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