1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Code which implements the kernel side of a minimal userspace
7 * interface to our DLM.
9 * Many of the functions here are pared down versions of dlmglue.c
12 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
14 * This program is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU General Public
16 * License as published by the Free Software Foundation; either
17 * version 2 of the License, or (at your option) any later version.
19 * This program is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 * General Public License for more details.
24 * You should have received a copy of the GNU General Public
25 * License along with this program; if not, write to the
26 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
27 * Boston, MA 021110-1307, USA.
30 #include <linux/signal.h>
32 #include <linux/module.h>
34 #include <linux/types.h>
35 #include <linux/crc32.h>
38 #include "cluster/nodemanager.h"
39 #include "cluster/heartbeat.h"
40 #include "cluster/tcp.h"
46 #define MLOG_MASK_PREFIX ML_DLMFS
47 #include "cluster/masklog.h"
49 static inline int user_check_wait_flag(struct user_lock_res
*lockres
,
54 spin_lock(&lockres
->l_lock
);
55 ret
= lockres
->l_flags
& flag
;
56 spin_unlock(&lockres
->l_lock
);
61 static inline void user_wait_on_busy_lock(struct user_lock_res
*lockres
)
64 wait_event(lockres
->l_event
,
65 !user_check_wait_flag(lockres
, USER_LOCK_BUSY
));
68 static inline void user_wait_on_blocked_lock(struct user_lock_res
*lockres
)
71 wait_event(lockres
->l_event
,
72 !user_check_wait_flag(lockres
, USER_LOCK_BLOCKED
));
75 /* I heart container_of... */
76 static inline struct dlm_ctxt
*
77 dlm_ctxt_from_user_lockres(struct user_lock_res
*lockres
)
79 struct dlmfs_inode_private
*ip
;
81 ip
= container_of(lockres
,
82 struct dlmfs_inode_private
,
88 user_dlm_inode_from_user_lockres(struct user_lock_res
*lockres
)
90 struct dlmfs_inode_private
*ip
;
92 ip
= container_of(lockres
,
93 struct dlmfs_inode_private
,
95 return &ip
->ip_vfs_inode
;
98 static inline void user_recover_from_dlm_error(struct user_lock_res
*lockres
)
100 spin_lock(&lockres
->l_lock
);
101 lockres
->l_flags
&= ~USER_LOCK_BUSY
;
102 spin_unlock(&lockres
->l_lock
);
105 #define user_log_dlm_error(_func, _stat, _lockres) do { \
106 mlog(ML_ERROR, "Dlm error \"%s\" while calling %s on " \
107 "resource %.*s: %s\n", dlm_errname(_stat), _func, \
108 _lockres->l_namelen, _lockres->l_name, dlm_errmsg(_stat)); \
111 /* WARNING: This function lives in a world where the only three lock
112 * levels are EX, PR, and NL. It *will* have to be adjusted when more
113 * lock types are added. */
114 static inline int user_highest_compat_lock_level(int level
)
116 int new_level
= LKM_EXMODE
;
118 if (level
== LKM_EXMODE
)
119 new_level
= LKM_NLMODE
;
120 else if (level
== LKM_PRMODE
)
121 new_level
= LKM_PRMODE
;
125 static void user_ast(void *opaque
)
127 struct user_lock_res
*lockres
= opaque
;
128 struct dlm_lockstatus
*lksb
;
130 mlog(0, "AST fired for lockres %.*s\n", lockres
->l_namelen
,
133 spin_lock(&lockres
->l_lock
);
135 lksb
= &(lockres
->l_lksb
);
136 if (lksb
->status
!= DLM_NORMAL
) {
137 mlog(ML_ERROR
, "lksb status value of %u on lockres %.*s\n",
138 lksb
->status
, lockres
->l_namelen
, lockres
->l_name
);
139 spin_unlock(&lockres
->l_lock
);
143 mlog_bug_on_msg(lockres
->l_requested
== LKM_IVMODE
,
144 "Lockres %.*s, requested ivmode. flags 0x%x\n",
145 lockres
->l_namelen
, lockres
->l_name
, lockres
->l_flags
);
147 /* we're downconverting. */
148 if (lockres
->l_requested
< lockres
->l_level
) {
149 if (lockres
->l_requested
<=
150 user_highest_compat_lock_level(lockres
->l_blocking
)) {
151 lockres
->l_blocking
= LKM_NLMODE
;
152 lockres
->l_flags
&= ~USER_LOCK_BLOCKED
;
156 lockres
->l_level
= lockres
->l_requested
;
157 lockres
->l_requested
= LKM_IVMODE
;
158 lockres
->l_flags
|= USER_LOCK_ATTACHED
;
159 lockres
->l_flags
&= ~USER_LOCK_BUSY
;
161 spin_unlock(&lockres
->l_lock
);
163 wake_up(&lockres
->l_event
);
166 static inline void user_dlm_grab_inode_ref(struct user_lock_res
*lockres
)
169 inode
= user_dlm_inode_from_user_lockres(lockres
);
174 static void user_dlm_unblock_lock(struct work_struct
*work
);
176 static void __user_dlm_queue_lockres(struct user_lock_res
*lockres
)
178 if (!(lockres
->l_flags
& USER_LOCK_QUEUED
)) {
179 user_dlm_grab_inode_ref(lockres
);
181 INIT_WORK(&lockres
->l_work
, user_dlm_unblock_lock
);
183 queue_work(user_dlm_worker
, &lockres
->l_work
);
184 lockres
->l_flags
|= USER_LOCK_QUEUED
;
188 static void __user_dlm_cond_queue_lockres(struct user_lock_res
*lockres
)
192 if (!(lockres
->l_flags
& USER_LOCK_BLOCKED
))
195 switch (lockres
->l_blocking
) {
197 if (!lockres
->l_ex_holders
&& !lockres
->l_ro_holders
)
201 if (!lockres
->l_ex_holders
)
209 __user_dlm_queue_lockres(lockres
);
212 static void user_bast(void *opaque
, int level
)
214 struct user_lock_res
*lockres
= opaque
;
216 mlog(0, "Blocking AST fired for lockres %.*s. Blocking level %d\n",
217 lockres
->l_namelen
, lockres
->l_name
, level
);
219 spin_lock(&lockres
->l_lock
);
220 lockres
->l_flags
|= USER_LOCK_BLOCKED
;
221 if (level
> lockres
->l_blocking
)
222 lockres
->l_blocking
= level
;
224 __user_dlm_queue_lockres(lockres
);
225 spin_unlock(&lockres
->l_lock
);
227 wake_up(&lockres
->l_event
);
230 static void user_unlock_ast(void *opaque
, enum dlm_status status
)
232 struct user_lock_res
*lockres
= opaque
;
234 mlog(0, "UNLOCK AST called on lock %.*s\n", lockres
->l_namelen
,
237 if (status
!= DLM_NORMAL
&& status
!= DLM_CANCELGRANT
)
238 mlog(ML_ERROR
, "Dlm returns status %d\n", status
);
240 spin_lock(&lockres
->l_lock
);
241 /* The teardown flag gets set early during the unlock process,
242 * so test the cancel flag to make sure that this ast isn't
243 * for a concurrent cancel. */
244 if (lockres
->l_flags
& USER_LOCK_IN_TEARDOWN
245 && !(lockres
->l_flags
& USER_LOCK_IN_CANCEL
)) {
246 lockres
->l_level
= LKM_IVMODE
;
247 } else if (status
== DLM_CANCELGRANT
) {
248 /* We tried to cancel a convert request, but it was
249 * already granted. Don't clear the busy flag - the
250 * ast should've done this already. */
251 BUG_ON(!(lockres
->l_flags
& USER_LOCK_IN_CANCEL
));
252 lockres
->l_flags
&= ~USER_LOCK_IN_CANCEL
;
255 BUG_ON(!(lockres
->l_flags
& USER_LOCK_IN_CANCEL
));
256 /* Cancel succeeded, we want to re-queue */
257 lockres
->l_requested
= LKM_IVMODE
; /* cancel an
260 lockres
->l_flags
&= ~USER_LOCK_IN_CANCEL
;
261 /* we want the unblock thread to look at it again
263 if (lockres
->l_flags
& USER_LOCK_BLOCKED
)
264 __user_dlm_queue_lockres(lockres
);
267 lockres
->l_flags
&= ~USER_LOCK_BUSY
;
269 spin_unlock(&lockres
->l_lock
);
271 wake_up(&lockres
->l_event
);
274 static inline void user_dlm_drop_inode_ref(struct user_lock_res
*lockres
)
277 inode
= user_dlm_inode_from_user_lockres(lockres
);
281 static void user_dlm_unblock_lock(struct work_struct
*work
)
283 int new_level
, status
;
284 struct user_lock_res
*lockres
=
285 container_of(work
, struct user_lock_res
, l_work
);
286 struct dlm_ctxt
*dlm
= dlm_ctxt_from_user_lockres(lockres
);
288 mlog(0, "processing lockres %.*s\n", lockres
->l_namelen
,
291 spin_lock(&lockres
->l_lock
);
293 mlog_bug_on_msg(!(lockres
->l_flags
& USER_LOCK_QUEUED
),
294 "Lockres %.*s, flags 0x%x\n",
295 lockres
->l_namelen
, lockres
->l_name
, lockres
->l_flags
);
297 /* notice that we don't clear USER_LOCK_BLOCKED here. If it's
298 * set, we want user_ast clear it. */
299 lockres
->l_flags
&= ~USER_LOCK_QUEUED
;
301 /* It's valid to get here and no longer be blocked - if we get
302 * several basts in a row, we might be queued by the first
303 * one, the unblock thread might run and clear the queued
304 * flag, and finally we might get another bast which re-queues
305 * us before our ast for the downconvert is called. */
306 if (!(lockres
->l_flags
& USER_LOCK_BLOCKED
)) {
307 spin_unlock(&lockres
->l_lock
);
311 if (lockres
->l_flags
& USER_LOCK_IN_TEARDOWN
) {
312 spin_unlock(&lockres
->l_lock
);
316 if (lockres
->l_flags
& USER_LOCK_BUSY
) {
317 if (lockres
->l_flags
& USER_LOCK_IN_CANCEL
) {
318 spin_unlock(&lockres
->l_lock
);
322 lockres
->l_flags
|= USER_LOCK_IN_CANCEL
;
323 spin_unlock(&lockres
->l_lock
);
325 status
= dlmunlock(dlm
,
330 if (status
!= DLM_NORMAL
)
331 user_log_dlm_error("dlmunlock", status
, lockres
);
335 /* If there are still incompat holders, we can exit safely
336 * without worrying about re-queueing this lock as that will
337 * happen on the last call to user_cluster_unlock. */
338 if ((lockres
->l_blocking
== LKM_EXMODE
)
339 && (lockres
->l_ex_holders
|| lockres
->l_ro_holders
)) {
340 spin_unlock(&lockres
->l_lock
);
341 mlog(0, "can't downconvert for ex: ro = %u, ex = %u\n",
342 lockres
->l_ro_holders
, lockres
->l_ex_holders
);
346 if ((lockres
->l_blocking
== LKM_PRMODE
)
347 && lockres
->l_ex_holders
) {
348 spin_unlock(&lockres
->l_lock
);
349 mlog(0, "can't downconvert for pr: ex = %u\n",
350 lockres
->l_ex_holders
);
354 /* yay, we can downconvert now. */
355 new_level
= user_highest_compat_lock_level(lockres
->l_blocking
);
356 lockres
->l_requested
= new_level
;
357 lockres
->l_flags
|= USER_LOCK_BUSY
;
358 mlog(0, "Downconvert lock from %d to %d\n",
359 lockres
->l_level
, new_level
);
360 spin_unlock(&lockres
->l_lock
);
362 /* need lock downconvert request now... */
363 status
= dlmlock(dlm
,
366 LKM_CONVERT
|LKM_VALBLK
,
372 if (status
!= DLM_NORMAL
) {
373 user_log_dlm_error("dlmlock", status
, lockres
);
374 user_recover_from_dlm_error(lockres
);
378 user_dlm_drop_inode_ref(lockres
);
381 static inline void user_dlm_inc_holders(struct user_lock_res
*lockres
,
386 lockres
->l_ex_holders
++;
389 lockres
->l_ro_holders
++;
396 /* predict what lock level we'll be dropping down to on behalf
397 * of another node, and return true if the currently wanted
398 * level will be compatible with it. */
400 user_may_continue_on_blocked_lock(struct user_lock_res
*lockres
,
403 BUG_ON(!(lockres
->l_flags
& USER_LOCK_BLOCKED
));
405 return wanted
<= user_highest_compat_lock_level(lockres
->l_blocking
);
408 int user_dlm_cluster_lock(struct user_lock_res
*lockres
,
412 int status
, local_flags
;
413 struct dlm_ctxt
*dlm
= dlm_ctxt_from_user_lockres(lockres
);
415 if (level
!= LKM_EXMODE
&&
416 level
!= LKM_PRMODE
) {
417 mlog(ML_ERROR
, "lockres %.*s: invalid request!\n",
418 lockres
->l_namelen
, lockres
->l_name
);
423 mlog(0, "lockres %.*s: asking for %s lock, passed flags = 0x%x\n",
424 lockres
->l_namelen
, lockres
->l_name
,
425 (level
== LKM_EXMODE
) ? "LKM_EXMODE" : "LKM_PRMODE",
429 if (signal_pending(current
)) {
430 status
= -ERESTARTSYS
;
434 spin_lock(&lockres
->l_lock
);
436 /* We only compare against the currently granted level
437 * here. If the lock is blocked waiting on a downconvert,
438 * we'll get caught below. */
439 if ((lockres
->l_flags
& USER_LOCK_BUSY
) &&
440 (level
> lockres
->l_level
)) {
441 /* is someone sitting in dlm_lock? If so, wait on
443 spin_unlock(&lockres
->l_lock
);
445 user_wait_on_busy_lock(lockres
);
449 if ((lockres
->l_flags
& USER_LOCK_BLOCKED
) &&
450 (!user_may_continue_on_blocked_lock(lockres
, level
))) {
451 /* is the lock is currently blocked on behalf of
453 spin_unlock(&lockres
->l_lock
);
455 user_wait_on_blocked_lock(lockres
);
459 if (level
> lockres
->l_level
) {
460 local_flags
= lkm_flags
| LKM_VALBLK
;
461 if (lockres
->l_level
!= LKM_IVMODE
)
462 local_flags
|= LKM_CONVERT
;
464 lockres
->l_requested
= level
;
465 lockres
->l_flags
|= USER_LOCK_BUSY
;
466 spin_unlock(&lockres
->l_lock
);
468 BUG_ON(level
== LKM_IVMODE
);
469 BUG_ON(level
== LKM_NLMODE
);
471 /* call dlm_lock to upgrade lock now */
472 status
= dlmlock(dlm
,
481 if (status
!= DLM_NORMAL
) {
482 if ((lkm_flags
& LKM_NOQUEUE
) &&
483 (status
== DLM_NOTQUEUED
))
486 user_log_dlm_error("dlmlock", status
, lockres
);
489 user_recover_from_dlm_error(lockres
);
493 user_wait_on_busy_lock(lockres
);
497 user_dlm_inc_holders(lockres
, level
);
498 spin_unlock(&lockres
->l_lock
);
505 static inline void user_dlm_dec_holders(struct user_lock_res
*lockres
,
510 BUG_ON(!lockres
->l_ex_holders
);
511 lockres
->l_ex_holders
--;
514 BUG_ON(!lockres
->l_ro_holders
);
515 lockres
->l_ro_holders
--;
522 void user_dlm_cluster_unlock(struct user_lock_res
*lockres
,
525 if (level
!= LKM_EXMODE
&&
526 level
!= LKM_PRMODE
) {
527 mlog(ML_ERROR
, "lockres %.*s: invalid request!\n",
528 lockres
->l_namelen
, lockres
->l_name
);
532 spin_lock(&lockres
->l_lock
);
533 user_dlm_dec_holders(lockres
, level
);
534 __user_dlm_cond_queue_lockres(lockres
);
535 spin_unlock(&lockres
->l_lock
);
538 void user_dlm_write_lvb(struct inode
*inode
,
542 struct user_lock_res
*lockres
= &DLMFS_I(inode
)->ip_lockres
;
543 char *lvb
= lockres
->l_lksb
.lvb
;
545 BUG_ON(len
> DLM_LVB_LEN
);
547 spin_lock(&lockres
->l_lock
);
549 BUG_ON(lockres
->l_level
< LKM_EXMODE
);
550 memcpy(lvb
, val
, len
);
552 spin_unlock(&lockres
->l_lock
);
555 void user_dlm_read_lvb(struct inode
*inode
,
559 struct user_lock_res
*lockres
= &DLMFS_I(inode
)->ip_lockres
;
560 char *lvb
= lockres
->l_lksb
.lvb
;
562 BUG_ON(len
> DLM_LVB_LEN
);
564 spin_lock(&lockres
->l_lock
);
566 BUG_ON(lockres
->l_level
< LKM_PRMODE
);
567 memcpy(val
, lvb
, len
);
569 spin_unlock(&lockres
->l_lock
);
572 void user_dlm_lock_res_init(struct user_lock_res
*lockres
,
573 struct dentry
*dentry
)
575 memset(lockres
, 0, sizeof(*lockres
));
577 spin_lock_init(&lockres
->l_lock
);
578 init_waitqueue_head(&lockres
->l_event
);
579 lockres
->l_level
= LKM_IVMODE
;
580 lockres
->l_requested
= LKM_IVMODE
;
581 lockres
->l_blocking
= LKM_IVMODE
;
583 /* should have been checked before getting here. */
584 BUG_ON(dentry
->d_name
.len
>= USER_DLM_LOCK_ID_MAX_LEN
);
586 memcpy(lockres
->l_name
,
589 lockres
->l_namelen
= dentry
->d_name
.len
;
592 int user_dlm_destroy_lock(struct user_lock_res
*lockres
)
595 struct dlm_ctxt
*dlm
= dlm_ctxt_from_user_lockres(lockres
);
597 mlog(0, "asked to destroy %.*s\n", lockres
->l_namelen
, lockres
->l_name
);
599 spin_lock(&lockres
->l_lock
);
600 if (lockres
->l_flags
& USER_LOCK_IN_TEARDOWN
) {
601 spin_unlock(&lockres
->l_lock
);
605 lockres
->l_flags
|= USER_LOCK_IN_TEARDOWN
;
607 while (lockres
->l_flags
& USER_LOCK_BUSY
) {
608 spin_unlock(&lockres
->l_lock
);
610 user_wait_on_busy_lock(lockres
);
612 spin_lock(&lockres
->l_lock
);
615 if (lockres
->l_ro_holders
|| lockres
->l_ex_holders
) {
616 spin_unlock(&lockres
->l_lock
);
621 if (!(lockres
->l_flags
& USER_LOCK_ATTACHED
)) {
622 spin_unlock(&lockres
->l_lock
);
626 lockres
->l_flags
&= ~USER_LOCK_ATTACHED
;
627 lockres
->l_flags
|= USER_LOCK_BUSY
;
628 spin_unlock(&lockres
->l_lock
);
630 status
= dlmunlock(dlm
,
635 if (status
!= DLM_NORMAL
) {
636 user_log_dlm_error("dlmunlock", status
, lockres
);
641 user_wait_on_busy_lock(lockres
);
648 struct dlm_ctxt
*user_dlm_register_context(struct qstr
*name
,
649 struct dlm_protocol_version
*proto
)
651 struct dlm_ctxt
*dlm
;
655 domain
= kmalloc(name
->len
+ 1, GFP_NOFS
);
658 return ERR_PTR(-ENOMEM
);
661 dlm_key
= crc32_le(0, name
->name
, name
->len
);
663 snprintf(domain
, name
->len
+ 1, "%.*s", name
->len
, name
->name
);
665 dlm
= dlm_register_domain(domain
, dlm_key
, proto
);
667 mlog_errno(PTR_ERR(dlm
));
673 void user_dlm_unregister_context(struct dlm_ctxt
*dlm
)
675 dlm_unregister_domain(dlm
);