GFS2: Processes waiting on inode glock that no processes are holding
[linux-2.6-block.git] / fs / gfs2 / glock.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
4  *
5  * This copyrighted material is made available to anyone wishing to use,
6  * modify, copy, or redistribute it subject to the terms and conditions
7  * of the GNU General Public License version 2.
8  */
9
10 #include <linux/sched.h>
11 #include <linux/slab.h>
12 #include <linux/spinlock.h>
13 #include <linux/buffer_head.h>
14 #include <linux/delay.h>
15 #include <linux/sort.h>
16 #include <linux/jhash.h>
17 #include <linux/kallsyms.h>
18 #include <linux/gfs2_ondisk.h>
19 #include <linux/list.h>
20 #include <linux/wait.h>
21 #include <linux/module.h>
22 #include <asm/uaccess.h>
23 #include <linux/seq_file.h>
24 #include <linux/debugfs.h>
25 #include <linux/kthread.h>
26 #include <linux/freezer.h>
27 #include <linux/workqueue.h>
28 #include <linux/jiffies.h>
29 #include <linux/rcupdate.h>
30 #include <linux/rculist_bl.h>
31 #include <linux/bit_spinlock.h>
32
33 #include "gfs2.h"
34 #include "incore.h"
35 #include "glock.h"
36 #include "glops.h"
37 #include "inode.h"
38 #include "lops.h"
39 #include "meta_io.h"
40 #include "quota.h"
41 #include "super.h"
42 #include "util.h"
43 #include "bmap.h"
44 #define CREATE_TRACE_POINTS
45 #include "trace_gfs2.h"
46
47 struct gfs2_glock_iter {
48         int hash;                       /* hash bucket index         */
49         struct gfs2_sbd *sdp;           /* incore superblock         */
50         struct gfs2_glock *gl;          /* current glock struct      */
51         char string[512];               /* scratch space             */
52 };
53
54 typedef void (*glock_examiner) (struct gfs2_glock * gl);
55
56 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl);
57 #define GLOCK_BUG_ON(gl,x) do { if (unlikely(x)) { __dump_glock(NULL, gl); BUG(); } } while(0)
58 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
59
60 static struct dentry *gfs2_root;
61 static struct workqueue_struct *glock_workqueue;
62 struct workqueue_struct *gfs2_delete_workqueue;
63 static LIST_HEAD(lru_list);
64 static atomic_t lru_count = ATOMIC_INIT(0);
65 static DEFINE_SPINLOCK(lru_lock);
66
67 #define GFS2_GL_HASH_SHIFT      15
68 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
69 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
70
71 static struct hlist_bl_head gl_hash_table[GFS2_GL_HASH_SIZE];
72 static struct dentry *gfs2_root;
73
74 /**
75  * gl_hash() - Turn glock number into hash bucket number
76  * @lock: The glock number
77  *
78  * Returns: The number of the corresponding hash bucket
79  */
80
81 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
82                             const struct lm_lockname *name)
83 {
84         unsigned int h;
85
86         h = jhash(&name->ln_number, sizeof(u64), 0);
87         h = jhash(&name->ln_type, sizeof(unsigned int), h);
88         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
89         h &= GFS2_GL_HASH_MASK;
90
91         return h;
92 }
93
94 static inline void spin_lock_bucket(unsigned int hash)
95 {
96         hlist_bl_lock(&gl_hash_table[hash]);
97 }
98
99 static inline void spin_unlock_bucket(unsigned int hash)
100 {
101         hlist_bl_unlock(&gl_hash_table[hash]);
102 }
103
104 static void gfs2_glock_dealloc(struct rcu_head *rcu)
105 {
106         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
107
108         if (gl->gl_ops->go_flags & GLOF_ASPACE)
109                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
110         else
111                 kmem_cache_free(gfs2_glock_cachep, gl);
112 }
113
114 void gfs2_glock_free(struct gfs2_glock *gl)
115 {
116         struct gfs2_sbd *sdp = gl->gl_sbd;
117
118         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
119         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
120                 wake_up(&sdp->sd_glock_wait);
121 }
122
123 /**
124  * gfs2_glock_hold() - increment reference count on glock
125  * @gl: The glock to hold
126  *
127  */
128
129 void gfs2_glock_hold(struct gfs2_glock *gl)
130 {
131         GLOCK_BUG_ON(gl, atomic_read(&gl->gl_ref) == 0);
132         atomic_inc(&gl->gl_ref);
133 }
134
135 /**
136  * demote_ok - Check to see if it's ok to unlock a glock
137  * @gl: the glock
138  *
139  * Returns: 1 if it's ok
140  */
141
142 static int demote_ok(const struct gfs2_glock *gl)
143 {
144         const struct gfs2_glock_operations *glops = gl->gl_ops;
145
146         if (gl->gl_state == LM_ST_UNLOCKED)
147                 return 0;
148         if (!list_empty(&gl->gl_holders))
149                 return 0;
150         if (glops->go_demote_ok)
151                 return glops->go_demote_ok(gl);
152         return 1;
153 }
154
155
156 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
157 {
158         spin_lock(&lru_lock);
159
160         if (!list_empty(&gl->gl_lru))
161                 list_del_init(&gl->gl_lru);
162         else
163                 atomic_inc(&lru_count);
164
165         list_add_tail(&gl->gl_lru, &lru_list);
166         set_bit(GLF_LRU, &gl->gl_flags);
167         spin_unlock(&lru_lock);
168 }
169
170 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
171 {
172         spin_lock(&lru_lock);
173         if (!list_empty(&gl->gl_lru)) {
174                 list_del_init(&gl->gl_lru);
175                 atomic_dec(&lru_count);
176                 clear_bit(GLF_LRU, &gl->gl_flags);
177         }
178         spin_unlock(&lru_lock);
179 }
180
181 /**
182  * __gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
183  * @gl: the glock
184  *
185  * If the glock is demotable, then we add it (or move it) to the end
186  * of the glock LRU list.
187  */
188
189 static void __gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
190 {
191         if (demote_ok(gl))
192                 gfs2_glock_add_to_lru(gl);
193 }
194
195 /**
196  * gfs2_glock_put_nolock() - Decrement reference count on glock
197  * @gl: The glock to put
198  *
199  * This function should only be used if the caller has its own reference
200  * to the glock, in addition to the one it is dropping.
201  */
202
203 void gfs2_glock_put_nolock(struct gfs2_glock *gl)
204 {
205         if (atomic_dec_and_test(&gl->gl_ref))
206                 GLOCK_BUG_ON(gl, 1);
207 }
208
209 /**
210  * gfs2_glock_put() - Decrement reference count on glock
211  * @gl: The glock to put
212  *
213  */
214
215 void gfs2_glock_put(struct gfs2_glock *gl)
216 {
217         struct gfs2_sbd *sdp = gl->gl_sbd;
218         struct address_space *mapping = gfs2_glock2aspace(gl);
219
220         if (atomic_dec_and_test(&gl->gl_ref)) {
221                 spin_lock_bucket(gl->gl_hash);
222                 hlist_bl_del_rcu(&gl->gl_list);
223                 spin_unlock_bucket(gl->gl_hash);
224                 gfs2_glock_remove_from_lru(gl);
225                 GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
226                 GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
227                 trace_gfs2_glock_put(gl);
228                 sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
229         }
230 }
231
232 /**
233  * search_bucket() - Find struct gfs2_glock by lock number
234  * @bucket: the bucket to search
235  * @name: The lock name
236  *
237  * Returns: NULL, or the struct gfs2_glock with the requested number
238  */
239
240 static struct gfs2_glock *search_bucket(unsigned int hash,
241                                         const struct gfs2_sbd *sdp,
242                                         const struct lm_lockname *name)
243 {
244         struct gfs2_glock *gl;
245         struct hlist_bl_node *h;
246
247         hlist_bl_for_each_entry_rcu(gl, h, &gl_hash_table[hash], gl_list) {
248                 if (!lm_name_equal(&gl->gl_name, name))
249                         continue;
250                 if (gl->gl_sbd != sdp)
251                         continue;
252                 if (atomic_inc_not_zero(&gl->gl_ref))
253                         return gl;
254         }
255
256         return NULL;
257 }
258
259 /**
260  * may_grant - check if its ok to grant a new lock
261  * @gl: The glock
262  * @gh: The lock request which we wish to grant
263  *
264  * Returns: true if its ok to grant the lock
265  */
266
267 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
268 {
269         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
270         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
271              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
272                 return 0;
273         if (gl->gl_state == gh->gh_state)
274                 return 1;
275         if (gh->gh_flags & GL_EXACT)
276                 return 0;
277         if (gl->gl_state == LM_ST_EXCLUSIVE) {
278                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
279                         return 1;
280                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
281                         return 1;
282         }
283         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
284                 return 1;
285         return 0;
286 }
287
288 static void gfs2_holder_wake(struct gfs2_holder *gh)
289 {
290         clear_bit(HIF_WAIT, &gh->gh_iflags);
291         smp_mb__after_clear_bit();
292         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
293 }
294
295 /**
296  * do_error - Something unexpected has happened during a lock request
297  *
298  */
299
300 static inline void do_error(struct gfs2_glock *gl, const int ret)
301 {
302         struct gfs2_holder *gh, *tmp;
303
304         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
305                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
306                         continue;
307                 if (ret & LM_OUT_ERROR)
308                         gh->gh_error = -EIO;
309                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
310                         gh->gh_error = GLR_TRYFAILED;
311                 else
312                         continue;
313                 list_del_init(&gh->gh_list);
314                 trace_gfs2_glock_queue(gh, 0);
315                 gfs2_holder_wake(gh);
316         }
317 }
318
319 /**
320  * do_promote - promote as many requests as possible on the current queue
321  * @gl: The glock
322  * 
323  * Returns: 1 if there is a blocked holder at the head of the list, or 2
324  *          if a type specific operation is underway.
325  */
326
327 static int do_promote(struct gfs2_glock *gl)
328 __releases(&gl->gl_spin)
329 __acquires(&gl->gl_spin)
330 {
331         const struct gfs2_glock_operations *glops = gl->gl_ops;
332         struct gfs2_holder *gh, *tmp;
333         int ret;
334
335 restart:
336         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
337                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
338                         continue;
339                 if (may_grant(gl, gh)) {
340                         if (gh->gh_list.prev == &gl->gl_holders &&
341                             glops->go_lock) {
342                                 spin_unlock(&gl->gl_spin);
343                                 /* FIXME: eliminate this eventually */
344                                 ret = glops->go_lock(gh);
345                                 spin_lock(&gl->gl_spin);
346                                 if (ret) {
347                                         if (ret == 1)
348                                                 return 2;
349                                         gh->gh_error = ret;
350                                         list_del_init(&gh->gh_list);
351                                         trace_gfs2_glock_queue(gh, 0);
352                                         gfs2_holder_wake(gh);
353                                         goto restart;
354                                 }
355                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
356                                 trace_gfs2_promote(gh, 1);
357                                 gfs2_holder_wake(gh);
358                                 goto restart;
359                         }
360                         set_bit(HIF_HOLDER, &gh->gh_iflags);
361                         trace_gfs2_promote(gh, 0);
362                         gfs2_holder_wake(gh);
363                         continue;
364                 }
365                 if (gh->gh_list.prev == &gl->gl_holders)
366                         return 1;
367                 do_error(gl, 0);
368                 break;
369         }
370         return 0;
371 }
372
373 /**
374  * find_first_waiter - find the first gh that's waiting for the glock
375  * @gl: the glock
376  */
377
378 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
379 {
380         struct gfs2_holder *gh;
381
382         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
383                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
384                         return gh;
385         }
386         return NULL;
387 }
388
389 /**
390  * state_change - record that the glock is now in a different state
391  * @gl: the glock
392  * @new_state the new state
393  *
394  */
395
396 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
397 {
398         int held1, held2;
399
400         held1 = (gl->gl_state != LM_ST_UNLOCKED);
401         held2 = (new_state != LM_ST_UNLOCKED);
402
403         if (held1 != held2) {
404                 if (held2)
405                         gfs2_glock_hold(gl);
406                 else
407                         gfs2_glock_put_nolock(gl);
408         }
409         if (held1 && held2 && list_empty(&gl->gl_holders))
410                 clear_bit(GLF_QUEUED, &gl->gl_flags);
411
412         gl->gl_state = new_state;
413         gl->gl_tchange = jiffies;
414 }
415
416 static void gfs2_demote_wake(struct gfs2_glock *gl)
417 {
418         gl->gl_demote_state = LM_ST_EXCLUSIVE;
419         clear_bit(GLF_DEMOTE, &gl->gl_flags);
420         smp_mb__after_clear_bit();
421         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
422 }
423
424 /**
425  * finish_xmote - The DLM has replied to one of our lock requests
426  * @gl: The glock
427  * @ret: The status from the DLM
428  *
429  */
430
431 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
432 {
433         const struct gfs2_glock_operations *glops = gl->gl_ops;
434         struct gfs2_holder *gh;
435         unsigned state = ret & LM_OUT_ST_MASK;
436         int rv;
437
438         spin_lock(&gl->gl_spin);
439         trace_gfs2_glock_state_change(gl, state);
440         state_change(gl, state);
441         gh = find_first_waiter(gl);
442
443         /* Demote to UN request arrived during demote to SH or DF */
444         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
445             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
446                 gl->gl_target = LM_ST_UNLOCKED;
447
448         /* Check for state != intended state */
449         if (unlikely(state != gl->gl_target)) {
450                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
451                         /* move to back of queue and try next entry */
452                         if (ret & LM_OUT_CANCELED) {
453                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
454                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
455                                 gh = find_first_waiter(gl);
456                                 gl->gl_target = gh->gh_state;
457                                 goto retry;
458                         }
459                         /* Some error or failed "try lock" - report it */
460                         if ((ret & LM_OUT_ERROR) ||
461                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
462                                 gl->gl_target = gl->gl_state;
463                                 do_error(gl, ret);
464                                 goto out;
465                         }
466                 }
467                 switch(state) {
468                 /* Unlocked due to conversion deadlock, try again */
469                 case LM_ST_UNLOCKED:
470 retry:
471                         do_xmote(gl, gh, gl->gl_target);
472                         break;
473                 /* Conversion fails, unlock and try again */
474                 case LM_ST_SHARED:
475                 case LM_ST_DEFERRED:
476                         do_xmote(gl, gh, LM_ST_UNLOCKED);
477                         break;
478                 default: /* Everything else */
479                         printk(KERN_ERR "GFS2: wanted %u got %u\n", gl->gl_target, state);
480                         GLOCK_BUG_ON(gl, 1);
481                 }
482                 spin_unlock(&gl->gl_spin);
483                 return;
484         }
485
486         /* Fast path - we got what we asked for */
487         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
488                 gfs2_demote_wake(gl);
489         if (state != LM_ST_UNLOCKED) {
490                 if (glops->go_xmote_bh) {
491                         spin_unlock(&gl->gl_spin);
492                         rv = glops->go_xmote_bh(gl, gh);
493                         spin_lock(&gl->gl_spin);
494                         if (rv) {
495                                 do_error(gl, rv);
496                                 goto out;
497                         }
498                 }
499                 rv = do_promote(gl);
500                 if (rv == 2)
501                         goto out_locked;
502         }
503 out:
504         clear_bit(GLF_LOCK, &gl->gl_flags);
505 out_locked:
506         spin_unlock(&gl->gl_spin);
507 }
508
509 /**
510  * do_xmote - Calls the DLM to change the state of a lock
511  * @gl: The lock state
512  * @gh: The holder (only for promotes)
513  * @target: The target lock state
514  *
515  */
516
517 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
518 __releases(&gl->gl_spin)
519 __acquires(&gl->gl_spin)
520 {
521         const struct gfs2_glock_operations *glops = gl->gl_ops;
522         struct gfs2_sbd *sdp = gl->gl_sbd;
523         unsigned int lck_flags = gh ? gh->gh_flags : 0;
524         int ret;
525
526         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
527                       LM_FLAG_PRIORITY);
528         GLOCK_BUG_ON(gl, gl->gl_state == target);
529         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
530         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
531             glops->go_inval) {
532                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
533                 do_error(gl, 0); /* Fail queued try locks */
534         }
535         gl->gl_req = target;
536         spin_unlock(&gl->gl_spin);
537         if (glops->go_xmote_th)
538                 glops->go_xmote_th(gl);
539         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
540                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
541         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
542
543         gfs2_glock_hold(gl);
544         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
545                 /* lock_dlm */
546                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
547                 GLOCK_BUG_ON(gl, ret);
548         } else { /* lock_nolock */
549                 finish_xmote(gl, target);
550                 if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
551                         gfs2_glock_put(gl);
552         }
553
554         spin_lock(&gl->gl_spin);
555 }
556
557 /**
558  * find_first_holder - find the first "holder" gh
559  * @gl: the glock
560  */
561
562 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
563 {
564         struct gfs2_holder *gh;
565
566         if (!list_empty(&gl->gl_holders)) {
567                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
568                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
569                         return gh;
570         }
571         return NULL;
572 }
573
574 /**
575  * run_queue - do all outstanding tasks related to a glock
576  * @gl: The glock in question
577  * @nonblock: True if we must not block in run_queue
578  *
579  */
580
581 static void run_queue(struct gfs2_glock *gl, const int nonblock)
582 __releases(&gl->gl_spin)
583 __acquires(&gl->gl_spin)
584 {
585         struct gfs2_holder *gh = NULL;
586         int ret;
587
588         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
589                 return;
590
591         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
592
593         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
594             gl->gl_demote_state != gl->gl_state) {
595                 if (find_first_holder(gl))
596                         goto out_unlock;
597                 if (nonblock)
598                         goto out_sched;
599                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
600                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
601                 gl->gl_target = gl->gl_demote_state;
602         } else {
603                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
604                         gfs2_demote_wake(gl);
605                 ret = do_promote(gl);
606                 if (ret == 0)
607                         goto out_unlock;
608                 if (ret == 2)
609                         goto out;
610                 gh = find_first_waiter(gl);
611                 gl->gl_target = gh->gh_state;
612                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
613                         do_error(gl, 0); /* Fail queued try locks */
614         }
615         do_xmote(gl, gh, gl->gl_target);
616 out:
617         return;
618
619 out_sched:
620         clear_bit(GLF_LOCK, &gl->gl_flags);
621         smp_mb__after_clear_bit();
622         gfs2_glock_hold(gl);
623         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
624                 gfs2_glock_put_nolock(gl);
625         return;
626
627 out_unlock:
628         clear_bit(GLF_LOCK, &gl->gl_flags);
629         smp_mb__after_clear_bit();
630         return;
631 }
632
633 static void delete_work_func(struct work_struct *work)
634 {
635         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
636         struct gfs2_sbd *sdp = gl->gl_sbd;
637         struct gfs2_inode *ip;
638         struct inode *inode;
639         u64 no_addr = gl->gl_name.ln_number;
640
641         ip = gl->gl_object;
642         /* Note: Unsafe to dereference ip as we don't hold right refs/locks */
643
644         if (ip)
645                 inode = gfs2_ilookup(sdp->sd_vfs, no_addr, 1);
646         else
647                 inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
648         if (inode && !IS_ERR(inode)) {
649                 d_prune_aliases(inode);
650                 iput(inode);
651         }
652         gfs2_glock_put(gl);
653 }
654
655 static void glock_work_func(struct work_struct *work)
656 {
657         unsigned long delay = 0;
658         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
659         int drop_ref = 0;
660
661         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
662                 finish_xmote(gl, gl->gl_reply);
663                 drop_ref = 1;
664         }
665         spin_lock(&gl->gl_spin);
666         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
667             gl->gl_state != LM_ST_UNLOCKED &&
668             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
669                 unsigned long holdtime, now = jiffies;
670
671                 holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
672                 if (time_before(now, holdtime))
673                         delay = holdtime - now;
674
675                 if (!delay) {
676                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
677                         set_bit(GLF_DEMOTE, &gl->gl_flags);
678                 }
679         }
680         run_queue(gl, 0);
681         spin_unlock(&gl->gl_spin);
682         if (!delay ||
683             queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
684                 gfs2_glock_put(gl);
685         if (drop_ref)
686                 gfs2_glock_put(gl);
687 }
688
689 /**
690  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
691  * @sdp: The GFS2 superblock
692  * @number: the lock number
693  * @glops: The glock_operations to use
694  * @create: If 0, don't create the glock if it doesn't exist
695  * @glp: the glock is returned here
696  *
697  * This does not lock a glock, just finds/creates structures for one.
698  *
699  * Returns: errno
700  */
701
702 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
703                    const struct gfs2_glock_operations *glops, int create,
704                    struct gfs2_glock **glp)
705 {
706         struct super_block *s = sdp->sd_vfs;
707         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
708         struct gfs2_glock *gl, *tmp;
709         unsigned int hash = gl_hash(sdp, &name);
710         struct address_space *mapping;
711         struct kmem_cache *cachep;
712
713         rcu_read_lock();
714         gl = search_bucket(hash, sdp, &name);
715         rcu_read_unlock();
716
717         *glp = gl;
718         if (gl)
719                 return 0;
720         if (!create)
721                 return -ENOENT;
722
723         if (glops->go_flags & GLOF_ASPACE)
724                 cachep = gfs2_glock_aspace_cachep;
725         else
726                 cachep = gfs2_glock_cachep;
727         gl = kmem_cache_alloc(cachep, GFP_KERNEL);
728         if (!gl)
729                 return -ENOMEM;
730
731         atomic_inc(&sdp->sd_glock_disposal);
732         gl->gl_flags = 0;
733         gl->gl_name = name;
734         atomic_set(&gl->gl_ref, 1);
735         gl->gl_state = LM_ST_UNLOCKED;
736         gl->gl_target = LM_ST_UNLOCKED;
737         gl->gl_demote_state = LM_ST_EXCLUSIVE;
738         gl->gl_hash = hash;
739         gl->gl_ops = glops;
740         snprintf(gl->gl_strname, GDLM_STRNAME_BYTES, "%8x%16llx", name.ln_type, (unsigned long long)number);
741         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
742         gl->gl_lksb.sb_lvbptr = gl->gl_lvb;
743         gl->gl_tchange = jiffies;
744         gl->gl_object = NULL;
745         gl->gl_sbd = sdp;
746         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
747         INIT_WORK(&gl->gl_delete, delete_work_func);
748
749         mapping = gfs2_glock2aspace(gl);
750         if (mapping) {
751                 mapping->a_ops = &gfs2_meta_aops;
752                 mapping->host = s->s_bdev->bd_inode;
753                 mapping->flags = 0;
754                 mapping_set_gfp_mask(mapping, GFP_NOFS);
755                 mapping->assoc_mapping = NULL;
756                 mapping->backing_dev_info = s->s_bdi;
757                 mapping->writeback_index = 0;
758         }
759
760         spin_lock_bucket(hash);
761         tmp = search_bucket(hash, sdp, &name);
762         if (tmp) {
763                 spin_unlock_bucket(hash);
764                 kmem_cache_free(cachep, gl);
765                 atomic_dec(&sdp->sd_glock_disposal);
766                 gl = tmp;
767         } else {
768                 hlist_bl_add_head_rcu(&gl->gl_list, &gl_hash_table[hash]);
769                 spin_unlock_bucket(hash);
770         }
771
772         *glp = gl;
773
774         return 0;
775 }
776
777 /**
778  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
779  * @gl: the glock
780  * @state: the state we're requesting
781  * @flags: the modifier flags
782  * @gh: the holder structure
783  *
784  */
785
786 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
787                       struct gfs2_holder *gh)
788 {
789         INIT_LIST_HEAD(&gh->gh_list);
790         gh->gh_gl = gl;
791         gh->gh_ip = (unsigned long)__builtin_return_address(0);
792         gh->gh_owner_pid = get_pid(task_pid(current));
793         gh->gh_state = state;
794         gh->gh_flags = flags;
795         gh->gh_error = 0;
796         gh->gh_iflags = 0;
797         gfs2_glock_hold(gl);
798 }
799
800 /**
801  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
802  * @state: the state we're requesting
803  * @flags: the modifier flags
804  * @gh: the holder structure
805  *
806  * Don't mess with the glock.
807  *
808  */
809
810 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
811 {
812         gh->gh_state = state;
813         gh->gh_flags = flags;
814         gh->gh_iflags = 0;
815         gh->gh_ip = (unsigned long)__builtin_return_address(0);
816         if (gh->gh_owner_pid)
817                 put_pid(gh->gh_owner_pid);
818         gh->gh_owner_pid = get_pid(task_pid(current));
819 }
820
821 /**
822  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
823  * @gh: the holder structure
824  *
825  */
826
827 void gfs2_holder_uninit(struct gfs2_holder *gh)
828 {
829         put_pid(gh->gh_owner_pid);
830         gfs2_glock_put(gh->gh_gl);
831         gh->gh_gl = NULL;
832         gh->gh_ip = 0;
833 }
834
835 /**
836  * gfs2_glock_holder_wait
837  * @word: unused
838  *
839  * This function and gfs2_glock_demote_wait both show up in the WCHAN
840  * field. Thus I've separated these otherwise identical functions in
841  * order to be more informative to the user.
842  */
843
844 static int gfs2_glock_holder_wait(void *word)
845 {
846         schedule();
847         return 0;
848 }
849
850 static int gfs2_glock_demote_wait(void *word)
851 {
852         schedule();
853         return 0;
854 }
855
856 static void wait_on_holder(struct gfs2_holder *gh)
857 {
858         might_sleep();
859         wait_on_bit(&gh->gh_iflags, HIF_WAIT, gfs2_glock_holder_wait, TASK_UNINTERRUPTIBLE);
860 }
861
862 static void wait_on_demote(struct gfs2_glock *gl)
863 {
864         might_sleep();
865         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, gfs2_glock_demote_wait, TASK_UNINTERRUPTIBLE);
866 }
867
868 /**
869  * handle_callback - process a demote request
870  * @gl: the glock
871  * @state: the state the caller wants us to change to
872  *
873  * There are only two requests that we are going to see in actual
874  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
875  */
876
877 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
878                             unsigned long delay)
879 {
880         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
881
882         set_bit(bit, &gl->gl_flags);
883         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
884                 gl->gl_demote_state = state;
885                 gl->gl_demote_time = jiffies;
886         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
887                         gl->gl_demote_state != state) {
888                 gl->gl_demote_state = LM_ST_UNLOCKED;
889         }
890         if (gl->gl_ops->go_callback)
891                 gl->gl_ops->go_callback(gl);
892         trace_gfs2_demote_rq(gl);
893 }
894
895 /**
896  * gfs2_glock_wait - wait on a glock acquisition
897  * @gh: the glock holder
898  *
899  * Returns: 0 on success
900  */
901
902 int gfs2_glock_wait(struct gfs2_holder *gh)
903 {
904         wait_on_holder(gh);
905         return gh->gh_error;
906 }
907
908 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
909 {
910         struct va_format vaf;
911         va_list args;
912
913         va_start(args, fmt);
914
915         if (seq) {
916                 struct gfs2_glock_iter *gi = seq->private;
917                 vsprintf(gi->string, fmt, args);
918                 seq_printf(seq, gi->string);
919         } else {
920                 vaf.fmt = fmt;
921                 vaf.va = &args;
922
923                 printk(KERN_ERR " %pV", &vaf);
924         }
925
926         va_end(args);
927 }
928
929 /**
930  * add_to_queue - Add a holder to the wait queue (but look for recursion)
931  * @gh: the holder structure to add
932  *
933  * Eventually we should move the recursive locking trap to a
934  * debugging option or something like that. This is the fast
935  * path and needs to have the minimum number of distractions.
936  * 
937  */
938
939 static inline void add_to_queue(struct gfs2_holder *gh)
940 __releases(&gl->gl_spin)
941 __acquires(&gl->gl_spin)
942 {
943         struct gfs2_glock *gl = gh->gh_gl;
944         struct gfs2_sbd *sdp = gl->gl_sbd;
945         struct list_head *insert_pt = NULL;
946         struct gfs2_holder *gh2;
947         int try_lock = 0;
948
949         BUG_ON(gh->gh_owner_pid == NULL);
950         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
951                 BUG();
952
953         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
954                 if (test_bit(GLF_LOCK, &gl->gl_flags))
955                         try_lock = 1;
956                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
957                         goto fail;
958         }
959
960         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
961                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
962                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
963                         goto trap_recursive;
964                 if (try_lock &&
965                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) &&
966                     !may_grant(gl, gh)) {
967 fail:
968                         gh->gh_error = GLR_TRYFAILED;
969                         gfs2_holder_wake(gh);
970                         return;
971                 }
972                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
973                         continue;
974                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
975                         insert_pt = &gh2->gh_list;
976         }
977         set_bit(GLF_QUEUED, &gl->gl_flags);
978         trace_gfs2_glock_queue(gh, 1);
979         if (likely(insert_pt == NULL)) {
980                 list_add_tail(&gh->gh_list, &gl->gl_holders);
981                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
982                         goto do_cancel;
983                 return;
984         }
985         list_add_tail(&gh->gh_list, insert_pt);
986 do_cancel:
987         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
988         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
989                 spin_unlock(&gl->gl_spin);
990                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
991                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
992                 spin_lock(&gl->gl_spin);
993         }
994         return;
995
996 trap_recursive:
997         print_symbol(KERN_ERR "original: %s\n", gh2->gh_ip);
998         printk(KERN_ERR "pid: %d\n", pid_nr(gh2->gh_owner_pid));
999         printk(KERN_ERR "lock type: %d req lock state : %d\n",
1000                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1001         print_symbol(KERN_ERR "new: %s\n", gh->gh_ip);
1002         printk(KERN_ERR "pid: %d\n", pid_nr(gh->gh_owner_pid));
1003         printk(KERN_ERR "lock type: %d req lock state : %d\n",
1004                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1005         __dump_glock(NULL, gl);
1006         BUG();
1007 }
1008
1009 /**
1010  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1011  * @gh: the holder structure
1012  *
1013  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1014  *
1015  * Returns: 0, GLR_TRYFAILED, or errno on failure
1016  */
1017
1018 int gfs2_glock_nq(struct gfs2_holder *gh)
1019 {
1020         struct gfs2_glock *gl = gh->gh_gl;
1021         struct gfs2_sbd *sdp = gl->gl_sbd;
1022         int error = 0;
1023
1024         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1025                 return -EIO;
1026
1027         if (test_bit(GLF_LRU, &gl->gl_flags))
1028                 gfs2_glock_remove_from_lru(gl);
1029
1030         spin_lock(&gl->gl_spin);
1031         add_to_queue(gh);
1032         if ((LM_FLAG_NOEXP & gh->gh_flags) &&
1033             test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1034                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1035         run_queue(gl, 1);
1036         spin_unlock(&gl->gl_spin);
1037
1038         if (!(gh->gh_flags & GL_ASYNC))
1039                 error = gfs2_glock_wait(gh);
1040
1041         return error;
1042 }
1043
1044 /**
1045  * gfs2_glock_poll - poll to see if an async request has been completed
1046  * @gh: the holder
1047  *
1048  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1049  */
1050
1051 int gfs2_glock_poll(struct gfs2_holder *gh)
1052 {
1053         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1054 }
1055
1056 /**
1057  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1058  * @gh: the glock holder
1059  *
1060  */
1061
1062 void gfs2_glock_dq(struct gfs2_holder *gh)
1063 {
1064         struct gfs2_glock *gl = gh->gh_gl;
1065         const struct gfs2_glock_operations *glops = gl->gl_ops;
1066         unsigned delay = 0;
1067         int fast_path = 0;
1068
1069         spin_lock(&gl->gl_spin);
1070         if (gh->gh_flags & GL_NOCACHE)
1071                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1072
1073         list_del_init(&gh->gh_list);
1074         if (find_first_holder(gl) == NULL) {
1075                 if (glops->go_unlock) {
1076                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1077                         spin_unlock(&gl->gl_spin);
1078                         glops->go_unlock(gh);
1079                         spin_lock(&gl->gl_spin);
1080                         clear_bit(GLF_LOCK, &gl->gl_flags);
1081                 }
1082                 if (list_empty(&gl->gl_holders) &&
1083                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1084                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1085                         fast_path = 1;
1086         }
1087         if (!test_bit(GLF_LFLUSH, &gl->gl_flags))
1088                 __gfs2_glock_schedule_for_reclaim(gl);
1089         trace_gfs2_glock_queue(gh, 0);
1090         spin_unlock(&gl->gl_spin);
1091         if (likely(fast_path))
1092                 return;
1093
1094         gfs2_glock_hold(gl);
1095         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1096             !test_bit(GLF_DEMOTE, &gl->gl_flags))
1097                 delay = gl->gl_ops->go_min_hold_time;
1098         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1099                 gfs2_glock_put(gl);
1100 }
1101
1102 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1103 {
1104         struct gfs2_glock *gl = gh->gh_gl;
1105         gfs2_glock_dq(gh);
1106         wait_on_demote(gl);
1107 }
1108
1109 /**
1110  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1111  * @gh: the holder structure
1112  *
1113  */
1114
1115 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1116 {
1117         gfs2_glock_dq(gh);
1118         gfs2_holder_uninit(gh);
1119 }
1120
1121 /**
1122  * gfs2_glock_nq_num - acquire a glock based on lock number
1123  * @sdp: the filesystem
1124  * @number: the lock number
1125  * @glops: the glock operations for the type of glock
1126  * @state: the state to acquire the glock in
1127  * @flags: modifier flags for the acquisition
1128  * @gh: the struct gfs2_holder
1129  *
1130  * Returns: errno
1131  */
1132
1133 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1134                       const struct gfs2_glock_operations *glops,
1135                       unsigned int state, int flags, struct gfs2_holder *gh)
1136 {
1137         struct gfs2_glock *gl;
1138         int error;
1139
1140         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1141         if (!error) {
1142                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1143                 gfs2_glock_put(gl);
1144         }
1145
1146         return error;
1147 }
1148
1149 /**
1150  * glock_compare - Compare two struct gfs2_glock structures for sorting
1151  * @arg_a: the first structure
1152  * @arg_b: the second structure
1153  *
1154  */
1155
1156 static int glock_compare(const void *arg_a, const void *arg_b)
1157 {
1158         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1159         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1160         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1161         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1162
1163         if (a->ln_number > b->ln_number)
1164                 return 1;
1165         if (a->ln_number < b->ln_number)
1166                 return -1;
1167         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1168         return 0;
1169 }
1170
1171 /**
1172  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1173  * @num_gh: the number of structures
1174  * @ghs: an array of struct gfs2_holder structures
1175  *
1176  * Returns: 0 on success (all glocks acquired),
1177  *          errno on failure (no glocks acquired)
1178  */
1179
1180 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1181                      struct gfs2_holder **p)
1182 {
1183         unsigned int x;
1184         int error = 0;
1185
1186         for (x = 0; x < num_gh; x++)
1187                 p[x] = &ghs[x];
1188
1189         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1190
1191         for (x = 0; x < num_gh; x++) {
1192                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1193
1194                 error = gfs2_glock_nq(p[x]);
1195                 if (error) {
1196                         while (x--)
1197                                 gfs2_glock_dq(p[x]);
1198                         break;
1199                 }
1200         }
1201
1202         return error;
1203 }
1204
1205 /**
1206  * gfs2_glock_nq_m - acquire multiple glocks
1207  * @num_gh: the number of structures
1208  * @ghs: an array of struct gfs2_holder structures
1209  *
1210  *
1211  * Returns: 0 on success (all glocks acquired),
1212  *          errno on failure (no glocks acquired)
1213  */
1214
1215 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1216 {
1217         struct gfs2_holder *tmp[4];
1218         struct gfs2_holder **pph = tmp;
1219         int error = 0;
1220
1221         switch(num_gh) {
1222         case 0:
1223                 return 0;
1224         case 1:
1225                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1226                 return gfs2_glock_nq(ghs);
1227         default:
1228                 if (num_gh <= 4)
1229                         break;
1230                 pph = kmalloc(num_gh * sizeof(struct gfs2_holder *), GFP_NOFS);
1231                 if (!pph)
1232                         return -ENOMEM;
1233         }
1234
1235         error = nq_m_sync(num_gh, ghs, pph);
1236
1237         if (pph != tmp)
1238                 kfree(pph);
1239
1240         return error;
1241 }
1242
1243 /**
1244  * gfs2_glock_dq_m - release multiple glocks
1245  * @num_gh: the number of structures
1246  * @ghs: an array of struct gfs2_holder structures
1247  *
1248  */
1249
1250 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1251 {
1252         while (num_gh--)
1253                 gfs2_glock_dq(&ghs[num_gh]);
1254 }
1255
1256 /**
1257  * gfs2_glock_dq_uninit_m - release multiple glocks
1258  * @num_gh: the number of structures
1259  * @ghs: an array of struct gfs2_holder structures
1260  *
1261  */
1262
1263 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1264 {
1265         while (num_gh--)
1266                 gfs2_glock_dq_uninit(&ghs[num_gh]);
1267 }
1268
1269 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1270 {
1271         unsigned long delay = 0;
1272         unsigned long holdtime;
1273         unsigned long now = jiffies;
1274
1275         gfs2_glock_hold(gl);
1276         holdtime = gl->gl_tchange + gl->gl_ops->go_min_hold_time;
1277         if (test_bit(GLF_QUEUED, &gl->gl_flags)) {
1278                 if (time_before(now, holdtime))
1279                         delay = holdtime - now;
1280                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1281                         delay = gl->gl_ops->go_min_hold_time;
1282         }
1283
1284         spin_lock(&gl->gl_spin);
1285         handle_callback(gl, state, delay);
1286         spin_unlock(&gl->gl_spin);
1287         if (queue_delayed_work(glock_workqueue, &gl->gl_work, delay) == 0)
1288                 gfs2_glock_put(gl);
1289 }
1290
1291 /**
1292  * gfs2_should_freeze - Figure out if glock should be frozen
1293  * @gl: The glock in question
1294  *
1295  * Glocks are not frozen if (a) the result of the dlm operation is
1296  * an error, (b) the locking operation was an unlock operation or
1297  * (c) if there is a "noexp" flagged request anywhere in the queue
1298  *
1299  * Returns: 1 if freezing should occur, 0 otherwise
1300  */
1301
1302 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1303 {
1304         const struct gfs2_holder *gh;
1305
1306         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1307                 return 0;
1308         if (gl->gl_target == LM_ST_UNLOCKED)
1309                 return 0;
1310
1311         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1312                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1313                         continue;
1314                 if (LM_FLAG_NOEXP & gh->gh_flags)
1315                         return 0;
1316         }
1317
1318         return 1;
1319 }
1320
1321 /**
1322  * gfs2_glock_complete - Callback used by locking
1323  * @gl: Pointer to the glock
1324  * @ret: The return value from the dlm
1325  *
1326  * The gl_reply field is under the gl_spin lock so that it is ok
1327  * to use a bitfield shared with other glock state fields.
1328  */
1329
1330 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1331 {
1332         struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
1333
1334         spin_lock(&gl->gl_spin);
1335         gl->gl_reply = ret;
1336
1337         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_flags))) {
1338                 if (gfs2_should_freeze(gl)) {
1339                         set_bit(GLF_FROZEN, &gl->gl_flags);
1340                         spin_unlock(&gl->gl_spin);
1341                         return;
1342                 }
1343         }
1344
1345         spin_unlock(&gl->gl_spin);
1346         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1347         smp_wmb();
1348         gfs2_glock_hold(gl);
1349         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1350                 gfs2_glock_put(gl);
1351 }
1352
1353
1354 static int gfs2_shrink_glock_memory(struct shrinker *shrink, int nr, gfp_t gfp_mask)
1355 {
1356         struct gfs2_glock *gl;
1357         int may_demote;
1358         int nr_skipped = 0;
1359         LIST_HEAD(skipped);
1360
1361         if (nr == 0)
1362                 goto out;
1363
1364         if (!(gfp_mask & __GFP_FS))
1365                 return -1;
1366
1367         spin_lock(&lru_lock);
1368         while(nr && !list_empty(&lru_list)) {
1369                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1370                 list_del_init(&gl->gl_lru);
1371                 clear_bit(GLF_LRU, &gl->gl_flags);
1372                 atomic_dec(&lru_count);
1373
1374                 /* Test for being demotable */
1375                 if (!test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1376                         gfs2_glock_hold(gl);
1377                         spin_unlock(&lru_lock);
1378                         spin_lock(&gl->gl_spin);
1379                         may_demote = demote_ok(gl);
1380                         if (may_demote) {
1381                                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1382                                 nr--;
1383                         }
1384                         clear_bit(GLF_LOCK, &gl->gl_flags);
1385                         smp_mb__after_clear_bit();
1386                         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1387                                 gfs2_glock_put_nolock(gl);
1388                         spin_unlock(&gl->gl_spin);
1389                         spin_lock(&lru_lock);
1390                         continue;
1391                 }
1392                 nr_skipped++;
1393                 list_add(&gl->gl_lru, &skipped);
1394                 set_bit(GLF_LRU, &gl->gl_flags);
1395         }
1396         list_splice(&skipped, &lru_list);
1397         atomic_add(nr_skipped, &lru_count);
1398         spin_unlock(&lru_lock);
1399 out:
1400         return (atomic_read(&lru_count) / 100) * sysctl_vfs_cache_pressure;
1401 }
1402
1403 static struct shrinker glock_shrinker = {
1404         .shrink = gfs2_shrink_glock_memory,
1405         .seeks = DEFAULT_SEEKS,
1406 };
1407
1408 /**
1409  * examine_bucket - Call a function for glock in a hash bucket
1410  * @examiner: the function
1411  * @sdp: the filesystem
1412  * @bucket: the bucket
1413  *
1414  */
1415
1416 static void examine_bucket(glock_examiner examiner, const struct gfs2_sbd *sdp,
1417                           unsigned int hash)
1418 {
1419         struct gfs2_glock *gl;
1420         struct hlist_bl_head *head = &gl_hash_table[hash];
1421         struct hlist_bl_node *pos;
1422
1423         rcu_read_lock();
1424         hlist_bl_for_each_entry_rcu(gl, pos, head, gl_list) {
1425                 if ((gl->gl_sbd == sdp) && atomic_read(&gl->gl_ref))
1426                         examiner(gl);
1427         }
1428         rcu_read_unlock();
1429         cond_resched();
1430 }
1431
1432 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1433 {
1434         unsigned x;
1435
1436         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1437                 examine_bucket(examiner, sdp, x);
1438 }
1439
1440
1441 /**
1442  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1443  * @gl: The glock to thaw
1444  *
1445  * N.B. When we freeze a glock, we leave a ref to the glock outstanding,
1446  * so this has to result in the ref count being dropped by one.
1447  */
1448
1449 static void thaw_glock(struct gfs2_glock *gl)
1450 {
1451         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))
1452                 return;
1453         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1454         gfs2_glock_hold(gl);
1455         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1456                 gfs2_glock_put(gl);
1457 }
1458
1459 /**
1460  * clear_glock - look at a glock and see if we can free it from glock cache
1461  * @gl: the glock to look at
1462  *
1463  */
1464
1465 static void clear_glock(struct gfs2_glock *gl)
1466 {
1467         gfs2_glock_remove_from_lru(gl);
1468
1469         spin_lock(&gl->gl_spin);
1470         if (gl->gl_state != LM_ST_UNLOCKED)
1471                 handle_callback(gl, LM_ST_UNLOCKED, 0);
1472         spin_unlock(&gl->gl_spin);
1473         gfs2_glock_hold(gl);
1474         if (queue_delayed_work(glock_workqueue, &gl->gl_work, 0) == 0)
1475                 gfs2_glock_put(gl);
1476 }
1477
1478 /**
1479  * gfs2_glock_thaw - Thaw any frozen glocks
1480  * @sdp: The super block
1481  *
1482  */
1483
1484 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1485 {
1486         glock_hash_walk(thaw_glock, sdp);
1487 }
1488
1489 static int dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1490 {
1491         int ret;
1492         spin_lock(&gl->gl_spin);
1493         ret = __dump_glock(seq, gl);
1494         spin_unlock(&gl->gl_spin);
1495         return ret;
1496 }
1497
1498 static void dump_glock_func(struct gfs2_glock *gl)
1499 {
1500         dump_glock(NULL, gl);
1501 }
1502
1503 /**
1504  * gfs2_gl_hash_clear - Empty out the glock hash table
1505  * @sdp: the filesystem
1506  * @wait: wait until it's all gone
1507  *
1508  * Called when unmounting the filesystem.
1509  */
1510
1511 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1512 {
1513         glock_hash_walk(clear_glock, sdp);
1514         flush_workqueue(glock_workqueue);
1515         wait_event(sdp->sd_glock_wait, atomic_read(&sdp->sd_glock_disposal) == 0);
1516         glock_hash_walk(dump_glock_func, sdp);
1517 }
1518
1519 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1520 {
1521         struct gfs2_glock *gl = ip->i_gl;
1522         int ret;
1523
1524         ret = gfs2_truncatei_resume(ip);
1525         gfs2_assert_withdraw(gl->gl_sbd, ret == 0);
1526
1527         spin_lock(&gl->gl_spin);
1528         clear_bit(GLF_LOCK, &gl->gl_flags);
1529         run_queue(gl, 1);
1530         spin_unlock(&gl->gl_spin);
1531 }
1532
1533 static const char *state2str(unsigned state)
1534 {
1535         switch(state) {
1536         case LM_ST_UNLOCKED:
1537                 return "UN";
1538         case LM_ST_SHARED:
1539                 return "SH";
1540         case LM_ST_DEFERRED:
1541                 return "DF";
1542         case LM_ST_EXCLUSIVE:
1543                 return "EX";
1544         }
1545         return "??";
1546 }
1547
1548 static const char *hflags2str(char *buf, unsigned flags, unsigned long iflags)
1549 {
1550         char *p = buf;
1551         if (flags & LM_FLAG_TRY)
1552                 *p++ = 't';
1553         if (flags & LM_FLAG_TRY_1CB)
1554                 *p++ = 'T';
1555         if (flags & LM_FLAG_NOEXP)
1556                 *p++ = 'e';
1557         if (flags & LM_FLAG_ANY)
1558                 *p++ = 'A';
1559         if (flags & LM_FLAG_PRIORITY)
1560                 *p++ = 'p';
1561         if (flags & GL_ASYNC)
1562                 *p++ = 'a';
1563         if (flags & GL_EXACT)
1564                 *p++ = 'E';
1565         if (flags & GL_NOCACHE)
1566                 *p++ = 'c';
1567         if (test_bit(HIF_HOLDER, &iflags))
1568                 *p++ = 'H';
1569         if (test_bit(HIF_WAIT, &iflags))
1570                 *p++ = 'W';
1571         if (test_bit(HIF_FIRST, &iflags))
1572                 *p++ = 'F';
1573         *p = 0;
1574         return buf;
1575 }
1576
1577 /**
1578  * dump_holder - print information about a glock holder
1579  * @seq: the seq_file struct
1580  * @gh: the glock holder
1581  *
1582  * Returns: 0 on success, -ENOBUFS when we run out of space
1583  */
1584
1585 static int dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1586 {
1587         struct task_struct *gh_owner = NULL;
1588         char flags_buf[32];
1589
1590         if (gh->gh_owner_pid)
1591                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1592         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1593                        state2str(gh->gh_state),
1594                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1595                        gh->gh_error,
1596                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1597                        gh_owner ? gh_owner->comm : "(ended)",
1598                        (void *)gh->gh_ip);
1599         return 0;
1600 }
1601
1602 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1603 {
1604         const unsigned long *gflags = &gl->gl_flags;
1605         char *p = buf;
1606
1607         if (test_bit(GLF_LOCK, gflags))
1608                 *p++ = 'l';
1609         if (test_bit(GLF_DEMOTE, gflags))
1610                 *p++ = 'D';
1611         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1612                 *p++ = 'd';
1613         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1614                 *p++ = 'p';
1615         if (test_bit(GLF_DIRTY, gflags))
1616                 *p++ = 'y';
1617         if (test_bit(GLF_LFLUSH, gflags))
1618                 *p++ = 'f';
1619         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1620                 *p++ = 'i';
1621         if (test_bit(GLF_REPLY_PENDING, gflags))
1622                 *p++ = 'r';
1623         if (test_bit(GLF_INITIAL, gflags))
1624                 *p++ = 'I';
1625         if (test_bit(GLF_FROZEN, gflags))
1626                 *p++ = 'F';
1627         if (test_bit(GLF_QUEUED, gflags))
1628                 *p++ = 'q';
1629         if (test_bit(GLF_LRU, gflags))
1630                 *p++ = 'L';
1631         if (gl->gl_object)
1632                 *p++ = 'o';
1633         *p = 0;
1634         return buf;
1635 }
1636
1637 /**
1638  * __dump_glock - print information about a glock
1639  * @seq: The seq_file struct
1640  * @gl: the glock
1641  *
1642  * The file format is as follows:
1643  * One line per object, capital letters are used to indicate objects
1644  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1645  * other objects are indented by a single space and follow the glock to
1646  * which they are related. Fields are indicated by lower case letters
1647  * followed by a colon and the field value, except for strings which are in
1648  * [] so that its possible to see if they are composed of spaces for
1649  * example. The field's are n = number (id of the object), f = flags,
1650  * t = type, s = state, r = refcount, e = error, p = pid.
1651  *
1652  * Returns: 0 on success, -ENOBUFS when we run out of space
1653  */
1654
1655 static int __dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1656 {
1657         const struct gfs2_glock_operations *glops = gl->gl_ops;
1658         unsigned long long dtime;
1659         const struct gfs2_holder *gh;
1660         char gflags_buf[32];
1661         int error = 0;
1662
1663         dtime = jiffies - gl->gl_demote_time;
1664         dtime *= 1000000/HZ; /* demote time in uSec */
1665         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1666                 dtime = 0;
1667         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d v:%d r:%d\n",
1668                   state2str(gl->gl_state),
1669                   gl->gl_name.ln_type,
1670                   (unsigned long long)gl->gl_name.ln_number,
1671                   gflags2str(gflags_buf, gl),
1672                   state2str(gl->gl_target),
1673                   state2str(gl->gl_demote_state), dtime,
1674                   atomic_read(&gl->gl_ail_count),
1675                   atomic_read(&gl->gl_revokes),
1676                   atomic_read(&gl->gl_ref));
1677
1678         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1679                 error = dump_holder(seq, gh);
1680                 if (error)
1681                         goto out;
1682         }
1683         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1684                 error = glops->go_dump(seq, gl);
1685 out:
1686         return error;
1687 }
1688
1689
1690
1691
1692 int __init gfs2_glock_init(void)
1693 {
1694         unsigned i;
1695         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
1696                 INIT_HLIST_BL_HEAD(&gl_hash_table[i]);
1697         }
1698
1699         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1700                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1701         if (IS_ERR(glock_workqueue))
1702                 return PTR_ERR(glock_workqueue);
1703         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1704                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1705                                                 0);
1706         if (IS_ERR(gfs2_delete_workqueue)) {
1707                 destroy_workqueue(glock_workqueue);
1708                 return PTR_ERR(gfs2_delete_workqueue);
1709         }
1710
1711         register_shrinker(&glock_shrinker);
1712
1713         return 0;
1714 }
1715
1716 void gfs2_glock_exit(void)
1717 {
1718         unregister_shrinker(&glock_shrinker);
1719         destroy_workqueue(glock_workqueue);
1720         destroy_workqueue(gfs2_delete_workqueue);
1721 }
1722
1723 static inline struct gfs2_glock *glock_hash_chain(unsigned hash)
1724 {
1725         return hlist_bl_entry(hlist_bl_first_rcu(&gl_hash_table[hash]),
1726                               struct gfs2_glock, gl_list);
1727 }
1728
1729 static inline struct gfs2_glock *glock_hash_next(struct gfs2_glock *gl)
1730 {
1731         return hlist_bl_entry(rcu_dereference(gl->gl_list.next),
1732                               struct gfs2_glock, gl_list);
1733 }
1734
1735 static int gfs2_glock_iter_next(struct gfs2_glock_iter *gi)
1736 {
1737         struct gfs2_glock *gl;
1738
1739         do {
1740                 gl = gi->gl;
1741                 if (gl) {
1742                         gi->gl = glock_hash_next(gl);
1743                 } else {
1744                         gi->gl = glock_hash_chain(gi->hash);
1745                 }
1746                 while (gi->gl == NULL) {
1747                         gi->hash++;
1748                         if (gi->hash >= GFS2_GL_HASH_SIZE) {
1749                                 rcu_read_unlock();
1750                                 return 1;
1751                         }
1752                         gi->gl = glock_hash_chain(gi->hash);
1753                 }
1754         /* Skip entries for other sb and dead entries */
1755         } while (gi->sdp != gi->gl->gl_sbd || atomic_read(&gi->gl->gl_ref) == 0);
1756
1757         return 0;
1758 }
1759
1760 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1761 {
1762         struct gfs2_glock_iter *gi = seq->private;
1763         loff_t n = *pos;
1764
1765         gi->hash = 0;
1766         rcu_read_lock();
1767
1768         do {
1769                 if (gfs2_glock_iter_next(gi))
1770                         return NULL;
1771         } while (n--);
1772
1773         return gi->gl;
1774 }
1775
1776 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1777                                  loff_t *pos)
1778 {
1779         struct gfs2_glock_iter *gi = seq->private;
1780
1781         (*pos)++;
1782
1783         if (gfs2_glock_iter_next(gi))
1784                 return NULL;
1785
1786         return gi->gl;
1787 }
1788
1789 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1790 {
1791         struct gfs2_glock_iter *gi = seq->private;
1792
1793         if (gi->gl)
1794                 rcu_read_unlock();
1795         gi->gl = NULL;
1796 }
1797
1798 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
1799 {
1800         return dump_glock(seq, iter_ptr);
1801 }
1802
1803 static const struct seq_operations gfs2_glock_seq_ops = {
1804         .start = gfs2_glock_seq_start,
1805         .next  = gfs2_glock_seq_next,
1806         .stop  = gfs2_glock_seq_stop,
1807         .show  = gfs2_glock_seq_show,
1808 };
1809
1810 static int gfs2_debugfs_open(struct inode *inode, struct file *file)
1811 {
1812         int ret = seq_open_private(file, &gfs2_glock_seq_ops,
1813                                    sizeof(struct gfs2_glock_iter));
1814         if (ret == 0) {
1815                 struct seq_file *seq = file->private_data;
1816                 struct gfs2_glock_iter *gi = seq->private;
1817                 gi->sdp = inode->i_private;
1818         }
1819         return ret;
1820 }
1821
1822 static const struct file_operations gfs2_debug_fops = {
1823         .owner   = THIS_MODULE,
1824         .open    = gfs2_debugfs_open,
1825         .read    = seq_read,
1826         .llseek  = seq_lseek,
1827         .release = seq_release_private,
1828 };
1829
1830 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
1831 {
1832         sdp->debugfs_dir = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
1833         if (!sdp->debugfs_dir)
1834                 return -ENOMEM;
1835         sdp->debugfs_dentry_glocks = debugfs_create_file("glocks",
1836                                                          S_IFREG | S_IRUGO,
1837                                                          sdp->debugfs_dir, sdp,
1838                                                          &gfs2_debug_fops);
1839         if (!sdp->debugfs_dentry_glocks)
1840                 return -ENOMEM;
1841
1842         return 0;
1843 }
1844
1845 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
1846 {
1847         if (sdp && sdp->debugfs_dir) {
1848                 if (sdp->debugfs_dentry_glocks) {
1849                         debugfs_remove(sdp->debugfs_dentry_glocks);
1850                         sdp->debugfs_dentry_glocks = NULL;
1851                 }
1852                 debugfs_remove(sdp->debugfs_dir);
1853                 sdp->debugfs_dir = NULL;
1854         }
1855 }
1856
1857 int gfs2_register_debugfs(void)
1858 {
1859         gfs2_root = debugfs_create_dir("gfs2", NULL);
1860         return gfs2_root ? 0 : -ENOMEM;
1861 }
1862
1863 void gfs2_unregister_debugfs(void)
1864 {
1865         debugfs_remove(gfs2_root);
1866         gfs2_root = NULL;
1867 }