[GFS2] Remove all_list from lock_dlm
[linux-2.6/verdex.git] / fs / gfs2 / locking / dlm / lock.c
blob2482c9047505fcc0b2439b0693142f3646793f1a
1 /*
2 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This copyrighted material is made available to anyone wishing to use,
6 * modify, copy, or redistribute it subject to the terms and conditions
7 * of the GNU General Public License version 2.
8 */
10 #include "lock_dlm.h"
12 static char junk_lvb[GDLM_LVB_SIZE];
15 /* convert dlm lock-mode to gfs lock-state */
17 static s16 gdlm_make_lmstate(s16 dlmmode)
19 switch (dlmmode) {
20 case DLM_LOCK_IV:
21 case DLM_LOCK_NL:
22 return LM_ST_UNLOCKED;
23 case DLM_LOCK_EX:
24 return LM_ST_EXCLUSIVE;
25 case DLM_LOCK_CW:
26 return LM_ST_DEFERRED;
27 case DLM_LOCK_PR:
28 return LM_ST_SHARED;
30 gdlm_assert(0, "unknown DLM mode %d", dlmmode);
31 return -1;
34 /* A lock placed on this queue is re-submitted to DLM as soon as the lock_dlm
35 thread gets to it. */
37 static void queue_submit(struct gdlm_lock *lp)
39 struct gdlm_ls *ls = lp->ls;
41 spin_lock(&ls->async_lock);
42 list_add_tail(&lp->delay_list, &ls->submit);
43 spin_unlock(&ls->async_lock);
44 wake_up(&ls->thread_wait);
47 static void wake_up_ast(struct gdlm_lock *lp)
49 clear_bit(LFL_AST_WAIT, &lp->flags);
50 smp_mb__after_clear_bit();
51 wake_up_bit(&lp->flags, LFL_AST_WAIT);
54 static void gdlm_delete_lp(struct gdlm_lock *lp)
56 struct gdlm_ls *ls = lp->ls;
58 spin_lock(&ls->async_lock);
59 if (!list_empty(&lp->delay_list))
60 list_del_init(&lp->delay_list);
61 ls->all_locks_count--;
62 spin_unlock(&ls->async_lock);
64 kfree(lp);
67 static void gdlm_queue_delayed(struct gdlm_lock *lp)
69 struct gdlm_ls *ls = lp->ls;
71 spin_lock(&ls->async_lock);
72 list_add_tail(&lp->delay_list, &ls->delayed);
73 spin_unlock(&ls->async_lock);
76 static void process_complete(struct gdlm_lock *lp)
78 struct gdlm_ls *ls = lp->ls;
79 struct lm_async_cb acb;
81 memset(&acb, 0, sizeof(acb));
83 if (lp->lksb.sb_status == -DLM_ECANCEL) {
84 log_info("complete dlm cancel %x,%llx flags %lx",
85 lp->lockname.ln_type,
86 (unsigned long long)lp->lockname.ln_number,
87 lp->flags);
89 lp->req = lp->cur;
90 acb.lc_ret |= LM_OUT_CANCELED;
91 if (lp->cur == DLM_LOCK_IV)
92 lp->lksb.sb_lkid = 0;
93 goto out;
96 if (test_and_clear_bit(LFL_DLM_UNLOCK, &lp->flags)) {
97 if (lp->lksb.sb_status != -DLM_EUNLOCK) {
98 log_info("unlock sb_status %d %x,%llx flags %lx",
99 lp->lksb.sb_status, lp->lockname.ln_type,
100 (unsigned long long)lp->lockname.ln_number,
101 lp->flags);
102 return;
105 lp->cur = DLM_LOCK_IV;
106 lp->req = DLM_LOCK_IV;
107 lp->lksb.sb_lkid = 0;
109 if (test_and_clear_bit(LFL_UNLOCK_DELETE, &lp->flags)) {
110 gdlm_delete_lp(lp);
111 return;
113 goto out;
116 if (lp->lksb.sb_flags & DLM_SBF_VALNOTVALID)
117 memset(lp->lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
119 if (lp->lksb.sb_flags & DLM_SBF_ALTMODE) {
120 if (lp->req == DLM_LOCK_PR)
121 lp->req = DLM_LOCK_CW;
122 else if (lp->req == DLM_LOCK_CW)
123 lp->req = DLM_LOCK_PR;
127 * A canceled lock request. The lock was just taken off the delayed
128 * list and was never even submitted to dlm.
131 if (test_and_clear_bit(LFL_CANCEL, &lp->flags)) {
132 log_info("complete internal cancel %x,%llx",
133 lp->lockname.ln_type,
134 (unsigned long long)lp->lockname.ln_number);
135 lp->req = lp->cur;
136 acb.lc_ret |= LM_OUT_CANCELED;
137 goto out;
141 * An error occured.
144 if (lp->lksb.sb_status) {
145 /* a "normal" error */
146 if ((lp->lksb.sb_status == -EAGAIN) &&
147 (lp->lkf & DLM_LKF_NOQUEUE)) {
148 lp->req = lp->cur;
149 if (lp->cur == DLM_LOCK_IV)
150 lp->lksb.sb_lkid = 0;
151 goto out;
154 /* this could only happen with cancels I think */
155 log_info("ast sb_status %d %x,%llx flags %lx",
156 lp->lksb.sb_status, lp->lockname.ln_type,
157 (unsigned long long)lp->lockname.ln_number,
158 lp->flags);
159 return;
163 * This is an AST for an EX->EX conversion for sync_lvb from GFS.
166 if (test_and_clear_bit(LFL_SYNC_LVB, &lp->flags)) {
167 wake_up_ast(lp);
168 return;
172 * A lock has been demoted to NL because it initially completed during
173 * BLOCK_LOCKS. Now it must be requested in the originally requested
174 * mode.
177 if (test_and_clear_bit(LFL_REREQUEST, &lp->flags)) {
178 gdlm_assert(lp->req == DLM_LOCK_NL, "%x,%llx",
179 lp->lockname.ln_type,
180 (unsigned long long)lp->lockname.ln_number);
181 gdlm_assert(lp->prev_req > DLM_LOCK_NL, "%x,%llx",
182 lp->lockname.ln_type,
183 (unsigned long long)lp->lockname.ln_number);
185 lp->cur = DLM_LOCK_NL;
186 lp->req = lp->prev_req;
187 lp->prev_req = DLM_LOCK_IV;
188 lp->lkf &= ~DLM_LKF_CONVDEADLK;
190 set_bit(LFL_NOCACHE, &lp->flags);
192 if (test_bit(DFL_BLOCK_LOCKS, &ls->flags) &&
193 !test_bit(LFL_NOBLOCK, &lp->flags))
194 gdlm_queue_delayed(lp);
195 else
196 queue_submit(lp);
197 return;
201 * A request is granted during dlm recovery. It may be granted
202 * because the locks of a failed node were cleared. In that case,
203 * there may be inconsistent data beneath this lock and we must wait
204 * for recovery to complete to use it. When gfs recovery is done this
205 * granted lock will be converted to NL and then reacquired in this
206 * granted state.
209 if (test_bit(DFL_BLOCK_LOCKS, &ls->flags) &&
210 !test_bit(LFL_NOBLOCK, &lp->flags) &&
211 lp->req != DLM_LOCK_NL) {
213 lp->cur = lp->req;
214 lp->prev_req = lp->req;
215 lp->req = DLM_LOCK_NL;
216 lp->lkf |= DLM_LKF_CONVERT;
217 lp->lkf &= ~DLM_LKF_CONVDEADLK;
219 log_debug("rereq %x,%llx id %x %d,%d",
220 lp->lockname.ln_type,
221 (unsigned long long)lp->lockname.ln_number,
222 lp->lksb.sb_lkid, lp->cur, lp->req);
224 set_bit(LFL_REREQUEST, &lp->flags);
225 queue_submit(lp);
226 return;
230 * DLM demoted the lock to NL before it was granted so GFS must be
231 * told it cannot cache data for this lock.
234 if (lp->lksb.sb_flags & DLM_SBF_DEMOTED)
235 set_bit(LFL_NOCACHE, &lp->flags);
237 out:
239 * This is an internal lock_dlm lock
242 if (test_bit(LFL_INLOCK, &lp->flags)) {
243 clear_bit(LFL_NOBLOCK, &lp->flags);
244 lp->cur = lp->req;
245 wake_up_ast(lp);
246 return;
250 * Normal completion of a lock request. Tell GFS it now has the lock.
253 clear_bit(LFL_NOBLOCK, &lp->flags);
254 lp->cur = lp->req;
256 acb.lc_name = lp->lockname;
257 acb.lc_ret |= gdlm_make_lmstate(lp->cur);
259 ls->fscb(ls->sdp, LM_CB_ASYNC, &acb);
262 static void gdlm_ast(void *astarg)
264 struct gdlm_lock *lp = astarg;
265 clear_bit(LFL_ACTIVE, &lp->flags);
266 process_complete(lp);
269 static void process_blocking(struct gdlm_lock *lp, int bast_mode)
271 struct gdlm_ls *ls = lp->ls;
272 unsigned int cb = 0;
274 switch (gdlm_make_lmstate(bast_mode)) {
275 case LM_ST_EXCLUSIVE:
276 cb = LM_CB_NEED_E;
277 break;
278 case LM_ST_DEFERRED:
279 cb = LM_CB_NEED_D;
280 break;
281 case LM_ST_SHARED:
282 cb = LM_CB_NEED_S;
283 break;
284 default:
285 gdlm_assert(0, "unknown bast mode %u", bast_mode);
288 ls->fscb(ls->sdp, cb, &lp->lockname);
292 static void gdlm_bast(void *astarg, int mode)
294 struct gdlm_lock *lp = astarg;
296 if (!mode) {
297 printk(KERN_INFO "lock_dlm: bast mode zero %x,%llx\n",
298 lp->lockname.ln_type,
299 (unsigned long long)lp->lockname.ln_number);
300 return;
303 process_blocking(lp, mode);
306 /* convert gfs lock-state to dlm lock-mode */
308 static s16 make_mode(s16 lmstate)
310 switch (lmstate) {
311 case LM_ST_UNLOCKED:
312 return DLM_LOCK_NL;
313 case LM_ST_EXCLUSIVE:
314 return DLM_LOCK_EX;
315 case LM_ST_DEFERRED:
316 return DLM_LOCK_CW;
317 case LM_ST_SHARED:
318 return DLM_LOCK_PR;
320 gdlm_assert(0, "unknown LM state %d", lmstate);
321 return -1;
325 /* verify agreement with GFS on the current lock state, NB: DLM_LOCK_NL and
326 DLM_LOCK_IV are both considered LM_ST_UNLOCKED by GFS. */
328 static void check_cur_state(struct gdlm_lock *lp, unsigned int cur_state)
330 s16 cur = make_mode(cur_state);
331 if (lp->cur != DLM_LOCK_IV)
332 gdlm_assert(lp->cur == cur, "%d, %d", lp->cur, cur);
335 static inline unsigned int make_flags(struct gdlm_lock *lp,
336 unsigned int gfs_flags,
337 s16 cur, s16 req)
339 unsigned int lkf = 0;
341 if (gfs_flags & LM_FLAG_TRY)
342 lkf |= DLM_LKF_NOQUEUE;
344 if (gfs_flags & LM_FLAG_TRY_1CB) {
345 lkf |= DLM_LKF_NOQUEUE;
346 lkf |= DLM_LKF_NOQUEUEBAST;
349 if (gfs_flags & LM_FLAG_PRIORITY) {
350 lkf |= DLM_LKF_NOORDER;
351 lkf |= DLM_LKF_HEADQUE;
354 if (gfs_flags & LM_FLAG_ANY) {
355 if (req == DLM_LOCK_PR)
356 lkf |= DLM_LKF_ALTCW;
357 else if (req == DLM_LOCK_CW)
358 lkf |= DLM_LKF_ALTPR;
361 if (lp->lksb.sb_lkid != 0) {
362 lkf |= DLM_LKF_CONVERT;
365 if (lp->lvb)
366 lkf |= DLM_LKF_VALBLK;
368 return lkf;
371 /* make_strname - convert GFS lock numbers to a string */
373 static inline void make_strname(const struct lm_lockname *lockname,
374 struct gdlm_strname *str)
376 sprintf(str->name, "%8x%16llx", lockname->ln_type,
377 (unsigned long long)lockname->ln_number);
378 str->namelen = GDLM_STRNAME_BYTES;
381 static int gdlm_create_lp(struct gdlm_ls *ls, struct lm_lockname *name,
382 struct gdlm_lock **lpp)
384 struct gdlm_lock *lp;
386 lp = kzalloc(sizeof(struct gdlm_lock), GFP_NOFS);
387 if (!lp)
388 return -ENOMEM;
390 lp->lockname = *name;
391 make_strname(name, &lp->strname);
392 lp->ls = ls;
393 lp->cur = DLM_LOCK_IV;
394 INIT_LIST_HEAD(&lp->delay_list);
396 spin_lock(&ls->async_lock);
397 ls->all_locks_count++;
398 spin_unlock(&ls->async_lock);
400 *lpp = lp;
401 return 0;
404 int gdlm_get_lock(void *lockspace, struct lm_lockname *name,
405 void **lockp)
407 struct gdlm_lock *lp;
408 int error;
410 error = gdlm_create_lp(lockspace, name, &lp);
412 *lockp = lp;
413 return error;
416 void gdlm_put_lock(void *lock)
418 gdlm_delete_lp(lock);
421 unsigned int gdlm_do_lock(struct gdlm_lock *lp)
423 struct gdlm_ls *ls = lp->ls;
424 int error, bast = 1;
427 * When recovery is in progress, delay lock requests for submission
428 * once recovery is done. Requests for recovery (NOEXP) and unlocks
429 * can pass.
432 if (test_bit(DFL_BLOCK_LOCKS, &ls->flags) &&
433 !test_bit(LFL_NOBLOCK, &lp->flags) && lp->req != DLM_LOCK_NL) {
434 gdlm_queue_delayed(lp);
435 return LM_OUT_ASYNC;
439 * Submit the actual lock request.
442 if (test_bit(LFL_NOBAST, &lp->flags))
443 bast = 0;
445 set_bit(LFL_ACTIVE, &lp->flags);
447 log_debug("lk %x,%llx id %x %d,%d %x", lp->lockname.ln_type,
448 (unsigned long long)lp->lockname.ln_number, lp->lksb.sb_lkid,
449 lp->cur, lp->req, lp->lkf);
451 error = dlm_lock(ls->dlm_lockspace, lp->req, &lp->lksb, lp->lkf,
452 lp->strname.name, lp->strname.namelen, 0, gdlm_ast,
453 lp, bast ? gdlm_bast : NULL);
455 if ((error == -EAGAIN) && (lp->lkf & DLM_LKF_NOQUEUE)) {
456 lp->lksb.sb_status = -EAGAIN;
457 gdlm_ast(lp);
458 error = 0;
461 if (error) {
462 log_error("%s: gdlm_lock %x,%llx err=%d cur=%d req=%d lkf=%x "
463 "flags=%lx", ls->fsname, lp->lockname.ln_type,
464 (unsigned long long)lp->lockname.ln_number, error,
465 lp->cur, lp->req, lp->lkf, lp->flags);
466 return LM_OUT_ERROR;
468 return LM_OUT_ASYNC;
471 static unsigned int gdlm_do_unlock(struct gdlm_lock *lp)
473 struct gdlm_ls *ls = lp->ls;
474 unsigned int lkf = 0;
475 int error;
477 set_bit(LFL_DLM_UNLOCK, &lp->flags);
478 set_bit(LFL_ACTIVE, &lp->flags);
480 if (lp->lvb)
481 lkf = DLM_LKF_VALBLK;
483 log_debug("un %x,%llx %x %d %x", lp->lockname.ln_type,
484 (unsigned long long)lp->lockname.ln_number,
485 lp->lksb.sb_lkid, lp->cur, lkf);
487 error = dlm_unlock(ls->dlm_lockspace, lp->lksb.sb_lkid, lkf, NULL, lp);
489 if (error) {
490 log_error("%s: gdlm_unlock %x,%llx err=%d cur=%d req=%d lkf=%x "
491 "flags=%lx", ls->fsname, lp->lockname.ln_type,
492 (unsigned long long)lp->lockname.ln_number, error,
493 lp->cur, lp->req, lp->lkf, lp->flags);
494 return LM_OUT_ERROR;
496 return LM_OUT_ASYNC;
499 unsigned int gdlm_lock(void *lock, unsigned int cur_state,
500 unsigned int req_state, unsigned int flags)
502 struct gdlm_lock *lp = lock;
504 if (req_state == LM_ST_UNLOCKED)
505 return gdlm_unlock(lock, cur_state);
507 if (req_state == LM_ST_UNLOCKED)
508 return gdlm_unlock(lock, cur_state);
510 clear_bit(LFL_DLM_CANCEL, &lp->flags);
511 if (flags & LM_FLAG_NOEXP)
512 set_bit(LFL_NOBLOCK, &lp->flags);
514 check_cur_state(lp, cur_state);
515 lp->req = make_mode(req_state);
516 lp->lkf = make_flags(lp, flags, lp->cur, lp->req);
518 return gdlm_do_lock(lp);
521 unsigned int gdlm_unlock(void *lock, unsigned int cur_state)
523 struct gdlm_lock *lp = lock;
525 clear_bit(LFL_DLM_CANCEL, &lp->flags);
526 if (lp->cur == DLM_LOCK_IV)
527 return 0;
528 return gdlm_do_unlock(lp);
531 void gdlm_cancel(void *lock)
533 struct gdlm_lock *lp = lock;
534 struct gdlm_ls *ls = lp->ls;
535 int error, delay_list = 0;
537 if (test_bit(LFL_DLM_CANCEL, &lp->flags))
538 return;
540 log_info("gdlm_cancel %x,%llx flags %lx", lp->lockname.ln_type,
541 (unsigned long long)lp->lockname.ln_number, lp->flags);
543 spin_lock(&ls->async_lock);
544 if (!list_empty(&lp->delay_list)) {
545 list_del_init(&lp->delay_list);
546 delay_list = 1;
548 spin_unlock(&ls->async_lock);
550 if (delay_list) {
551 set_bit(LFL_CANCEL, &lp->flags);
552 set_bit(LFL_ACTIVE, &lp->flags);
553 gdlm_ast(lp);
554 return;
557 if (!test_bit(LFL_ACTIVE, &lp->flags) ||
558 test_bit(LFL_DLM_UNLOCK, &lp->flags)) {
559 log_info("gdlm_cancel skip %x,%llx flags %lx",
560 lp->lockname.ln_type,
561 (unsigned long long)lp->lockname.ln_number, lp->flags);
562 return;
565 /* the lock is blocked in the dlm */
567 set_bit(LFL_DLM_CANCEL, &lp->flags);
568 set_bit(LFL_ACTIVE, &lp->flags);
570 error = dlm_unlock(ls->dlm_lockspace, lp->lksb.sb_lkid, DLM_LKF_CANCEL,
571 NULL, lp);
573 log_info("gdlm_cancel rv %d %x,%llx flags %lx", error,
574 lp->lockname.ln_type,
575 (unsigned long long)lp->lockname.ln_number, lp->flags);
577 if (error == -EBUSY)
578 clear_bit(LFL_DLM_CANCEL, &lp->flags);
581 static int gdlm_add_lvb(struct gdlm_lock *lp)
583 char *lvb;
585 lvb = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
586 if (!lvb)
587 return -ENOMEM;
589 lp->lksb.sb_lvbptr = lvb;
590 lp->lvb = lvb;
591 return 0;
594 static void gdlm_del_lvb(struct gdlm_lock *lp)
596 kfree(lp->lvb);
597 lp->lvb = NULL;
598 lp->lksb.sb_lvbptr = NULL;
601 static int gdlm_ast_wait(void *word)
603 schedule();
604 return 0;
607 /* This can do a synchronous dlm request (requiring a lock_dlm thread to get
608 the completion) because gfs won't call hold_lvb() during a callback (from
609 the context of a lock_dlm thread). */
611 static int hold_null_lock(struct gdlm_lock *lp)
613 struct gdlm_lock *lpn = NULL;
614 int error;
616 if (lp->hold_null) {
617 printk(KERN_INFO "lock_dlm: lvb already held\n");
618 return 0;
621 error = gdlm_create_lp(lp->ls, &lp->lockname, &lpn);
622 if (error)
623 goto out;
625 lpn->lksb.sb_lvbptr = junk_lvb;
626 lpn->lvb = junk_lvb;
628 lpn->req = DLM_LOCK_NL;
629 lpn->lkf = DLM_LKF_VALBLK | DLM_LKF_EXPEDITE;
630 set_bit(LFL_NOBAST, &lpn->flags);
631 set_bit(LFL_INLOCK, &lpn->flags);
632 set_bit(LFL_AST_WAIT, &lpn->flags);
634 gdlm_do_lock(lpn);
635 wait_on_bit(&lpn->flags, LFL_AST_WAIT, gdlm_ast_wait, TASK_UNINTERRUPTIBLE);
636 error = lpn->lksb.sb_status;
637 if (error) {
638 printk(KERN_INFO "lock_dlm: hold_null_lock dlm error %d\n",
639 error);
640 gdlm_delete_lp(lpn);
641 lpn = NULL;
643 out:
644 lp->hold_null = lpn;
645 return error;
648 /* This cannot do a synchronous dlm request (requiring a lock_dlm thread to get
649 the completion) because gfs may call unhold_lvb() during a callback (from
650 the context of a lock_dlm thread) which could cause a deadlock since the
651 other lock_dlm thread could be engaged in recovery. */
653 static void unhold_null_lock(struct gdlm_lock *lp)
655 struct gdlm_lock *lpn = lp->hold_null;
657 gdlm_assert(lpn, "%x,%llx", lp->lockname.ln_type,
658 (unsigned long long)lp->lockname.ln_number);
659 lpn->lksb.sb_lvbptr = NULL;
660 lpn->lvb = NULL;
661 set_bit(LFL_UNLOCK_DELETE, &lpn->flags);
662 gdlm_do_unlock(lpn);
663 lp->hold_null = NULL;
666 /* Acquire a NL lock because gfs requires the value block to remain
667 intact on the resource while the lvb is "held" even if it's holding no locks
668 on the resource. */
670 int gdlm_hold_lvb(void *lock, char **lvbp)
672 struct gdlm_lock *lp = lock;
673 int error;
675 error = gdlm_add_lvb(lp);
676 if (error)
677 return error;
679 *lvbp = lp->lvb;
681 error = hold_null_lock(lp);
682 if (error)
683 gdlm_del_lvb(lp);
685 return error;
688 void gdlm_unhold_lvb(void *lock, char *lvb)
690 struct gdlm_lock *lp = lock;
692 unhold_null_lock(lp);
693 gdlm_del_lvb(lp);
696 void gdlm_submit_delayed(struct gdlm_ls *ls)
698 struct gdlm_lock *lp, *safe;
700 spin_lock(&ls->async_lock);
701 list_for_each_entry_safe(lp, safe, &ls->delayed, delay_list) {
702 list_del_init(&lp->delay_list);
703 list_add_tail(&lp->delay_list, &ls->submit);
705 spin_unlock(&ls->async_lock);
706 wake_up(&ls->thread_wait);