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(NULL, _n, _l)
51
52 enum dlm_mle_type {
53 DLM_MLE_BLOCK = 0,
54 DLM_MLE_MASTER = 1,
55 DLM_MLE_MIGRATION = 2,
56 DLM_MLE_NUM_TYPES = 3,
57 };
58
59 struct dlm_master_list_entry {
60 struct hlist_node master_hash_node;
61 struct list_head hb_events;
62 struct dlm_ctxt *dlm;
63 spinlock_t spinlock;
64 wait_queue_head_t wq;
65 atomic_t woken;
66 struct kref mle_refs;
67 int inuse;
68 unsigned long maybe_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
69 unsigned long vote_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
70 unsigned long response_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
71 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
72 u8 master;
73 u8 new_master;
74 enum dlm_mle_type type;
75 struct o2hb_callback_func mle_hb_up;
76 struct o2hb_callback_func mle_hb_down;
77 struct dlm_lock_resource *mleres;
78 unsigned char mname[DLM_LOCKID_NAME_MAX];
79 unsigned int mnamelen;
80 unsigned int mnamehash;
81 };
82
83 enum dlm_ast_type {
84 DLM_AST = 0,
85 DLM_BAST = 1,
86 DLM_ASTUNLOCK = 2,
87 };
88
89
90 #define LKM_VALID_FLAGS (LKM_VALBLK | LKM_CONVERT | LKM_UNLOCK | \
91 LKM_CANCEL | LKM_INVVALBLK | LKM_FORCE | \
92 LKM_RECOVERY | LKM_LOCAL | LKM_NOQUEUE)
93
94 #define DLM_RECOVERY_LOCK_NAME "$RECOVERY"
95 #define DLM_RECOVERY_LOCK_NAME_LEN 9
96
dlm_is_recovery_lock(const char * lock_name,int name_len)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 node_data;
112 u8 new_master;
113 u8 dead_node;
114 u16 state;
115 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
116 wait_queue_head_t event;
117 };
118
119 enum dlm_ctxt_state {
120 DLM_CTXT_NEW = 0,
121 DLM_CTXT_JOINED = 1,
122 DLM_CTXT_IN_SHUTDOWN = 2,
123 DLM_CTXT_LEAVING = 3,
124 };
125
126 struct dlm_ctxt
127 {
128 struct list_head list;
129 struct hlist_head **lockres_hash;
130 struct list_head dirty_list;
131 struct list_head purge_list;
132 struct list_head pending_asts;
133 struct list_head pending_basts;
134 struct list_head tracking_list;
135 unsigned int purge_count;
136 spinlock_t spinlock;
137 spinlock_t ast_lock;
138 spinlock_t track_lock;
139 char *name;
140 u8 node_num;
141 u32 key;
142 u8 joining_node;
143 wait_queue_head_t dlm_join_events;
144 unsigned long live_nodes_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
145 unsigned long domain_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
146 unsigned long exit_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
dlm_lockres_hash(struct dlm_ctxt * dlm,unsigned i)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
dlm_master_hash(struct dlm_ctxt * dlm,unsigned i)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
dlm_init_work_item(struct dlm_ctxt * dlm,struct dlm_work_item * i,dlm_workfunc_t * f,void * data)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
__dlm_set_joining_node(struct dlm_ctxt * dlm,u8 node)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 #define DLM_LOCK_RES_RECOVERY_WAITING 0x00004000
286
287 /* max milliseconds to wait to sync up a network failure with a node death */
288 #define DLM_NODE_DEATH_WAIT_MAX (5 * 1000)
289
290 #define DLM_PURGE_INTERVAL_MS (8 * 1000)
291
292 struct dlm_lock_resource
293 {
294 /* WARNING: Please see the comment in dlm_init_lockres before
295 * adding fields here. */
296 struct hlist_node hash_node;
297 struct qstr lockname;
298 struct kref refs;
299
300 /*
301 * Please keep granted, converting, and blocked in this order,
302 * as some funcs want to iterate over all lists.
303 *
304 * All four lists are protected by the hash's reference.
305 */
306 struct list_head granted;
307 struct list_head converting;
308 struct list_head blocked;
309 struct list_head purge;
310
311 /*
312 * These two lists require you to hold an additional reference
313 * while they are on the list.
314 */
315 struct list_head dirty;
316 struct list_head recovering; // dlm_recovery_ctxt.resources list
317
318 /* Added during init and removed during release */
319 struct list_head tracking; /* dlm->tracking_list */
320
321 /* unused lock resources have their last_used stamped and are
322 * put on a list for the dlm thread to run. */
323 unsigned long last_used;
324
325 struct dlm_ctxt *dlm;
326
327 unsigned migration_pending:1;
328 atomic_t asts_reserved;
329 spinlock_t spinlock;
330 wait_queue_head_t wq;
331 u8 owner; //node which owns the lock resource, or unknown
332 u16 state;
333 char lvb[DLM_LVB_LEN];
334 unsigned int inflight_locks;
335 unsigned int inflight_assert_workers;
336 unsigned long refmap[BITS_TO_LONGS(O2NM_MAX_NODES)];
337 };
338
339 struct dlm_migratable_lock
340 {
341 __be64 cookie;
342
343 /* these 3 are just padding for the in-memory structure, but
344 * list and flags are actually used when sent over the wire */
345 __be16 pad1;
346 u8 list; // 0=granted, 1=converting, 2=blocked
347 u8 flags;
348
349 s8 type;
350 s8 convert_type;
351 s8 highest_blocked;
352 u8 node;
353 }; // 16 bytes
354
355 struct dlm_lock
356 {
357 struct dlm_migratable_lock ml;
358
359 struct list_head list;
360 struct list_head ast_list;
361 struct list_head bast_list;
362 struct dlm_lock_resource *lockres;
363 spinlock_t spinlock;
364 struct kref lock_refs;
365
366 // ast and bast must be callable while holding a spinlock!
367 dlm_astlockfunc_t *ast;
368 dlm_bastlockfunc_t *bast;
369 void *astdata;
370 struct dlm_lockstatus *lksb;
371 unsigned ast_pending:1,
372 bast_pending:1,
373 convert_pending:1,
374 lock_pending:1,
375 cancel_pending:1,
376 unlock_pending:1,
377 lksb_kernel_allocated:1;
378 };
379
380 enum dlm_lockres_list {
381 DLM_GRANTED_LIST = 0,
382 DLM_CONVERTING_LIST = 1,
383 DLM_BLOCKED_LIST = 2,
384 };
385
dlm_lvb_is_empty(char * lvb)386 static inline int dlm_lvb_is_empty(char *lvb)
387 {
388 int i;
389 for (i=0; i<DLM_LVB_LEN; i++)
390 if (lvb[i])
391 return 0;
392 return 1;
393 }
394
dlm_list_in_text(enum dlm_lockres_list idx)395 static inline char *dlm_list_in_text(enum dlm_lockres_list idx)
396 {
397 if (idx == DLM_GRANTED_LIST)
398 return "granted";
399 else if (idx == DLM_CONVERTING_LIST)
400 return "converting";
401 else if (idx == DLM_BLOCKED_LIST)
402 return "blocked";
403 else
404 return "unknown";
405 }
406
407 static inline struct list_head *
dlm_list_idx_to_ptr(struct dlm_lock_resource * res,enum dlm_lockres_list idx)408 dlm_list_idx_to_ptr(struct dlm_lock_resource *res, enum dlm_lockres_list idx)
409 {
410 struct list_head *ret = NULL;
411 if (idx == DLM_GRANTED_LIST)
412 ret = &res->granted;
413 else if (idx == DLM_CONVERTING_LIST)
414 ret = &res->converting;
415 else if (idx == DLM_BLOCKED_LIST)
416 ret = &res->blocked;
417 else
418 BUG();
419 return ret;
420 }
421
422
423
424
425 struct dlm_node_iter
426 {
427 unsigned long node_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
428 int curnode;
429 };
430
431
432 enum {
433 DLM_MASTER_REQUEST_MSG = 500,
434 DLM_UNUSED_MSG1 = 501,
435 DLM_ASSERT_MASTER_MSG = 502,
436 DLM_CREATE_LOCK_MSG = 503,
437 DLM_CONVERT_LOCK_MSG = 504,
438 DLM_PROXY_AST_MSG = 505,
439 DLM_UNLOCK_LOCK_MSG = 506,
440 DLM_DEREF_LOCKRES_MSG = 507,
441 DLM_MIGRATE_REQUEST_MSG = 508,
442 DLM_MIG_LOCKRES_MSG = 509,
443 DLM_QUERY_JOIN_MSG = 510,
444 DLM_ASSERT_JOINED_MSG = 511,
445 DLM_CANCEL_JOIN_MSG = 512,
446 DLM_EXIT_DOMAIN_MSG = 513,
447 DLM_MASTER_REQUERY_MSG = 514,
448 DLM_LOCK_REQUEST_MSG = 515,
449 DLM_RECO_DATA_DONE_MSG = 516,
450 DLM_BEGIN_RECO_MSG = 517,
451 DLM_FINALIZE_RECO_MSG = 518,
452 DLM_QUERY_REGION = 519,
453 DLM_QUERY_NODEINFO = 520,
454 DLM_BEGIN_EXIT_DOMAIN_MSG = 521,
455 DLM_DEREF_LOCKRES_DONE = 522,
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 = 1,
469 DLM_RECO_NODE_DATA_REQUESTED = 2,
470 DLM_RECO_NODE_DATA_RECEIVING = 3,
471 DLM_RECO_NODE_DATA_DONE = 4,
472 DLM_RECO_NODE_DATA_FINALIZE_SENT = 5,
473 };
474
475
476 enum {
477 DLM_MASTER_RESP_NO = 0,
478 DLM_MASTER_RESP_YES = 1,
479 DLM_MASTER_RESP_MAYBE = 2,
480 DLM_MASTER_RESP_ERROR = 3,
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 * O2NET_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 (O2NET_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 = 1,
658 JOIN_OK_NO_MAP = 2,
659 JOIN_PROTOCOL_MISMATCH = 3,
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 __be32 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_query_region {
738 u8 qr_node;
739 u8 qr_numregions;
740 u8 qr_namelen;
741 u8 pad1;
742 u8 qr_domain[O2NM_MAX_NAME_LEN];
743 u8 qr_regions[O2HB_MAX_REGION_NAME_LEN * O2NM_MAX_REGIONS];
744 };
745
746 struct dlm_node_info {
747 u8 ni_nodenum;
748 u8 pad1;
749 __be16 ni_ipv4_port;
750 __be32 ni_ipv4_address;
751 };
752
753 struct dlm_query_nodeinfo {
754 u8 qn_nodenum;
755 u8 qn_numnodes;
756 u8 qn_namelen;
757 u8 pad1;
758 u8 qn_domain[O2NM_MAX_NAME_LEN];
759 struct dlm_node_info qn_nodes[O2NM_MAX_NODES];
760 };
761
762 struct dlm_exit_domain
763 {
764 u8 node_idx;
765 u8 pad1[3];
766 };
767
768 struct dlm_finalize_reco
769 {
770 u8 node_idx;
771 u8 dead_node;
772 u8 flags;
773 u8 pad1;
774 __be32 pad2;
775 };
776
777 struct dlm_deref_lockres
778 {
779 u32 pad1;
780 u16 pad2;
781 u8 node_idx;
782 u8 namelen;
783
784 u8 name[O2NM_MAX_NAME_LEN];
785 };
786
787 enum {
788 DLM_DEREF_RESPONSE_DONE = 0,
789 DLM_DEREF_RESPONSE_INPROG = 1,
790 };
791
792 struct dlm_deref_lockres_done {
793 u32 pad1;
794 u16 pad2;
795 u8 node_idx;
796 u8 namelen;
797
798 u8 name[O2NM_MAX_NAME_LEN];
799 };
800
801 static inline enum dlm_status
__dlm_lockres_state_to_status(struct dlm_lock_resource * res)802 __dlm_lockres_state_to_status(struct dlm_lock_resource *res)
803 {
804 enum dlm_status status = DLM_NORMAL;
805
806 assert_spin_locked(&res->spinlock);
807
808 if (res->state & (DLM_LOCK_RES_RECOVERING|
809 DLM_LOCK_RES_RECOVERY_WAITING))
810 status = DLM_RECOVERING;
811 else if (res->state & DLM_LOCK_RES_MIGRATING)
812 status = DLM_MIGRATING;
813 else if (res->state & DLM_LOCK_RES_IN_PROGRESS)
814 status = DLM_FORWARD;
815
816 return status;
817 }
818
dlm_get_lock_cookie_node(u64 cookie)819 static inline u8 dlm_get_lock_cookie_node(u64 cookie)
820 {
821 u8 ret;
822 cookie >>= 56;
823 ret = (u8)(cookie & 0xffULL);
824 return ret;
825 }
826
dlm_get_lock_cookie_seq(u64 cookie)827 static inline unsigned long long dlm_get_lock_cookie_seq(u64 cookie)
828 {
829 unsigned long long ret;
830 ret = ((unsigned long long)cookie) & 0x00ffffffffffffffULL;
831 return ret;
832 }
833
834 struct dlm_lock * dlm_new_lock(int type, u8 node, u64 cookie,
835 struct dlm_lockstatus *lksb);
836 void dlm_lock_get(struct dlm_lock *lock);
837 void dlm_lock_put(struct dlm_lock *lock);
838
839 void dlm_lock_attach_lockres(struct dlm_lock *lock,
840 struct dlm_lock_resource *res);
841
842 int dlm_create_lock_handler(struct o2net_msg *msg, u32 len, void *data,
843 void **ret_data);
844 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
845 void **ret_data);
846 int dlm_proxy_ast_handler(struct o2net_msg *msg, u32 len, void *data,
847 void **ret_data);
848
849 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
850 struct dlm_lock *lock);
851 void dlm_revert_pending_lock(struct dlm_lock_resource *res,
852 struct dlm_lock *lock);
853
854 int dlm_unlock_lock_handler(struct o2net_msg *msg, u32 len, void *data,
855 void **ret_data);
856 void dlm_commit_pending_cancel(struct dlm_lock_resource *res,
857 struct dlm_lock *lock);
858 void dlm_commit_pending_unlock(struct dlm_lock_resource *res,
859 struct dlm_lock *lock);
860
861 int dlm_launch_thread(struct dlm_ctxt *dlm);
862 void dlm_complete_thread(struct dlm_ctxt *dlm);
863 int dlm_launch_recovery_thread(struct dlm_ctxt *dlm);
864 void dlm_complete_recovery_thread(struct dlm_ctxt *dlm);
865 void dlm_wait_for_recovery(struct dlm_ctxt *dlm);
866 void dlm_kick_recovery_thread(struct dlm_ctxt *dlm);
867 int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node);
868 void dlm_wait_for_node_death(struct dlm_ctxt *dlm, u8 node, int timeout);
869 void dlm_wait_for_node_recovery(struct dlm_ctxt *dlm, u8 node, int timeout);
870
871 void dlm_put(struct dlm_ctxt *dlm);
872 struct dlm_ctxt *dlm_grab(struct dlm_ctxt *dlm);
873 int dlm_domain_fully_joined(struct dlm_ctxt *dlm);
874
875 void __dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
876 struct dlm_lock_resource *res);
877 void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
878 struct dlm_lock_resource *res);
dlm_lockres_get(struct dlm_lock_resource * res)879 static inline void dlm_lockres_get(struct dlm_lock_resource *res)
880 {
881 /* This is called on every lookup, so it might be worth
882 * inlining. */
883 kref_get(&res->refs);
884 }
885 void dlm_lockres_put(struct dlm_lock_resource *res);
886 void __dlm_unhash_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
887 void __dlm_insert_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
888 struct dlm_lock_resource * __dlm_lookup_lockres_full(struct dlm_ctxt *dlm,
889 const char *name,
890 unsigned int len,
891 unsigned int hash);
892 struct dlm_lock_resource * __dlm_lookup_lockres(struct dlm_ctxt *dlm,
893 const char *name,
894 unsigned int len,
895 unsigned int hash);
896 struct dlm_lock_resource * dlm_lookup_lockres(struct dlm_ctxt *dlm,
897 const char *name,
898 unsigned int len);
899
900 int dlm_is_host_down(int errno);
901
902 struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
903 const char *lockid,
904 int namelen,
905 int flags);
906 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
907 const char *name,
908 unsigned int namelen);
909
910 void dlm_lockres_set_refmap_bit(struct dlm_ctxt *dlm,
911 struct dlm_lock_resource *res, int bit);
912 void dlm_lockres_clear_refmap_bit(struct dlm_ctxt *dlm,
913 struct dlm_lock_resource *res, int bit);
914
915 void dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
916 struct dlm_lock_resource *res);
917 void dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
918 struct dlm_lock_resource *res);
919
920 void __dlm_lockres_grab_inflight_worker(struct dlm_ctxt *dlm,
921 struct dlm_lock_resource *res);
922
923 void dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
924 void dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
925 void __dlm_queue_ast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
926 void __dlm_queue_bast(struct dlm_ctxt *dlm, struct dlm_lock *lock);
927 void dlm_do_local_ast(struct dlm_ctxt *dlm,
928 struct dlm_lock_resource *res,
929 struct dlm_lock *lock);
930 int dlm_do_remote_ast(struct dlm_ctxt *dlm,
931 struct dlm_lock_resource *res,
932 struct dlm_lock *lock);
933 void dlm_do_local_bast(struct dlm_ctxt *dlm,
934 struct dlm_lock_resource *res,
935 struct dlm_lock *lock,
936 int blocked_type);
937 int dlm_send_proxy_ast_msg(struct dlm_ctxt *dlm,
938 struct dlm_lock_resource *res,
939 struct dlm_lock *lock,
940 int msg_type,
941 int blocked_type, int flags);
dlm_send_proxy_bast(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_lock * lock,int blocked_type)942 static inline int dlm_send_proxy_bast(struct dlm_ctxt *dlm,
943 struct dlm_lock_resource *res,
944 struct dlm_lock *lock,
945 int blocked_type)
946 {
947 return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_BAST,
948 blocked_type, 0);
949 }
950
dlm_send_proxy_ast(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,struct dlm_lock * lock,int flags)951 static inline int dlm_send_proxy_ast(struct dlm_ctxt *dlm,
952 struct dlm_lock_resource *res,
953 struct dlm_lock *lock,
954 int flags)
955 {
956 return dlm_send_proxy_ast_msg(dlm, res, lock, DLM_AST,
957 0, flags);
958 }
959
960 void dlm_print_one_lock_resource(struct dlm_lock_resource *res);
961 void __dlm_print_one_lock_resource(struct dlm_lock_resource *res);
962
963 u8 dlm_nm_this_node(struct dlm_ctxt *dlm);
964 void dlm_kick_thread(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
965 void __dlm_dirty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
966
967
968 int dlm_nm_init(struct dlm_ctxt *dlm);
969 int dlm_heartbeat_init(struct dlm_ctxt *dlm);
970 void dlm_hb_node_down_cb(struct o2nm_node *node, int idx, void *data);
971 void dlm_hb_node_up_cb(struct o2nm_node *node, int idx, void *data);
972
973 int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res);
974 int dlm_finish_migration(struct dlm_ctxt *dlm,
975 struct dlm_lock_resource *res,
976 u8 old_master);
977 void dlm_lockres_release_ast(struct dlm_ctxt *dlm,
978 struct dlm_lock_resource *res);
979 void __dlm_lockres_reserve_ast(struct dlm_lock_resource *res);
980
981 int dlm_master_request_handler(struct o2net_msg *msg, u32 len, void *data,
982 void **ret_data);
983 int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data,
984 void **ret_data);
985 void dlm_assert_master_post_handler(int status, void *data, void *ret_data);
986 int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
987 void **ret_data);
988 int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
989 void **ret_data);
990 int dlm_migrate_request_handler(struct o2net_msg *msg, u32 len, void *data,
991 void **ret_data);
992 int dlm_mig_lockres_handler(struct o2net_msg *msg, u32 len, void *data,
993 void **ret_data);
994 int dlm_master_requery_handler(struct o2net_msg *msg, u32 len, void *data,
995 void **ret_data);
996 int dlm_request_all_locks_handler(struct o2net_msg *msg, u32 len, void *data,
997 void **ret_data);
998 int dlm_reco_data_done_handler(struct o2net_msg *msg, u32 len, void *data,
999 void **ret_data);
1000 int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data,
1001 void **ret_data);
1002 int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
1003 void **ret_data);
1004 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
1005 u8 nodenum, u8 *real_master);
1006
1007 void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
1008 struct dlm_lock_resource *res);
1009
1010 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
1011 struct dlm_lock_resource *res,
1012 int ignore_higher,
1013 u8 request_from,
1014 u32 flags);
1015
1016
1017 int dlm_send_one_lockres(struct dlm_ctxt *dlm,
1018 struct dlm_lock_resource *res,
1019 struct dlm_migratable_lockres *mres,
1020 u8 send_to,
1021 u8 flags);
1022 void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
1023 struct dlm_lock_resource *res);
1024
1025 /* will exit holding res->spinlock, but may drop in function */
1026 void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags);
1027
1028 /* will exit holding res->spinlock, but may drop in function */
__dlm_wait_on_lockres(struct dlm_lock_resource * res)1029 static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res)
1030 {
1031 __dlm_wait_on_lockres_flags(res, (DLM_LOCK_RES_IN_PROGRESS|
1032 DLM_LOCK_RES_RECOVERING|
1033 DLM_LOCK_RES_RECOVERY_WAITING|
1034 DLM_LOCK_RES_MIGRATING));
1035 }
1036
1037 void __dlm_unlink_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1038 void __dlm_insert_mle(struct dlm_ctxt *dlm, struct dlm_master_list_entry *mle);
1039
1040 /* create/destroy slab caches */
1041 int dlm_init_master_caches(void);
1042 void dlm_destroy_master_caches(void);
1043
1044 int dlm_init_lock_cache(void);
1045 void dlm_destroy_lock_cache(void);
1046
1047 int dlm_init_mle_cache(void);
1048 void dlm_destroy_mle_cache(void);
1049
1050 void dlm_hb_event_notify_attached(struct dlm_ctxt *dlm, int idx, int node_up);
1051 int dlm_drop_lockres_ref(struct dlm_ctxt *dlm,
1052 struct dlm_lock_resource *res);
1053 void dlm_clean_master_list(struct dlm_ctxt *dlm,
1054 u8 dead_node);
1055 void dlm_force_free_mles(struct dlm_ctxt *dlm);
1056 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock);
1057 int __dlm_lockres_has_locks(struct dlm_lock_resource *res);
1058 int __dlm_lockres_unused(struct dlm_lock_resource *res);
1059
dlm_lock_mode_name(int mode)1060 static inline const char * dlm_lock_mode_name(int mode)
1061 {
1062 switch (mode) {
1063 case LKM_EXMODE:
1064 return "EX";
1065 case LKM_PRMODE:
1066 return "PR";
1067 case LKM_NLMODE:
1068 return "NL";
1069 }
1070 return "UNKNOWN";
1071 }
1072
1073
dlm_lock_compatible(int existing,int request)1074 static inline int dlm_lock_compatible(int existing, int request)
1075 {
1076 /* NO_LOCK compatible with all */
1077 if (request == LKM_NLMODE ||
1078 existing == LKM_NLMODE)
1079 return 1;
1080
1081 /* EX incompatible with all non-NO_LOCK */
1082 if (request == LKM_EXMODE)
1083 return 0;
1084
1085 /* request must be PR, which is compatible with PR */
1086 if (existing == LKM_PRMODE)
1087 return 1;
1088
1089 return 0;
1090 }
1091
dlm_lock_on_list(struct list_head * head,struct dlm_lock * lock)1092 static inline int dlm_lock_on_list(struct list_head *head,
1093 struct dlm_lock *lock)
1094 {
1095 struct dlm_lock *tmplock;
1096
1097 list_for_each_entry(tmplock, head, list) {
1098 if (tmplock == lock)
1099 return 1;
1100 }
1101 return 0;
1102 }
1103
1104
dlm_err_to_dlm_status(int err)1105 static inline enum dlm_status dlm_err_to_dlm_status(int err)
1106 {
1107 enum dlm_status ret;
1108 if (err == -ENOMEM)
1109 ret = DLM_SYSERR;
1110 else if (err == -ETIMEDOUT || o2net_link_down(err, NULL))
1111 ret = DLM_NOLOCKMGR;
1112 else if (err == -EINVAL)
1113 ret = DLM_BADPARAM;
1114 else if (err == -ENAMETOOLONG)
1115 ret = DLM_IVBUFLEN;
1116 else
1117 ret = DLM_BADARGS;
1118 return ret;
1119 }
1120
1121
dlm_node_iter_init(unsigned long * map,struct dlm_node_iter * iter)1122 static inline void dlm_node_iter_init(unsigned long *map,
1123 struct dlm_node_iter *iter)
1124 {
1125 memcpy(iter->node_map, map, sizeof(iter->node_map));
1126 iter->curnode = -1;
1127 }
1128
dlm_node_iter_next(struct dlm_node_iter * iter)1129 static inline int dlm_node_iter_next(struct dlm_node_iter *iter)
1130 {
1131 int bit;
1132 bit = find_next_bit(iter->node_map, O2NM_MAX_NODES, iter->curnode+1);
1133 if (bit >= O2NM_MAX_NODES) {
1134 iter->curnode = O2NM_MAX_NODES;
1135 return -ENOENT;
1136 }
1137 iter->curnode = bit;
1138 return bit;
1139 }
1140
dlm_set_lockres_owner(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 owner)1141 static inline void dlm_set_lockres_owner(struct dlm_ctxt *dlm,
1142 struct dlm_lock_resource *res,
1143 u8 owner)
1144 {
1145 assert_spin_locked(&res->spinlock);
1146
1147 res->owner = owner;
1148 }
1149
dlm_change_lockres_owner(struct dlm_ctxt * dlm,struct dlm_lock_resource * res,u8 owner)1150 static inline void dlm_change_lockres_owner(struct dlm_ctxt *dlm,
1151 struct dlm_lock_resource *res,
1152 u8 owner)
1153 {
1154 assert_spin_locked(&res->spinlock);
1155
1156 if (owner != res->owner)
1157 dlm_set_lockres_owner(dlm, res, owner);
1158 }
1159
1160 #endif /* DLMCOMMON_H */
1161