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