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,
  54         DLM_MLE_MASTER,
  55         DLM_MLE_MIGRATION,
  56         DLM_MLE_NUM_TYPES
  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,
  86         DLM_ASTUNLOCK
  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,
 123         DLM_CTXT_IN_SHUTDOWN,
 124         DLM_CTXT_LEAVING,
 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,
 392         DLM_BLOCKED_LIST
 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 struct list_head *
 405 dlm_list_idx_to_ptr(struct dlm_lock_resource *res, enum dlm_lockres_list idx)
 406 {
 407         struct list_head *ret = NULL;
 408         if (idx == DLM_GRANTED_LIST)
 409                 ret = &res->granted;
 410         else if (idx == DLM_CONVERTING_LIST)
 411                 ret = &res->converting;
 412         else if (idx == DLM_BLOCKED_LIST)
 413                 ret = &res->blocked;
 414         else
 415                 BUG();
 416         return ret;
 417 }
 418
 419
 420
 421
 422 struct dlm_node_iter
 423 {
 424         unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
 425         int curnode;
 426 };
 427
 428
 429 enum {
 430         DLM_MASTER_REQUEST_MSG    = 500,
 431         DLM_UNUSED_MSG1,         /* 501 */
 432         DLM_ASSERT_MASTER_MSG,   /* 502 */
 433         DLM_CREATE_LOCK_MSG,     /* 503 */
 434         DLM_CONVERT_LOCK_MSG,    /* 504 */
 435         DLM_PROXY_AST_MSG,       /* 505 */
 436         DLM_UNLOCK_LOCK_MSG,     /* 506 */
 437         DLM_DEREF_LOCKRES_MSG,   /* 507 */
 438         DLM_MIGRATE_REQUEST_MSG, /* 508 */
 439         DLM_MIG_LOCKRES_MSG,     /* 509 */
 440         DLM_QUERY_JOIN_MSG,      /* 510 */
 441         DLM_ASSERT_JOINED_MSG,   /* 511 */
 442         DLM_CANCEL_JOIN_MSG,     /* 512 */
 443         DLM_EXIT_DOMAIN_MSG,     /* 513 */
 444         DLM_MASTER_REQUERY_MSG,  /* 514 */
 445         DLM_LOCK_REQUEST_MSG,    /* 515 */
 446         DLM_RECO_DATA_DONE_MSG,  /* 516 */
 447         DLM_BEGIN_RECO_MSG,      /* 517 */
 448         DLM_FINALIZE_RECO_MSG,   /* 518 */
 449         DLM_QUERY_REGION,        /* 519 */
 450         DLM_QUERY_NODEINFO,      /* 520 */
 451 };
 452
 453 struct dlm_reco_node_data
 454 {
 455         int state;
 456         u8 node_num;
 457         struct list_head list;
 458 };
 459
 460 enum {
 461         DLM_RECO_NODE_DATA_DEAD = -1,
 462         DLM_RECO_NODE_DATA_INIT = 0,
 463         DLM_RECO_NODE_DATA_REQUESTING,
 464         DLM_RECO_NODE_DATA_REQUESTED,
 465         DLM_RECO_NODE_DATA_RECEIVING,
 466         DLM_RECO_NODE_DATA_DONE,
 467         DLM_RECO_NODE_DATA_FINALIZE_SENT,
 468 };
 469
 470
 471 enum {
 472         DLM_MASTER_RESP_NO = 0,
 473         DLM_MASTER_RESP_YES,
 474         DLM_MASTER_RESP_MAYBE,
 475         DLM_MASTER_RESP_ERROR
 476 };
 477
 478
 479 struct dlm_master_request
 480 {
 481         u8 node_idx;
 482         u8 namelen;
 483         __be16 pad1;
 484         __be32 flags;
 485
 486         u8 name[O2NM_MAX_NAME_LEN];
 487 };
 488
 489 #define DLM_ASSERT_RESPONSE_REASSERT       0x00000001
 490 #define DLM_ASSERT_RESPONSE_MASTERY_REF    0x00000002
 491
 492 #define DLM_ASSERT_MASTER_MLE_CLEANUP      0x00000001
 493 #define DLM_ASSERT_MASTER_REQUERY          0x00000002
 494 #define DLM_ASSERT_MASTER_FINISH_MIGRATION 0x00000004
 495 struct dlm_assert_master
 496 {
 497         u8 node_idx;
 498         u8 namelen;
 499         __be16 pad1;
 500         __be32 flags;
 501
 502         u8 name[O2NM_MAX_NAME_LEN];
 503 };
 504
 505 #define DLM_MIGRATE_RESPONSE_MASTERY_REF   0x00000001
 506
 507 struct dlm_migrate_request
 508 {
 509         u8 master;
 510         u8 new_master;
 511         u8 namelen;
 512         u8 pad1;
 513         __be32 pad2;
 514         u8 name[O2NM_MAX_NAME_LEN];
 515 };
 516
 517 struct dlm_master_requery
 518 {
 519         u8 pad1;
 520         u8 pad2;
 521         u8 node_idx;
 522         u8 namelen;
 523         __be32 pad3;
 524         u8 name[O2NM_MAX_NAME_LEN];
 525 };
 526
 527 #define DLM_MRES_RECOVERY   0x01
 528 #define DLM_MRES_MIGRATION  0x02
 529 #define DLM_MRES_ALL_DONE   0x04
 530
 531 /*
 532  * We would like to get one whole lockres into a single network
 533  * message whenever possible.  Generally speaking, there will be
 534  * at most one dlm_lock on a lockres for each node in the cluster,
 535  * plus (infrequently) any additional locks coming in from userdlm.
 536  *
 537  * struct _dlm_lockres_page
 538  * {
 539  *      dlm_migratable_lockres mres;
 540  *      dlm_migratable_lock ml[DLM_MAX_MIGRATABLE_LOCKS];
 541  *      u8 pad[DLM_MIG_LOCKRES_RESERVED];
 542  * };
 543  *
 544  * from ../cluster/tcp.h
 545  *    NET_MAX_PAYLOAD_BYTES  (4096 - sizeof(net_msg))
 546  *    (roughly 4080 bytes)
 547  * and sizeof(dlm_migratable_lockres) = 112 bytes
 548  * and sizeof(dlm_migratable_lock) = 16 bytes
 549  *
 550  * Choosing DLM_MAX_MIGRATABLE_LOCKS=240 and
 551  * DLM_MIG_LOCKRES_RESERVED=128 means we have this:
 552  *
 553  *  (DLM_MAX_MIGRATABLE_LOCKS * sizeof(dlm_migratable_lock)) +
 554  *     sizeof(dlm_migratable_lockres) + DLM_MIG_LOCKRES_RESERVED =
 555  *        NET_MAX_PAYLOAD_BYTES
 556  *  (240 * 16) + 112 + 128 = 4080
 557  *
 558  * So a lockres would need more than 240 locks before it would
 559  * use more than one network packet to recover.  Not too bad.
 560  */
 561 #define DLM_MAX_MIGRATABLE_LOCKS   240
 562
 563 struct dlm_migratable_lockres
 564 {
 565         u8 master;
 566         u8 lockname_len;
 567         u8 num_locks;    // locks sent in this structure
 568         u8 flags;
 569         __be32 total_locks; // locks to be sent for this migration cookie
 570         __be64 mig_cookie;  // cookie for this lockres migration
 571                          // or zero if not needed
 572         // 16 bytes
 573         u8 lockname[DLM_LOCKID_NAME_MAX];
 574         // 48 bytes
 575         u8 lvb[DLM_LVB_LEN];
 576         // 112 bytes
 577         struct dlm_migratable_lock ml[0];  // 16 bytes each, begins at byte 112
 578 };
 579 #define DLM_MIG_LOCKRES_MAX_LEN  \
 580         (sizeof(struct dlm_migratable_lockres) + \
 581          (sizeof(struct dlm_migratable_lock) * \
 582           DLM_MAX_MIGRATABLE_LOCKS) )
 583
 584 /* from above, 128 bytes
 585  * for some undetermined future use */
 586 #define DLM_MIG_LOCKRES_RESERVED   (NET_MAX_PAYLOAD_BYTES - \
 587                                     DLM_MIG_LOCKRES_MAX_LEN)
 588
 589 struct dlm_create_lock
 590 {
 591         __be64 cookie;
 592
 593         __be32 flags;
 594         u8 pad1;
 595         u8 node_idx;
 596         s8 requested_type;
 597         u8 namelen;
 598
 599         u8 name[O2NM_MAX_NAME_LEN];
 600 };
 601
 602 struct dlm_convert_lock
 603 {
 604         __be64 cookie;
 605
 606         __be32 flags;
 607         u8 pad1;
 608         u8 node_idx;
 609         s8 requested_type;
 610         u8 namelen;
 611
 612         u8 name[O2NM_MAX_NAME_LEN];
 613
 614         s8 lvb[0];
 615 };
 616 #define DLM_CONVERT_LOCK_MAX_LEN  (sizeof(struct dlm_convert_lock)+DLM_LVB_LEN)
 617
 618 struct dlm_unlock_lock
 619 {
 620         __be64 cookie;
 621
 622         __be32 flags;
 623         __be16 pad1;
 624         u8 node_idx;
 625         u8 namelen;
 626
 627         u8 name[O2NM_MAX_NAME_LEN];
 628
 629         s8 lvb[0];
 630 };
 631 #define DLM_UNLOCK_LOCK_MAX_LEN  (sizeof(struct dlm_unlock_lock)+DLM_LVB_LEN)
 632
 633 struct dlm_proxy_ast
 634 {
 635         __be64 cookie;
 636
 637         __be32 flags;
 638         u8 node_idx;
 639         u8 type;
 640         u8 blocked_type;
 641         u8 namelen;
 642
 643         u8 name[O2NM_MAX_NAME_LEN];
 644
 645         s8 lvb[0];
 646 };
 647 #define DLM_PROXY_AST_MAX_LEN  (sizeof(struct dlm_proxy_ast)+DLM_LVB_LEN)
 648
 649 #define DLM_MOD_KEY (0x666c6172)
 650 enum dlm_query_join_response_code {
 651         JOIN_DISALLOW = 0,
 652         JOIN_OK,
 653         JOIN_OK_NO_MAP,
 654         JOIN_PROTOCOL_MISMATCH,
 655 };
 656
 657 struct dlm_query_join_packet {
 658         u8 code;        /* Response code.  dlm_minor and fs_minor
 659                            are only valid if this is JOIN_OK */
 660         u8 dlm_minor;   /* The minor version of the protocol the
 661                            dlm is speaking. */
 662         u8 fs_minor;    /* The minor version of the protocol the
 663                            filesystem is speaking. */
 664         u8 reserved;
 665 };
 666
 667 union dlm_query_join_response {
 668         u32 intval;
 669         struct dlm_query_join_packet packet;
 670 };
 671
 672 struct dlm_lock_request
 673 {
 674         u8 node_idx;
 675         u8 dead_node;
 676         __be16 pad1;
 677         __be32 pad2;
 678 };
 679
 680 struct dlm_reco_data_done
 681 {
 682         u8 node_idx;
 683         u8 dead_node;
 684         __be16 pad1;
 685         __be32 pad2;
 686
 687         /* unused for now */
 688         /* eventually we can use this to attempt
 689          * lvb recovery based on each node's info */
 690         u8 reco_lvb[DLM_LVB_LEN];
 691 };
 692
 693 struct dlm_begin_reco
 694 {
 695         u8 node_idx;
 696         u8 dead_node;
 697         __be16 pad1;
 698         __be32 pad2;
 699 };
 700
 701
 702 #define BITS_PER_BYTE 8
 703 #define BITS_TO_BYTES(bits) (((bits)+BITS_PER_BYTE-1)/BITS_PER_BYTE)
 704
 705 struct dlm_query_join_request
 706 {
 707         u8 node_idx;
 708         u8 pad1[2];
 709         u8 name_len;
 710         struct dlm_protocol_version dlm_proto;
 711         struct dlm_protocol_version fs_proto;
 712         u8 domain[O2NM_MAX_NAME_LEN];
 713         u8 node_map[BITS_TO_BYTES(O2NM_MAX_NODES)];
 714 };
 715
 716 struct dlm_assert_joined
 717 {
 718         u8 node_idx;
 719         u8 pad1[2];
 720         u8 name_len;
 721         u8 domain[O2NM_MAX_NAME_LEN];
 722 };
 723
 724 struct dlm_cancel_join
 725 {
 726         u8 node_idx;
 727         u8 pad1[2];
 728         u8 name_len;
 729         u8 domain[O2NM_MAX_NAME_LEN];
 730 };
 731
 732 struct dlm_query_region {
 733         u8 qr_node;
 734         u8 qr_numregions;
 735         u8 qr_namelen;
 736         u8 pad1;
 737         u8 qr_domain[O2NM_MAX_NAME_LEN];
 738         u8 qr_regions[O2HB_MAX_REGION_NAME_LEN * O2NM_MAX_REGIONS];
 739 };
 740
 741 struct dlm_node_info {
 742         u8 ni_nodenum;
 743         u8 pad1;
 744         u16 ni_ipv4_port;
 745         u32 ni_ipv4_address;
 746 };
 747
 748 struct dlm_query_nodeinfo {
 749         u8 qn_nodenum;
 750         u8 qn_numnodes;
 751         u8 qn_namelen;
 752         u8 pad1;
 753         u8 qn_domain[O2NM_MAX_NAME_LEN];
 754         struct dlm_node_info qn_nodes[O2NM_MAX_NODES];
 755 };
 756
 757 struct dlm_exit_domain
 758 {
 759         u8 node_idx;
 760         u8 pad1[3];
 761 };
 762
 763 struct dlm_finalize_reco
 764 {
 765         u8 node_idx;
 766         u8 dead_node;
 767         u8 flags;
 768         u8 pad1;
 769         __be32 pad2;
 770 };
 771
 772 struct dlm_deref_lockres
 773 {
 774         u32 pad1;
 775         u16 pad2;
 776         u8 node_idx;
 777         u8 namelen;
 778
 779         u8 name[O2NM_MAX_NAME_LEN];
 780 };
 781
 782 static inline enum dlm_status
 783 __dlm_lockres_state_to_status(struct dlm_lock_resource *res)
 784 {
 785         enum dlm_status status = DLM_NORMAL;
 786
 787         assert_spin_locked(&res->spinlock);
 788
 789         if (res->state & DLM_LOCK_RES_RECOVERING)
 790                 status = DLM_RECOVERING;
 791         else if (res->state & DLM_LOCK_RES_MIGRATING)
 792                 status = DLM_MIGRATING;
 793         else if (res->state & DLM_LOCK_RES_IN_PROGRESS)
 794                 status = DLM_FORWARD;
 795
 796         return status;
 797 }
 798
 799 static inline u8 dlm_get_lock_cookie_node(u64 cookie)
 800 {
 801         u8 ret;
 802         cookie >>= 56;
 803         ret = (u8)(cookie & 0xffULL);
 804         return ret;
 805 }
 806
 807 static inline unsigned long long dlm_get_lock_cookie_seq(u64 cookie)
 808 {
 809         unsigned long long ret;
 810         ret = ((unsigned long long)cookie) & 0x00ffffffffffffffULL;
 811         return ret;
 812 }
 813
 814 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
 815                                struct dlm_lockstatus *lksb);
 816 void dlm_lock_get(struct dlm_lock *lock);
 817 void dlm_lock_put(struct dlm_lock *lock);
 818
 819 void dlm_lock_attach_lockres(struct dlm_lock *lock,
 820                              struct dlm_lock_resource *res);
 821
 822 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 823                             void **ret_data);
 824 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 825                              void **ret_data);
 826 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
 827                           void **ret_data);
 828
 829 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
 830                                 struct dlm_lock *lock);
 831 void dlm_revert_pending_lock(struct dlm_lock_resource *res,
 832                              struct dlm_lock *lock);
 833
 834 int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 835                             void **ret_data);
 836 void dlm_commit_pending_cancel(struct dlm_lock_resource *res,
 837                                struct dlm_lock *lock);
 838 void dlm_commit_pending_unlock(struct dlm_lock_resource *res,
 839                                struct dlm_lock *lock);
 840
 841 int dlm_launch_thread(struct dlm_ctxt *dlm);
 842 void dlm_complete_thread(struct dlm_ctxt *dlm);
 843 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
 844 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
 845 void dlm_wait_for_recovery(struct dlm_ctxt *dlm);
 846 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
 847 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node);
 848 int dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout);
 849 int dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout);
 850
 851 void dlm_put(struct dlm_ctxt *dlm);
 852 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm);
 853 int dlm_domain_fully_joined(struct dlm_ctxt *dlm);
 854
 855 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 856                               struct dlm_lock_resource *res);
 857 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
 858                             struct dlm_lock_resource *res);
 859 static inline void dlm_lockres_get(struct dlm_lock_resource *res)
 860 {
 861         /* This is called on every lookup, so it might be worth
 862          * inlining. */
 863         kref_get(&res->refs);
 864 }
 865 void dlm_lockres_put(struct dlm_lock_resource *res);
 866 void __dlm_unhash_lockres(struct dlm_lock_resource *res);
 867 void __dlm_insert_lockres(struct dlm_ctxt *dlm,
 868                           struct dlm_lock_resource *res);
 869 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
 870                                                      const char *name,
 871                                                      unsigned int len,
 872                                                      unsigned int hash);
 873 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
 874                                                 const char *name,
 875                                                 unsigned int len,
 876                                                 unsigned int hash);
 877 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
 878                                               const char *name,
 879                                               unsigned int len);
 880
 881 int dlm_is_host_down(int errno);
 882
 883 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 884                                                  const char *lockid,
 885                                                  int namelen,
 886                                                  int flags);
 887 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
 888                                           const char *name,
 889                                           unsigned int namelen);
 890
 891 #define dlm_lockres_set_refmap_bit(bit,res)  \
 892         __dlm_lockres_set_refmap_bit(bit,res,__FILE__,__LINE__)
 893 #define dlm_lockres_clear_refmap_bit(bit,res)  \
 894         __dlm_lockres_clear_refmap_bit(bit,res,__FILE__,__LINE__)
 895
 896 static inline void __dlm_lockres_set_refmap_bit(int bit,
 897                                                 struct dlm_lock_resource *res,
 898                                                 const char *file,
 899                                                 int line)
 900 {
 901         //printk("%s:%d:%.*s: setting bit %d\n", file, line,
 902         //     res->lockname.len, res->lockname.name, bit);
 903         set_bit(bit, res->refmap);
 904 }
 905
 906 static inline void __dlm_lockres_clear_refmap_bit(int bit,
 907                                                   struct dlm_lock_resource *res,
 908                                                   const char *file,
 909                                                   int line)
 910 {
 911         //printk("%s:%d:%.*s: clearing bit %d\n", file, line,
 912         //     res->lockname.len, res->lockname.name, bit);
 913         clear_bit(bit, res->refmap);
 914 }
 915
 916 void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
 917                                    struct dlm_lock_resource *res,
 918                                    const char *file,
 919                                    int line);
 920 void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
 921                                    struct dlm_lock_resource *res,
 922                                    int new_lockres,
 923                                    const char *file,
 924                                    int line);
 925 #define dlm_lockres_drop_inflight_ref(d,r)  \
 926         __dlm_lockres_drop_inflight_ref(d,r,__FILE__,__LINE__)
 927 #define dlm_lockres_grab_inflight_ref(d,r)  \
 928         __dlm_lockres_grab_inflight_ref(d,r,0,__FILE__,__LINE__)
 929 #define dlm_lockres_grab_inflight_ref_new(d,r)  \
 930         __dlm_lockres_grab_inflight_ref(d,r,1,__FILE__,__LINE__)
 931
 932 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 933 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 934 void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 935 void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
 936 void dlm_do_local_ast(struct dlm_ctxt *dlm,
 937                       struct dlm_lock_resource *res,
 938                       struct dlm_lock *lock);
 939 int dlm_do_remote_ast(struct dlm_ctxt *dlm,
 940                       struct dlm_lock_resource *res,
 941                       struct dlm_lock *lock);
 942 void dlm_do_local_bast(struct dlm_ctxt *dlm,
 943                        struct dlm_lock_resource *res,
 944                        struct dlm_lock *lock,
 945                        int blocked_type);
 946 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm,
 947                            struct dlm_lock_resource *res,
 948                            struct dlm_lock *lock,
 949                            int msg_type,
 950                            int blocked_type, int flags);
 951 static inline int dlm_send_proxy_bast(struct dlm_ctxt *dlm,
 952                                       struct dlm_lock_resource *res,
 953                                       struct dlm_lock *lock,
 954                                       int blocked_type)
 955 {
 956         return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_BAST,
 957                                       blocked_type, 0);
 958 }
 959
 960 static inline int dlm_send_proxy_ast(struct dlm_ctxt *dlm,
 961                                      struct dlm_lock_resource *res,
 962                                      struct dlm_lock *lock,
 963                                      int flags)
 964 {
 965         return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_AST,
 966                                       0, flags);
 967 }
 968
 969 void dlm_print_one_lock_resource(struct dlm_lock_resource *res);
 970 void __dlm_print_one_lock_resource(struct dlm_lock_resource *res);
 971
 972 u8 dlm_nm_this_node(struct dlm_ctxt *dlm);
 973 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 974 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 975
 976
 977 int dlm_nm_init(struct dlm_ctxt *dlm);
 978 int dlm_heartbeat_init(struct dlm_ctxt *dlm);
 979 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data);
 980 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data);
 981
 982 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
 983 int dlm_finish_migration(struct dlm_ctxt *dlm,
 984                          struct dlm_lock_resource *res,
 985                          u8 old_master);
 986 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
 987                              struct dlm_lock_resource *res);
 988 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res);
 989
 990 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
 991                                void **ret_data);
 992 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
 993                               void **ret_data);
 994 void dlm_assert_master_post_handler(int status, void *data, void *ret_data);
 995 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
 996                               void **ret_data);
 997 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
 998                                 void **ret_data);
 999 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
1000                             void **ret_data);
1001 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
1002                                void **ret_data);
1003 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
1004                                   void **ret_data);
1005 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
1006                                void **ret_data);
1007 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
1008                            void **ret_data);
1009 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
1010                               void **ret_data);
1011 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1012                           u8 nodenum, u8 *real_master);
1013
1014
1015 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
1016                                struct dlm_lock_resource *res,
1017                                int ignore_higher,
1018                                u8 request_from,
1019                                u32 flags);
1020
1021
1022 int dlm_send_one_lockres(struct dlm_ctxt *dlm,
1023                          struct dlm_lock_resource *res,
1024                          struct dlm_migratable_lockres *mres,
1025                          u8 send_to,
1026                          u8 flags);
1027 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1028                                        struct dlm_lock_resource *res);
1029
1030 /* will exit holding res->spinlock, but may drop in function */
1031 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags);
1032 void __dlm_wait_on_lockres_flags_set(struct dlm_lock_resource *res, int flags);
1033
1034 /* will exit holding res->spinlock, but may drop in function */
1035 static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res)
1036 {
1037         __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_IN_PROGRESS|
1038                                           DLM_LOCK_RES_RECOVERING|
1039                                           DLM_LOCK_RES_MIGRATING));
1040 }
1041
1042 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1043 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1044
1045 /* create/destroy slab caches */
1046 int dlm_init_master_caches(void);
1047 void dlm_destroy_master_caches(void);
1048
1049 int dlm_init_lock_cache(void);
1050 void dlm_destroy_lock_cache(void);
1051
1052 int dlm_init_mle_cache(void);
1053 void dlm_destroy_mle_cache(void);
1054
1055 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up);
1056 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm,
1057                          struct dlm_lock_resource *res);
1058 void dlm_clean_master_list(struct dlm_ctxt *dlm,
1059                            u8 dead_node);
1060 void dlm_force_free_mles(struct dlm_ctxt *dlm);
1061 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock);
1062 int __dlm_lockres_has_locks(struct dlm_lock_resource *res);
1063 int __dlm_lockres_unused(struct dlm_lock_resource *res);
1064
1065 static inline const char * dlm_lock_mode_name(int mode)
1066 {
1067         switch (mode) {
1068                 case LKM_EXMODE:
1069                         return "EX";
1070                 case LKM_PRMODE:
1071                         return "PR";
1072                 case LKM_NLMODE:
1073                         return "NL";
1074         }
1075         return "UNKNOWN";
1076 }
1077
1078
1079 static inline int dlm_lock_compatible(int existing, int request)
1080 {
1081         /* NO_LOCK compatible with all */
1082         if (request == LKM_NLMODE ||
1083             existing == LKM_NLMODE)
1084                 return 1;
1085
1086         /* EX incompatible with all non-NO_LOCK */
1087         if (request == LKM_EXMODE)
1088                 return 0;
1089
1090         /* request must be PR, which is compatible with PR */
1091         if (existing == LKM_PRMODE)
1092                 return 1;
1093
1094         return 0;
1095 }
1096
1097 static inline int dlm_lock_on_list(struct list_head *head,
1098                                    struct dlm_lock *lock)
1099 {
1100         struct list_head *iter;
1101         struct dlm_lock *tmplock;
1102
1103         list_for_each(iter, head) {
1104                 tmplock = list_entry(iter, struct dlm_lock, list);
1105                 if (tmplock == lock)
1106                         return 1;
1107         }
1108         return 0;
1109 }
1110
1111
1112 static inline enum dlm_status dlm_err_to_dlm_status(int err)
1113 {
1114         enum dlm_status ret;
1115         if (err == -ENOMEM)
1116                 ret = DLM_SYSERR;
1117         else if (err == -ETIMEDOUT || o2net_link_down(err, NULL))
1118                 ret = DLM_NOLOCKMGR;
1119         else if (err == -EINVAL)
1120                 ret = DLM_BADPARAM;
1121         else if (err == -ENAMETOOLONG)
1122                 ret = DLM_IVBUFLEN;
1123         else
1124                 ret = DLM_BADARGS;
1125         return ret;
1126 }
1127
1128
1129 static inline void dlm_node_iter_init(unsigned long *map,
1130                                       struct dlm_node_iter *iter)
1131 {
1132         memcpy(iter->node_map, map, sizeof(iter->node_map));
1133         iter->curnode = -1;
1134 }
1135
1136 static inline int dlm_node_iter_next(struct dlm_node_iter *iter)
1137 {
1138         int bit;
1139         bit = find_next_bit(iter->node_map, O2NM_MAX_NODES, iter->curnode+1);
1140         if (bit >= O2NM_MAX_NODES) {
1141                 iter->curnode = O2NM_MAX_NODES;
1142                 return -ENOENT;
1143         }
1144         iter->curnode = bit;
1145         return bit;
1146 }
1147
1148 static inline void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
1149                                          struct dlm_lock_resource *res,
1150                                          u8 owner)
1151 {
1152         assert_spin_locked(&res->spinlock);
1153
1154         res->owner = owner;
1155 }
1156
1157 static inline void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
1158                                             struct dlm_lock_resource *res,
1159                                             u8 owner)
1160 {
1161         assert_spin_locked(&res->spinlock);
1162
1163         if (owner != res->owner)
1164                 dlm_set_lockres_owner(dlm, res, owner);
1165 }
1166
1167 #endif /* DLMCOMMON_H */