Pull bugzilla-5452 into release branch
[deliverable/linux.git] / fs / ocfs2 / dlm / userdlm.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * userdlm.c
5 *
6 * Code which implements the kernel side of a minimal userspace
7 * interface to our DLM.
8 *
9 * Many of the functions here are pared down versions of dlmglue.c
10 * functions.
11 *
12 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
13 *
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.
18 *
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.
23 *
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.
28 */
29
30 #include <linux/signal.h>
31
32 #include <linux/module.h>
33 #include <linux/fs.h>
34 #include <linux/types.h>
35 #include <linux/crc32.h>
36
37
38 #include "cluster/nodemanager.h"
39 #include "cluster/heartbeat.h"
40 #include "cluster/tcp.h"
41
42 #include "dlmapi.h"
43
44 #include "userdlm.h"
45
46 #define MLOG_MASK_PREFIX ML_DLMFS
47 #include "cluster/masklog.h"
48
49 static inline int user_check_wait_flag(struct user_lock_res *lockres,
50 int flag)
51 {
52 int ret;
53
54 spin_lock(&lockres->l_lock);
55 ret = lockres->l_flags & flag;
56 spin_unlock(&lockres->l_lock);
57
58 return ret;
59 }
60
61 static inline void user_wait_on_busy_lock(struct user_lock_res *lockres)
62
63 {
64 wait_event(lockres->l_event,
65 !user_check_wait_flag(lockres, USER_LOCK_BUSY));
66 }
67
68 static inline void user_wait_on_blocked_lock(struct user_lock_res *lockres)
69
70 {
71 wait_event(lockres->l_event,
72 !user_check_wait_flag(lockres, USER_LOCK_BLOCKED));
73 }
74
75 /* I heart container_of... */
76 static inline struct dlm_ctxt *
77 dlm_ctxt_from_user_lockres(struct user_lock_res *lockres)
78 {
79 struct dlmfs_inode_private *ip;
80
81 ip = container_of(lockres,
82 struct dlmfs_inode_private,
83 ip_lockres);
84 return ip->ip_dlm;
85 }
86
87 static struct inode *
88 user_dlm_inode_from_user_lockres(struct user_lock_res *lockres)
89 {
90 struct dlmfs_inode_private *ip;
91
92 ip = container_of(lockres,
93 struct dlmfs_inode_private,
94 ip_lockres);
95 return &ip->ip_vfs_inode;
96 }
97
98 static inline void user_recover_from_dlm_error(struct user_lock_res *lockres)
99 {
100 spin_lock(&lockres->l_lock);
101 lockres->l_flags &= ~USER_LOCK_BUSY;
102 spin_unlock(&lockres->l_lock);
103 }
104
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_name, dlm_errmsg(_stat)); \
109 } while (0)
110
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)
115 {
116 int new_level = LKM_EXMODE;
117
118 if (level == LKM_EXMODE)
119 new_level = LKM_NLMODE;
120 else if (level == LKM_PRMODE)
121 new_level = LKM_PRMODE;
122 return new_level;
123 }
124
125 static void user_ast(void *opaque)
126 {
127 struct user_lock_res *lockres = opaque;
128 struct dlm_lockstatus *lksb;
129
130 mlog(0, "AST fired for lockres %s\n", lockres->l_name);
131
132 spin_lock(&lockres->l_lock);
133
134 lksb = &(lockres->l_lksb);
135 if (lksb->status != DLM_NORMAL) {
136 mlog(ML_ERROR, "lksb status value of %u on lockres %s\n",
137 lksb->status, lockres->l_name);
138 spin_unlock(&lockres->l_lock);
139 return;
140 }
141
142 mlog_bug_on_msg(lockres->l_requested == LKM_IVMODE,
143 "Lockres %s, requested ivmode. flags 0x%x\n",
144 lockres->l_name, lockres->l_flags);
145
146 /* we're downconverting. */
147 if (lockres->l_requested < lockres->l_level) {
148 if (lockres->l_requested <=
149 user_highest_compat_lock_level(lockres->l_blocking)) {
150 lockres->l_blocking = LKM_NLMODE;
151 lockres->l_flags &= ~USER_LOCK_BLOCKED;
152 }
153 }
154
155 lockres->l_level = lockres->l_requested;
156 lockres->l_requested = LKM_IVMODE;
157 lockres->l_flags |= USER_LOCK_ATTACHED;
158 lockres->l_flags &= ~USER_LOCK_BUSY;
159
160 spin_unlock(&lockres->l_lock);
161
162 wake_up(&lockres->l_event);
163 }
164
165 static inline void user_dlm_grab_inode_ref(struct user_lock_res *lockres)
166 {
167 struct inode *inode;
168 inode = user_dlm_inode_from_user_lockres(lockres);
169 if (!igrab(inode))
170 BUG();
171 }
172
173 static void user_dlm_unblock_lock(void *opaque);
174
175 static void __user_dlm_queue_lockres(struct user_lock_res *lockres)
176 {
177 if (!(lockres->l_flags & USER_LOCK_QUEUED)) {
178 user_dlm_grab_inode_ref(lockres);
179
180 INIT_WORK(&lockres->l_work, user_dlm_unblock_lock,
181 lockres);
182
183 queue_work(user_dlm_worker, &lockres->l_work);
184 lockres->l_flags |= USER_LOCK_QUEUED;
185 }
186 }
187
188 static void __user_dlm_cond_queue_lockres(struct user_lock_res *lockres)
189 {
190 int queue = 0;
191
192 if (!(lockres->l_flags & USER_LOCK_BLOCKED))
193 return;
194
195 switch (lockres->l_blocking) {
196 case LKM_EXMODE:
197 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
198 queue = 1;
199 break;
200 case LKM_PRMODE:
201 if (!lockres->l_ex_holders)
202 queue = 1;
203 break;
204 default:
205 BUG();
206 }
207
208 if (queue)
209 __user_dlm_queue_lockres(lockres);
210 }
211
212 static void user_bast(void *opaque, int level)
213 {
214 struct user_lock_res *lockres = opaque;
215
216 mlog(0, "Blocking AST fired for lockres %s. Blocking level %d\n",
217 lockres->l_name, level);
218
219 spin_lock(&lockres->l_lock);
220 lockres->l_flags |= USER_LOCK_BLOCKED;
221 if (level > lockres->l_blocking)
222 lockres->l_blocking = level;
223
224 __user_dlm_queue_lockres(lockres);
225 spin_unlock(&lockres->l_lock);
226
227 wake_up(&lockres->l_event);
228 }
229
230 static void user_unlock_ast(void *opaque, enum dlm_status status)
231 {
232 struct user_lock_res *lockres = opaque;
233
234 mlog(0, "UNLOCK AST called on lock %s\n", lockres->l_name);
235
236 if (status != DLM_NORMAL && status != DLM_CANCELGRANT)
237 mlog(ML_ERROR, "Dlm returns status %d\n", status);
238
239 spin_lock(&lockres->l_lock);
240 /* The teardown flag gets set early during the unlock process,
241 * so test the cancel flag to make sure that this ast isn't
242 * for a concurrent cancel. */
243 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN
244 && !(lockres->l_flags & USER_LOCK_IN_CANCEL)) {
245 lockres->l_level = LKM_IVMODE;
246 } else if (status == DLM_CANCELGRANT) {
247 mlog(0, "Lock %s, cancel fails, flags 0x%x\n",
248 lockres->l_name, lockres->l_flags);
249 /* We tried to cancel a convert request, but it was
250 * already granted. Don't clear the busy flag - the
251 * ast should've done this already. */
252 BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL));
253 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
254 goto out_noclear;
255 } else {
256 BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL));
257 /* Cancel succeeded, we want to re-queue */
258 mlog(0, "Lock %s, cancel succeeds, flags 0x%x\n",
259 lockres->l_name, lockres->l_flags);
260 lockres->l_requested = LKM_IVMODE; /* cancel an
261 * upconvert
262 * request. */
263 lockres->l_flags &= ~USER_LOCK_IN_CANCEL;
264 /* we want the unblock thread to look at it again
265 * now. */
266 if (lockres->l_flags & USER_LOCK_BLOCKED)
267 __user_dlm_queue_lockres(lockres);
268 }
269
270 lockres->l_flags &= ~USER_LOCK_BUSY;
271 out_noclear:
272 spin_unlock(&lockres->l_lock);
273
274 wake_up(&lockres->l_event);
275 }
276
277 static inline void user_dlm_drop_inode_ref(struct user_lock_res *lockres)
278 {
279 struct inode *inode;
280 inode = user_dlm_inode_from_user_lockres(lockres);
281 iput(inode);
282 }
283
284 static void user_dlm_unblock_lock(void *opaque)
285 {
286 int new_level, status;
287 struct user_lock_res *lockres = (struct user_lock_res *) opaque;
288 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
289
290 mlog(0, "processing lockres %s\n", lockres->l_name);
291
292 spin_lock(&lockres->l_lock);
293
294 mlog_bug_on_msg(!(lockres->l_flags & USER_LOCK_QUEUED),
295 "Lockres %s, flags 0x%x\n",
296 lockres->l_name, lockres->l_flags);
297
298 /* notice that we don't clear USER_LOCK_BLOCKED here. If it's
299 * set, we want user_ast clear it. */
300 lockres->l_flags &= ~USER_LOCK_QUEUED;
301
302 /* It's valid to get here and no longer be blocked - if we get
303 * several basts in a row, we might be queued by the first
304 * one, the unblock thread might run and clear the queued
305 * flag, and finally we might get another bast which re-queues
306 * us before our ast for the downconvert is called. */
307 if (!(lockres->l_flags & USER_LOCK_BLOCKED)) {
308 mlog(0, "Lockres %s, flags 0x%x: queued but not blocking\n",
309 lockres->l_name, lockres->l_flags);
310 spin_unlock(&lockres->l_lock);
311 goto drop_ref;
312 }
313
314 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
315 mlog(0, "lock is in teardown so we do nothing\n");
316 spin_unlock(&lockres->l_lock);
317 goto drop_ref;
318 }
319
320 if (lockres->l_flags & USER_LOCK_BUSY) {
321 mlog(0, "Cancel lock %s, flags 0x%x\n",
322 lockres->l_name, lockres->l_flags);
323
324 if (lockres->l_flags & USER_LOCK_IN_CANCEL) {
325 spin_unlock(&lockres->l_lock);
326 goto drop_ref;
327 }
328
329 lockres->l_flags |= USER_LOCK_IN_CANCEL;
330 spin_unlock(&lockres->l_lock);
331
332 status = dlmunlock(dlm,
333 &lockres->l_lksb,
334 LKM_CANCEL,
335 user_unlock_ast,
336 lockres);
337 if (status != DLM_NORMAL)
338 user_log_dlm_error("dlmunlock", status, lockres);
339 goto drop_ref;
340 }
341
342 /* If there are still incompat holders, we can exit safely
343 * without worrying about re-queueing this lock as that will
344 * happen on the last call to user_cluster_unlock. */
345 if ((lockres->l_blocking == LKM_EXMODE)
346 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
347 spin_unlock(&lockres->l_lock);
348 mlog(0, "can't downconvert for ex: ro = %u, ex = %u\n",
349 lockres->l_ro_holders, lockres->l_ex_holders);
350 goto drop_ref;
351 }
352
353 if ((lockres->l_blocking == LKM_PRMODE)
354 && lockres->l_ex_holders) {
355 spin_unlock(&lockres->l_lock);
356 mlog(0, "can't downconvert for pr: ex = %u\n",
357 lockres->l_ex_holders);
358 goto drop_ref;
359 }
360
361 /* yay, we can downconvert now. */
362 new_level = user_highest_compat_lock_level(lockres->l_blocking);
363 lockres->l_requested = new_level;
364 lockres->l_flags |= USER_LOCK_BUSY;
365 mlog(0, "Downconvert lock from %d to %d\n",
366 lockres->l_level, new_level);
367 spin_unlock(&lockres->l_lock);
368
369 /* need lock downconvert request now... */
370 status = dlmlock(dlm,
371 new_level,
372 &lockres->l_lksb,
373 LKM_CONVERT|LKM_VALBLK,
374 lockres->l_name,
375 user_ast,
376 lockres,
377 user_bast);
378 if (status != DLM_NORMAL) {
379 user_log_dlm_error("dlmlock", status, lockres);
380 user_recover_from_dlm_error(lockres);
381 }
382
383 drop_ref:
384 user_dlm_drop_inode_ref(lockres);
385 }
386
387 static inline void user_dlm_inc_holders(struct user_lock_res *lockres,
388 int level)
389 {
390 switch(level) {
391 case LKM_EXMODE:
392 lockres->l_ex_holders++;
393 break;
394 case LKM_PRMODE:
395 lockres->l_ro_holders++;
396 break;
397 default:
398 BUG();
399 }
400 }
401
402 /* predict what lock level we'll be dropping down to on behalf
403 * of another node, and return true if the currently wanted
404 * level will be compatible with it. */
405 static inline int
406 user_may_continue_on_blocked_lock(struct user_lock_res *lockres,
407 int wanted)
408 {
409 BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED));
410
411 return wanted <= user_highest_compat_lock_level(lockres->l_blocking);
412 }
413
414 int user_dlm_cluster_lock(struct user_lock_res *lockres,
415 int level,
416 int lkm_flags)
417 {
418 int status, local_flags;
419 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
420
421 if (level != LKM_EXMODE &&
422 level != LKM_PRMODE) {
423 mlog(ML_ERROR, "lockres %s: invalid request!\n",
424 lockres->l_name);
425 status = -EINVAL;
426 goto bail;
427 }
428
429 mlog(0, "lockres %s: asking for %s lock, passed flags = 0x%x\n",
430 lockres->l_name,
431 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE",
432 lkm_flags);
433
434 again:
435 if (signal_pending(current)) {
436 status = -ERESTARTSYS;
437 goto bail;
438 }
439
440 spin_lock(&lockres->l_lock);
441
442 /* We only compare against the currently granted level
443 * here. If the lock is blocked waiting on a downconvert,
444 * we'll get caught below. */
445 if ((lockres->l_flags & USER_LOCK_BUSY) &&
446 (level > lockres->l_level)) {
447 /* is someone sitting in dlm_lock? If so, wait on
448 * them. */
449 spin_unlock(&lockres->l_lock);
450
451 user_wait_on_busy_lock(lockres);
452 goto again;
453 }
454
455 if ((lockres->l_flags & USER_LOCK_BLOCKED) &&
456 (!user_may_continue_on_blocked_lock(lockres, level))) {
457 /* is the lock is currently blocked on behalf of
458 * another node */
459 spin_unlock(&lockres->l_lock);
460
461 user_wait_on_blocked_lock(lockres);
462 goto again;
463 }
464
465 if (level > lockres->l_level) {
466 local_flags = lkm_flags | LKM_VALBLK;
467 if (lockres->l_level != LKM_IVMODE)
468 local_flags |= LKM_CONVERT;
469
470 lockres->l_requested = level;
471 lockres->l_flags |= USER_LOCK_BUSY;
472 spin_unlock(&lockres->l_lock);
473
474 BUG_ON(level == LKM_IVMODE);
475 BUG_ON(level == LKM_NLMODE);
476
477 mlog(0, "lock %s, get lock from %d to level = %d\n",
478 lockres->l_name, lockres->l_level, level);
479
480 /* call dlm_lock to upgrade lock now */
481 status = dlmlock(dlm,
482 level,
483 &lockres->l_lksb,
484 local_flags,
485 lockres->l_name,
486 user_ast,
487 lockres,
488 user_bast);
489 if (status != DLM_NORMAL) {
490 if ((lkm_flags & LKM_NOQUEUE) &&
491 (status == DLM_NOTQUEUED))
492 status = -EAGAIN;
493 else {
494 user_log_dlm_error("dlmlock", status, lockres);
495 status = -EINVAL;
496 }
497 user_recover_from_dlm_error(lockres);
498 goto bail;
499 }
500
501 mlog(0, "lock %s, successfull return from dlmlock\n",
502 lockres->l_name);
503
504 user_wait_on_busy_lock(lockres);
505 goto again;
506 }
507
508 user_dlm_inc_holders(lockres, level);
509 spin_unlock(&lockres->l_lock);
510
511 mlog(0, "lockres %s: Got %s lock!\n", lockres->l_name,
512 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE");
513
514 status = 0;
515 bail:
516 return status;
517 }
518
519 static inline void user_dlm_dec_holders(struct user_lock_res *lockres,
520 int level)
521 {
522 switch(level) {
523 case LKM_EXMODE:
524 BUG_ON(!lockres->l_ex_holders);
525 lockres->l_ex_holders--;
526 break;
527 case LKM_PRMODE:
528 BUG_ON(!lockres->l_ro_holders);
529 lockres->l_ro_holders--;
530 break;
531 default:
532 BUG();
533 }
534 }
535
536 void user_dlm_cluster_unlock(struct user_lock_res *lockres,
537 int level)
538 {
539 if (level != LKM_EXMODE &&
540 level != LKM_PRMODE) {
541 mlog(ML_ERROR, "lockres %s: invalid request!\n", lockres->l_name);
542 return;
543 }
544
545 mlog(0, "lockres %s: dropping %s lock\n", lockres->l_name,
546 (level == LKM_EXMODE) ? "LKM_EXMODE" : "LKM_PRMODE");
547
548 spin_lock(&lockres->l_lock);
549 user_dlm_dec_holders(lockres, level);
550 __user_dlm_cond_queue_lockres(lockres);
551 spin_unlock(&lockres->l_lock);
552 }
553
554 void user_dlm_write_lvb(struct inode *inode,
555 const char *val,
556 unsigned int len)
557 {
558 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
559 char *lvb = lockres->l_lksb.lvb;
560
561 BUG_ON(len > DLM_LVB_LEN);
562
563 spin_lock(&lockres->l_lock);
564
565 BUG_ON(lockres->l_level < LKM_EXMODE);
566 memcpy(lvb, val, len);
567
568 spin_unlock(&lockres->l_lock);
569 }
570
571 void user_dlm_read_lvb(struct inode *inode,
572 char *val,
573 unsigned int len)
574 {
575 struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres;
576 char *lvb = lockres->l_lksb.lvb;
577
578 BUG_ON(len > DLM_LVB_LEN);
579
580 spin_lock(&lockres->l_lock);
581
582 BUG_ON(lockres->l_level < LKM_PRMODE);
583 memcpy(val, lvb, len);
584
585 spin_unlock(&lockres->l_lock);
586 }
587
588 void user_dlm_lock_res_init(struct user_lock_res *lockres,
589 struct dentry *dentry)
590 {
591 memset(lockres, 0, sizeof(*lockres));
592
593 spin_lock_init(&lockres->l_lock);
594 init_waitqueue_head(&lockres->l_event);
595 lockres->l_level = LKM_IVMODE;
596 lockres->l_requested = LKM_IVMODE;
597 lockres->l_blocking = LKM_IVMODE;
598
599 /* should have been checked before getting here. */
600 BUG_ON(dentry->d_name.len >= USER_DLM_LOCK_ID_MAX_LEN);
601
602 memcpy(lockres->l_name,
603 dentry->d_name.name,
604 dentry->d_name.len);
605 }
606
607 int user_dlm_destroy_lock(struct user_lock_res *lockres)
608 {
609 int status = -EBUSY;
610 struct dlm_ctxt *dlm = dlm_ctxt_from_user_lockres(lockres);
611
612 mlog(0, "asked to destroy %s\n", lockres->l_name);
613
614 spin_lock(&lockres->l_lock);
615 if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) {
616 mlog(0, "Lock is already torn down\n");
617 spin_unlock(&lockres->l_lock);
618 return 0;
619 }
620
621 lockres->l_flags |= USER_LOCK_IN_TEARDOWN;
622
623 while (lockres->l_flags & USER_LOCK_BUSY) {
624 spin_unlock(&lockres->l_lock);
625
626 mlog(0, "lock %s is busy\n", lockres->l_name);
627
628 user_wait_on_busy_lock(lockres);
629
630 spin_lock(&lockres->l_lock);
631 }
632
633 if (lockres->l_ro_holders || lockres->l_ex_holders) {
634 spin_unlock(&lockres->l_lock);
635 mlog(0, "lock %s has holders\n", lockres->l_name);
636 goto bail;
637 }
638
639 status = 0;
640 if (!(lockres->l_flags & USER_LOCK_ATTACHED)) {
641 spin_unlock(&lockres->l_lock);
642 mlog(0, "lock %s is not attached\n", lockres->l_name);
643 goto bail;
644 }
645
646 lockres->l_flags &= ~USER_LOCK_ATTACHED;
647 lockres->l_flags |= USER_LOCK_BUSY;
648 spin_unlock(&lockres->l_lock);
649
650 mlog(0, "unlocking lockres %s\n", lockres->l_name);
651 status = dlmunlock(dlm,
652 &lockres->l_lksb,
653 LKM_VALBLK,
654 user_unlock_ast,
655 lockres);
656 if (status != DLM_NORMAL) {
657 user_log_dlm_error("dlmunlock", status, lockres);
658 status = -EINVAL;
659 goto bail;
660 }
661
662 user_wait_on_busy_lock(lockres);
663
664 status = 0;
665 bail:
666 return status;
667 }
668
669 struct dlm_ctxt *user_dlm_register_context(struct qstr *name)
670 {
671 struct dlm_ctxt *dlm;
672 u32 dlm_key;
673 char *domain;
674
675 domain = kmalloc(name->len + 1, GFP_KERNEL);
676 if (!domain) {
677 mlog_errno(-ENOMEM);
678 return ERR_PTR(-ENOMEM);
679 }
680
681 dlm_key = crc32_le(0, name->name, name->len);
682
683 snprintf(domain, name->len + 1, "%.*s", name->len, name->name);
684
685 dlm = dlm_register_domain(domain, dlm_key);
686 if (IS_ERR(dlm))
687 mlog_errno(PTR_ERR(dlm));
688
689 kfree(domain);
690 return dlm;
691 }
692
693 void user_dlm_unregister_context(struct dlm_ctxt *dlm)
694 {
695 dlm_unregister_domain(dlm);
696 }
This page took 0.04471 seconds and 6 git commands to generate.