SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 << 14)
  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,
  54         DLM_MLE_MASTER,
  55         DLM_MLE_MIGRATION,
  56         DLM_MLE_NUM_TYPES
  57 };
  58
  59 struct dlm_lock_name {
  60         unsigned int hash;
  61         unsigned int len;
  62         unsigned char name[DLM_LOCKID_NAME_MAX];
  63 };
  64
  65 struct dlm_master_list_entry {
  66         struct hlist_node master_hash_node;
  67         struct list_head hb_events;
  68         struct dlm_ctxt *dlm;
  69         spinlock_t spinlock;
  70         wait_queue_head_t wq;
  71         atomic_t woken;
  72         struct kref mle_refs;
  73         int inuse;
  74         unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  75         unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  76         unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  77         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
  78         u8 master;
  79         u8 new_master;
  80         enum dlm_mle_type type;
  81         struct o2hb_callback_func mle_hb_up;
  82         struct o2hb_callback_func mle_hb_down;
  83         union {
  84                 struct dlm_lock_resource *mleres;
  85                 struct dlm_lock_name mlename;
  86         } u;
  87 };
  88
  89 enum dlm_ast_type {
  90         DLM_AST = 0,
  91         DLM_BAST,
  92         DLM_ASTUNLOCK
  93 };
  94
  95
  96 #define LKM_VALID_FLAGS (LKM_VALBLK | LKM_CONVERT | LKM_UNLOCK | \
  97                          LKM_CANCEL | LKM_INVVALBLK | LKM_FORCE | \
  98                          LKM_RECOVERY | LKM_LOCAL | LKM_NOQUEUE)
  99
 100 #define DLM_RECOVERY_LOCK_NAME       "$RECOVERY"
 101 #define DLM_RECOVERY_LOCK_NAME_LEN   9
 102
 103 static inline int dlm_is_recovery_lock(const char *lock_name, int name_len)
 104 {
 105         if (name_len == DLM_RECOVERY_LOCK_NAME_LEN &&
 106             memcmp(lock_name, DLM_RECOVERY_LOCK_NAME, name_len)==0)
 107                 return 1;
 108         return 0;
 109 }
 110
 111 #define DLM_RECO_STATE_ACTIVE    0x0001
 112 #define DLM_RECO_STATE_FINALIZE  0x0002
 113
 114 struct dlm_recovery_ctxt
 115 {
 116         struct list_head resources;
 117         struct list_head received;
 118         struct list_head node_data;
 119         u8  new_master;
 120         u8  dead_node;
 121         u16 state;
 122         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 123         wait_queue_head_t event;
 124 };
 125
 126 enum dlm_ctxt_state {
 127         DLM_CTXT_NEW = 0,
 128         DLM_CTXT_JOINED,
 129         DLM_CTXT_IN_SHUTDOWN,
 130         DLM_CTXT_LEAVING,
 131 };
 132
 133 struct dlm_ctxt
 134 {
 135         struct list_head list;
 136         struct hlist_head **lockres_hash;
 137         struct list_head dirty_list;
 138         struct list_head purge_list;
 139         struct list_head pending_asts;
 140         struct list_head pending_basts;
 141         struct list_head tracking_list;
 142         unsigned int purge_count;
 143         spinlock_t spinlock;
 144         spinlock_t ast_lock;
 145         spinlock_t track_lock;
 146         char *name;
 147         u8 node_num;
 148         u32 key;
 149         u8  joining_node;
 150         wait_queue_head_t dlm_join_events;
 151         unsigned long live_nodes_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 152         unsigned long domain_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 153         unsigned long recovery_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 154         struct dlm_recovery_ctxt reco;
 155         spinlock_t master_lock;
 156         struct hlist_head **master_hash;
 157         struct list_head mle_hb_events;
 158
 159         /* these give a really vague idea of the system load */
 160         atomic_t mle_tot_count[DLM_MLE_NUM_TYPES];
 161         atomic_t mle_cur_count[DLM_MLE_NUM_TYPES];
 162         atomic_t local_resources;
 163         atomic_t remote_resources;
 164         atomic_t unknown_resources;
 165
 166         struct dlm_debug_ctxt *dlm_debug_ctxt;
 167         struct dentry *dlm_debugfs_subroot;
 168
 169         /* NOTE: Next three are protected by dlm_domain_lock */
 170         struct kref dlm_refs;
 171         enum dlm_ctxt_state dlm_state;
 172         unsigned int num_joins;
 173
 174         struct o2hb_callback_func dlm_hb_up;
 175         struct o2hb_callback_func dlm_hb_down;
 176         struct task_struct *dlm_thread_task;
 177         struct task_struct *dlm_reco_thread_task;
 178         struct workqueue_struct *dlm_worker;
 179         wait_queue_head_t dlm_thread_wq;
 180         wait_queue_head_t dlm_reco_thread_wq;
 181         wait_queue_head_t ast_wq;
 182         wait_queue_head_t migration_wq;
 183
 184         struct work_struct dispatched_work;
 185         struct list_head work_list;
 186         spinlock_t work_lock;
 187         struct list_head dlm_domain_handlers;
 188         struct list_head        dlm_eviction_callbacks;
 189
 190         /* The filesystem specifies this at domain registration.  We
 191          * cache it here to know what to tell other nodes. */
 192         struct dlm_protocol_version fs_locking_proto;
 193         /* This is the inter-dlm communication version */
 194         struct dlm_protocol_version dlm_locking_proto;
 195 };
 196
 197 static inline struct hlist_head *dlm_lockres_hash(struct dlm_ctxt *dlm, unsigned i)
 198 {
 199         return dlm->lockres_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES] + (i % DLM_BUCKETS_PER_PAGE);
 200 }
 201
 202 static inline struct hlist_head *dlm_master_hash(struct dlm_ctxt *dlm,
 203                                                  unsigned i)
 204 {
 205         return dlm->master_hash[(i / DLM_BUCKETS_PER_PAGE) % DLM_HASH_PAGES] +
 206                         (i % DLM_BUCKETS_PER_PAGE);
 207 }
 208
 209 /* these keventd work queue items are for less-frequently
 210  * called functions that cannot be directly called from the
 211  * net message handlers for some reason, usually because
 212  * they need to send net messages of their own. */
 213 void dlm_dispatch_work(struct work_struct *work);
 214
 215 struct dlm_lock_resource;
 216 struct dlm_work_item;
 217
 218 typedef void (dlm_workfunc_t)(struct dlm_work_item *, void *);
 219
 220 struct dlm_request_all_locks_priv
 221 {
 222         u8 reco_master;
 223         u8 dead_node;
 224 };
 225
 226 struct dlm_mig_lockres_priv
 227 {
 228         struct dlm_lock_resource *lockres;
 229         u8 real_master;
 230         u8 extra_ref;
 231 };
 232
 233 struct dlm_assert_master_priv
 234 {
 235         struct dlm_lock_resource *lockres;
 236         u8 request_from;
 237         u32 flags;
 238         unsigned ignore_higher:1;
 239 };
 240
 241 struct dlm_deref_lockres_priv
 242 {
 243         struct dlm_lock_resource *deref_res;
 244         u8 deref_node;
 245 };
 246
 247 struct dlm_work_item
 248 {
 249         struct list_head list;
 250         dlm_workfunc_t *func;
 251         struct dlm_ctxt *dlm;
 252         void *data;
 253         union {
 254                 struct dlm_request_all_locks_priv ral;
 255                 struct dlm_mig_lockres_priv ml;
 256                 struct dlm_assert_master_priv am;
 257                 struct dlm_deref_lockres_priv dl;
 258         } u;
 259 };
 260
 261 static inline void dlm_init_work_item(struct dlm_ctxt *dlm,
 262                                       struct dlm_work_item *i,
 263                                       dlm_workfunc_t *f, void *data)
 264 {
 265         memset(i, 0, sizeof(*i));
 266         i->func = f;
 267         INIT_LIST_HEAD(&i->list);
 268         i->data = data;
 269         i->dlm = dlm;  /* must have already done a dlm_grab on this! */
 270 }
 271
 272
 273
 274 static inline void __dlm_set_joining_node(struct dlm_ctxt *dlm,
 275                                           u8 node)
 276 {
 277         assert_spin_locked(&dlm->spinlock);
 278
 279         dlm->joining_node = node;
 280         wake_up(&dlm->dlm_join_events);
 281 }
 282
 283 #define DLM_LOCK_RES_UNINITED             0x00000001
 284 #define DLM_LOCK_RES_RECOVERING           0x00000002
 285 #define DLM_LOCK_RES_READY                0x00000004
 286 #define DLM_LOCK_RES_DIRTY                0x00000008
 287 #define DLM_LOCK_RES_IN_PROGRESS          0x00000010
 288 #define DLM_LOCK_RES_MIGRATING            0x00000020
 289 #define DLM_LOCK_RES_DROPPING_REF         0x00000040
 290 #define DLM_LOCK_RES_BLOCK_DIRTY          0x00001000
 291 #define DLM_LOCK_RES_SETREF_INPROG        0x00002000
 292
 293 /* max milliseconds to wait to sync up a network failure with a node death */
 294 #define DLM_NODE_DEATH_WAIT_MAX (5 * 1000)
 295
 296 #define DLM_PURGE_INTERVAL_MS   (8 * 1000)
 297
 298 struct dlm_lock_resource
 299 {
 300         /* WARNING: Please see the comment in dlm_init_lockres before
 301          * adding fields here. */
 302         struct hlist_node hash_node;
 303         struct qstr lockname;
 304         struct kref      refs;
 305
 306         /*
 307          * Please keep granted, converting, and blocked in this order,
 308          * as some funcs want to iterate over all lists.
 309          *
 310          * All four lists are protected by the hash's reference.
 311          */
 312         struct list_head granted;
 313         struct list_head converting;
 314         struct list_head blocked;
 315         struct list_head purge;
 316
 317         /*
 318          * These two lists require you to hold an additional reference
 319          * while they are on the list.
 320          */
 321         struct list_head dirty;
 322         struct list_head recovering; // dlm_recovery_ctxt.resources list
 323
 324         /* Added during init and removed during release */
 325         struct list_head tracking;      /* dlm->tracking_list */
 326
 327         /* unused lock resources have their last_used stamped and are
 328          * put on a list for the dlm thread to run. */
 329         unsigned long    last_used;
 330
 331         struct dlm_ctxt *dlm;
 332
 333         unsigned migration_pending:1;
 334         atomic_t asts_reserved;
 335         spinlock_t spinlock;
 336         wait_queue_head_t wq;
 337         u8  owner;              //node which owns the lock resource, or unknown
 338         u16 state;
 339         char lvb[DLM_LVB_LEN];
 340         unsigned int inflight_locks;
 341         unsigned long refmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
 342 };
 343
 344 struct dlm_migratable_lock
 345 {
 346         __be64 cookie;
 347
 348         /* these 3 are just padding for the in-memory structure, but
 349          * list and flags are actually used when sent over the wire */
 350         __be16 pad1;
 351         u8 list;  // 0=granted, 1=converting, 2=blocked
 352         u8 flags;
 353
 354         s8 type;
 355         s8 convert_type;
 356         s8 highest_blocked;
 357         u8 node;
 358 };  // 16 bytes
 359
 360 struct dlm_lock
 361 {
 362         struct dlm_migratable_lock ml;
 363
 364         struct list_head list;
 365         struct list_head ast_list;
 366         struct list_head bast_list;
 367         struct dlm_lock_resource *lockres;
 368         spinlock_t spinlock;
 369         struct kref lock_refs;
 370
 371         // ast and bast must be callable while holding a spinlock!
 372         dlm_astlockfunc_t *ast;
 373         dlm_bastlockfunc_t *bast;
 374         void *astdata;
 375         struct dlm_lockstatus *lksb;
 376         unsigned ast_pending:1,
 377                  bast_pending:1,
 378                  convert_pending:1,
 379                  lock_pending:1,
 380                  cancel_pending:1,
 381                  unlock_pending:1,
 382                  lksb_kernel_allocated:1;
 383 };
 384
 385
 386 #define DLM_LKSB_UNUSED1           0x01
 387 #define DLM_LKSB_PUT_LVB           0x02
 388 #define DLM_LKSB_GET_LVB           0x04
 389 #define DLM_LKSB_UNUSED2           0x08
 390 #define DLM_LKSB_UNUSED3           0x10
 391 #define DLM_LKSB_UNUSED4           0x20
 392 #define DLM_LKSB_UNUSED5           0x40
 393 #define DLM_LKSB_UNUSED6           0x80
 394
 395
 396 enum dlm_lockres_list {
 397         DLM_GRANTED_LIST = 0,
 398         DLM_CONVERTING_LIST,
 399         DLM_BLOCKED_LIST
 400 };
 401
 402 static inline int dlm_lvb_is_empty(char *lvb)
 403 {
 404         int i;
 405         for (i=0; i<DLM_LVB_LEN; i++)
 406                 if (lvb[i])
 407                         return 0;
 408         return 1;
 409 }
 410
 411 static inline struct list_head *
 412 dlm_list_idx_to_ptr(struct dlm_lock_resource *res, enum dlm_lockres_list idx)
 413 {
 414         struct list_head *ret = NULL;
 415         if (idx == DLM_GRANTED_LIST)
 416                 ret = &res->granted;
 417         else if (idx == DLM_CONVERTING_LIST)
 418                 ret = &res->converting;
 419         else if (idx == DLM_BLOCKED_LIST)
 420                 ret = &res->blocked;
 421         else
 422                 BUG();
 423         return ret;
 424 }
 425
 426
 427
 428
 429 struct dlm_node_iter
 430 {
 431         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 432         int curnode;
 433 };
 434
 435
 436 enum {
 437         DLM_MASTER_REQUEST_MSG    = 500,
 438         DLM_UNUSED_MSG1,         /* 501 */
 439         DLM_ASSERT_MASTER_MSG,   /* 502 */
 440         DLM_CREATE_LOCK_MSG,     /* 503 */
 441         DLM_CONVERT_LOCK_MSG,    /* 504 */
 442         DLM_PROXY_AST_MSG,       /* 505 */
 443         DLM_UNLOCK_LOCK_MSG,     /* 506 */
 444         DLM_DEREF_LOCKRES_MSG,   /* 507 */
 445         DLM_MIGRATE_REQUEST_MSG, /* 508 */
 446         DLM_MIG_LOCKRES_MSG,     /* 509 */
 447         DLM_QUERY_JOIN_MSG,      /* 510 */
 448         DLM_ASSERT_JOINED_MSG,   /* 511 */
 449         DLM_CANCEL_JOIN_MSG,     /* 512 */
 450         DLM_EXIT_DOMAIN_MSG,     /* 513 */
 451         DLM_MASTER_REQUERY_MSG,  /* 514 */
 452         DLM_LOCK_REQUEST_MSG,    /* 515 */
 453         DLM_RECO_DATA_DONE_MSG,  /* 516 */
 454         DLM_BEGIN_RECO_MSG,      /* 517 */
 455         DLM_FINALIZE_RECO_MSG    /* 518 */
 456 };
 457
 458 struct dlm_reco_node_data
 459 {
 460         int state;
 461         u8 node_num;
 462         struct list_head list;
 463 };
 464
 465 enum {
 466         DLM_RECO_NODE_DATA_DEAD = -1,
 467         DLM_RECO_NODE_DATA_INIT = 0,
 468         DLM_RECO_NODE_DATA_REQUESTING,
 469         DLM_RECO_NODE_DATA_REQUESTED,
 470         DLM_RECO_NODE_DATA_RECEIVING,
 471         DLM_RECO_NODE_DATA_DONE,
 472         DLM_RECO_NODE_DATA_FINALIZE_SENT,
 473 };
 474
 475
 476 enum {
 477         DLM_MASTER_RESP_NO = 0,
 478         DLM_MASTER_RESP_YES,
 479         DLM_MASTER_RESP_MAYBE,
 480         DLM_MASTER_RESP_ERROR
 481 };
 482
 483
 484 struct dlm_master_request
 485 {
 486         u8 node_idx;
 487         u8 namelen;
 488         __be16 pad1;
 489         __be32 flags;
 490
 491         u8 name[O2NM_MAX_NAME_LEN];
 492 };
 493
 494 #define DLM_ASSERT_RESPONSE_REASSERT       0x00000001
 495 #define DLM_ASSERT_RESPONSE_MASTERY_REF    0x00000002
 496
 497 #define DLM_ASSERT_MASTER_MLE_CLEANUP      0x00000001
 498 #define DLM_ASSERT_MASTER_REQUERY          0x00000002
 499 #define DLM_ASSERT_MASTER_FINISH_MIGRATION 0x00000004
 500 struct dlm_assert_master
 501 {
 502         u8 node_idx;
 503         u8 namelen;
 504         __be16 pad1;
 505         __be32 flags;
 506
 507         u8 name[O2NM_MAX_NAME_LEN];
 508 };
 509
 510 #define DLM_MIGRATE_RESPONSE_MASTERY_REF   0x00000001
 511
 512 struct dlm_migrate_request
 513 {
 514         u8 master;
 515         u8 new_master;
 516         u8 namelen;
 517         u8 pad1;
 518         __be32 pad2;
 519         u8 name[O2NM_MAX_NAME_LEN];
 520 };
 521
 522 struct dlm_master_requery
 523 {
 524         u8 pad1;
 525         u8 pad2;
 526         u8 node_idx;
 527         u8 namelen;
 528         __be32 pad3;
 529         u8 name[O2NM_MAX_NAME_LEN];
 530 };
 531
 532 #define DLM_MRES_RECOVERY   0x01
 533 #define DLM_MRES_MIGRATION  0x02
 534 #define DLM_MRES_ALL_DONE   0x04
 535
 536 /*
 537  * We would like to get one whole lockres into a single network
 538  * message whenever possible.  Generally speaking, there will be
 539  * at most one dlm_lock on a lockres for each node in the cluster,
 540  * plus (infrequently) any additional locks coming in from userdlm.
 541  *
 542  * struct _dlm_lockres_page
 543  * {
 544  *      dlm_migratable_lockres mres;
 545  *      dlm_migratable_lock ml[DLM_MAX_MIGRATABLE_LOCKS];
 546  *      u8 pad[DLM_MIG_LOCKRES_RESERVED];
 547  * };
 548  *
 549  * from ../cluster/tcp.h
 550  *    NET_MAX_PAYLOAD_BYTES  (4096 - sizeof(net_msg))
 551  *    (roughly 4080 bytes)
 552  * and sizeof(dlm_migratable_lockres) = 112 bytes
 553  * and sizeof(dlm_migratable_lock) = 16 bytes
 554  *
 555  * Choosing DLM_MAX_MIGRATABLE_LOCKS=240 and
 556  * DLM_MIG_LOCKRES_RESERVED=128 means we have this:
 557  *
 558  *  (DLM_MAX_MIGRATABLE_LOCKS * sizeof(dlm_migratable_lock)) +
 559  *     sizeof(dlm_migratable_lockres) + DLM_MIG_LOCKRES_RESERVED =
 560  *        NET_MAX_PAYLOAD_BYTES
 561  *  (240 * 16) + 112 + 128 = 4080
 562  *
 563  * So a lockres would need more than 240 locks before it would
 564  * use more than one network packet to recover.  Not too bad.
 565  */
 566 #define DLM_MAX_MIGRATABLE_LOCKS   240
 567
 568 struct dlm_migratable_lockres
 569 {
 570         u8 master;
 571         u8 lockname_len;
 572         u8 num_locks;    // locks sent in this structure
 573         u8 flags;
 574         __be32 total_locks; // locks to be sent for this migration cookie
 575         __be64 mig_cookie;  // cookie for this lockres migration
 576                          // or zero if not needed
 577         // 16 bytes
 578         u8 lockname[DLM_LOCKID_NAME_MAX];
 579         // 48 bytes
 580         u8 lvb[DLM_LVB_LEN];
 581         // 112 bytes
 582         struct dlm_migratable_lock ml[0];  // 16 bytes each, begins at byte 112
 583 };
 584 #define DLM_MIG_LOCKRES_MAX_LEN  \
 585         (sizeof(struct dlm_migratable_lockres) + \
 586          (sizeof(struct dlm_migratable_lock) * \
 587           DLM_MAX_MIGRATABLE_LOCKS) )
 588
 589 /* from above, 128 bytes
 590  * for some undetermined future use */
 591 #define DLM_MIG_LOCKRES_RESERVED   (NET_MAX_PAYLOAD_BYTES - \
 592                                     DLM_MIG_LOCKRES_MAX_LEN)
 593
 594 struct dlm_create_lock
 595 {
 596         __be64 cookie;
 597
 598         __be32 flags;
 599         u8 pad1;
 600         u8 node_idx;
 601         s8 requested_type;
 602         u8 namelen;
 603
 604         u8 name[O2NM_MAX_NAME_LEN];
 605 };
 606
 607 struct dlm_convert_lock
 608 {
 609         __be64 cookie;
 610
 611         __be32 flags;
 612         u8 pad1;
 613         u8 node_idx;
 614         s8 requested_type;
 615         u8 namelen;
 616
 617         u8 name[O2NM_MAX_NAME_LEN];
 618
 619         s8 lvb[0];
 620 };
 621 #define DLM_CONVERT_LOCK_MAX_LEN  (sizeof(struct dlm_convert_lock)+DLM_LVB_LEN)
 622
 623 struct dlm_unlock_lock
 624 {
 625         __be64 cookie;
 626
 627         __be32 flags;
 628         __be16 pad1;
 629         u8 node_idx;
 630         u8 namelen;
 631
 632         u8 name[O2NM_MAX_NAME_LEN];
 633
 634         s8 lvb[0];
 635 };
 636 #define DLM_UNLOCK_LOCK_MAX_LEN  (sizeof(struct dlm_unlock_lock)+DLM_LVB_LEN)
 637
 638 struct dlm_proxy_ast
 639 {
 640         __be64 cookie;
 641
 642         __be32 flags;
 643         u8 node_idx;
 644         u8 type;
 645         u8 blocked_type;
 646         u8 namelen;
 647
 648         u8 name[O2NM_MAX_NAME_LEN];
 649
 650         s8 lvb[0];
 651 };
 652 #define DLM_PROXY_AST_MAX_LEN  (sizeof(struct dlm_proxy_ast)+DLM_LVB_LEN)
 653
 654 #define DLM_MOD_KEY (0x666c6172)
 655 enum dlm_query_join_response_code {
 656         JOIN_DISALLOW = 0,
 657         JOIN_OK,
 658         JOIN_OK_NO_MAP,
 659         JOIN_PROTOCOL_MISMATCH,
 660 };
 661
 662 struct dlm_query_join_packet {
 663         u8 code;        /* Response code.  dlm_minor and fs_minor
 664                            are only valid if this is JOIN_OK */
 665         u8 dlm_minor;   /* The minor version of the protocol the
 666                            dlm is speaking. */
 667         u8 fs_minor;    /* The minor version of the protocol the
 668                            filesystem is speaking. */
 669         u8 reserved;
 670 };
 671
 672 union dlm_query_join_response {
 673         u32 intval;
 674         struct dlm_query_join_packet packet;
 675 };
 676
 677 struct dlm_lock_request
 678 {
 679         u8 node_idx;
 680         u8 dead_node;
 681         __be16 pad1;
 682         __be32 pad2;
 683 };
 684
 685 struct dlm_reco_data_done
 686 {
 687         u8 node_idx;
 688         u8 dead_node;
 689         __be16 pad1;
 690         __be32 pad2;
 691
 692         /* unused for now */
 693         /* eventually we can use this to attempt
 694          * lvb recovery based on each node's info */
 695         u8 reco_lvb[DLM_LVB_LEN];
 696 };
 697
 698 struct dlm_begin_reco
 699 {
 700         u8 node_idx;
 701         u8 dead_node;
 702         __be16 pad1;
 703         __be32 pad2;
 704 };
 705
 706
 707 #define BITS_PER_BYTE 8
 708 #define BITS_TO_BYTES(bits) (((bits)+BITS_PER_BYTE-1)/BITS_PER_BYTE)
 709
 710 struct dlm_query_join_request
 711 {
 712         u8 node_idx;
 713         u8 pad1[2];
 714         u8 name_len;
 715         struct dlm_protocol_version dlm_proto;
 716         struct dlm_protocol_version fs_proto;
 717         u8 domain[O2NM_MAX_NAME_LEN];
 718         u8 node_map[BITS_TO_BYTES(O2NM_MAX_NODES)];
 719 };
 720
 721 struct dlm_assert_joined
 722 {
 723         u8 node_idx;
 724         u8 pad1[2];
 725         u8 name_len;
 726         u8 domain[O2NM_MAX_NAME_LEN];
 727 };
 728
 729 struct dlm_cancel_join
 730 {
 731         u8 node_idx;
 732         u8 pad1[2];
 733         u8 name_len;
 734         u8 domain[O2NM_MAX_NAME_LEN];
 735 };
 736
 737 struct dlm_exit_domain
 738 {
 739         u8 node_idx;
 740         u8 pad1[3];
 741 };
 742
 743 struct dlm_finalize_reco
 744 {
 745         u8 node_idx;
 746         u8 dead_node;
 747         u8 flags;
 748         u8 pad1;
 749         __be32 pad2;
 750 };
 751
 752 struct dlm_deref_lockres
 753 {
 754         u32 pad1;
 755         u16 pad2;
 756         u8 node_idx;
 757         u8 namelen;
 758
 759         u8 name[O2NM_MAX_NAME_LEN];
 760 };
 761
 762 static inline enum dlm_status
 763 __dlm_lockres_state_to_status(struct dlm_lock_resource *res)
 764 {
 765         enum dlm_status status = DLM_NORMAL;
 766
 767         assert_spin_locked(&res->spinlock);
 768
 769         if (res->state & DLM_LOCK_RES_RECOVERING)
 770                 status = DLM_RECOVERING;
 771         else if (res->state & DLM_LOCK_RES_MIGRATING)
 772                 status = DLM_MIGRATING;
 773         else if (res->state & DLM_LOCK_RES_IN_PROGRESS)
 774                 status = DLM_FORWARD;
 775
 776         return status;
 777 }
 778
 779 static inline u8 dlm_get_lock_cookie_node(u64 cookie)
 780 {
 781         u8 ret;
 782         cookie >>= 56;
 783         ret = (u8)(cookie & 0xffULL);
 784         return ret;
 785 }
 786
 787 static inline unsigned long long dlm_get_lock_cookie_seq(u64 cookie)
 788 {
 789         unsigned long long ret;
 790         ret = ((unsigned long long)cookie) & 0x00ffffffffffffffULL;
 791         return ret;
 792 }
 793
 794 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
 795                                struct dlm_lockstatus *lksb);
 796 void dlm_lock_get(struct dlm_lock *lock);
 797 void dlm_lock_put(struct dlm_lock *lock);
 798
 799 void dlm_lock_attach_lockres(struct dlm_lock *lock,
 800                              struct dlm_lock_resource *res);
 801
 802 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 803                             void **ret_data);
 804 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 805                              void **ret_data);
 806 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
 807                           void **ret_data);
 808
 809 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
 810                                 struct dlm_lock *lock);
 811 void dlm_revert_pending_lock(struct dlm_lock_resource *res,
 812                              struct dlm_lock *lock);
 813
 814 int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 815                             void **ret_data);
 816 void dlm_commit_pending_cancel(struct dlm_lock_resource *res,
 817                                struct dlm_lock *lock);
 818 void dlm_commit_pending_unlock(struct dlm_lock_resource *res,
 819                                struct dlm_lock *lock);
 820
 821 int dlm_launch_thread(struct dlm_ctxt *dlm);
 822 void dlm_complete_thread(struct dlm_ctxt *dlm);
 823 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
 824 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
 825 void dlm_wait_for_recovery(struct dlm_ctxt *dlm);
 826 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
 827 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node);
 828 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout);
 829 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout);
 830
 831 void dlm_put(struct dlm_ctxt *dlm);
 832 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm);
 833 int dlm_domain_fully_joined(struct dlm_ctxt *dlm);
 834
 835 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 836                               struct dlm_lock_resource *res);
 837 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 838                             struct dlm_lock_resource *res);
 839 static inline void dlm_lockres_get(struct dlm_lock_resource *res)
 840 {
 841         /* This is called on every lookup, so it might be worth
 842          * inlining. */
 843         kref_get(&res->refs);
 844 }
 845 void dlm_lockres_put(struct dlm_lock_resource *res);
 846 void __dlm_unhash_lockres(struct dlm_lock_resource *res);
 847 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 848                           struct dlm_lock_resource *res);
 849 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 850                                                      const char *name,
 851                                                      unsigned int len,
 852                                                      unsigned int hash);
 853 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 854                                                 const char *name,
 855                                                 unsigned int len,
 856                                                 unsigned int hash);
 857 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 858                                               const char *name,
 859                                               unsigned int len);
 860
 861 int dlm_is_host_down(int errno);
 862 void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
 863                               struct dlm_lock_resource *res,
 864                               u8 owner);
 865 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 866                                                  const char *lockid,
 867                                                  int namelen,
 868                                                  int flags);
 869 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 870                                           const char *name,
 871                                           unsigned int namelen);
 872
 873 #define dlm_lockres_set_refmap_bit(bit,res)  \
 874         __dlm_lockres_set_refmap_bit(bit,res,__FILE__,__LINE__)
 875 #define dlm_lockres_clear_refmap_bit(bit,res)  \
 876         __dlm_lockres_clear_refmap_bit(bit,res,__FILE__,__LINE__)
 877
 878 static inline void __dlm_lockres_set_refmap_bit(int bit,
 879                                                 struct dlm_lock_resource *res,
 880                                                 const char *file,
 881                                                 int line)
 882 {
 883         //printk("%s:%d:%.*s: setting bit %d\n", file, line,
 884         //     res->lockname.len, res->lockname.name, bit);
 885         set_bit(bit, res->refmap);
 886 }
 887
 888 static inline void __dlm_lockres_clear_refmap_bit(int bit,
 889                                                   struct dlm_lock_resource *res,
 890                                                   const char *file,
 891                                                   int line)
 892 {
 893         //printk("%s:%d:%.*s: clearing bit %d\n", file, line,
 894         //     res->lockname.len, res->lockname.name, bit);
 895         clear_bit(bit, res->refmap);
 896 }
 897
 898 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 899                                    struct dlm_lock_resource *res,
 900                                    const char *file,
 901                                    int line);
 902 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 903                                    struct dlm_lock_resource *res,
 904                                    int new_lockres,
 905                                    const char *file,
 906                                    int line);
 907 #define dlm_lockres_drop_inflight_ref(d,r)  \
 908         __dlm_lockres_drop_inflight_ref(d,r,__FILE__,__LINE__)
 909 #define dlm_lockres_grab_inflight_ref(d,r)  \
 910         __dlm_lockres_grab_inflight_ref(d,r,0,__FILE__,__LINE__)
 911 #define dlm_lockres_grab_inflight_ref_new(d,r)  \
 912         __dlm_lockres_grab_inflight_ref(d,r,1,__FILE__,__LINE__)
 913
 914 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 915 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 916 void dlm_do_local_ast(struct dlm_ctxt *dlm,
 917                       struct dlm_lock_resource *res,
 918                       struct dlm_lock *lock);
 919 int dlm_do_remote_ast(struct dlm_ctxt *dlm,
 920                       struct dlm_lock_resource *res,
 921                       struct dlm_lock *lock);
 922 void dlm_do_local_bast(struct dlm_ctxt *dlm,
 923                        struct dlm_lock_resource *res,
 924                        struct dlm_lock *lock,
 925                        int blocked_type);
 926 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm,
 927                            struct dlm_lock_resource *res,
 928                            struct dlm_lock *lock,
 929                            int msg_type,
 930                            int blocked_type, int flags);
 931 static inline int dlm_send_proxy_bast(struct dlm_ctxt *dlm,
 932                                       struct dlm_lock_resource *res,
 933                                       struct dlm_lock *lock,
 934                                       int blocked_type)
 935 {
 936         return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_BAST,
 937                                       blocked_type, 0);
 938 }
 939
 940 static inline int dlm_send_proxy_ast(struct dlm_ctxt *dlm,
 941                                      struct dlm_lock_resource *res,
 942                                      struct dlm_lock *lock,
 943                                      int flags)
 944 {
 945         return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_AST,
 946                                       0, flags);
 947 }
 948
 949 void dlm_print_one_lock_resource(struct dlm_lock_resource *res);
 950 void __dlm_print_one_lock_resource(struct dlm_lock_resource *res);
 951
 952 u8 dlm_nm_this_node(struct dlm_ctxt *dlm);
 953 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 954 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 955
 956
 957 int dlm_nm_init(struct dlm_ctxt *dlm);
 958 int dlm_heartbeat_init(struct dlm_ctxt *dlm);
 959 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data);
 960 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data);
 961
 962 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 963 int dlm_finish_migration(struct dlm_ctxt *dlm,
 964                          struct dlm_lock_resource *res,
 965                          u8 old_master);
 966 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
 967                              struct dlm_lock_resource *res);
 968 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res);
 969
 970 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
 971                                void **ret_data);
 972 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
 973                               void **ret_data);
 974 void dlm_assert_master_post_handler(int status, void *data, void *ret_data);
 975 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
 976                               void **ret_data);
 977 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
 978                                 void **ret_data);
 979 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
 980                             void **ret_data);
 981 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
 982                                void **ret_data);
 983 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
 984                                   void **ret_data);
 985 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
 986                                void **ret_data);
 987 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
 988                            void **ret_data);
 989 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
 990                               void **ret_data);
 991 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
 992                           u8 nodenum, u8 *real_master);
 993
 994
 995 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
 996                                struct dlm_lock_resource *res,
 997                                int ignore_higher,
 998                                u8 request_from,
 999                                u32 flags);
1000
1001
1002 int dlm_send_one_lockres(struct dlm_ctxt *dlm,
1003                          struct dlm_lock_resource *res,
1004                          struct dlm_migratable_lockres *mres,
1005                          u8 send_to,
1006                          u8 flags);
1007 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1008                                        struct dlm_lock_resource *res);
1009
1010 /* will exit holding res->spinlock, but may drop in function */
1011 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags);
1012 void __dlm_wait_on_lockres_flags_set(struct dlm_lock_resource *res, int flags);
1013
1014 /* will exit holding res->spinlock, but may drop in function */
1015 static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res)
1016 {
1017         __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_IN_PROGRESS|
1018                                           DLM_LOCK_RES_RECOVERING|
1019                                           DLM_LOCK_RES_MIGRATING));
1020 }
1021
1022 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1023 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1024
1025 /* create/destroy slab caches */
1026 int dlm_init_master_caches(void);
1027 void dlm_destroy_master_caches(void);
1028
1029 int dlm_init_lock_cache(void);
1030 void dlm_destroy_lock_cache(void);
1031
1032 int dlm_init_mle_cache(void);
1033 void dlm_destroy_mle_cache(void);
1034
1035 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up);
1036 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm,
1037                          struct dlm_lock_resource *res);
1038 void dlm_clean_master_list(struct dlm_ctxt *dlm,
1039                            u8 dead_node);
1040 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock);
1041 int __dlm_lockres_has_locks(struct dlm_lock_resource *res);
1042 int __dlm_lockres_unused(struct dlm_lock_resource *res);
1043
1044 static inline const char * dlm_lock_mode_name(int mode)
1045 {
1046         switch (mode) {
1047                 case LKM_EXMODE:
1048                         return "EX";
1049                 case LKM_PRMODE:
1050                         return "PR";
1051                 case LKM_NLMODE:
1052                         return "NL";
1053         }
1054         return "UNKNOWN";
1055 }
1056
1057
1058 static inline int dlm_lock_compatible(int existing, int request)
1059 {
1060         /* NO_LOCK compatible with all */
1061         if (request == LKM_NLMODE ||
1062             existing == LKM_NLMODE)
1063                 return 1;
1064
1065         /* EX incompatible with all non-NO_LOCK */
1066         if (request == LKM_EXMODE)
1067                 return 0;
1068
1069         /* request must be PR, which is compatible with PR */
1070         if (existing == LKM_PRMODE)
1071                 return 1;
1072
1073         return 0;
1074 }
1075
1076 static inline int dlm_lock_on_list(struct list_head *head,
1077                                    struct dlm_lock *lock)
1078 {
1079         struct list_head *iter;
1080         struct dlm_lock *tmplock;
1081
1082         list_for_each(iter, head) {
1083                 tmplock = list_entry(iter, struct dlm_lock, list);
1084                 if (tmplock == lock)
1085                         return 1;
1086         }
1087         return 0;
1088 }
1089
1090
1091 static inline enum dlm_status dlm_err_to_dlm_status(int err)
1092 {
1093         enum dlm_status ret;
1094         if (err == -ENOMEM)
1095                 ret = DLM_SYSERR;
1096         else if (err == -ETIMEDOUT || o2net_link_down(err, NULL))
1097                 ret = DLM_NOLOCKMGR;
1098         else if (err == -EINVAL)
1099                 ret = DLM_BADPARAM;
1100         else if (err == -ENAMETOOLONG)
1101                 ret = DLM_IVBUFLEN;
1102         else
1103                 ret = DLM_BADARGS;
1104         return ret;
1105 }
1106
1107
1108 static inline void dlm_node_iter_init(unsigned long *map,
1109                                       struct dlm_node_iter *iter)
1110 {
1111         memcpy(iter->node_map, map, sizeof(iter->node_map));
1112         iter->curnode = -1;
1113 }
1114
1115 static inline int dlm_node_iter_next(struct dlm_node_iter *iter)
1116 {
1117         int bit;
1118         bit = find_next_bit(iter->node_map, O2NM_MAX_NODES, iter->curnode+1);
1119         if (bit >= O2NM_MAX_NODES) {
1120                 iter->curnode = O2NM_MAX_NODES;
1121                 return -ENOENT;
1122         }
1123         iter->curnode = bit;
1124         return bit;
1125 }
1126
1127
1128
1129 #endif /* DLMCOMMON_H */