS_TO_BYTES(bits) (((bits)+BITS_PER_BYTE-1)/BITS_PER_BYTE)
 716
 717 struct dlm_query_join_request
 718 {
 719         u8 node_idx;
 720         u8 pad1[2];
 721         u8 name_len;
 722         struct dlm_protocol_version dlm_proto;
 723         struct dlm_protocol_version fs_proto;
 724         u8 domain[O2NM_MAX_NAME_LEN];
 725         u8 node_map[BITS_TO_BYTES(O2NM_MAX_NODES)];
 726 };
 727
 728 struct dlm_assert_joined
 729 {
 730         u8 node_idx;
 731         u8 pad1[2];
 732         u8 name_len;
 733         u8 domain[O2NM_MAX_NAME_LEN];
 734 };
 735
 736 struct dlm_cancel_join
 737 {
 738         u8 node_idx;
 739         u8 pad1[2];
 740         u8 name_len;
 741         u8 domain[O2NM_MAX_NAME_LEN];
 742 };
 743
 744 struct dlm_query_region {
 745         u8 qr_node;
 746         u8 qr_numregions;
 747         u8 qr_namelen;
 748         u8 pad1;
 749         u8 qr_domain[O2NM_MAX_NAME_LEN];
 750         u8 qr_regions[O2HB_MAX_REGION_NAME_LEN * O2NM_MAX_REGIONS];
 751 };
 752
 753 struct dlm_node_info {
 754         u8 ni_nodenum;
 755         u8 pad1;
 756         u16 ni_ipv4_port;
 757         u32 ni_ipv4_address;
 758 };
 759
 760 struct dlm_query_nodeinfo {
 761         u8 qn_nodenum;
 762         u8 qn_numnodes;
 763         u8 qn_namelen;
 764         u8 pad1;
 765         u8 qn_domain[O2NM_MAX_NAME_LEN];
 766         struct dlm_node_info qn_nodes[O2NM_MAX_NODES];
 767 };
 768
 769 struct dlm_exit_domain
 770 {
 771         u8 node_idx;
 772         u8 pad1[3];
 773 };
 774
 775 struct dlm_finalize_reco
 776 {
 777         u8 node_idx;
 778         u8 dead_node;
 779         u8 flags;
 780         u8 pad1;
 781         __be32 pad2;
 782 };
 783
 784 struct dlm_deref_lockres
 785 {
 786         u32 pad1;
 787         u16 pad2;
 788         u8 node_idx;
 789         u8 namelen;
 790
 791         u8 name[O2NM_MAX_NAME_LEN];
 792 };
 793
 794 static inline enum dlm_status
 795 __dlm_lockres_state_to_status(struct dlm_lock_resource *res)
 796 {
 797         enum dlm_status status = DLM_NORMAL;
 798
 799         assert_spin_locked(&res->spinlock);
 800
 801         if (res->state & DLM_LOCK_RES_RECOVERING)
 802                 status = DLM_RECOVERING;
 803         else if (res->state & DLM_LOCK_RES_MIGRATING)
 804                 status = DLM_MIGRATING;
 805         else if (res->state & DLM_LOCK_RES_IN_PROGRESS)
 806                 status = DLM_FORWARD;
 807
 808         return status;
 809 }
 810
 811 static inline u8 dlm_get_lock_cookie_node(u64 cookie)
 812 {
 813         u8 ret;
 814         cookie >>= 56;
 815         ret = (u8)(cookie & 0xffULL);
 816         return ret;
 817 }
 818
 819 static inline unsigned long long dlm_get_lock_cookie_seq(u64 cookie)
 820 {
 821         unsigned long long ret;
 822         ret = ((unsigned long long)cookie) & 0x00ffffffffffffffULL;
 823         return ret;
 824 }
 825
 826 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
 827                                struct dlm_lockstatus *lksb);
 828 void dlm_lock_get(struct dlm_lock *lock);
 829 void dlm_lock_put(struct dlm_lock *lock);
 830
 831 void dlm_lock_attach_lockres(struct dlm_lock *lock,
 832                              struct dlm_lock_resource *res);
 833
 834 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 835                             void **ret_data);
 836 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 837                              void **ret_data);
 838 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
 839                           void **ret_data);
 840
 841 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
 842                                 struct dlm_lock *lock);
 843 void dlm_revert_pending_lock(struct dlm_lock_resource *res,
 844                              struct dlm_lock *lock);
 845
 846 int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 847                             void **ret_data);
 848 void dlm_commit_pending_cancel(struct dlm_lock_resource *res,
 849                                struct dlm_lock *lock);
 850 void dlm_commit_pending_unlock(struct dlm_lock_resource *res,
 851                                struct dlm_lock *lock);
 852
 853 int dlm_launch_thread(struct dlm_ctxt *dlm);
 854 void dlm_complete_thread(struct dlm_ctxt *dlm);
 855 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
 856 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
 857 void dlm_wait_for_recovery(struct dlm_ctxt *dlm);
 858 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
 859 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node);
 860 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout);
 861 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout);
 862
 863 void dlm_put(struct dlm_ctxt *dlm);
 864 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm);
 865 int dlm_domain_fully_joined(struct dlm_ctxt *dlm);
 866
 867 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 868                               struct dlm_lock_resource *res);
 869 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 870                             struct dlm_lock_resource *res);
 871 static inline void dlm_lockres_get(struct dlm_lock_resource *res)
 872 {
 873         /* This is called on every lookup, so it might be worth
 874          * inlining. */
 875         kref_get(&res->refs);
 876 }
 877 void dlm_lockres_put(struct dlm_lock_resource *res);
 878 void __dlm_unhash_lockres(struct dlm_lock_resource *res);
 879 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 880                           struct dlm_lock_resource *res);
 881 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 882                                                      const char *name,
 883                                                      unsigned int len,
 884                                                      unsigned int hash);
 885 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 886                                                 const char *name,
 887                                                 unsigned int len,
 888                                                 unsigned int hash);
 889 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 890                                               const char *name,
 891                                               unsigned int len);
 892
 893 int dlm_is_host_down(int errno);
 894
 895 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 896                                                  const char *lockid,
 897                                                  int namelen,
 898                                                  int flags);
 899 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 900                                           const char *name,
 901                                           unsigned int namelen);
 902
 903 #define dlm_lockres_set_refmap_bit(bit,res)  \
 904         __dlm_lockres_set_refmap_bit(bit,res,__FILE__,__LINE__)
 905 #define dlm_lockres_clear_refmap_bit(bit,res)  \
 906         __dlm_lockres_clear_refmap_bit(bit,res,__FILE__,__LINE__)
 907
 908 static inline void __dlm_lockres_set_refmap_bit(int bit,
 909                                                 struct dlm_lock_resource *res,
 910                                                 const char *file,
 911                                                 int line)
 912 {
 913         //printk("%s:%d:%.*s: setting bit %d\n", file, line,
 914         //     res->lockname.len, res->lockname.name, bit);
 915         set_bit(bit, res->refmap);
 916 }
 917
 918 static inline void __dlm_lockres_clear_refmap_bit(int bit,
 919                                                   struct dlm_lock_resource *res,
 920                                                   const char *file,
 921                                                   int line)
 922 {
 923         //printk("%s:%d:%.*s: clearing bit %d\n", file, line,
 924         //     res->lockname.len, res->lockname.name, bit);
 925         clear_bit(bit, res->refmap);
 926 }
 927
 928 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 929                                    struct dlm_lock_resource *res,
 930                                    const char *file,
 931                                    int line);
 932 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 933                                    struct dlm_lock_resource *res,
 934                                    int new_lockres,
 935                                    const char *file,
 936                                    int line);
 937 #define dlm_lockres_drop_inflight_ref(d,r)  \
 938         __dlm_lockres_drop_inflight_ref(d,r,__FILE__,__LINE__)
 939 #define dlm_lockres_grab_inflight_ref(d,r)  \
 940         __dlm_lockres_grab_inflight_ref(d,r,0,__FILE__,__LINE__)
 941 #define dlm_lockres_grab_inflight_ref_new(d,r)  \
 942         __dlm_lockres_grab_inflight_ref(d,r,1,__FILE__,__LINE__)
 943
 944 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 945 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 946 void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 947 void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 948 void dlm_do_local_ast(struct dlm_ctxt *dlm,
 949                       struct dlm_lock_resource *res,
 950                       struct dlm_lock *lock);
 951 int dlm_do_remote_ast(struct dlm_ctxt *dlm,
 952                       struct dlm_lock_resource *res,
 953                       struct dlm_lock *lock);
 954 void dlm_do_local_bast(struct dlm_ctxt *dlm,
 955                        struct dlm_lock_resource *res,
 956                        struct dlm_lock *lock,
 957                        int blocked_type);
 958 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm,
 959                            struct dlm_lock_resource *res,
 960                            struct dlm_lock *lock,
 961                            int msg_type,
 962                            int blocked_type, int flags);
 963 static inline int dlm_send_proxy_bast(struct dlm_ctxt *dlm,
 964                                       struct dlm_lock_resource *res,
 965                                       struct dlm_lock *lock,
 966                                       int blocked_type)
 967 {
 968         return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_BAST,
 969                                       blocked_type, 0);
 970 }
 971
 972 static inline int dlm_send_proxy_ast(struct dlm_ctxt *dlm,
 973                                      struct dlm_lock_resource *res,
 974                                      struct dlm_lock *lock,
 975                                      int flags)
 976 {
 977         return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_AST,
 978                                       0, flags);
 979 }
 980
 981 void dlm_print_one_lock_resource(struct dlm_lock_resource *res);
 982 void __dlm_print_one_lock_resource(struct dlm_lock_resource *res);
 983
 984 u8 dlm_nm_this_node(struct dlm_ctxt *dlm);
 985 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 986 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 987
 988
 989 int dlm_nm_init(struct dlm_ctxt *dlm);
 990 int dlm_heartbeat_init(struct dlm_ctxt *dlm);
 991 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data);
 992 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data);
 993
 994 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 995 int dlm_finish_migration(struct dlm_ctxt *dlm,
 996                          struct dlm_lock_resource *res,
 997                          u8 old_master);
 998 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
 999                              struct dlm_lock_resource *res);
1000 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res);
1001
1002 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
1003                                void **ret_data);
1004 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
1005                               void **ret_data);
1006 void dlm_assert_master_post_handler(int status, void *data, void *ret_data);
1007 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1008                               void **ret_data);
1009 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
1010                                 void **ret_data);
1011 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1012                             void **ret_data);
1013 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1014                                void **ret_data);
1015 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
1016                                   void **ret_data);
1017 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
1018                                void **ret_data);
1019 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
1020                            void **ret_data);
1021 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
1022                               void **ret_data);
1023 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1024                           u8 nodenum, u8 *real_master);
1025
1026
1027 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
1028                                struct dlm_lock_resource *res,
1029                                int ignore_higher,
1030                                u8 request_from,
1031                                u32 flags);
1032
1033
1034 int dlm_send_one_lockres(struct dlm_ctxt *dlm,
1035                          struct dlm_lock_resource *res,
1036                          struct dlm_migratable_lockres *mres,
1037                          u8 send_to,
1038                          u8 flags);
1039 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1040                                        struct dlm_lock_resource *res);
1041
1042 /* will exit holding res->spinlock, but may drop in function */
1043 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags);
1044 void __dlm_wait_on_lockres_flags_set(struct dlm_lock_resource *res, int flags);
1045
1046 /* will exit holding res->spinlock, but may drop in function */
1047 static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res)
1048 {
1049         __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_IN_PROGRESS|
1050                                           DLM_LOCK_RES_RECOVERING|
1051                                           DLM_LOCK_RES_MIGRATING));
1052 }
1053
1054 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1055 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1056
1057 /* create/destroy slab caches */
1058 int dlm_init_master_caches(void);
1059 void dlm_destroy_master_caches(void);
1060
1061 int dlm_init_lock_cache(void);
1062 void dlm_destroy_lock_cache(void);
1063
1064 int dlm_init_mle_cache(void);
1065 void dlm_destroy_mle_cache(void);
1066
1067 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up);
1068 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm,
1069                          struct dlm_lock_resource *res);
1070 void dlm_clean_master_list(struct dlm_ctxt *dlm,
1071                            u8 dead_node);
1072 void dlm_force_free_mles(struct dlm_ctxt *dlm);
1073 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock);
1074 int __dlm_lockres_has_locks(struct dlm_lock_resource *res);
1075 int __dlm_lockres_unused(struct dlm_lock_resource *res);
1076
1077 static inline const char * dlm_lock_mode_name(int mode)
1078 {
1079         switch (mode) {
1080                 case LKM_EXMODE:
1081                         return "EX";
1082                 case LKM_PRMODE:
1083                         return "PR";
1084                 case LKM_NLMODE:
1085                         return "NL";
1086         }
1087         return "UNKNOWN";
1088 }
1089
1090
1091 static inline int dlm_lock_compatible(int existing, int request)
1092 {
1093         /* NO_LOCK compatible with all */
1094         if (request == LKM_NLMODE ||
1095             existing == LKM_NLMODE)
1096                 return 1;
1097
1098         /* EX incompatible with all non-NO_LOCK */
1099         if (request == LKM_EXMODE)
1100                 return 0;
1101
1102         /* request must be PR, which is compatible with PR */
1103         if (existing == LKM_PRMODE)
1104                 return 1;
1105
1106         return 0;
1107 }
1108
1109 static inline int dlm_lock_on_list(struct list_head *head,
1110                                    struct dlm_lock *lock)
1111 {
1112         struct list_head *iter;
1113         struct dlm_lock *tmplock;
1114
1115         list_for_each(iter, head) {
1116                 tmplock = list_entry(iter, struct dlm_lock, list);
1117                 if (tmplock == lock)
1118                         return 1;
1119         }
1120         return 0;
1121 }
1122
1123
1124 static inline enum dlm_status dlm_err_to_dlm_status(int err)
1125 {
1126         enum dlm_status ret;
1127         if (err == -ENOMEM)
1128                 ret = DLM_SYSERR;
1129         else if (err == -ETIMEDOUT || o2net_link_down(err, NULL))
1130                 ret = DLM_NOLOCKMGR;
1131         else if (err == -EINVAL)
1132                 ret = DLM_BADPARAM;
1133         else if (err == -ENAMETOOLONG)
1134                 ret = DLM_IVBUFLEN;
1135         else
1136                 ret = DLM_BADARGS;
1137         return ret;
1138 }
1139
1140
1141 static inline void dlm_node_iter_init(unsigned long *map,
1142                                       struct dlm_node_iter *iter)
1143 {
1144         memcpy(iter->node_map, map, sizeof(iter->node_map));
1145         iter->curnode = -1;
1146 }
1147
1148 static inline int dlm_node_iter_next(struct dlm_node_iter *iter)
1149 {
1150         int bit;
1151         bit = find_next_bit(iter->node_map, O2NM_MAX_NODES, iter->curnode+1);
1152         if (bit >= O2NM_MAX_NODES) {
1153                 iter->curnode = O2NM_MAX_NODES;
1154                 return -ENOENT;
1155         }
1156         iter->curnode = bit;
1157         return bit;
1158 }
1159
1160 static inline void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
1161                                          struct dlm_lock_resource *res,
1162                                          u8 owner)
1163 {
1164         assert_spin_locked(&res->spinlock);
1165
1166         res->owner = owner;
1167 }
1168
1169 static inline void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
1170                                             struct dlm_lock_resource *res,
1171                                             u8 owner)
1172 {
1173         assert_spin_locked(&res->spinlock);
1174
1175         if (owner != res->owner)
1176                 dlm_set_lockres_owner(dlm, res, owner);
1177 }
1178
1179 #endif /* DLMCOMMON_H */