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