GFS2: Wait for unlock completion on umount
[linux-2.6.git] / fs / gfs2 / glock.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
4  *
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  */
9
10 #include <linux/sched.h>
11 #include <linux/slab.h>
12 #include <linux/spinlock.h>
13 #include <linux/buffer_head.h>
14 #include <linux/delay.h>
15 #include <linux/sort.h>
16 #include <linux/jhash.h>
17 #include <linux/kallsyms.h>
18 #include <linux/gfs2_ondisk.h>
19 #include <linux/list.h>
20 #include <linux/wait.h>
21 #include <linux/module.h>
22 #include <linux/rwsem.h>
23 #include <asm/uaccess.h>
24 #include <linux/seq_file.h>
25 #include <linux/debugfs.h>
26 #include <linux/kthread.h>
27 #include <linux/freezer.h>
28 #include <linux/workqueue.h>
29 #include <linux/jiffies.h>
30
31 #include "gfs2.h"
32 #include "incore.h"
33 #include "glock.h"
34 #include "glops.h"
35 #include "inode.h"
36 #include "lops.h"
37 #include "meta_io.h"
38 #include "quota.h"
39 #include "super.h"
40 #include "util.h"
41 #include "bmap.h"
42 #define CREATE_TRACE_POINTS
43 #include "trace_gfs2.h"
44
45 struct gfs2_gl_hash_bucket {
46         struct hlist_head hb_list;
47 };
48
49 struct gfs2_glock_iter {
50         int hash;                       /* hash bucket index         */
51         struct gfs2_sbd *sdp;           /* incore superblock         */
52         struct gfs2_glock *gl;          /* current glock struct      */
53         char string[512];               /* scratch space             */
54 };
55
56 typedef void (*glock_examiner) (struct gfs2_glock * gl);
57
58 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp);
59 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl);
60 #define GLOCK_BUG_ON(gl,x) do { if (unlikely(x)) { __dump_glock(NULL, gl); BUG(); } } while(0)
61 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
62
63 static DECLARE_RWSEM(gfs2_umount_flush_sem);
64 static struct dentry *gfs2_root;
65 static struct workqueue_struct *glock_workqueue;
66 struct workqueue_struct *gfs2_delete_workqueue;
67 static LIST_HEAD(lru_list);
68 static atomic_t lru_count = ATOMIC_INIT(0);
69 static DEFINE_SPINLOCK(lru_lock);
70
71 #define GFS2_GL_HASH_SHIFT      15
72 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
73 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
74
75 static struct gfs2_gl_hash_bucket gl_hash_table[GFS2_GL_HASH_SIZE];
76 static struct dentry *gfs2_root;
77
78 /*
79  * Despite what you might think, the numbers below are not arbitrary :-)
80  * They are taken from the ipv4 routing hash code, which is well tested
81  * and thus should be nearly optimal. Later on we might tweek the numbers
82  * but for now this should be fine.
83  *
84  * The reason for putting the locks in a separate array from the list heads
85  * is that we can have fewer locks than list heads and save memory. We use
86  * the same hash function for both, but with a different hash mask.
87  */
88 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \
89         defined(CONFIG_PROVE_LOCKING)
90
91 #ifdef CONFIG_LOCKDEP
92 # define GL_HASH_LOCK_SZ        256
93 #else
94 # if NR_CPUS >= 32
95 #  define GL_HASH_LOCK_SZ       4096
96 # elif NR_CPUS >= 16
97 #  define GL_HASH_LOCK_SZ       2048
98 # elif NR_CPUS >= 8
99 #  define GL_HASH_LOCK_SZ       1024
100 # elif NR_CPUS >= 4
101 #  define GL_HASH_LOCK_SZ       512
102 # else
103 #  define GL_HASH_LOCK_SZ       256
104 # endif
105 #endif
106
107 /* We never want more locks than chains */
108 #if GFS2_GL_HASH_SIZE < GL_HASH_LOCK_SZ
109 # undef GL_HASH_LOCK_SZ
110 # define GL_HASH_LOCK_SZ GFS2_GL_HASH_SIZE
111 #endif
112
113 static rwlock_t gl_hash_locks[GL_HASH_LOCK_SZ];
114
115 static inline rwlock_t *gl_lock_addr(unsigned int x)
116 {
117         return &gl_hash_locks[x & (GL_HASH_LOCK_SZ-1)];
118 }
119 #else /* not SMP, so no spinlocks required */
120 static inline rwlock_t *gl_lock_addr(unsigned int x)
121 {
122         return NULL;
123 }
124 #endif
125
126 /**
127  * gl_hash() - Turn glock number into hash bucket number
128  * @lock: The glock number
129  *
130  * Returns: The number of the corresponding hash bucket
131  */
132
133 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
134                             const struct lm_lockname *name)
135 {
136         unsigned int h;
137
138         h = jhash(&name->ln_number, sizeof(u64), 0);
139         h = jhash(&name->ln_type, sizeof(unsigned int), h);
140         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
141         h &= GFS2_GL_HASH_MASK;
142
143         return h;
144 }
145
146 /**
147  * glock_free() - Perform a few checks and then release struct gfs2_glock
148  * @gl: The glock to release
149  *
150  * Also calls lock module to release its internal structure for this glock.
151  *
152  */
153
154 static void glock_free(struct gfs2_glock *gl)
155 {
156         struct gfs2_sbd *sdp = gl->gl_sbd;
157         struct inode *aspace = gl->gl_aspace;
158
159         if (aspace)
160                 gfs2_aspace_put(aspace);
161         trace_gfs2_glock_put(gl);
162         sdp->sd_lockstruct.ls_ops->lm_put_lock(gfs2_glock_cachep, gl);
163 }
164
165 /**
166  * gfs2_glock_hold() - increment reference count on glock
167  * @gl: The glock to hold
168  *
169  */
170
171 void gfs2_glock_hold(struct gfs2_glock *gl)
172 {
173         GLOCK_BUG_ON(gl, atomic_read(&gl->gl_ref) == 0);
174         atomic_inc(&gl->gl_ref);
175 }
176
177 /**
178  * demote_ok - Check to see if it's ok to unlock a glock
179  * @gl: the glock
180  *
181  * Returns: 1 if it's ok
182  */
183
184 static int demote_ok(const struct gfs2_glock *gl)
185 {
186         const struct gfs2_glock_operations *glops = gl->gl_ops;
187
188         if (gl->gl_state == LM_ST_UNLOCKED)
189                 return 0;
190         if (!list_empty(&gl->gl_holders))
191                 return 0;
192         if (glops->go_demote_ok)
193                 return glops->go_demote_ok(gl);
194         return 1;
195 }
196
197 /**
198  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
199  * @gl: the glock
200  *
201  */
202
203 static void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
204 {
205         int may_reclaim;
206         may_reclaim = (demote_ok(gl) &&
207                        (atomic_read(&gl->gl_ref) == 1 ||
208                         (gl->gl_name.ln_type == LM_TYPE_INODE &&
209                          atomic_read(&gl->gl_ref) <= 2)));
210         spin_lock(&lru_lock);
211         if (list_empty(&gl->gl_lru) && may_reclaim) {
212                 list_add_tail(&gl->gl_lru, &lru_list);
213                 atomic_inc(&lru_count);
214         }
215         spin_unlock(&lru_lock);
216 }
217
218 /**
219  * gfs2_glock_put_nolock() - Decrement reference count on glock
220  * @gl: The glock to put
221  *
222  * This function should only be used if the caller has its own reference
223  * to the glock, in addition to the one it is dropping.
224  */
225
226 void gfs2_glock_put_nolock(struct gfs2_glock *gl)
227 {
228         if (atomic_dec_and_test(&gl->gl_ref))
229                 GLOCK_BUG_ON(gl, 1);
230         gfs2_glock_schedule_for_reclaim(gl);
231 }
232
233 /**
234  * gfs2_glock_put() - Decrement reference count on glock
235  * @gl: The glock to put
236  *
237  */
238
239 int gfs2_glock_put(struct gfs2_glock *gl)
240 {
241         int rv = 0;
242
243         write_lock(gl_lock_addr(gl->gl_hash));
244         if (atomic_dec_and_lock(&gl->gl_ref, &lru_lock)) {
245                 hlist_del(&gl->gl_list);
246                 if (!list_empty(&gl->gl_lru)) {
247                         list_del_init(&gl->gl_lru);
248                         atomic_dec(&lru_count);
249                 }
250                 spin_unlock(&lru_lock);
251                 write_unlock(gl_lock_addr(gl->gl_hash));
252                 GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
253                 glock_free(gl);
254                 rv = 1;
255                 goto out;
256         }
257         spin_lock(&gl->gl_spin);
258         gfs2_glock_schedule_for_reclaim(gl);
259         spin_unlock(&gl->gl_spin);
260         write_unlock(gl_lock_addr(gl->gl_hash));
261 out:
262         return rv;
263 }
264
265 /**
266  * search_bucket() - Find struct gfs2_glock by lock number
267  * @bucket: the bucket to search
268  * @name: The lock name
269  *
270  * Returns: NULL, or the struct gfs2_glock with the requested number
271  */
272
273 static struct gfs2_glock *search_bucket(unsigned int hash,
274                                         const struct gfs2_sbd *sdp,
275                                         const struct lm_lockname *name)
276 {
277         struct gfs2_glock *gl;
278         struct hlist_node *h;
279
280         hlist_for_each_entry(gl, h, &gl_hash_table[hash].hb_list, gl_list) {
281                 if (!lm_name_equal(&gl->gl_name, name))
282                         continue;
283                 if (gl->gl_sbd != sdp)
284                         continue;
285
286                 atomic_inc(&gl->gl_ref);
287
288                 return gl;
289         }
290
291         return NULL;
292 }
293
294 /**
295  * may_grant - check if its ok to grant a new lock
296  * @gl: The glock
297  * @gh: The lock request which we wish to grant
298  *
299  * Returns: true if its ok to grant the lock
300  */
301
302 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
303 {
304         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
305         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
306              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
307                 return 0;
308         if (gl->gl_state == gh->gh_state)
309                 return 1;
310         if (gh->gh_flags & GL_EXACT)
311                 return 0;
312         if (gl->gl_state == LM_ST_EXCLUSIVE) {
313                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
314                         return 1;
315                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
316                         return 1;
317         }
318         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
319                 return 1;
320         return 0;
321 }
322
323 static void gfs2_holder_wake(struct gfs2_holder *gh)
324 {
325         clear_bit(HIF_WAIT, &gh->gh_iflags);
326         smp_mb__after_clear_bit();
327         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
328 }
329
330 /**
331  * do_promote - promote as many requests as possible on the current queue
332  * @gl: The glock
333  * 
334  * Returns: 1 if there is a blocked holder at the head of the list, or 2
335  *          if a type specific operation is underway.
336  */
337
338 static int do_promote(struct gfs2_glock *gl)
339 __releases(&gl->gl_spin)
340 __acquires(&gl->gl_spin)
341 {
342         const struct gfs2_glock_operations *glops = gl->gl_ops;
343         struct gfs2_holder *gh, *tmp;
344         int ret;
345
346 restart:
347         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
348                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
349                         continue;
350                 if (may_grant(gl, gh)) {
351                         if (gh->gh_list.prev == &gl->gl_holders &&
352                             glops->go_lock) {
353                                 spin_unlock(&gl->gl_spin);
354                                 /* FIXME: eliminate this eventually */
355                                 ret = glops->go_lock(gh);
356                                 spin_lock(&gl->gl_spin);
357                                 if (ret) {
358                                         if (ret == 1)
359                                                 return 2;
360                                         gh->gh_error = ret;
361                                         list_del_init(&gh->gh_list);
362                                         trace_gfs2_glock_queue(gh, 0);
363                                         gfs2_holder_wake(gh);
364                                         goto restart;
365                                 }
366                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
367                                 trace_gfs2_promote(gh, 1);
368                                 gfs2_holder_wake(gh);
369                                 goto restart;
370                         }
371                         set_bit(HIF_HOLDER, &gh->gh_iflags);
372                         trace_gfs2_promote(gh, 0);
373                         gfs2_holder_wake(gh);
374                         continue;
375                 }
376                 if (gh->gh_list.prev == &gl->gl_holders)
377                         return 1;
378                 break;
379         }
380         return 0;
381 }
382
383 /**
384  * do_error - Something unexpected has happened during a lock request
385  *
386  */
387
388 static inline void do_error(struct gfs2_glock *gl, const int ret)
389 {
390         struct gfs2_holder *gh, *tmp;
391
392         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
393                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
394                         continue;
395                 if (ret & LM_OUT_ERROR)
396                         gh->gh_error = -EIO;
397                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
398                         gh->gh_error = GLR_TRYFAILED;
399                 else
400                         continue;
401                 list_del_init(&gh->gh_list);
402                 trace_gfs2_glock_queue(gh, 0);
403                 gfs2_holder_wake(gh);
404         }
405 }
406
407 /**
408  * find_first_waiter - find the first gh that's waiting for the glock
409  * @gl: the glock
410  */
411
412 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
413 {
414         struct gfs2_holder *gh;
415
416         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
417                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
418                         return gh;
419         }
420         return NULL;
421 }
422
423 /**
424  * state_change - record that the glock is now in a different state
425  * @gl: the glock
426  * @new_state the new state
427  *
428  */
429
430 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
431 {
432         int held1, held2;
433
434         held1 = (gl->gl_state != LM_ST_UNLOCKED);
435         held2 = (new_state != LM_ST_UNLOCKED);
436
437         if (held1 != held2) {
438                 if (held2)
439                         gfs2_glock_hold(gl);
440                 else
441                         gfs2_glock_put_nolock(gl);
442         }
443
444         gl->gl_state = new_state;
445         gl->gl_tchange = jiffies;
446 }
447
448 static void gfs2_demote_wake(struct gfs2_glock *gl)
449 {
450         gl->gl_demote_state = LM_ST_EXCLUSIVE;
451         clear_bit(GLF_DEMOTE, &gl->gl_flags);
452         smp_mb__after_clear_bit();
453         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
454 }
455
456 /**
457  * finish_xmote - The DLM has replied to one of our lock requests
458  * @gl: The glock
459  * @ret: The status from the DLM
460  *
461  */
462
463 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
464 {
465         const struct gfs2_glock_operations *glops = gl->gl_ops;
466         struct gfs2_holder *gh;
467         unsigned state = ret & LM_OUT_ST_MASK;
468         int rv;
469
470         spin_lock(&gl->gl_spin);
471         trace_gfs2_glock_state_change(gl, state);
472         state_change(gl, state);
473         gh = find_first_waiter(gl);
474
475         /* Demote to UN request arrived during demote to SH or DF */
476         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
477             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
478                 gl->gl_target = LM_ST_UNLOCKED;
479
480         /* Check for state != intended state */
481         if (unlikely(state != gl->gl_target)) {
482                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
483                         /* move to back of queue and try next entry */
484                         if (ret & LM_OUT_CANCELED) {
485                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
486                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
487                                 gh = find_first_waiter(gl);
488                                 gl->gl_target = gh->gh_state;
489                                 goto retry;
490                         }
491                         /* Some error or failed "try lock" - report it */
492                         if ((ret & LM_OUT_ERROR) ||
493                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
494                                 gl->gl_target = gl->gl_state;
495                                 do_error(gl, ret);
496                                 goto out;
497                         }
498                 }
499                 switch(state) {
500                 /* Unlocked due to conversion deadlock, try again */
501                 case LM_ST_UNLOCKED:
502 retry:
503                         do_xmote(gl, gh, gl->gl_target);
504                         break;
505                 /* Conversion fails, unlock and try again */
506                 case LM_ST_SHARED:
507                 case LM_ST_DEFERRED:
508                         do_xmote(gl, gh, LM_ST_UNLOCKED);
509                         break;
510                 default: /* Everything else */
511                         printk(KERN_ERR "GFS2: wanted %u got %u\n", gl->gl_target, state);
512                         GLOCK_BUG_ON(gl, 1);
513                 }
514                 spin_unlock(&gl->gl_spin);
515                 return;
516         }
517
518         /* Fast path - we got what we asked for */
519         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
520                 gfs2_demote_wake(gl);
521         if (state != LM_ST_UNLOCKED) {
522                 if (glops->go_xmote_bh) {
523                         spin_unlock(&gl->gl_spin);
524                         rv = glops->go_xmote_bh(gl, gh);
525                         spin_lock(&gl->gl_spin);
526                         if (rv) {
527                                 do_error(gl, rv);
528                                 goto out;
529                         }
530                 }
531                 rv = do_promote(gl);
532                 if (rv == 2)
533                         goto out_locked;
534         }
535 out:
536         clear_bit(GLF_LOCK, &gl->gl_flags);
537 out_locked:
538         spin_unlock(&gl->gl_spin);
539 }
540
541 static unsigned int gfs2_lm_lock(struct gfs2_sbd *sdp, void *lock,
542                                  unsigned int req_state,
543                                  unsigned int flags)
544 {
545         int ret = LM_OUT_ERROR;
546
547         if (!sdp->sd_lockstruct.ls_ops->lm_lock)
548                 return req_state == LM_ST_UNLOCKED ? 0 : req_state;
549
550         if (likely(!test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
551                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(lock,
552                                                          req_state, flags);
553         return ret;
554 }
555
556 /**
557  * do_xmote - Calls the DLM to change the state of a lock
558  * @gl: The lock state
559  * @gh: The holder (only for promotes)
560  * @target: The target lock state
561  *
562  */
563
564 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
565 __releases(&gl->gl_spin)
566 __acquires(&gl->gl_spin)
567 {
568         const struct gfs2_glock_operations *glops = gl->gl_ops;
569         struct gfs2_sbd *sdp = gl->gl_sbd;
570         unsigned int lck_flags = gh ? gh->gh_flags : 0;
571         int ret;
572
573         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
574                       LM_FLAG_PRIORITY);
575         BUG_ON(gl->gl_state == target);
576         BUG_ON(gl->gl_state == gl->gl_target);
577         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
578             glops->go_inval) {
579                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
580                 do_error(gl, 0); /* Fail queued try locks */
581         }
582         spin_unlock(&gl->gl_spin);
583         if (glops->go_xmote_th)
584                 glops->go_xmote_th(gl);
585         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
586                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
587         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
588
589         gfs2_glock_hold(gl);
590         if (target != LM_ST_UNLOCKED && (gl->gl_state == LM_ST_SHARED ||
591             gl->gl_state == LM_ST_DEFERRED) &&
592             !(lck_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
593                 lck_flags |= LM_FLAG_TRY_1CB;
594         ret = gfs2_lm_lock(sdp, gl, target, lck_flags);
595
596         if (!(ret & LM_OUT_ASYNC)) {
597                 finish_xmote(gl, ret);
598                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
599                         gfs2_glock_put(gl);
600         } else {
601                 GLOCK_BUG_ON(gl, ret != LM_OUT_ASYNC);
602         }
603         spin_lock(&gl->gl_spin);
604 }
605
606 /**
607  * find_first_holder - find the first "holder" gh
608  * @gl: the glock
609  */
610
611 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
612 {
613         struct gfs2_holder *gh;
614
615         if (!list_empty(&gl->gl_holders)) {
616                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
617                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
618                         return gh;
619         }
620         return NULL;
621 }
622
623 /**
624  * run_queue - do all outstanding tasks related to a glock
625  * @gl: The glock in question
626  * @nonblock: True if we must not block in run_queue
627  *
628  */
629
630 static void run_queue(struct gfs2_glock *gl, const int nonblock)
631 __releases(&gl->gl_spin)
632 __acquires(&gl->gl_spin)
633 {
634         struct gfs2_holder *gh = NULL;
635         int ret;
636
637         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
638                 return;
639
640         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
641
642         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
643             gl->gl_demote_state != gl->gl_state) {
644                 if (find_first_holder(gl))
645                         goto out_unlock;
646                 if (nonblock)
647                         goto out_sched;
648                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
649                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
650                 gl->gl_target = gl->gl_demote_state;
651         } else {
652                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
653                         gfs2_demote_wake(gl);
654                 ret = do_promote(gl);
655                 if (ret == 0)
656                         goto out_unlock;
657                 if (ret == 2)
658                         goto out;
659                 gh = find_first_waiter(gl);
660                 gl->gl_target = gh->gh_state;
661                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
662                         do_error(gl, 0); /* Fail queued try locks */
663         }
664         do_xmote(gl, gh, gl->gl_target);
665 out:
666         return;
667
668 out_sched:
669         clear_bit(GLF_LOCK, &gl->gl_flags);
670         smp_mb__after_clear_bit();
671         gfs2_glock_hold(gl);
672         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
673                 gfs2_glock_put_nolock(gl);
674         return;
675
676 out_unlock:
677         clear_bit(GLF_LOCK, &gl->gl_flags);
678         smp_mb__after_clear_bit();
679         return;
680 }
681
682 static void delete_work_func(struct work_struct *work)
683 {
684         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
685         struct gfs2_sbd *sdp = gl->gl_sbd;
686         struct gfs2_inode *ip = NULL;
687         struct inode *inode;
688         u64 no_addr = 0;
689
690         spin_lock(&gl->gl_spin);
691         ip = (struct gfs2_inode *)gl->gl_object;
692         if (ip)
693                 no_addr = ip->i_no_addr;
694         spin_unlock(&gl->gl_spin);
695         if (ip) {
696                 inode = gfs2_ilookup(sdp->sd_vfs, no_addr);
697                 if (inode) {
698                         d_prune_aliases(inode);
699                         iput(inode);
700                 }
701         }
702         gfs2_glock_put(gl);
703 }
704
705 static void glock_work_func(struct work_struct *work)
706 {
707         unsigned long delay = 0;
708         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
709         int drop_ref = 0;
710
711         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
712                 finish_xmote(gl, gl->gl_reply);
713                 drop_ref = 1;
714         }
715         down_read(&gfs2_umount_flush_sem);
716         spin_lock(&gl->gl_spin);
717         if (test_and_clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
718             gl->gl_state != LM_ST_UNLOCKED &&
719             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
720                 unsigned long holdtime, now = jiffies;
721                 holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
722                 if (time_before(now, holdtime))
723                         delay = holdtime - now;
724                 set_bit(delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE, &gl->gl_flags);
725         }
726         run_queue(gl, 0);
727         spin_unlock(&gl->gl_spin);
728         up_read(&gfs2_umount_flush_sem);
729         if (!delay ||
730             queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
731                 gfs2_glock_put(gl);
732         if (drop_ref)
733                 gfs2_glock_put(gl);
734 }
735
736 /**
737  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
738  * @sdp: The GFS2 superblock
739  * @number: the lock number
740  * @glops: The glock_operations to use
741  * @create: If 0, don't create the glock if it doesn't exist
742  * @glp: the glock is returned here
743  *
744  * This does not lock a glock, just finds/creates structures for one.
745  *
746  * Returns: errno
747  */
748
749 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
750                    const struct gfs2_glock_operations *glops, int create,
751                    struct gfs2_glock **glp)
752 {
753         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
754         struct gfs2_glock *gl, *tmp;
755         unsigned int hash = gl_hash(sdp, &name);
756         int error;
757
758         read_lock(gl_lock_addr(hash));
759         gl = search_bucket(hash, sdp, &name);
760         read_unlock(gl_lock_addr(hash));
761
762         *glp = gl;
763         if (gl)
764                 return 0;
765         if (!create)
766                 return -ENOENT;
767
768         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
769         if (!gl)
770                 return -ENOMEM;
771
772         gl->gl_flags = 0;
773         gl->gl_name = name;
774         atomic_set(&gl->gl_ref, 1);
775         gl->gl_state = LM_ST_UNLOCKED;
776         gl->gl_target = LM_ST_UNLOCKED;
777         gl->gl_demote_state = LM_ST_EXCLUSIVE;
778         gl->gl_hash = hash;
779         gl->gl_ops = glops;
780         snprintf(gl->gl_strname, GDLM_STRNAME_BYTES, "%8x%16llx", name.ln_type, (unsigned long long)number);
781         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
782         gl->gl_lksb.sb_lvbptr = gl->gl_lvb;
783         gl->gl_tchange = jiffies;
784         gl->gl_object = NULL;
785         gl->gl_sbd = sdp;
786         gl->gl_aspace = NULL;
787         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
788         INIT_WORK(&gl->gl_delete, delete_work_func);
789
790         /* If this glock protects actual on-disk data or metadata blocks,
791            create a VFS inode to manage the pages/buffers holding them. */
792         if (glops == &gfs2_inode_glops || glops == &gfs2_rgrp_glops) {
793                 gl->gl_aspace = gfs2_aspace_get(sdp);
794                 if (!gl->gl_aspace) {
795                         error = -ENOMEM;
796                         goto fail;
797                 }
798         }
799
800         write_lock(gl_lock_addr(hash));
801         tmp = search_bucket(hash, sdp, &name);
802         if (tmp) {
803                 write_unlock(gl_lock_addr(hash));
804                 glock_free(gl);
805                 gl = tmp;
806         } else {
807                 hlist_add_head(&gl->gl_list, &gl_hash_table[hash].hb_list);
808                 write_unlock(gl_lock_addr(hash));
809         }
810
811         *glp = gl;
812
813         return 0;
814
815 fail:
816         kmem_cache_free(gfs2_glock_cachep, gl);
817         return error;
818 }
819
820 /**
821  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
822  * @gl: the glock
823  * @state: the state we're requesting
824  * @flags: the modifier flags
825  * @gh: the holder structure
826  *
827  */
828
829 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
830                       struct gfs2_holder *gh)
831 {
832         INIT_LIST_HEAD(&gh->gh_list);
833         gh->gh_gl = gl;
834         gh->gh_ip = (unsigned long)__builtin_return_address(0);
835         gh->gh_owner_pid = get_pid(task_pid(current));
836         gh->gh_state = state;
837         gh->gh_flags = flags;
838         gh->gh_error = 0;
839         gh->gh_iflags = 0;
840         gfs2_glock_hold(gl);
841 }
842
843 /**
844  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
845  * @state: the state we're requesting
846  * @flags: the modifier flags
847  * @gh: the holder structure
848  *
849  * Don't mess with the glock.
850  *
851  */
852
853 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
854 {
855         gh->gh_state = state;
856         gh->gh_flags = flags;
857         gh->gh_iflags = 0;
858         gh->gh_ip = (unsigned long)__builtin_return_address(0);
859 }
860
861 /**
862  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
863  * @gh: the holder structure
864  *
865  */
866
867 void gfs2_holder_uninit(struct gfs2_holder *gh)
868 {
869         put_pid(gh->gh_owner_pid);
870         gfs2_glock_put(gh->gh_gl);
871         gh->gh_gl = NULL;
872         gh->gh_ip = 0;
873 }
874
875 /**
876  * gfs2_glock_holder_wait
877  * @word: unused
878  *
879  * This function and gfs2_glock_demote_wait both show up in the WCHAN
880  * field. Thus I've separated these otherwise identical functions in
881  * order to be more informative to the user.
882  */
883
884 static int gfs2_glock_holder_wait(void *word)
885 {
886         schedule();
887         return 0;
888 }
889
890 static int gfs2_glock_demote_wait(void *word)
891 {
892         schedule();
893         return 0;
894 }
895
896 static void wait_on_holder(struct gfs2_holder *gh)
897 {
898         might_sleep();
899         wait_on_bit(&gh->gh_iflags, HIF_WAIT, gfs2_glock_holder_wait, TASK_UNINTERRUPTIBLE);
900 }
901
902 static void wait_on_demote(struct gfs2_glock *gl)
903 {
904         might_sleep();
905         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, gfs2_glock_demote_wait, TASK_UNINTERRUPTIBLE);
906 }
907
908 /**
909  * handle_callback - process a demote request
910  * @gl: the glock
911  * @state: the state the caller wants us to change to
912  *
913  * There are only two requests that we are going to see in actual
914  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
915  */
916
917 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
918                             unsigned long delay)
919 {
920         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
921
922         set_bit(bit, &gl->gl_flags);
923         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
924                 gl->gl_demote_state = state;
925                 gl->gl_demote_time = jiffies;
926         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
927                         gl->gl_demote_state != state) {
928                 gl->gl_demote_state = LM_ST_UNLOCKED;
929         }
930         if (gl->gl_ops->go_callback)
931                 gl->gl_ops->go_callback(gl);
932         trace_gfs2_demote_rq(gl);
933 }
934
935 /**
936  * gfs2_glock_wait - wait on a glock acquisition
937  * @gh: the glock holder
938  *
939  * Returns: 0 on success
940  */
941
942 int gfs2_glock_wait(struct gfs2_holder *gh)
943 {
944         wait_on_holder(gh);
945         return gh->gh_error;
946 }
947
948 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
949 {
950         va_list args;
951
952         va_start(args, fmt);
953         if (seq) {
954                 struct gfs2_glock_iter *gi = seq->private;
955                 vsprintf(gi->string, fmt, args);
956                 seq_printf(seq, gi->string);
957         } else {
958                 printk(KERN_ERR " ");
959                 vprintk(fmt, args);
960         }
961         va_end(args);
962 }
963
964 /**
965  * add_to_queue - Add a holder to the wait queue (but look for recursion)
966  * @gh: the holder structure to add
967  *
968  * Eventually we should move the recursive locking trap to a
969  * debugging option or something like that. This is the fast
970  * path and needs to have the minimum number of distractions.
971  * 
972  */
973
974 static inline void add_to_queue(struct gfs2_holder *gh)
975 __releases(&gl->gl_spin)
976 __acquires(&gl->gl_spin)
977 {
978         struct gfs2_glock *gl = gh->gh_gl;
979         struct gfs2_sbd *sdp = gl->gl_sbd;
980         struct list_head *insert_pt = NULL;
981         struct gfs2_holder *gh2;
982         int try_lock = 0;
983
984         BUG_ON(gh->gh_owner_pid == NULL);
985         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
986                 BUG();
987
988         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
989                 if (test_bit(GLF_LOCK, &gl->gl_flags))
990                         try_lock = 1;
991                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
992                         goto fail;
993         }
994
995         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
996                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
997                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
998                         goto trap_recursive;
999                 if (try_lock &&
1000                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) &&
1001                     !may_grant(gl, gh)) {
1002 fail:
1003                         gh->gh_error = GLR_TRYFAILED;
1004                         gfs2_holder_wake(gh);
1005                         return;
1006                 }
1007                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1008                         continue;
1009                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1010                         insert_pt = &gh2->gh_list;
1011         }
1012         if (likely(insert_pt == NULL)) {
1013                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1014                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1015                         goto do_cancel;
1016                 return;
1017         }
1018         trace_gfs2_glock_queue(gh, 1);
1019         list_add_tail(&gh->gh_list, insert_pt);
1020 do_cancel:
1021         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1022         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1023                 spin_unlock(&gl->gl_spin);
1024                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1025                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1026                 spin_lock(&gl->gl_spin);
1027         }
1028         return;
1029
1030 trap_recursive:
1031         print_symbol(KERN_ERR "original: %s\n", gh2->gh_ip);
1032         printk(KERN_ERR "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1033         printk(KERN_ERR "lock type: %d req lock state : %d\n",
1034                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1035         print_symbol(KERN_ERR "new: %s\n", gh->gh_ip);
1036         printk(KERN_ERR "pid: %d\n", pid_nr(gh->gh_owner_pid));
1037         printk(KERN_ERR "lock type: %d req lock state : %d\n",
1038                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1039         __dump_glock(NULL, gl);
1040         BUG();
1041 }
1042
1043 /**
1044  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1045  * @gh: the holder structure
1046  *
1047  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1048  *
1049  * Returns: 0, GLR_TRYFAILED, or errno on failure
1050  */
1051
1052 int gfs2_glock_nq(struct gfs2_holder *gh)
1053 {
1054         struct gfs2_glock *gl = gh->gh_gl;
1055         struct gfs2_sbd *sdp = gl->gl_sbd;
1056         int error = 0;
1057
1058         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1059                 return -EIO;
1060
1061         spin_lock(&gl->gl_spin);
1062         add_to_queue(gh);
1063         run_queue(gl, 1);
1064         spin_unlock(&gl->gl_spin);
1065
1066         if (!(gh->gh_flags & GL_ASYNC))
1067                 error = gfs2_glock_wait(gh);
1068
1069         return error;
1070 }
1071
1072 /**
1073  * gfs2_glock_poll - poll to see if an async request has been completed
1074  * @gh: the holder
1075  *
1076  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1077  */
1078
1079 int gfs2_glock_poll(struct gfs2_holder *gh)
1080 {
1081         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1082 }
1083
1084 /**
1085  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1086  * @gh: the glock holder
1087  *
1088  */
1089
1090 void gfs2_glock_dq(struct gfs2_holder *gh)
1091 {
1092         struct gfs2_glock *gl = gh->gh_gl;
1093         const struct gfs2_glock_operations *glops = gl->gl_ops;
1094         unsigned delay = 0;
1095         int fast_path = 0;
1096
1097         spin_lock(&gl->gl_spin);
1098         if (gh->gh_flags & GL_NOCACHE)
1099                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1100
1101         list_del_init(&gh->gh_list);
1102         if (find_first_holder(gl) == NULL) {
1103                 if (glops->go_unlock) {
1104                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1105                         spin_unlock(&gl->gl_spin);
1106                         glops->go_unlock(gh);
1107                         spin_lock(&gl->gl_spin);
1108                         clear_bit(GLF_LOCK, &gl->gl_flags);
1109                 }
1110                 if (list_empty(&gl->gl_holders) &&
1111                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1112                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1113                         fast_path = 1;
1114         }
1115         trace_gfs2_glock_queue(gh, 0);
1116         spin_unlock(&gl->gl_spin);
1117         if (likely(fast_path))
1118                 return;
1119
1120         gfs2_glock_hold(gl);
1121         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1122             !test_bit(GLF_DEMOTE, &gl->gl_flags))
1123                 delay = gl->gl_ops->go_min_hold_time;
1124         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1125                 gfs2_glock_put(gl);
1126 }
1127
1128 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1129 {
1130         struct gfs2_glock *gl = gh->gh_gl;
1131         gfs2_glock_dq(gh);
1132         wait_on_demote(gl);
1133 }
1134
1135 /**
1136  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1137  * @gh: the holder structure
1138  *
1139  */
1140
1141 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1142 {
1143         gfs2_glock_dq(gh);
1144         gfs2_holder_uninit(gh);
1145 }
1146
1147 /**
1148  * gfs2_glock_nq_num - acquire a glock based on lock number
1149  * @sdp: the filesystem
1150  * @number: the lock number
1151  * @glops: the glock operations for the type of glock
1152  * @state: the state to acquire the glock in
1153  * @flags: modifier flags for the aquisition
1154  * @gh: the struct gfs2_holder
1155  *
1156  * Returns: errno
1157  */
1158
1159 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1160                       const struct gfs2_glock_operations *glops,
1161                       unsigned int state, int flags, struct gfs2_holder *gh)
1162 {
1163         struct gfs2_glock *gl;
1164         int error;
1165
1166         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1167         if (!error) {
1168                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1169                 gfs2_glock_put(gl);
1170         }
1171
1172         return error;
1173 }
1174
1175 /**
1176  * glock_compare - Compare two struct gfs2_glock structures for sorting
1177  * @arg_a: the first structure
1178  * @arg_b: the second structure
1179  *
1180  */
1181
1182 static int glock_compare(const void *arg_a, const void *arg_b)
1183 {
1184         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1185         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1186         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1187         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1188
1189         if (a->ln_number > b->ln_number)
1190                 return 1;
1191         if (a->ln_number < b->ln_number)
1192                 return -1;
1193         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1194         return 0;
1195 }
1196
1197 /**
1198  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1199  * @num_gh: the number of structures
1200  * @ghs: an array of struct gfs2_holder structures
1201  *
1202  * Returns: 0 on success (all glocks acquired),
1203  *          errno on failure (no glocks acquired)
1204  */
1205
1206 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1207                      struct gfs2_holder **p)
1208 {
1209         unsigned int x;
1210         int error = 0;
1211
1212         for (x = 0; x < num_gh; x++)
1213                 p[x] = &ghs[x];
1214
1215         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1216
1217         for (x = 0; x < num_gh; x++) {
1218                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1219
1220                 error = gfs2_glock_nq(p[x]);
1221                 if (error) {
1222                         while (x--)
1223                                 gfs2_glock_dq(p[x]);
1224                         break;
1225                 }
1226         }
1227
1228         return error;
1229 }
1230
1231 /**
1232  * gfs2_glock_nq_m - acquire multiple glocks
1233  * @num_gh: the number of structures
1234  * @ghs: an array of struct gfs2_holder structures
1235  *
1236  *
1237  * Returns: 0 on success (all glocks acquired),
1238  *          errno on failure (no glocks acquired)
1239  */
1240
1241 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1242 {
1243         struct gfs2_holder *tmp[4];
1244         struct gfs2_holder **pph = tmp;
1245         int error = 0;
1246
1247         switch(num_gh) {
1248         case 0:
1249                 return 0;
1250         case 1:
1251                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1252                 return gfs2_glock_nq(ghs);
1253         default:
1254                 if (num_gh <= 4)
1255                         break;
1256                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1257                 if (!pph)
1258                         return -ENOMEM;
1259         }
1260
1261         error = nq_m_sync(num_gh, ghs, pph);
1262
1263         if (pph != tmp)
1264                 kfree(pph);
1265
1266         return error;
1267 }
1268
1269 /**
1270  * gfs2_glock_dq_m - release multiple glocks
1271  * @num_gh: the number of structures
1272  * @ghs: an array of struct gfs2_holder structures
1273  *
1274  */
1275
1276 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1277 {
1278         unsigned int x;
1279
1280         for (x = 0; x < num_gh; x++)
1281                 gfs2_glock_dq(&ghs[x]);
1282 }
1283
1284 /**
1285  * gfs2_glock_dq_uninit_m - release multiple glocks
1286  * @num_gh: the number of structures
1287  * @ghs: an array of struct gfs2_holder structures
1288  *
1289  */
1290
1291 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1292 {
1293         unsigned int x;
1294
1295         for (x = 0; x < num_gh; x++)
1296                 gfs2_glock_dq_uninit(&ghs[x]);
1297 }
1298
1299 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1300 {
1301         unsigned long delay = 0;
1302         unsigned long holdtime;
1303         unsigned long now = jiffies;
1304
1305         gfs2_glock_hold(gl);
1306         holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
1307         if (time_before(now, holdtime))
1308                 delay = holdtime - now;
1309         if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1310                 delay = gl->gl_ops->go_min_hold_time;
1311
1312         spin_lock(&gl->gl_spin);
1313         handle_callback(gl, state, delay);
1314         spin_unlock(&gl->gl_spin);
1315         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1316                 gfs2_glock_put(gl);
1317 }
1318
1319 /**
1320  * gfs2_glock_complete - Callback used by locking
1321  * @gl: Pointer to the glock
1322  * @ret: The return value from the dlm
1323  *
1324  */
1325
1326 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1327 {
1328         struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
1329         gl->gl_reply = ret;
1330         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_flags))) {
1331                 struct gfs2_holder *gh;
1332                 spin_lock(&gl->gl_spin);
1333                 gh = find_first_waiter(gl);
1334                 if ((!(gh && (gh->gh_flags & LM_FLAG_NOEXP)) &&
1335                      (gl->gl_target != LM_ST_UNLOCKED)) ||
1336                     ((ret & ~LM_OUT_ST_MASK) != 0))
1337                         set_bit(GLF_FROZEN, &gl->gl_flags);
1338                 spin_unlock(&gl->gl_spin);
1339                 if (test_bit(GLF_FROZEN, &gl->gl_flags))
1340                         return;
1341         }
1342         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1343         gfs2_glock_hold(gl);
1344         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1345                 gfs2_glock_put(gl);
1346 }
1347
1348
1349 static int gfs2_shrink_glock_memory(int nr, gfp_t gfp_mask)
1350 {
1351         struct gfs2_glock *gl;
1352         int may_demote;
1353         int nr_skipped = 0;
1354         LIST_HEAD(skipped);
1355
1356         if (nr == 0)
1357                 goto out;
1358
1359         if (!(gfp_mask & __GFP_FS))
1360                 return -1;
1361
1362         spin_lock(&lru_lock);
1363         while(nr && !list_empty(&lru_list)) {
1364                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1365                 list_del_init(&gl->gl_lru);
1366                 atomic_dec(&lru_count);
1367
1368                 /* Test for being demotable */
1369                 if (!test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1370                         gfs2_glock_hold(gl);
1371                         spin_unlock(&lru_lock);
1372                         spin_lock(&gl->gl_spin);
1373                         may_demote = demote_ok(gl);
1374                         if (may_demote) {
1375                                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1376                                 nr--;
1377                         }
1378                         clear_bit(GLF_LOCK, &gl->gl_flags);
1379                         smp_mb__after_clear_bit();
1380                         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1381                                 gfs2_glock_put_nolock(gl);
1382                         spin_unlock(&gl->gl_spin);
1383                         spin_lock(&lru_lock);
1384                         continue;
1385                 }
1386                 nr_skipped++;
1387                 list_add(&gl->gl_lru, &skipped);
1388         }
1389         list_splice(&skipped, &lru_list);
1390         atomic_add(nr_skipped, &lru_count);
1391         spin_unlock(&lru_lock);
1392 out:
1393         return (atomic_read(&lru_count) / 100) * sysctl_vfs_cache_pressure;
1394 }
1395
1396 static struct shrinker glock_shrinker = {
1397         .shrink = gfs2_shrink_glock_memory,
1398         .seeks = DEFAULT_SEEKS,
1399 };
1400
1401 /**
1402  * examine_bucket - Call a function for glock in a hash bucket
1403  * @examiner: the function
1404  * @sdp: the filesystem
1405  * @bucket: the bucket
1406  *
1407  * Returns: 1 if the bucket has entries
1408  */
1409
1410 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
1411                           unsigned int hash)
1412 {
1413         struct gfs2_glock *gl, *prev = NULL;
1414         int has_entries = 0;
1415         struct hlist_head *head = &gl_hash_table[hash].hb_list;
1416
1417         read_lock(gl_lock_addr(hash));
1418         /* Can't use hlist_for_each_entry - don't want prefetch here */
1419         if (hlist_empty(head))
1420                 goto out;
1421         gl = list_entry(head->first, struct gfs2_glock, gl_list);
1422         while(1) {
1423                 if (!sdp || gl->gl_sbd == sdp) {
1424                         gfs2_glock_hold(gl);
1425                         read_unlock(gl_lock_addr(hash));
1426                         if (prev)
1427                                 gfs2_glock_put(prev);
1428                         prev = gl;
1429                         examiner(gl);
1430                         has_entries = 1;
1431                         read_lock(gl_lock_addr(hash));
1432                 }
1433                 if (gl->gl_list.next == NULL)
1434                         break;
1435                 gl = list_entry(gl->gl_list.next, struct gfs2_glock, gl_list);
1436         }
1437 out:
1438         read_unlock(gl_lock_addr(hash));
1439         if (prev)
1440                 gfs2_glock_put(prev);
1441         cond_resched();
1442         return has_entries;
1443 }
1444
1445
1446 /**
1447  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1448  * @gl: The glock to thaw
1449  *
1450  * N.B. When we freeze a glock, we leave a ref to the glock outstanding,
1451  * so this has to result in the ref count being dropped by one.
1452  */
1453
1454 static void thaw_glock(struct gfs2_glock *gl)
1455 {
1456         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1457                 return;
1458         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1459         gfs2_glock_hold(gl);
1460         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1461                 gfs2_glock_put(gl);
1462 }
1463
1464 /**
1465  * clear_glock - look at a glock and see if we can free it from glock cache
1466  * @gl: the glock to look at
1467  *
1468  */
1469
1470 static void clear_glock(struct gfs2_glock *gl)
1471 {
1472         spin_lock(&lru_lock);
1473         if (!list_empty(&gl->gl_lru)) {
1474                 list_del_init(&gl->gl_lru);
1475                 atomic_dec(&lru_count);
1476         }
1477         spin_unlock(&lru_lock);
1478
1479         spin_lock(&gl->gl_spin);
1480         if (find_first_holder(gl) == NULL && gl->gl_state != LM_ST_UNLOCKED)
1481                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1482         spin_unlock(&gl->gl_spin);
1483         gfs2_glock_hold(gl);
1484         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1485                 gfs2_glock_put(gl);
1486 }
1487
1488 /**
1489  * gfs2_glock_thaw - Thaw any frozen glocks
1490  * @sdp: The super block
1491  *
1492  */
1493
1494 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1495 {
1496         unsigned x;
1497
1498         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1499                 examine_bucket(thaw_glock, sdp, x);
1500 }
1501
1502 /**
1503  * gfs2_gl_hash_clear - Empty out the glock hash table
1504  * @sdp: the filesystem
1505  * @wait: wait until it's all gone
1506  *
1507  * Called when unmounting the filesystem.
1508  */
1509
1510 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1511 {
1512         unsigned long t;
1513         unsigned int x;
1514         int cont;
1515
1516         t = jiffies;
1517
1518         for (;;) {
1519                 cont = 0;
1520                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1521                         if (examine_bucket(clear_glock, sdp, x))
1522                                 cont = 1;
1523                 }
1524
1525                 if (!cont)
1526                         break;
1527
1528                 if (time_after_eq(jiffies,
1529                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
1530                         fs_warn(sdp, "Unmount seems to be stalled. "
1531                                      "Dumping lock state...\n");
1532                         gfs2_dump_lockstate(sdp);
1533                         t = jiffies;
1534                 }
1535
1536                 down_write(&gfs2_umount_flush_sem);
1537                 invalidate_inodes(sdp->sd_vfs);
1538                 up_write(&gfs2_umount_flush_sem);
1539                 msleep(10);
1540         }
1541 }
1542
1543 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1544 {
1545         struct gfs2_glock *gl = ip->i_gl;
1546         int ret;
1547
1548         ret = gfs2_truncatei_resume(ip);
1549         gfs2_assert_withdraw(gl->gl_sbd, ret == 0);
1550
1551         spin_lock(&gl->gl_spin);
1552         clear_bit(GLF_LOCK, &gl->gl_flags);
1553         run_queue(gl, 1);
1554         spin_unlock(&gl->gl_spin);
1555 }
1556
1557 static const char *state2str(unsigned state)
1558 {
1559         switch(state) {
1560         case LM_ST_UNLOCKED:
1561                 return "UN";
1562         case LM_ST_SHARED:
1563                 return "SH";
1564         case LM_ST_DEFERRED:
1565                 return "DF";
1566         case LM_ST_EXCLUSIVE:
1567                 return "EX";
1568         }
1569         return "??";
1570 }
1571
1572 static const char *hflags2str(char *buf, unsigned flags, unsigned long iflags)
1573 {
1574         char *p = buf;
1575         if (flags & LM_FLAG_TRY)
1576                 *p++ = 't';
1577         if (flags & LM_FLAG_TRY_1CB)
1578                 *p++ = 'T';
1579         if (flags & LM_FLAG_NOEXP)
1580                 *p++ = 'e';
1581         if (flags & LM_FLAG_ANY)
1582                 *p++ = 'A';
1583         if (flags & LM_FLAG_PRIORITY)
1584                 *p++ = 'p';
1585         if (flags & GL_ASYNC)
1586                 *p++ = 'a';
1587         if (flags & GL_EXACT)
1588                 *p++ = 'E';
1589         if (flags & GL_NOCACHE)
1590                 *p++ = 'c';
1591         if (test_bit(HIF_HOLDER, &iflags))
1592                 *p++ = 'H';
1593         if (test_bit(HIF_WAIT, &iflags))
1594                 *p++ = 'W';
1595         if (test_bit(HIF_FIRST, &iflags))
1596                 *p++ = 'F';
1597         *p = 0;
1598         return buf;
1599 }
1600
1601 /**
1602  * dump_holder - print information about a glock holder
1603  * @seq: the seq_file struct
1604  * @gh: the glock holder
1605  *
1606  * Returns: 0 on success, -ENOBUFS when we run out of space
1607  */
1608
1609 static int dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1610 {
1611         struct task_struct *gh_owner = NULL;
1612         char buffer[KSYM_SYMBOL_LEN];
1613         char flags_buf[32];
1614
1615         sprint_symbol(buffer, gh->gh_ip);
1616         if (gh->gh_owner_pid)
1617                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1618         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %s\n",
1619                   state2str(gh->gh_state),
1620                   hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1621                   gh->gh_error, 
1622                   gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1623                   gh_owner ? gh_owner->comm : "(ended)", buffer);
1624         return 0;
1625 }
1626
1627 static const char *gflags2str(char *buf, const unsigned long *gflags)
1628 {
1629         char *p = buf;
1630         if (test_bit(GLF_LOCK, gflags))
1631                 *p++ = 'l';
1632         if (test_bit(GLF_DEMOTE, gflags))
1633                 *p++ = 'D';
1634         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1635                 *p++ = 'd';
1636         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1637                 *p++ = 'p';
1638         if (test_bit(GLF_DIRTY, gflags))
1639                 *p++ = 'y';
1640         if (test_bit(GLF_LFLUSH, gflags))
1641                 *p++ = 'f';
1642         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1643                 *p++ = 'i';
1644         if (test_bit(GLF_REPLY_PENDING, gflags))
1645                 *p++ = 'r';
1646         if (test_bit(GLF_INITIAL, gflags))
1647                 *p++ = 'I';
1648         if (test_bit(GLF_FROZEN, gflags))
1649                 *p++ = 'F';
1650         *p = 0;
1651         return buf;
1652 }
1653
1654 /**
1655  * __dump_glock - print information about a glock
1656  * @seq: The seq_file struct
1657  * @gl: the glock
1658  *
1659  * The file format is as follows:
1660  * One line per object, capital letters are used to indicate objects
1661  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1662  * other objects are indented by a single space and follow the glock to
1663  * which they are related. Fields are indicated by lower case letters
1664  * followed by a colon and the field value, except for strings which are in
1665  * [] so that its possible to see if they are composed of spaces for
1666  * example. The field's are n = number (id of the object), f = flags,
1667  * t = type, s = state, r = refcount, e = error, p = pid.
1668  *
1669  * Returns: 0 on success, -ENOBUFS when we run out of space
1670  */
1671
1672 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1673 {
1674         const struct gfs2_glock_operations *glops = gl->gl_ops;
1675         unsigned long long dtime;
1676         const struct gfs2_holder *gh;
1677         char gflags_buf[32];
1678         int error = 0;
1679
1680         dtime = jiffies - gl->gl_demote_time;
1681         dtime *= 1000000/HZ; /* demote time in uSec */
1682         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1683                 dtime = 0;
1684         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llu f:%s t:%s d:%s/%llu a:%d r:%d\n",
1685                   state2str(gl->gl_state),
1686                   gl->gl_name.ln_type,
1687                   (unsigned long long)gl->gl_name.ln_number,
1688                   gflags2str(gflags_buf, &gl->gl_flags),
1689                   state2str(gl->gl_target),
1690                   state2str(gl->gl_demote_state), dtime,
1691                   atomic_read(&gl->gl_ail_count),
1692                   atomic_read(&gl->gl_ref));
1693
1694         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1695                 error = dump_holder(seq, gh);
1696                 if (error)
1697                         goto out;
1698         }
1699         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1700                 error = glops->go_dump(seq, gl);
1701 out:
1702         return error;
1703 }
1704
1705 static int dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1706 {
1707         int ret;
1708         spin_lock(&gl->gl_spin);
1709         ret = __dump_glock(seq, gl);
1710         spin_unlock(&gl->gl_spin);
1711         return ret;
1712 }
1713
1714 /**
1715  * gfs2_dump_lockstate - print out the current lockstate
1716  * @sdp: the filesystem
1717  * @ub: the buffer to copy the information into
1718  *
1719  * If @ub is NULL, dump the lockstate to the console.
1720  *
1721  */
1722
1723 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
1724 {
1725         struct gfs2_glock *gl;
1726         struct hlist_node *h;
1727         unsigned int x;
1728         int error = 0;
1729
1730         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1731
1732                 read_lock(gl_lock_addr(x));
1733
1734                 hlist_for_each_entry(gl, h, &gl_hash_table[x].hb_list, gl_list) {
1735                         if (gl->gl_sbd != sdp)
1736                                 continue;
1737
1738                         error = dump_glock(NULL, gl);
1739                         if (error)
1740                                 break;
1741                 }
1742
1743                 read_unlock(gl_lock_addr(x));
1744
1745                 if (error)
1746                         break;
1747         }
1748
1749
1750         return error;
1751 }
1752
1753
1754 int __init gfs2_glock_init(void)
1755 {
1756         unsigned i;
1757         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
1758                 INIT_HLIST_HEAD(&gl_hash_table[i].hb_list);
1759         }
1760 #ifdef GL_HASH_LOCK_SZ
1761         for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
1762                 rwlock_init(&gl_hash_locks[i]);
1763         }
1764 #endif
1765
1766         glock_workqueue = create_workqueue("glock_workqueue");
1767         if (IS_ERR(glock_workqueue))
1768                 return PTR_ERR(glock_workqueue);
1769         gfs2_delete_workqueue = create_workqueue("delete_workqueue");
1770         if (IS_ERR(gfs2_delete_workqueue)) {
1771                 destroy_workqueue(glock_workqueue);
1772                 return PTR_ERR(gfs2_delete_workqueue);
1773         }
1774
1775         register_shrinker(&glock_shrinker);
1776
1777         return 0;
1778 }
1779
1780 void gfs2_glock_exit(void)
1781 {
1782         unregister_shrinker(&glock_shrinker);
1783         destroy_workqueue(glock_workqueue);
1784         destroy_workqueue(gfs2_delete_workqueue);
1785 }
1786
1787 static int gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1788 {
1789         struct gfs2_glock *gl;
1790
1791 restart:
1792         read_lock(gl_lock_addr(gi->hash));
1793         gl = gi->gl;
1794         if (gl) {
1795                 gi->gl = hlist_entry(gl->gl_list.next,
1796                                      struct gfs2_glock, gl_list);
1797         } else {
1798                 gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
1799                                      struct gfs2_glock, gl_list);
1800         }
1801         if (gi->gl)
1802                 gfs2_glock_hold(gi->gl);
1803         read_unlock(gl_lock_addr(gi->hash));
1804         if (gl)
1805                 gfs2_glock_put(gl);
1806         while (gi->gl == NULL) {
1807                 gi->hash++;
1808                 if (gi->hash >= GFS2_GL_HASH_SIZE)
1809                         return 1;
1810                 read_lock(gl_lock_addr(gi->hash));
1811                 gi->gl = hlist_entry(gl_hash_table[gi->hash].hb_list.first,
1812                                      struct gfs2_glock, gl_list);
1813                 if (gi->gl)
1814                         gfs2_glock_hold(gi->gl);
1815                 read_unlock(gl_lock_addr(gi->hash));
1816         }
1817
1818         if (gi->sdp != gi->gl->gl_sbd)
1819                 goto restart;
1820
1821         return 0;
1822 }
1823
1824 static void gfs2_glock_iter_free(struct gfs2_glock_iter *gi)
1825 {
1826         if (gi->gl)
1827                 gfs2_glock_put(gi->gl);
1828         gi->gl = NULL;
1829 }
1830
1831 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1832 {
1833         struct gfs2_glock_iter *gi = seq->private;
1834         loff_t n = *pos;
1835
1836         gi->hash = 0;
1837
1838         do {
1839                 if (gfs2_glock_iter_next(gi)) {
1840                         gfs2_glock_iter_free(gi);
1841                         return NULL;
1842                 }
1843         } while (n--);
1844
1845         return gi->gl;
1846 }
1847
1848 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1849                                  loff_t *pos)
1850 {
1851         struct gfs2_glock_iter *gi = seq->private;
1852
1853         (*pos)++;
1854
1855         if (gfs2_glock_iter_next(gi)) {
1856                 gfs2_glock_iter_free(gi);
1857                 return NULL;
1858         }
1859
1860         return gi->gl;
1861 }
1862
1863 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1864 {
1865         struct gfs2_glock_iter *gi = seq->private;
1866         gfs2_glock_iter_free(gi);
1867 }
1868
1869 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1870 {
1871         return dump_glock(seq, iter_ptr);
1872 }
1873
1874 static const struct seq_operations gfs2_glock_seq_ops = {
1875         .start = gfs2_glock_seq_start,
1876         .next  = gfs2_glock_seq_next,
1877         .stop  = gfs2_glock_seq_stop,
1878         .show  = gfs2_glock_seq_show,
1879 };
1880
1881 static int gfs2_debugfs_open(struct inode *inode, struct file *file)
1882 {
1883         int ret = seq_open_private(file, &gfs2_glock_seq_ops,
1884                                    sizeof(struct gfs2_glock_iter));
1885         if (ret == 0) {
1886                 struct seq_file *seq = file->private_data;
1887                 struct gfs2_glock_iter *gi = seq->private;
1888                 gi->sdp = inode->i_private;
1889         }
1890         return ret;
1891 }
1892
1893 static const struct file_operations gfs2_debug_fops = {
1894         .owner   = THIS_MODULE,
1895         .open    = gfs2_debugfs_open,
1896         .read    = seq_read,
1897         .llseek  = seq_lseek,
1898         .release = seq_release_private,
1899 };
1900
1901 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
1902 {
1903         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
1904         if (!sdp->debugfs_dir)
1905                 return -ENOMEM;
1906         sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
1907                                                          S_IFREG | S_IRUGO,
1908                                                          sdp->debugfs_dir, sdp,
1909                                                          &gfs2_debug_fops);
1910         if (!sdp->debugfs_dentry_glocks)
1911                 return -ENOMEM;
1912
1913         return 0;
1914 }
1915
1916 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
1917 {
1918         if (sdp && sdp->debugfs_dir) {
1919                 if (sdp->debugfs_dentry_glocks) {
1920                         debugfs_remove(sdp->debugfs_dentry_glocks);
1921                         sdp->debugfs_dentry_glocks = NULL;
1922                 }
1923                 debugfs_remove(sdp->debugfs_dir);
1924                 sdp->debugfs_dir = NULL;
1925         }
1926 }
1927
1928 int gfs2_register_debugfs(void)
1929 {
1930         gfs2_root = debugfs_create_dir("gfs2", NULL);
1931         return gfs2_root ? 0 : -ENOMEM;
1932 }
1933
1934 void gfs2_unregister_debugfs(void)
1935 {
1936         debugfs_remove(gfs2_root);
1937         gfs2_root = NULL;
1938 }