1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * underlying calls for lock creation
8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
28 #include <linux/module.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/spinlock.h>
41 #include <linux/delay.h>
44 #include "cluster/heartbeat.h"
45 #include "cluster/nodemanager.h"
46 #include "cluster/tcp.h"
49 #include "dlmcommon.h"
51 #include "dlmconvert.h"
53 #define MLOG_MASK_PREFIX ML_DLM
54 #include "cluster/masklog.h"
56 static DEFINE_SPINLOCK(dlm_cookie_lock
);
57 static u64 dlm_next_cookie
= 1;
59 static enum dlm_status
dlm_send_remote_lock_request(struct dlm_ctxt
*dlm
,
60 struct dlm_lock_resource
*res
,
61 struct dlm_lock
*lock
, int flags
);
62 static void dlm_init_lock(struct dlm_lock
*newlock
, int type
,
64 static void dlm_lock_release(struct kref
*kref
);
65 static void dlm_lock_detach_lockres(struct dlm_lock
*lock
);
67 /* Tell us whether we can grant a new lock request.
69 * caller needs: res->spinlock
72 * returns: 1 if the lock can be granted, 0 otherwise.
74 static int dlm_can_grant_new_lock(struct dlm_lock_resource
*res
,
75 struct dlm_lock
*lock
)
77 struct list_head
*iter
;
78 struct dlm_lock
*tmplock
;
80 list_for_each(iter
, &res
->granted
) {
81 tmplock
= list_entry(iter
, struct dlm_lock
, list
);
83 if (!dlm_lock_compatible(tmplock
->ml
.type
, lock
->ml
.type
))
87 list_for_each(iter
, &res
->converting
) {
88 tmplock
= list_entry(iter
, struct dlm_lock
, list
);
90 if (!dlm_lock_compatible(tmplock
->ml
.type
, lock
->ml
.type
))
97 /* performs lock creation at the lockres master site
100 * taken: takes and drops res->spinlock
102 * returns: DLM_NORMAL, DLM_NOTQUEUED
104 static enum dlm_status
dlmlock_master(struct dlm_ctxt
*dlm
,
105 struct dlm_lock_resource
*res
,
106 struct dlm_lock
*lock
, int flags
)
108 int call_ast
= 0, kick_thread
= 0;
109 enum dlm_status status
= DLM_NORMAL
;
111 mlog_entry("type=%d\n", lock
->ml
.type
);
113 spin_lock(&res
->spinlock
);
114 /* if called from dlm_create_lock_handler, need to
115 * ensure it will not sleep in dlm_wait_on_lockres */
116 status
= __dlm_lockres_state_to_status(res
);
117 if (status
!= DLM_NORMAL
&&
118 lock
->ml
.node
!= dlm
->node_num
) {
119 /* erf. state changed after lock was dropped. */
120 spin_unlock(&res
->spinlock
);
124 __dlm_wait_on_lockres(res
);
125 __dlm_lockres_reserve_ast(res
);
127 if (dlm_can_grant_new_lock(res
, lock
)) {
128 mlog(0, "I can grant this lock right away\n");
129 /* got it right away */
130 lock
->lksb
->status
= DLM_NORMAL
;
133 list_add_tail(&lock
->list
, &res
->granted
);
135 /* for the recovery lock, we can't allow the ast
136 * to be queued since the dlmthread is already
137 * frozen. but the recovery lock is always locked
138 * with LKM_NOQUEUE so we do not need the ast in
139 * this special case */
140 if (!dlm_is_recovery_lock(res
->lockname
.name
,
141 res
->lockname
.len
)) {
145 mlog(0, "%s: returning DLM_NORMAL to "
146 "node %u for reco lock\n", dlm
->name
,
150 /* for NOQUEUE request, unless we get the
151 * lock right away, return DLM_NOTQUEUED */
152 if (flags
& LKM_NOQUEUE
) {
153 status
= DLM_NOTQUEUED
;
154 if (dlm_is_recovery_lock(res
->lockname
.name
,
155 res
->lockname
.len
)) {
156 mlog(0, "%s: returning NOTQUEUED to "
157 "node %u for reco lock\n", dlm
->name
,
162 list_add_tail(&lock
->list
, &res
->blocked
);
167 spin_unlock(&res
->spinlock
);
170 /* either queue the ast or release it */
172 dlm_queue_ast(dlm
, lock
);
174 dlm_lockres_release_ast(dlm
, res
);
176 dlm_lockres_calc_usage(dlm
, res
);
178 dlm_kick_thread(dlm
, res
);
183 void dlm_revert_pending_lock(struct dlm_lock_resource
*res
,
184 struct dlm_lock
*lock
)
186 /* remove from local queue if it failed */
187 list_del_init(&lock
->list
);
188 lock
->lksb
->flags
&= ~DLM_LKSB_GET_LVB
;
195 * taken: takes and drops res->spinlock
197 * returns: DLM_DENIED, DLM_RECOVERING, or net status
199 static enum dlm_status
dlmlock_remote(struct dlm_ctxt
*dlm
,
200 struct dlm_lock_resource
*res
,
201 struct dlm_lock
*lock
, int flags
)
203 enum dlm_status status
= DLM_DENIED
;
204 int lockres_changed
= 1;
206 mlog_entry("type=%d\n", lock
->ml
.type
);
207 mlog(0, "lockres %.*s, flags = 0x%x\n", res
->lockname
.len
,
208 res
->lockname
.name
, flags
);
210 spin_lock(&res
->spinlock
);
212 /* will exit this call with spinlock held */
213 __dlm_wait_on_lockres(res
);
214 res
->state
|= DLM_LOCK_RES_IN_PROGRESS
;
216 /* add lock to local (secondary) queue */
218 list_add_tail(&lock
->list
, &res
->blocked
);
219 lock
->lock_pending
= 1;
220 spin_unlock(&res
->spinlock
);
222 /* spec seems to say that you will get DLM_NORMAL when the lock
223 * has been queued, meaning we need to wait for a reply here. */
224 status
= dlm_send_remote_lock_request(dlm
, res
, lock
, flags
);
226 spin_lock(&res
->spinlock
);
227 res
->state
&= ~DLM_LOCK_RES_IN_PROGRESS
;
228 lock
->lock_pending
= 0;
229 if (status
!= DLM_NORMAL
) {
230 if (status
== DLM_RECOVERING
&&
231 dlm_is_recovery_lock(res
->lockname
.name
,
232 res
->lockname
.len
)) {
233 /* recovery lock was mastered by dead node.
234 * we need to have calc_usage shoot down this
235 * lockres and completely remaster it. */
236 mlog(0, "%s: recovery lock was owned by "
237 "dead node %u, remaster it now.\n",
238 dlm
->name
, res
->owner
);
239 } else if (status
!= DLM_NOTQUEUED
) {
241 * DO NOT call calc_usage, as this would unhash
242 * the remote lockres before we ever get to use
243 * it. treat as if we never made any change to
249 dlm_revert_pending_lock(res
, lock
);
251 } else if (dlm_is_recovery_lock(res
->lockname
.name
,
252 res
->lockname
.len
)) {
253 /* special case for the $RECOVERY lock.
254 * there will never be an AST delivered to put
255 * this lock on the proper secondary queue
256 * (granted), so do it manually. */
257 mlog(0, "%s: $RECOVERY lock for this node (%u) is "
258 "mastered by %u; got lock, manually granting (no ast)\n",
259 dlm
->name
, dlm
->node_num
, res
->owner
);
260 list_move_tail(&lock
->list
, &res
->granted
);
262 spin_unlock(&res
->spinlock
);
265 dlm_lockres_calc_usage(dlm
, res
);
272 /* for remote lock creation.
274 * caller needs: none, but need res->state & DLM_LOCK_RES_IN_PROGRESS
277 * returns: DLM_NOLOCKMGR, or net status
279 static enum dlm_status
dlm_send_remote_lock_request(struct dlm_ctxt
*dlm
,
280 struct dlm_lock_resource
*res
,
281 struct dlm_lock
*lock
, int flags
)
283 struct dlm_create_lock create
;
284 int tmpret
, status
= 0;
289 memset(&create
, 0, sizeof(create
));
290 create
.node_idx
= dlm
->node_num
;
291 create
.requested_type
= lock
->ml
.type
;
292 create
.cookie
= lock
->ml
.cookie
;
293 create
.namelen
= res
->lockname
.len
;
294 create
.flags
= cpu_to_be32(flags
);
295 memcpy(create
.name
, res
->lockname
.name
, create
.namelen
);
297 tmpret
= o2net_send_message(DLM_CREATE_LOCK_MSG
, dlm
->key
, &create
,
298 sizeof(create
), res
->owner
, &status
);
300 // successfully sent and received
301 ret
= status
; // this is already a dlm_status
302 if (ret
== DLM_REJECTED
) {
303 mlog(ML_ERROR
, "%s:%.*s: BUG. this is a stale lockres "
304 "no longer owned by %u. that node is coming back "
305 "up currently.\n", dlm
->name
, create
.namelen
,
306 create
.name
, res
->owner
);
307 dlm_print_one_lock_resource(res
);
312 if (dlm_is_host_down(tmpret
)) {
313 ret
= DLM_RECOVERING
;
314 mlog(0, "node %u died so returning DLM_RECOVERING "
315 "from lock message!\n", res
->owner
);
317 ret
= dlm_err_to_dlm_status(tmpret
);
324 void dlm_lock_get(struct dlm_lock
*lock
)
326 kref_get(&lock
->lock_refs
);
329 void dlm_lock_put(struct dlm_lock
*lock
)
331 kref_put(&lock
->lock_refs
, dlm_lock_release
);
334 static void dlm_lock_release(struct kref
*kref
)
336 struct dlm_lock
*lock
;
338 lock
= container_of(kref
, struct dlm_lock
, lock_refs
);
340 BUG_ON(!list_empty(&lock
->list
));
341 BUG_ON(!list_empty(&lock
->ast_list
));
342 BUG_ON(!list_empty(&lock
->bast_list
));
343 BUG_ON(lock
->ast_pending
);
344 BUG_ON(lock
->bast_pending
);
346 dlm_lock_detach_lockres(lock
);
348 if (lock
->lksb_kernel_allocated
) {
349 mlog(0, "freeing kernel-allocated lksb\n");
355 /* associate a lock with it's lockres, getting a ref on the lockres */
356 void dlm_lock_attach_lockres(struct dlm_lock
*lock
,
357 struct dlm_lock_resource
*res
)
359 dlm_lockres_get(res
);
363 /* drop ref on lockres, if there is still one associated with lock */
364 static void dlm_lock_detach_lockres(struct dlm_lock
*lock
)
366 struct dlm_lock_resource
*res
;
370 lock
->lockres
= NULL
;
371 mlog(0, "removing lock's lockres reference\n");
372 dlm_lockres_put(res
);
376 static void dlm_init_lock(struct dlm_lock
*newlock
, int type
,
379 INIT_LIST_HEAD(&newlock
->list
);
380 INIT_LIST_HEAD(&newlock
->ast_list
);
381 INIT_LIST_HEAD(&newlock
->bast_list
);
382 spin_lock_init(&newlock
->spinlock
);
383 newlock
->ml
.type
= type
;
384 newlock
->ml
.convert_type
= LKM_IVMODE
;
385 newlock
->ml
.highest_blocked
= LKM_IVMODE
;
386 newlock
->ml
.node
= node
;
387 newlock
->ml
.pad1
= 0;
388 newlock
->ml
.list
= 0;
389 newlock
->ml
.flags
= 0;
391 newlock
->bast
= NULL
;
392 newlock
->astdata
= NULL
;
393 newlock
->ml
.cookie
= cpu_to_be64(cookie
);
394 newlock
->ast_pending
= 0;
395 newlock
->bast_pending
= 0;
396 newlock
->convert_pending
= 0;
397 newlock
->lock_pending
= 0;
398 newlock
->unlock_pending
= 0;
399 newlock
->cancel_pending
= 0;
400 newlock
->lksb_kernel_allocated
= 0;
402 kref_init(&newlock
->lock_refs
);
405 struct dlm_lock
* dlm_new_lock(int type
, u8 node
, u64 cookie
,
406 struct dlm_lockstatus
*lksb
)
408 struct dlm_lock
*lock
;
409 int kernel_allocated
= 0;
411 lock
= kcalloc(1, sizeof(*lock
), GFP_NOFS
);
416 /* zero memory only if kernel-allocated */
417 lksb
= kcalloc(1, sizeof(*lksb
), GFP_NOFS
);
422 kernel_allocated
= 1;
425 dlm_init_lock(lock
, type
, node
, cookie
);
426 if (kernel_allocated
)
427 lock
->lksb_kernel_allocated
= 1;
433 /* handler for lock creation net message
436 * taken: takes and drops res->spinlock
438 * returns: DLM_NORMAL, DLM_SYSERR, DLM_IVLOCKID, DLM_NOTQUEUED
440 int dlm_create_lock_handler(struct o2net_msg
*msg
, u32 len
, void *data
)
442 struct dlm_ctxt
*dlm
= data
;
443 struct dlm_create_lock
*create
= (struct dlm_create_lock
*)msg
->buf
;
444 struct dlm_lock_resource
*res
= NULL
;
445 struct dlm_lock
*newlock
= NULL
;
446 struct dlm_lockstatus
*lksb
= NULL
;
447 enum dlm_status status
= DLM_NORMAL
;
449 unsigned int namelen
;
459 namelen
= create
->namelen
;
460 status
= DLM_REJECTED
;
461 if (!dlm_domain_fully_joined(dlm
)) {
462 mlog(ML_ERROR
, "Domain %s not fully joined, but node %u is "
463 "sending a create_lock message for lock %.*s!\n",
464 dlm
->name
, create
->node_idx
, namelen
, name
);
469 status
= DLM_IVBUFLEN
;
470 if (namelen
> DLM_LOCKID_NAME_MAX
) {
476 newlock
= dlm_new_lock(create
->requested_type
,
478 be64_to_cpu(create
->cookie
), NULL
);
484 lksb
= newlock
->lksb
;
486 if (be32_to_cpu(create
->flags
) & LKM_GET_LVB
) {
487 lksb
->flags
|= DLM_LKSB_GET_LVB
;
488 mlog(0, "set DLM_LKSB_GET_LVB flag\n");
491 status
= DLM_IVLOCKID
;
492 res
= dlm_lookup_lockres(dlm
, name
, namelen
);
498 spin_lock(&res
->spinlock
);
499 status
= __dlm_lockres_state_to_status(res
);
500 spin_unlock(&res
->spinlock
);
502 if (status
!= DLM_NORMAL
) {
503 mlog(0, "lockres recovering/migrating/in-progress\n");
507 dlm_lock_attach_lockres(newlock
, res
);
509 status
= dlmlock_master(dlm
, res
, newlock
, be32_to_cpu(create
->flags
));
511 if (status
!= DLM_NORMAL
)
513 dlm_lock_put(newlock
);
516 dlm_lockres_put(res
);
524 /* fetch next node-local (u8 nodenum + u56 cookie) into u64 */
525 static inline void dlm_get_next_cookie(u8 node_num
, u64
*cookie
)
527 u64 tmpnode
= node_num
;
529 /* shift single byte of node num into top 8 bits */
532 spin_lock(&dlm_cookie_lock
);
533 *cookie
= (dlm_next_cookie
| tmpnode
);
534 if (++dlm_next_cookie
& 0xff00000000000000ull
) {
535 mlog(0, "This node's cookie will now wrap!\n");
538 spin_unlock(&dlm_cookie_lock
);
541 enum dlm_status
dlmlock(struct dlm_ctxt
*dlm
, int mode
,
542 struct dlm_lockstatus
*lksb
, int flags
,
543 const char *name
, dlm_astlockfunc_t
*ast
, void *data
,
544 dlm_bastlockfunc_t
*bast
)
546 enum dlm_status status
;
547 struct dlm_lock_resource
*res
= NULL
;
548 struct dlm_lock
*lock
= NULL
;
549 int convert
= 0, recovery
= 0;
551 /* yes this function is a mess.
552 * TODO: clean this up. lots of common code in the
553 * lock and convert paths, especially in the retry blocks */
555 dlm_error(DLM_BADARGS
);
559 status
= DLM_BADPARAM
;
560 if (mode
!= LKM_EXMODE
&& mode
!= LKM_PRMODE
&& mode
!= LKM_NLMODE
) {
565 if (flags
& ~LKM_VALID_FLAGS
) {
570 convert
= (flags
& LKM_CONVERT
);
571 recovery
= (flags
& LKM_RECOVERY
);
574 (!dlm_is_recovery_lock(name
, strlen(name
)) || convert
) ) {
578 if (convert
&& (flags
& LKM_LOCAL
)) {
579 mlog(ML_ERROR
, "strange LOCAL convert request!\n");
584 /* CONVERT request */
586 /* if converting, must pass in a valid dlm_lock */
589 mlog(ML_ERROR
, "NULL lock pointer in convert "
596 mlog(ML_ERROR
, "NULL lockres pointer in convert "
600 dlm_lockres_get(res
);
602 /* XXX: for ocfs2 purposes, the ast/bast/astdata/lksb are
603 * static after the original lock call. convert requests will
604 * ensure that everything is the same, or return DLM_BADARGS.
605 * this means that DLM_DENIED_NOASTS will never be returned.
607 if (lock
->lksb
!= lksb
|| lock
->ast
!= ast
||
608 lock
->bast
!= bast
|| lock
->astdata
!= data
) {
609 status
= DLM_BADARGS
;
610 mlog(ML_ERROR
, "new args: lksb=%p, ast=%p, bast=%p, "
611 "astdata=%p\n", lksb
, ast
, bast
, data
);
612 mlog(ML_ERROR
, "orig args: lksb=%p, ast=%p, bast=%p, "
613 "astdata=%p\n", lock
->lksb
, lock
->ast
,
614 lock
->bast
, lock
->astdata
);
618 dlm_wait_for_recovery(dlm
);
620 if (res
->owner
== dlm
->node_num
)
621 status
= dlmconvert_master(dlm
, res
, lock
, flags
, mode
);
623 status
= dlmconvert_remote(dlm
, res
, lock
, flags
, mode
);
624 if (status
== DLM_RECOVERING
|| status
== DLM_MIGRATING
||
625 status
== DLM_FORWARD
) {
626 /* for now, see how this works without sleeping
627 * and just retry right away. I suspect the reco
628 * or migration will complete fast enough that
629 * no waiting will be necessary */
630 mlog(0, "retrying convert with migration/recovery/"
639 status
= DLM_BADARGS
;
645 status
= DLM_IVBUFLEN
;
646 if (strlen(name
) > DLM_LOCKID_NAME_MAX
|| strlen(name
) < 1) {
651 dlm_get_next_cookie(dlm
->node_num
, &tmpcookie
);
652 lock
= dlm_new_lock(mode
, dlm
->node_num
, tmpcookie
, lksb
);
659 dlm_wait_for_recovery(dlm
);
661 /* find or create the lock resource */
662 res
= dlm_get_lock_resource(dlm
, name
, flags
);
664 status
= DLM_IVLOCKID
;
669 mlog(0, "type=%d, flags = 0x%x\n", mode
, flags
);
670 mlog(0, "creating lock: lock=%p res=%p\n", lock
, res
);
672 dlm_lock_attach_lockres(lock
, res
);
675 lock
->astdata
= data
;
678 if (flags
& LKM_VALBLK
) {
679 mlog(0, "LKM_VALBLK passed by caller\n");
681 /* LVB requests for non PR, PW or EX locks are
683 if (mode
< LKM_PRMODE
)
684 flags
&= ~LKM_VALBLK
;
686 flags
|= LKM_GET_LVB
;
687 lock
->lksb
->flags
|= DLM_LKSB_GET_LVB
;
691 if (res
->owner
== dlm
->node_num
)
692 status
= dlmlock_master(dlm
, res
, lock
, flags
);
694 status
= dlmlock_remote(dlm
, res
, lock
, flags
);
696 if (status
== DLM_RECOVERING
|| status
== DLM_MIGRATING
||
697 status
== DLM_FORWARD
) {
698 mlog(0, "retrying lock with migration/"
699 "recovery/in progress\n");
701 /* no waiting for dlm_reco_thread */
703 if (status
!= DLM_RECOVERING
)
706 mlog(0, "%s: got RECOVERING "
707 "for $RECOVERY lock, master "
708 "was %u\n", dlm
->name
,
710 /* wait to see the node go down, then
711 * drop down and allow the lockres to
712 * get cleaned up. need to remaster. */
713 dlm_wait_for_node_death(dlm
, res
->owner
,
714 DLM_NODE_DEATH_WAIT_MAX
);
716 dlm_wait_for_recovery(dlm
);
721 if (status
!= DLM_NORMAL
) {
722 lock
->lksb
->flags
&= ~DLM_LKSB_GET_LVB
;
723 if (status
!= DLM_NOTQUEUED
)
730 if (status
!= DLM_NORMAL
) {
731 if (lock
&& !convert
)
733 // this is kind of unnecessary
734 lksb
->status
= status
;
737 /* put lockres ref from the convert path
738 * or from dlm_get_lock_resource */
740 dlm_lockres_put(res
);
744 EXPORT_SYMBOL_GPL(dlmlock
);