GNU Linux-libre 4.19.211-gnu1
[releases.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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/sched.h>
13 #include <linux/slab.h>
14 #include <linux/spinlock.h>
15 #include <linux/buffer_head.h>
16 #include <linux/delay.h>
17 #include <linux/sort.h>
18 #include <linux/hash.h>
19 #include <linux/jhash.h>
20 #include <linux/kallsyms.h>
21 #include <linux/gfs2_ondisk.h>
22 #include <linux/list.h>
23 #include <linux/wait.h>
24 #include <linux/module.h>
25 #include <linux/uaccess.h>
26 #include <linux/seq_file.h>
27 #include <linux/debugfs.h>
28 #include <linux/kthread.h>
29 #include <linux/freezer.h>
30 #include <linux/workqueue.h>
31 #include <linux/jiffies.h>
32 #include <linux/rcupdate.h>
33 #include <linux/rculist_bl.h>
34 #include <linux/bit_spinlock.h>
35 #include <linux/percpu.h>
36 #include <linux/list_sort.h>
37 #include <linux/lockref.h>
38 #include <linux/rhashtable.h>
39
40 #include "gfs2.h"
41 #include "incore.h"
42 #include "glock.h"
43 #include "glops.h"
44 #include "inode.h"
45 #include "lops.h"
46 #include "meta_io.h"
47 #include "quota.h"
48 #include "super.h"
49 #include "util.h"
50 #include "bmap.h"
51 #define CREATE_TRACE_POINTS
52 #include "trace_gfs2.h"
53
54 struct gfs2_glock_iter {
55         struct gfs2_sbd *sdp;           /* incore superblock           */
56         struct rhashtable_iter hti;     /* rhashtable iterator         */
57         struct gfs2_glock *gl;          /* current glock struct        */
58         loff_t last_pos;                /* last position               */
59 };
60
61 typedef void (*glock_examiner) (struct gfs2_glock * gl);
62
63 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
64
65 static struct dentry *gfs2_root;
66 static struct workqueue_struct *glock_workqueue;
67 struct workqueue_struct *gfs2_delete_workqueue;
68 static LIST_HEAD(lru_list);
69 static atomic_t lru_count = ATOMIC_INIT(0);
70 static DEFINE_SPINLOCK(lru_lock);
71
72 #define GFS2_GL_HASH_SHIFT      15
73 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
74
75 static const struct rhashtable_params ht_parms = {
76         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
77         .key_len = offsetofend(struct lm_lockname, ln_type),
78         .key_offset = offsetof(struct gfs2_glock, gl_name),
79         .head_offset = offsetof(struct gfs2_glock, gl_node),
80 };
81
82 static struct rhashtable gl_hash_table;
83
84 #define GLOCK_WAIT_TABLE_BITS 12
85 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
86 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
87
88 struct wait_glock_queue {
89         struct lm_lockname *name;
90         wait_queue_entry_t wait;
91 };
92
93 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
94                                int sync, void *key)
95 {
96         struct wait_glock_queue *wait_glock =
97                 container_of(wait, struct wait_glock_queue, wait);
98         struct lm_lockname *wait_name = wait_glock->name;
99         struct lm_lockname *wake_name = key;
100
101         if (wake_name->ln_sbd != wait_name->ln_sbd ||
102             wake_name->ln_number != wait_name->ln_number ||
103             wake_name->ln_type != wait_name->ln_type)
104                 return 0;
105         return autoremove_wake_function(wait, mode, sync, key);
106 }
107
108 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
109 {
110         u32 hash = jhash2((u32 *)name, ht_parms.key_len / 4, 0);
111
112         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
113 }
114
115 /**
116  * wake_up_glock  -  Wake up waiters on a glock
117  * @gl: the glock
118  */
119 static void wake_up_glock(struct gfs2_glock *gl)
120 {
121         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
122
123         if (waitqueue_active(wq))
124                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
125 }
126
127 static void gfs2_glock_dealloc(struct rcu_head *rcu)
128 {
129         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
130
131         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
132                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
133         } else {
134                 kfree(gl->gl_lksb.sb_lvbptr);
135                 kmem_cache_free(gfs2_glock_cachep, gl);
136         }
137 }
138
139 void gfs2_glock_free(struct gfs2_glock *gl)
140 {
141         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
142
143         BUG_ON(atomic_read(&gl->gl_revokes));
144         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
145         smp_mb();
146         wake_up_glock(gl);
147         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
148         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
149                 wake_up(&sdp->sd_glock_wait);
150 }
151
152 /**
153  * gfs2_glock_hold() - increment reference count on glock
154  * @gl: The glock to hold
155  *
156  */
157
158 void gfs2_glock_hold(struct gfs2_glock *gl)
159 {
160         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
161         lockref_get(&gl->gl_lockref);
162 }
163
164 /**
165  * demote_ok - Check to see if it's ok to unlock a glock
166  * @gl: the glock
167  *
168  * Returns: 1 if it's ok
169  */
170
171 static int demote_ok(const struct gfs2_glock *gl)
172 {
173         const struct gfs2_glock_operations *glops = gl->gl_ops;
174
175         if (gl->gl_state == LM_ST_UNLOCKED)
176                 return 0;
177         if (!list_empty(&gl->gl_holders))
178                 return 0;
179         if (glops->go_demote_ok)
180                 return glops->go_demote_ok(gl);
181         return 1;
182 }
183
184
185 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
186 {
187         if (!(gl->gl_ops->go_flags & GLOF_LRU))
188                 return;
189
190         spin_lock(&lru_lock);
191
192         list_del(&gl->gl_lru);
193         list_add_tail(&gl->gl_lru, &lru_list);
194
195         if (!test_bit(GLF_LRU, &gl->gl_flags)) {
196                 set_bit(GLF_LRU, &gl->gl_flags);
197                 atomic_inc(&lru_count);
198         }
199
200         spin_unlock(&lru_lock);
201 }
202
203 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
204 {
205         if (!(gl->gl_ops->go_flags & GLOF_LRU))
206                 return;
207
208         spin_lock(&lru_lock);
209         if (test_bit(GLF_LRU, &gl->gl_flags)) {
210                 list_del_init(&gl->gl_lru);
211                 atomic_dec(&lru_count);
212                 clear_bit(GLF_LRU, &gl->gl_flags);
213         }
214         spin_unlock(&lru_lock);
215 }
216
217 /*
218  * Enqueue the glock on the work queue.  Passes one glock reference on to the
219  * work queue.
220  */
221 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
222         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
223                 /*
224                  * We are holding the lockref spinlock, and the work was still
225                  * queued above.  The queued work (glock_work_func) takes that
226                  * spinlock before dropping its glock reference(s), so it
227                  * cannot have dropped them in the meantime.
228                  */
229                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
230                 gl->gl_lockref.count--;
231         }
232 }
233
234 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
235         spin_lock(&gl->gl_lockref.lock);
236         __gfs2_glock_queue_work(gl, delay);
237         spin_unlock(&gl->gl_lockref.lock);
238 }
239
240 static void __gfs2_glock_put(struct gfs2_glock *gl)
241 {
242         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
243         struct address_space *mapping = gfs2_glock2aspace(gl);
244
245         lockref_mark_dead(&gl->gl_lockref);
246
247         gfs2_glock_remove_from_lru(gl);
248         spin_unlock(&gl->gl_lockref.lock);
249         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
250         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
251         trace_gfs2_glock_put(gl);
252         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
253 }
254
255 /*
256  * Cause the glock to be put in work queue context.
257  */
258 void gfs2_glock_queue_put(struct gfs2_glock *gl)
259 {
260         gfs2_glock_queue_work(gl, 0);
261 }
262
263 /**
264  * gfs2_glock_put() - Decrement reference count on glock
265  * @gl: The glock to put
266  *
267  */
268
269 void gfs2_glock_put(struct gfs2_glock *gl)
270 {
271         if (lockref_put_or_lock(&gl->gl_lockref))
272                 return;
273
274         __gfs2_glock_put(gl);
275 }
276
277 /**
278  * may_grant - check if its ok to grant a new lock
279  * @gl: The glock
280  * @gh: The lock request which we wish to grant
281  *
282  * Returns: true if its ok to grant the lock
283  */
284
285 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
286 {
287         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
288         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
289              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
290                 return 0;
291         if (gl->gl_state == gh->gh_state)
292                 return 1;
293         if (gh->gh_flags & GL_EXACT)
294                 return 0;
295         if (gl->gl_state == LM_ST_EXCLUSIVE) {
296                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
297                         return 1;
298                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
299                         return 1;
300         }
301         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
302                 return 1;
303         return 0;
304 }
305
306 static void gfs2_holder_wake(struct gfs2_holder *gh)
307 {
308         clear_bit(HIF_WAIT, &gh->gh_iflags);
309         smp_mb__after_atomic();
310         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
311 }
312
313 /**
314  * do_error - Something unexpected has happened during a lock request
315  *
316  */
317
318 static void do_error(struct gfs2_glock *gl, const int ret)
319 {
320         struct gfs2_holder *gh, *tmp;
321
322         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
323                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
324                         continue;
325                 if (ret & LM_OUT_ERROR)
326                         gh->gh_error = -EIO;
327                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
328                         gh->gh_error = GLR_TRYFAILED;
329                 else
330                         continue;
331                 list_del_init(&gh->gh_list);
332                 trace_gfs2_glock_queue(gh, 0);
333                 gfs2_holder_wake(gh);
334         }
335 }
336
337 /**
338  * do_promote - promote as many requests as possible on the current queue
339  * @gl: The glock
340  * 
341  * Returns: 1 if there is a blocked holder at the head of the list, or 2
342  *          if a type specific operation is underway.
343  */
344
345 static int do_promote(struct gfs2_glock *gl)
346 __releases(&gl->gl_lockref.lock)
347 __acquires(&gl->gl_lockref.lock)
348 {
349         const struct gfs2_glock_operations *glops = gl->gl_ops;
350         struct gfs2_holder *gh, *tmp;
351         int ret;
352
353 restart:
354         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
355                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
356                         continue;
357                 if (may_grant(gl, gh)) {
358                         if (gh->gh_list.prev == &gl->gl_holders &&
359                             glops->go_lock) {
360                                 spin_unlock(&gl->gl_lockref.lock);
361                                 /* FIXME: eliminate this eventually */
362                                 ret = glops->go_lock(gh);
363                                 spin_lock(&gl->gl_lockref.lock);
364                                 if (ret) {
365                                         if (ret == 1)
366                                                 return 2;
367                                         gh->gh_error = ret;
368                                         list_del_init(&gh->gh_list);
369                                         trace_gfs2_glock_queue(gh, 0);
370                                         gfs2_holder_wake(gh);
371                                         goto restart;
372                                 }
373                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
374                                 trace_gfs2_promote(gh, 1);
375                                 gfs2_holder_wake(gh);
376                                 goto restart;
377                         }
378                         set_bit(HIF_HOLDER, &gh->gh_iflags);
379                         trace_gfs2_promote(gh, 0);
380                         gfs2_holder_wake(gh);
381                         continue;
382                 }
383                 if (gh->gh_list.prev == &gl->gl_holders)
384                         return 1;
385                 do_error(gl, 0);
386                 break;
387         }
388         return 0;
389 }
390
391 /**
392  * find_first_waiter - find the first gh that's waiting for the glock
393  * @gl: the glock
394  */
395
396 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
397 {
398         struct gfs2_holder *gh;
399
400         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
401                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
402                         return gh;
403         }
404         return NULL;
405 }
406
407 /**
408  * state_change - record that the glock is now in a different state
409  * @gl: the glock
410  * @new_state the new state
411  *
412  */
413
414 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
415 {
416         int held1, held2;
417
418         held1 = (gl->gl_state != LM_ST_UNLOCKED);
419         held2 = (new_state != LM_ST_UNLOCKED);
420
421         if (held1 != held2) {
422                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
423                 if (held2)
424                         gl->gl_lockref.count++;
425                 else
426                         gl->gl_lockref.count--;
427         }
428         if (held1 && held2 && list_empty(&gl->gl_holders))
429                 clear_bit(GLF_QUEUED, &gl->gl_flags);
430
431         if (new_state != gl->gl_target)
432                 /* shorten our minimum hold time */
433                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
434                                        GL_GLOCK_MIN_HOLD);
435         gl->gl_state = new_state;
436         gl->gl_tchange = jiffies;
437 }
438
439 static void gfs2_demote_wake(struct gfs2_glock *gl)
440 {
441         gl->gl_demote_state = LM_ST_EXCLUSIVE;
442         clear_bit(GLF_DEMOTE, &gl->gl_flags);
443         smp_mb__after_atomic();
444         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
445 }
446
447 /**
448  * finish_xmote - The DLM has replied to one of our lock requests
449  * @gl: The glock
450  * @ret: The status from the DLM
451  *
452  */
453
454 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
455 {
456         const struct gfs2_glock_operations *glops = gl->gl_ops;
457         struct gfs2_holder *gh;
458         unsigned state = ret & LM_OUT_ST_MASK;
459         int rv;
460
461         spin_lock(&gl->gl_lockref.lock);
462         trace_gfs2_glock_state_change(gl, state);
463         state_change(gl, state);
464         gh = find_first_waiter(gl);
465
466         /* Demote to UN request arrived during demote to SH or DF */
467         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
468             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
469                 gl->gl_target = LM_ST_UNLOCKED;
470
471         /* Check for state != intended state */
472         if (unlikely(state != gl->gl_target)) {
473                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
474                         /* move to back of queue and try next entry */
475                         if (ret & LM_OUT_CANCELED) {
476                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
477                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
478                                 gh = find_first_waiter(gl);
479                                 gl->gl_target = gh->gh_state;
480                                 goto retry;
481                         }
482                         /* Some error or failed "try lock" - report it */
483                         if ((ret & LM_OUT_ERROR) ||
484                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
485                                 gl->gl_target = gl->gl_state;
486                                 do_error(gl, ret);
487                                 goto out;
488                         }
489                 }
490                 switch(state) {
491                 /* Unlocked due to conversion deadlock, try again */
492                 case LM_ST_UNLOCKED:
493 retry:
494                         do_xmote(gl, gh, gl->gl_target);
495                         break;
496                 /* Conversion fails, unlock and try again */
497                 case LM_ST_SHARED:
498                 case LM_ST_DEFERRED:
499                         do_xmote(gl, gh, LM_ST_UNLOCKED);
500                         break;
501                 default: /* Everything else */
502                         pr_err("wanted %u got %u\n", gl->gl_target, state);
503                         GLOCK_BUG_ON(gl, 1);
504                 }
505                 spin_unlock(&gl->gl_lockref.lock);
506                 return;
507         }
508
509         /* Fast path - we got what we asked for */
510         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
511                 gfs2_demote_wake(gl);
512         if (state != LM_ST_UNLOCKED) {
513                 if (glops->go_xmote_bh) {
514                         spin_unlock(&gl->gl_lockref.lock);
515                         rv = glops->go_xmote_bh(gl, gh);
516                         spin_lock(&gl->gl_lockref.lock);
517                         if (rv) {
518                                 do_error(gl, rv);
519                                 goto out;
520                         }
521                 }
522                 rv = do_promote(gl);
523                 if (rv == 2)
524                         goto out_locked;
525         }
526 out:
527         clear_bit(GLF_LOCK, &gl->gl_flags);
528 out_locked:
529         spin_unlock(&gl->gl_lockref.lock);
530 }
531
532 /**
533  * do_xmote - Calls the DLM to change the state of a lock
534  * @gl: The lock state
535  * @gh: The holder (only for promotes)
536  * @target: The target lock state
537  *
538  */
539
540 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
541 __releases(&gl->gl_lockref.lock)
542 __acquires(&gl->gl_lockref.lock)
543 {
544         const struct gfs2_glock_operations *glops = gl->gl_ops;
545         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
546         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
547         int ret;
548
549         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)) &&
550             target != LM_ST_UNLOCKED)
551                 return;
552         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
553                       LM_FLAG_PRIORITY);
554         GLOCK_BUG_ON(gl, gl->gl_state == target);
555         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
556         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
557             glops->go_inval) {
558                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
559                 do_error(gl, 0); /* Fail queued try locks */
560         }
561         gl->gl_req = target;
562         set_bit(GLF_BLOCKING, &gl->gl_flags);
563         if ((gl->gl_req == LM_ST_UNLOCKED) ||
564             (gl->gl_state == LM_ST_EXCLUSIVE) ||
565             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
566                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
567         spin_unlock(&gl->gl_lockref.lock);
568         if (glops->go_sync)
569                 glops->go_sync(gl);
570         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
571                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
572         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
573
574         gfs2_glock_hold(gl);
575         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
576                 /* lock_dlm */
577                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
578                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
579                     target == LM_ST_UNLOCKED &&
580                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
581                         finish_xmote(gl, target);
582                         gfs2_glock_queue_work(gl, 0);
583                 }
584                 else if (ret) {
585                         pr_err("lm_lock ret %d\n", ret);
586                         GLOCK_BUG_ON(gl, !test_bit(SDF_SHUTDOWN,
587                                                    &sdp->sd_flags));
588                 }
589         } else { /* lock_nolock */
590                 finish_xmote(gl, target);
591                 gfs2_glock_queue_work(gl, 0);
592         }
593
594         spin_lock(&gl->gl_lockref.lock);
595 }
596
597 /**
598  * find_first_holder - find the first "holder" gh
599  * @gl: the glock
600  */
601
602 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
603 {
604         struct gfs2_holder *gh;
605
606         if (!list_empty(&gl->gl_holders)) {
607                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
608                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
609                         return gh;
610         }
611         return NULL;
612 }
613
614 /**
615  * run_queue - do all outstanding tasks related to a glock
616  * @gl: The glock in question
617  * @nonblock: True if we must not block in run_queue
618  *
619  */
620
621 static void run_queue(struct gfs2_glock *gl, const int nonblock)
622 __releases(&gl->gl_lockref.lock)
623 __acquires(&gl->gl_lockref.lock)
624 {
625         struct gfs2_holder *gh = NULL;
626         int ret;
627
628         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
629                 return;
630
631         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
632
633         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
634             gl->gl_demote_state != gl->gl_state) {
635                 if (find_first_holder(gl))
636                         goto out_unlock;
637                 if (nonblock)
638                         goto out_sched;
639                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
640                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
641                 gl->gl_target = gl->gl_demote_state;
642         } else {
643                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
644                         gfs2_demote_wake(gl);
645                 ret = do_promote(gl);
646                 if (ret == 0)
647                         goto out_unlock;
648                 if (ret == 2)
649                         goto out;
650                 gh = find_first_waiter(gl);
651                 gl->gl_target = gh->gh_state;
652                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
653                         do_error(gl, 0); /* Fail queued try locks */
654         }
655         do_xmote(gl, gh, gl->gl_target);
656 out:
657         return;
658
659 out_sched:
660         clear_bit(GLF_LOCK, &gl->gl_flags);
661         smp_mb__after_atomic();
662         gl->gl_lockref.count++;
663         __gfs2_glock_queue_work(gl, 0);
664         return;
665
666 out_unlock:
667         clear_bit(GLF_LOCK, &gl->gl_flags);
668         smp_mb__after_atomic();
669         return;
670 }
671
672 static void delete_work_func(struct work_struct *work)
673 {
674         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
675         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
676         struct inode *inode;
677         u64 no_addr = gl->gl_name.ln_number;
678
679         /* If someone's using this glock to create a new dinode, the block must
680            have been freed by another node, then re-used, in which case our
681            iopen callback is too late after the fact. Ignore it. */
682         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
683                 goto out;
684
685         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
686         if (inode && !IS_ERR(inode)) {
687                 d_prune_aliases(inode);
688                 iput(inode);
689         }
690 out:
691         gfs2_glock_put(gl);
692 }
693
694 static void glock_work_func(struct work_struct *work)
695 {
696         unsigned long delay = 0;
697         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
698         unsigned int drop_refs = 1;
699
700         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
701                 finish_xmote(gl, gl->gl_reply);
702                 drop_refs++;
703         }
704         spin_lock(&gl->gl_lockref.lock);
705         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
706             gl->gl_state != LM_ST_UNLOCKED &&
707             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
708                 unsigned long holdtime, now = jiffies;
709
710                 holdtime = gl->gl_tchange + gl->gl_hold_time;
711                 if (time_before(now, holdtime))
712                         delay = holdtime - now;
713
714                 if (!delay) {
715                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
716                         set_bit(GLF_DEMOTE, &gl->gl_flags);
717                 }
718         }
719         run_queue(gl, 0);
720         if (delay) {
721                 /* Keep one glock reference for the work we requeue. */
722                 drop_refs--;
723                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
724                         delay = 0;
725                 __gfs2_glock_queue_work(gl, delay);
726         }
727
728         /*
729          * Drop the remaining glock references manually here. (Mind that
730          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
731          * here as well.)
732          */
733         gl->gl_lockref.count -= drop_refs;
734         if (!gl->gl_lockref.count) {
735                 __gfs2_glock_put(gl);
736                 return;
737         }
738         spin_unlock(&gl->gl_lockref.lock);
739 }
740
741 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
742                                             struct gfs2_glock *new)
743 {
744         struct wait_glock_queue wait;
745         wait_queue_head_t *wq = glock_waitqueue(name);
746         struct gfs2_glock *gl;
747
748         wait.name = name;
749         init_wait(&wait.wait);
750         wait.wait.func = glock_wake_function;
751
752 again:
753         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
754         rcu_read_lock();
755         if (new) {
756                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
757                         &new->gl_node, ht_parms);
758                 if (IS_ERR(gl))
759                         goto out;
760         } else {
761                 gl = rhashtable_lookup_fast(&gl_hash_table,
762                         name, ht_parms);
763         }
764         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
765                 rcu_read_unlock();
766                 schedule();
767                 goto again;
768         }
769 out:
770         rcu_read_unlock();
771         finish_wait(wq, &wait.wait);
772         return gl;
773 }
774
775 /**
776  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
777  * @sdp: The GFS2 superblock
778  * @number: the lock number
779  * @glops: The glock_operations to use
780  * @create: If 0, don't create the glock if it doesn't exist
781  * @glp: the glock is returned here
782  *
783  * This does not lock a glock, just finds/creates structures for one.
784  *
785  * Returns: errno
786  */
787
788 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
789                    const struct gfs2_glock_operations *glops, int create,
790                    struct gfs2_glock **glp)
791 {
792         struct super_block *s = sdp->sd_vfs;
793         struct lm_lockname name = { .ln_number = number,
794                                     .ln_type = glops->go_type,
795                                     .ln_sbd = sdp };
796         struct gfs2_glock *gl, *tmp;
797         struct address_space *mapping;
798         struct kmem_cache *cachep;
799         int ret = 0;
800
801         gl = find_insert_glock(&name, NULL);
802         if (gl) {
803                 *glp = gl;
804                 return 0;
805         }
806         if (!create)
807                 return -ENOENT;
808
809         if (glops->go_flags & GLOF_ASPACE)
810                 cachep = gfs2_glock_aspace_cachep;
811         else
812                 cachep = gfs2_glock_cachep;
813         gl = kmem_cache_alloc(cachep, GFP_NOFS);
814         if (!gl)
815                 return -ENOMEM;
816
817         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
818
819         if (glops->go_flags & GLOF_LVB) {
820                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_NOFS);
821                 if (!gl->gl_lksb.sb_lvbptr) {
822                         kmem_cache_free(cachep, gl);
823                         return -ENOMEM;
824                 }
825         }
826
827         atomic_inc(&sdp->sd_glock_disposal);
828         gl->gl_node.next = NULL;
829         gl->gl_flags = 0;
830         gl->gl_name = name;
831         gl->gl_lockref.count = 1;
832         gl->gl_state = LM_ST_UNLOCKED;
833         gl->gl_target = LM_ST_UNLOCKED;
834         gl->gl_demote_state = LM_ST_EXCLUSIVE;
835         gl->gl_ops = glops;
836         gl->gl_dstamp = 0;
837         preempt_disable();
838         /* We use the global stats to estimate the initial per-glock stats */
839         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
840         preempt_enable();
841         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
842         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
843         gl->gl_tchange = jiffies;
844         gl->gl_object = NULL;
845         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
846         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
847         INIT_WORK(&gl->gl_delete, delete_work_func);
848
849         mapping = gfs2_glock2aspace(gl);
850         if (mapping) {
851                 mapping->a_ops = &gfs2_meta_aops;
852                 mapping->host = s->s_bdev->bd_inode;
853                 mapping->flags = 0;
854                 mapping_set_gfp_mask(mapping, GFP_NOFS);
855                 mapping->private_data = NULL;
856                 mapping->writeback_index = 0;
857         }
858
859         tmp = find_insert_glock(&name, gl);
860         if (!tmp) {
861                 *glp = gl;
862                 goto out;
863         }
864         if (IS_ERR(tmp)) {
865                 ret = PTR_ERR(tmp);
866                 goto out_free;
867         }
868         *glp = tmp;
869
870 out_free:
871         kfree(gl->gl_lksb.sb_lvbptr);
872         kmem_cache_free(cachep, gl);
873         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
874                 wake_up(&sdp->sd_glock_wait);
875
876 out:
877         return ret;
878 }
879
880 /**
881  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
882  * @gl: the glock
883  * @state: the state we're requesting
884  * @flags: the modifier flags
885  * @gh: the holder structure
886  *
887  */
888
889 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
890                       struct gfs2_holder *gh)
891 {
892         INIT_LIST_HEAD(&gh->gh_list);
893         gh->gh_gl = gl;
894         gh->gh_ip = _RET_IP_;
895         gh->gh_owner_pid = get_pid(task_pid(current));
896         gh->gh_state = state;
897         gh->gh_flags = flags;
898         gh->gh_error = 0;
899         gh->gh_iflags = 0;
900         gfs2_glock_hold(gl);
901 }
902
903 /**
904  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
905  * @state: the state we're requesting
906  * @flags: the modifier flags
907  * @gh: the holder structure
908  *
909  * Don't mess with the glock.
910  *
911  */
912
913 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
914 {
915         gh->gh_state = state;
916         gh->gh_flags = flags;
917         gh->gh_iflags = 0;
918         gh->gh_ip = _RET_IP_;
919         put_pid(gh->gh_owner_pid);
920         gh->gh_owner_pid = get_pid(task_pid(current));
921 }
922
923 /**
924  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
925  * @gh: the holder structure
926  *
927  */
928
929 void gfs2_holder_uninit(struct gfs2_holder *gh)
930 {
931         put_pid(gh->gh_owner_pid);
932         gfs2_glock_put(gh->gh_gl);
933         gfs2_holder_mark_uninitialized(gh);
934         gh->gh_ip = 0;
935 }
936
937 /**
938  * gfs2_glock_wait - wait on a glock acquisition
939  * @gh: the glock holder
940  *
941  * Returns: 0 on success
942  */
943
944 int gfs2_glock_wait(struct gfs2_holder *gh)
945 {
946         unsigned long time1 = jiffies;
947
948         might_sleep();
949         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
950         if (time_after(jiffies, time1 + HZ)) /* have we waited > a second? */
951                 /* Lengthen the minimum hold time. */
952                 gh->gh_gl->gl_hold_time = min(gh->gh_gl->gl_hold_time +
953                                               GL_GLOCK_HOLD_INCR,
954                                               GL_GLOCK_MAX_HOLD);
955         return gh->gh_error;
956 }
957
958 /**
959  * handle_callback - process a demote request
960  * @gl: the glock
961  * @state: the state the caller wants us to change to
962  *
963  * There are only two requests that we are going to see in actual
964  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
965  */
966
967 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
968                             unsigned long delay, bool remote)
969 {
970         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
971
972         set_bit(bit, &gl->gl_flags);
973         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
974                 gl->gl_demote_state = state;
975                 gl->gl_demote_time = jiffies;
976         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
977                         gl->gl_demote_state != state) {
978                 gl->gl_demote_state = LM_ST_UNLOCKED;
979         }
980         if (gl->gl_ops->go_callback)
981                 gl->gl_ops->go_callback(gl, remote);
982         trace_gfs2_demote_rq(gl, remote);
983 }
984
985 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
986 {
987         struct va_format vaf;
988         va_list args;
989
990         va_start(args, fmt);
991
992         if (seq) {
993                 seq_vprintf(seq, fmt, args);
994         } else {
995                 vaf.fmt = fmt;
996                 vaf.va = &args;
997
998                 pr_err("%pV", &vaf);
999         }
1000
1001         va_end(args);
1002 }
1003
1004 /**
1005  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1006  * @gh: the holder structure to add
1007  *
1008  * Eventually we should move the recursive locking trap to a
1009  * debugging option or something like that. This is the fast
1010  * path and needs to have the minimum number of distractions.
1011  * 
1012  */
1013
1014 static inline void add_to_queue(struct gfs2_holder *gh)
1015 __releases(&gl->gl_lockref.lock)
1016 __acquires(&gl->gl_lockref.lock)
1017 {
1018         struct gfs2_glock *gl = gh->gh_gl;
1019         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1020         struct list_head *insert_pt = NULL;
1021         struct gfs2_holder *gh2;
1022         int try_futile = 0;
1023
1024         BUG_ON(gh->gh_owner_pid == NULL);
1025         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1026                 BUG();
1027
1028         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1029                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1030                         try_futile = !may_grant(gl, gh);
1031                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1032                         goto fail;
1033         }
1034
1035         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1036                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1037                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1038                         goto trap_recursive;
1039                 if (try_futile &&
1040                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1041 fail:
1042                         gh->gh_error = GLR_TRYFAILED;
1043                         gfs2_holder_wake(gh);
1044                         return;
1045                 }
1046                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1047                         continue;
1048                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1049                         insert_pt = &gh2->gh_list;
1050         }
1051         set_bit(GLF_QUEUED, &gl->gl_flags);
1052         trace_gfs2_glock_queue(gh, 1);
1053         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1054         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1055         if (likely(insert_pt == NULL)) {
1056                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1057                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1058                         goto do_cancel;
1059                 return;
1060         }
1061         list_add_tail(&gh->gh_list, insert_pt);
1062 do_cancel:
1063         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1064         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1065                 spin_unlock(&gl->gl_lockref.lock);
1066                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1067                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1068                 spin_lock(&gl->gl_lockref.lock);
1069         }
1070         return;
1071
1072 trap_recursive:
1073         pr_err("original: %pSR\n", (void *)gh2->gh_ip);
1074         pr_err("pid: %d\n", pid_nr(gh2->gh_owner_pid));
1075         pr_err("lock type: %d req lock state : %d\n",
1076                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1077         pr_err("new: %pSR\n", (void *)gh->gh_ip);
1078         pr_err("pid: %d\n", pid_nr(gh->gh_owner_pid));
1079         pr_err("lock type: %d req lock state : %d\n",
1080                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1081         gfs2_dump_glock(NULL, gl);
1082         BUG();
1083 }
1084
1085 /**
1086  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1087  * @gh: the holder structure
1088  *
1089  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1090  *
1091  * Returns: 0, GLR_TRYFAILED, or errno on failure
1092  */
1093
1094 int gfs2_glock_nq(struct gfs2_holder *gh)
1095 {
1096         struct gfs2_glock *gl = gh->gh_gl;
1097         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1098         int error = 0;
1099
1100         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1101                 return -EIO;
1102
1103         if (test_bit(GLF_LRU, &gl->gl_flags))
1104                 gfs2_glock_remove_from_lru(gl);
1105
1106         spin_lock(&gl->gl_lockref.lock);
1107         add_to_queue(gh);
1108         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1109                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1110                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1111                 gl->gl_lockref.count++;
1112                 __gfs2_glock_queue_work(gl, 0);
1113         }
1114         run_queue(gl, 1);
1115         spin_unlock(&gl->gl_lockref.lock);
1116
1117         if (!(gh->gh_flags & GL_ASYNC))
1118                 error = gfs2_glock_wait(gh);
1119
1120         return error;
1121 }
1122
1123 /**
1124  * gfs2_glock_poll - poll to see if an async request has been completed
1125  * @gh: the holder
1126  *
1127  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1128  */
1129
1130 int gfs2_glock_poll(struct gfs2_holder *gh)
1131 {
1132         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1133 }
1134
1135 /**
1136  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1137  * @gh: the glock holder
1138  *
1139  */
1140
1141 void gfs2_glock_dq(struct gfs2_holder *gh)
1142 {
1143         struct gfs2_glock *gl = gh->gh_gl;
1144         const struct gfs2_glock_operations *glops = gl->gl_ops;
1145         unsigned delay = 0;
1146         int fast_path = 0;
1147
1148         spin_lock(&gl->gl_lockref.lock);
1149         if (gh->gh_flags & GL_NOCACHE)
1150                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1151
1152         list_del_init(&gh->gh_list);
1153         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1154         if (find_first_holder(gl) == NULL) {
1155                 if (glops->go_unlock) {
1156                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1157                         spin_unlock(&gl->gl_lockref.lock);
1158                         glops->go_unlock(gh);
1159                         spin_lock(&gl->gl_lockref.lock);
1160                         clear_bit(GLF_LOCK, &gl->gl_flags);
1161                 }
1162                 if (list_empty(&gl->gl_holders) &&
1163                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1164                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1165                         fast_path = 1;
1166         }
1167         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl))
1168                 gfs2_glock_add_to_lru(gl);
1169
1170         trace_gfs2_glock_queue(gh, 0);
1171         if (unlikely(!fast_path)) {
1172                 gl->gl_lockref.count++;
1173                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1174                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1175                     gl->gl_name.ln_type == LM_TYPE_INODE)
1176                         delay = gl->gl_hold_time;
1177                 __gfs2_glock_queue_work(gl, delay);
1178         }
1179         spin_unlock(&gl->gl_lockref.lock);
1180 }
1181
1182 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1183 {
1184         struct gfs2_glock *gl = gh->gh_gl;
1185         gfs2_glock_dq(gh);
1186         might_sleep();
1187         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1188 }
1189
1190 /**
1191  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1192  * @gh: the holder structure
1193  *
1194  */
1195
1196 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1197 {
1198         gfs2_glock_dq(gh);
1199         gfs2_holder_uninit(gh);
1200 }
1201
1202 /**
1203  * gfs2_glock_nq_num - acquire a glock based on lock number
1204  * @sdp: the filesystem
1205  * @number: the lock number
1206  * @glops: the glock operations for the type of glock
1207  * @state: the state to acquire the glock in
1208  * @flags: modifier flags for the acquisition
1209  * @gh: the struct gfs2_holder
1210  *
1211  * Returns: errno
1212  */
1213
1214 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1215                       const struct gfs2_glock_operations *glops,
1216                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1217 {
1218         struct gfs2_glock *gl;
1219         int error;
1220
1221         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1222         if (!error) {
1223                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1224                 gfs2_glock_put(gl);
1225         }
1226
1227         return error;
1228 }
1229
1230 /**
1231  * glock_compare - Compare two struct gfs2_glock structures for sorting
1232  * @arg_a: the first structure
1233  * @arg_b: the second structure
1234  *
1235  */
1236
1237 static int glock_compare(const void *arg_a, const void *arg_b)
1238 {
1239         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1240         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1241         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1242         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1243
1244         if (a->ln_number > b->ln_number)
1245                 return 1;
1246         if (a->ln_number < b->ln_number)
1247                 return -1;
1248         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1249         return 0;
1250 }
1251
1252 /**
1253  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1254  * @num_gh: the number of structures
1255  * @ghs: an array of struct gfs2_holder structures
1256  *
1257  * Returns: 0 on success (all glocks acquired),
1258  *          errno on failure (no glocks acquired)
1259  */
1260
1261 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1262                      struct gfs2_holder **p)
1263 {
1264         unsigned int x;
1265         int error = 0;
1266
1267         for (x = 0; x < num_gh; x++)
1268                 p[x] = &ghs[x];
1269
1270         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1271
1272         for (x = 0; x < num_gh; x++) {
1273                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1274
1275                 error = gfs2_glock_nq(p[x]);
1276                 if (error) {
1277                         while (x--)
1278                                 gfs2_glock_dq(p[x]);
1279                         break;
1280                 }
1281         }
1282
1283         return error;
1284 }
1285
1286 /**
1287  * gfs2_glock_nq_m - acquire multiple glocks
1288  * @num_gh: the number of structures
1289  * @ghs: an array of struct gfs2_holder structures
1290  *
1291  *
1292  * Returns: 0 on success (all glocks acquired),
1293  *          errno on failure (no glocks acquired)
1294  */
1295
1296 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1297 {
1298         struct gfs2_holder *tmp[4];
1299         struct gfs2_holder **pph = tmp;
1300         int error = 0;
1301
1302         switch(num_gh) {
1303         case 0:
1304                 return 0;
1305         case 1:
1306                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1307                 return gfs2_glock_nq(ghs);
1308         default:
1309                 if (num_gh <= 4)
1310                         break;
1311                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1312                                     GFP_NOFS);
1313                 if (!pph)
1314                         return -ENOMEM;
1315         }
1316
1317         error = nq_m_sync(num_gh, ghs, pph);
1318
1319         if (pph != tmp)
1320                 kfree(pph);
1321
1322         return error;
1323 }
1324
1325 /**
1326  * gfs2_glock_dq_m - release multiple glocks
1327  * @num_gh: the number of structures
1328  * @ghs: an array of struct gfs2_holder structures
1329  *
1330  */
1331
1332 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1333 {
1334         while (num_gh--)
1335                 gfs2_glock_dq(&ghs[num_gh]);
1336 }
1337
1338 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1339 {
1340         unsigned long delay = 0;
1341         unsigned long holdtime;
1342         unsigned long now = jiffies;
1343
1344         gfs2_glock_hold(gl);
1345         holdtime = gl->gl_tchange + gl->gl_hold_time;
1346         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1347             gl->gl_name.ln_type == LM_TYPE_INODE) {
1348                 if (time_before(now, holdtime))
1349                         delay = holdtime - now;
1350                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1351                         delay = gl->gl_hold_time;
1352         }
1353
1354         spin_lock(&gl->gl_lockref.lock);
1355         handle_callback(gl, state, delay, true);
1356         __gfs2_glock_queue_work(gl, delay);
1357         spin_unlock(&gl->gl_lockref.lock);
1358 }
1359
1360 /**
1361  * gfs2_should_freeze - Figure out if glock should be frozen
1362  * @gl: The glock in question
1363  *
1364  * Glocks are not frozen if (a) the result of the dlm operation is
1365  * an error, (b) the locking operation was an unlock operation or
1366  * (c) if there is a "noexp" flagged request anywhere in the queue
1367  *
1368  * Returns: 1 if freezing should occur, 0 otherwise
1369  */
1370
1371 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1372 {
1373         const struct gfs2_holder *gh;
1374
1375         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1376                 return 0;
1377         if (gl->gl_target == LM_ST_UNLOCKED)
1378                 return 0;
1379
1380         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1381                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1382                         continue;
1383                 if (LM_FLAG_NOEXP & gh->gh_flags)
1384                         return 0;
1385         }
1386
1387         return 1;
1388 }
1389
1390 /**
1391  * gfs2_glock_complete - Callback used by locking
1392  * @gl: Pointer to the glock
1393  * @ret: The return value from the dlm
1394  *
1395  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1396  * to use a bitfield shared with other glock state fields.
1397  */
1398
1399 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1400 {
1401         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1402
1403         spin_lock(&gl->gl_lockref.lock);
1404         gl->gl_reply = ret;
1405
1406         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1407                 if (gfs2_should_freeze(gl)) {
1408                         set_bit(GLF_FROZEN, &gl->gl_flags);
1409                         spin_unlock(&gl->gl_lockref.lock);
1410                         return;
1411                 }
1412         }
1413
1414         gl->gl_lockref.count++;
1415         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1416         __gfs2_glock_queue_work(gl, 0);
1417         spin_unlock(&gl->gl_lockref.lock);
1418 }
1419
1420 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1421 {
1422         struct gfs2_glock *gla, *glb;
1423
1424         gla = list_entry(a, struct gfs2_glock, gl_lru);
1425         glb = list_entry(b, struct gfs2_glock, gl_lru);
1426
1427         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1428                 return 1;
1429         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1430                 return -1;
1431
1432         return 0;
1433 }
1434
1435 /**
1436  * gfs2_dispose_glock_lru - Demote a list of glocks
1437  * @list: The list to dispose of
1438  *
1439  * Disposing of glocks may involve disk accesses, so that here we sort
1440  * the glocks by number (i.e. disk location of the inodes) so that if
1441  * there are any such accesses, they'll be sent in order (mostly).
1442  *
1443  * Must be called under the lru_lock, but may drop and retake this
1444  * lock. While the lru_lock is dropped, entries may vanish from the
1445  * list, but no new entries will appear on the list (since it is
1446  * private)
1447  */
1448
1449 static void gfs2_dispose_glock_lru(struct list_head *list)
1450 __releases(&lru_lock)
1451 __acquires(&lru_lock)
1452 {
1453         struct gfs2_glock *gl;
1454
1455         list_sort(NULL, list, glock_cmp);
1456
1457         while(!list_empty(list)) {
1458                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1459                 list_del_init(&gl->gl_lru);
1460                 clear_bit(GLF_LRU, &gl->gl_flags);
1461                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1462 add_back_to_lru:
1463                         list_add(&gl->gl_lru, &lru_list);
1464                         set_bit(GLF_LRU, &gl->gl_flags);
1465                         atomic_inc(&lru_count);
1466                         continue;
1467                 }
1468                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1469                         spin_unlock(&gl->gl_lockref.lock);
1470                         goto add_back_to_lru;
1471                 }
1472                 gl->gl_lockref.count++;
1473                 if (demote_ok(gl))
1474                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1475                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1476                 __gfs2_glock_queue_work(gl, 0);
1477                 spin_unlock(&gl->gl_lockref.lock);
1478                 cond_resched_lock(&lru_lock);
1479         }
1480 }
1481
1482 /**
1483  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1484  * @nr: The number of entries to scan
1485  *
1486  * This function selects the entries on the LRU which are able to
1487  * be demoted, and then kicks off the process by calling
1488  * gfs2_dispose_glock_lru() above.
1489  */
1490
1491 static long gfs2_scan_glock_lru(int nr)
1492 {
1493         struct gfs2_glock *gl;
1494         LIST_HEAD(skipped);
1495         LIST_HEAD(dispose);
1496         long freed = 0;
1497
1498         spin_lock(&lru_lock);
1499         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1500                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1501
1502                 /* Test for being demotable */
1503                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1504                         list_move(&gl->gl_lru, &dispose);
1505                         atomic_dec(&lru_count);
1506                         freed++;
1507                         continue;
1508                 }
1509
1510                 list_move(&gl->gl_lru, &skipped);
1511         }
1512         list_splice(&skipped, &lru_list);
1513         if (!list_empty(&dispose))
1514                 gfs2_dispose_glock_lru(&dispose);
1515         spin_unlock(&lru_lock);
1516
1517         return freed;
1518 }
1519
1520 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1521                                             struct shrink_control *sc)
1522 {
1523         if (!(sc->gfp_mask & __GFP_FS))
1524                 return SHRINK_STOP;
1525         return gfs2_scan_glock_lru(sc->nr_to_scan);
1526 }
1527
1528 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1529                                              struct shrink_control *sc)
1530 {
1531         return vfs_pressure_ratio(atomic_read(&lru_count));
1532 }
1533
1534 static struct shrinker glock_shrinker = {
1535         .seeks = DEFAULT_SEEKS,
1536         .count_objects = gfs2_glock_shrink_count,
1537         .scan_objects = gfs2_glock_shrink_scan,
1538 };
1539
1540 /**
1541  * examine_bucket - Call a function for glock in a hash bucket
1542  * @examiner: the function
1543  * @sdp: the filesystem
1544  * @bucket: the bucket
1545  *
1546  * Note that the function can be called multiple times on the same
1547  * object.  So the user must ensure that the function can cope with
1548  * that.
1549  */
1550
1551 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1552 {
1553         struct gfs2_glock *gl;
1554         struct rhashtable_iter iter;
1555
1556         rhashtable_walk_enter(&gl_hash_table, &iter);
1557
1558         do {
1559                 rhashtable_walk_start(&iter);
1560
1561                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1562                         if (gl->gl_name.ln_sbd == sdp &&
1563                             lockref_get_not_dead(&gl->gl_lockref))
1564                                 examiner(gl);
1565
1566                 rhashtable_walk_stop(&iter);
1567         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1568
1569         rhashtable_walk_exit(&iter);
1570 }
1571
1572 /**
1573  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1574  * @gl: The glock to thaw
1575  *
1576  */
1577
1578 static void thaw_glock(struct gfs2_glock *gl)
1579 {
1580         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1581                 gfs2_glock_put(gl);
1582                 return;
1583         }
1584         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1585         gfs2_glock_queue_work(gl, 0);
1586 }
1587
1588 /**
1589  * clear_glock - look at a glock and see if we can free it from glock cache
1590  * @gl: the glock to look at
1591  *
1592  */
1593
1594 static void clear_glock(struct gfs2_glock *gl)
1595 {
1596         gfs2_glock_remove_from_lru(gl);
1597
1598         spin_lock(&gl->gl_lockref.lock);
1599         if (gl->gl_state != LM_ST_UNLOCKED)
1600                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1601         __gfs2_glock_queue_work(gl, 0);
1602         spin_unlock(&gl->gl_lockref.lock);
1603 }
1604
1605 /**
1606  * gfs2_glock_thaw - Thaw any frozen glocks
1607  * @sdp: The super block
1608  *
1609  */
1610
1611 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1612 {
1613         glock_hash_walk(thaw_glock, sdp);
1614 }
1615
1616 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1617 {
1618         spin_lock(&gl->gl_lockref.lock);
1619         gfs2_dump_glock(seq, gl);
1620         spin_unlock(&gl->gl_lockref.lock);
1621 }
1622
1623 static void dump_glock_func(struct gfs2_glock *gl)
1624 {
1625         dump_glock(NULL, gl);
1626 }
1627
1628 /**
1629  * gfs2_gl_hash_clear - Empty out the glock hash table
1630  * @sdp: the filesystem
1631  * @wait: wait until it's all gone
1632  *
1633  * Called when unmounting the filesystem.
1634  */
1635
1636 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1637 {
1638         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1639         flush_workqueue(glock_workqueue);
1640         glock_hash_walk(clear_glock, sdp);
1641         flush_workqueue(glock_workqueue);
1642         wait_event_timeout(sdp->sd_glock_wait,
1643                            atomic_read(&sdp->sd_glock_disposal) == 0,
1644                            HZ * 600);
1645         glock_hash_walk(dump_glock_func, sdp);
1646 }
1647
1648 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1649 {
1650         struct gfs2_glock *gl = ip->i_gl;
1651         int ret;
1652
1653         ret = gfs2_truncatei_resume(ip);
1654         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1655
1656         spin_lock(&gl->gl_lockref.lock);
1657         clear_bit(GLF_LOCK, &gl->gl_flags);
1658         run_queue(gl, 1);
1659         spin_unlock(&gl->gl_lockref.lock);
1660 }
1661
1662 static const char *state2str(unsigned state)
1663 {
1664         switch(state) {
1665         case LM_ST_UNLOCKED:
1666                 return "UN";
1667         case LM_ST_SHARED:
1668                 return "SH";
1669         case LM_ST_DEFERRED:
1670                 return "DF";
1671         case LM_ST_EXCLUSIVE:
1672                 return "EX";
1673         }
1674         return "??";
1675 }
1676
1677 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1678 {
1679         char *p = buf;
1680         if (flags & LM_FLAG_TRY)
1681                 *p++ = 't';
1682         if (flags & LM_FLAG_TRY_1CB)
1683                 *p++ = 'T';
1684         if (flags & LM_FLAG_NOEXP)
1685                 *p++ = 'e';
1686         if (flags & LM_FLAG_ANY)
1687                 *p++ = 'A';
1688         if (flags & LM_FLAG_PRIORITY)
1689                 *p++ = 'p';
1690         if (flags & GL_ASYNC)
1691                 *p++ = 'a';
1692         if (flags & GL_EXACT)
1693                 *p++ = 'E';
1694         if (flags & GL_NOCACHE)
1695                 *p++ = 'c';
1696         if (test_bit(HIF_HOLDER, &iflags))
1697                 *p++ = 'H';
1698         if (test_bit(HIF_WAIT, &iflags))
1699                 *p++ = 'W';
1700         if (test_bit(HIF_FIRST, &iflags))
1701                 *p++ = 'F';
1702         *p = 0;
1703         return buf;
1704 }
1705
1706 /**
1707  * dump_holder - print information about a glock holder
1708  * @seq: the seq_file struct
1709  * @gh: the glock holder
1710  *
1711  */
1712
1713 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1714 {
1715         struct task_struct *gh_owner = NULL;
1716         char flags_buf[32];
1717
1718         rcu_read_lock();
1719         if (gh->gh_owner_pid)
1720                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1721         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1722                        state2str(gh->gh_state),
1723                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1724                        gh->gh_error,
1725                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1726                        gh_owner ? gh_owner->comm : "(ended)",
1727                        (void *)gh->gh_ip);
1728         rcu_read_unlock();
1729 }
1730
1731 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1732 {
1733         const unsigned long *gflags = &gl->gl_flags;
1734         char *p = buf;
1735
1736         if (test_bit(GLF_LOCK, gflags))
1737                 *p++ = 'l';
1738         if (test_bit(GLF_DEMOTE, gflags))
1739                 *p++ = 'D';
1740         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1741                 *p++ = 'd';
1742         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1743                 *p++ = 'p';
1744         if (test_bit(GLF_DIRTY, gflags))
1745                 *p++ = 'y';
1746         if (test_bit(GLF_LFLUSH, gflags))
1747                 *p++ = 'f';
1748         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1749                 *p++ = 'i';
1750         if (test_bit(GLF_REPLY_PENDING, gflags))
1751                 *p++ = 'r';
1752         if (test_bit(GLF_INITIAL, gflags))
1753                 *p++ = 'I';
1754         if (test_bit(GLF_FROZEN, gflags))
1755                 *p++ = 'F';
1756         if (test_bit(GLF_QUEUED, gflags))
1757                 *p++ = 'q';
1758         if (test_bit(GLF_LRU, gflags))
1759                 *p++ = 'L';
1760         if (gl->gl_object)
1761                 *p++ = 'o';
1762         if (test_bit(GLF_BLOCKING, gflags))
1763                 *p++ = 'b';
1764         *p = 0;
1765         return buf;
1766 }
1767
1768 /**
1769  * gfs2_dump_glock - print information about a glock
1770  * @seq: The seq_file struct
1771  * @gl: the glock
1772  *
1773  * The file format is as follows:
1774  * One line per object, capital letters are used to indicate objects
1775  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1776  * other objects are indented by a single space and follow the glock to
1777  * which they are related. Fields are indicated by lower case letters
1778  * followed by a colon and the field value, except for strings which are in
1779  * [] so that its possible to see if they are composed of spaces for
1780  * example. The field's are n = number (id of the object), f = flags,
1781  * t = type, s = state, r = refcount, e = error, p = pid.
1782  *
1783  */
1784
1785 void gfs2_dump_glock(struct seq_file *seq, const struct gfs2_glock *gl)
1786 {
1787         const struct gfs2_glock_operations *glops = gl->gl_ops;
1788         unsigned long long dtime;
1789         const struct gfs2_holder *gh;
1790         char gflags_buf[32];
1791
1792         dtime = jiffies - gl->gl_demote_time;
1793         dtime *= 1000000/HZ; /* demote time in uSec */
1794         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1795                 dtime = 0;
1796         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d v:%d r:%d m:%ld\n",
1797                   state2str(gl->gl_state),
1798                   gl->gl_name.ln_type,
1799                   (unsigned long long)gl->gl_name.ln_number,
1800                   gflags2str(gflags_buf, gl),
1801                   state2str(gl->gl_target),
1802                   state2str(gl->gl_demote_state), dtime,
1803                   atomic_read(&gl->gl_ail_count),
1804                   atomic_read(&gl->gl_revokes),
1805                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1806
1807         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1808                 dump_holder(seq, gh);
1809
1810         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1811                 glops->go_dump(seq, gl);
1812 }
1813
1814 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1815 {
1816         struct gfs2_glock *gl = iter_ptr;
1817
1818         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
1819                    gl->gl_name.ln_type,
1820                    (unsigned long long)gl->gl_name.ln_number,
1821                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1822                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1823                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1824                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1825                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1826                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1827                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1828                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1829         return 0;
1830 }
1831
1832 static const char *gfs2_gltype[] = {
1833         "type",
1834         "reserved",
1835         "nondisk",
1836         "inode",
1837         "rgrp",
1838         "meta",
1839         "iopen",
1840         "flock",
1841         "plock",
1842         "quota",
1843         "journal",
1844 };
1845
1846 static const char *gfs2_stype[] = {
1847         [GFS2_LKS_SRTT]         = "srtt",
1848         [GFS2_LKS_SRTTVAR]      = "srttvar",
1849         [GFS2_LKS_SRTTB]        = "srttb",
1850         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1851         [GFS2_LKS_SIRT]         = "sirt",
1852         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1853         [GFS2_LKS_DCOUNT]       = "dlm",
1854         [GFS2_LKS_QCOUNT]       = "queue",
1855 };
1856
1857 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1858
1859 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1860 {
1861         struct gfs2_sbd *sdp = seq->private;
1862         loff_t pos = *(loff_t *)iter_ptr;
1863         unsigned index = pos >> 3;
1864         unsigned subindex = pos & 0x07;
1865         int i;
1866
1867         if (index == 0 && subindex != 0)
1868                 return 0;
1869
1870         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1871                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1872
1873         for_each_possible_cpu(i) {
1874                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1875
1876                 if (index == 0)
1877                         seq_printf(seq, " %15u", i);
1878                 else
1879                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
1880                                    lkstats[index - 1].stats[subindex]);
1881         }
1882         seq_putc(seq, '\n');
1883         return 0;
1884 }
1885
1886 int __init gfs2_glock_init(void)
1887 {
1888         int i, ret;
1889
1890         ret = rhashtable_init(&gl_hash_table, &ht_parms);
1891         if (ret < 0)
1892                 return ret;
1893
1894         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1895                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1896         if (!glock_workqueue) {
1897                 rhashtable_destroy(&gl_hash_table);
1898                 return -ENOMEM;
1899         }
1900         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1901                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1902                                                 0);
1903         if (!gfs2_delete_workqueue) {
1904                 destroy_workqueue(glock_workqueue);
1905                 rhashtable_destroy(&gl_hash_table);
1906                 return -ENOMEM;
1907         }
1908
1909         ret = register_shrinker(&glock_shrinker);
1910         if (ret) {
1911                 destroy_workqueue(gfs2_delete_workqueue);
1912                 destroy_workqueue(glock_workqueue);
1913                 rhashtable_destroy(&gl_hash_table);
1914                 return ret;
1915         }
1916
1917         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
1918                 init_waitqueue_head(glock_wait_table + i);
1919
1920         return 0;
1921 }
1922
1923 void gfs2_glock_exit(void)
1924 {
1925         unregister_shrinker(&glock_shrinker);
1926         rhashtable_destroy(&gl_hash_table);
1927         destroy_workqueue(glock_workqueue);
1928         destroy_workqueue(gfs2_delete_workqueue);
1929 }
1930
1931 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
1932 {
1933         struct gfs2_glock *gl = gi->gl;
1934
1935         if (gl) {
1936                 if (n == 0)
1937                         return;
1938                 if (!lockref_put_not_zero(&gl->gl_lockref))
1939                         gfs2_glock_queue_put(gl);
1940         }
1941         for (;;) {
1942                 gl = rhashtable_walk_next(&gi->hti);
1943                 if (IS_ERR_OR_NULL(gl)) {
1944                         if (gl == ERR_PTR(-EAGAIN)) {
1945                                 n = 1;
1946                                 continue;
1947                         }
1948                         gl = NULL;
1949                         break;
1950                 }
1951                 if (gl->gl_name.ln_sbd != gi->sdp)
1952                         continue;
1953                 if (n <= 1) {
1954                         if (!lockref_get_not_dead(&gl->gl_lockref))
1955                                 continue;
1956                         break;
1957                 } else {
1958                         if (__lockref_is_dead(&gl->gl_lockref))
1959                                 continue;
1960                         n--;
1961                 }
1962         }
1963         gi->gl = gl;
1964 }
1965
1966 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1967         __acquires(RCU)
1968 {
1969         struct gfs2_glock_iter *gi = seq->private;
1970         loff_t n;
1971
1972         /*
1973          * We can either stay where we are, skip to the next hash table
1974          * entry, or start from the beginning.
1975          */
1976         if (*pos < gi->last_pos) {
1977                 rhashtable_walk_exit(&gi->hti);
1978                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
1979                 n = *pos + 1;
1980         } else {
1981                 n = *pos - gi->last_pos;
1982         }
1983
1984         rhashtable_walk_start(&gi->hti);
1985
1986         gfs2_glock_iter_next(gi, n);
1987         gi->last_pos = *pos;
1988         return gi->gl;
1989 }
1990
1991 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1992                                  loff_t *pos)
1993 {
1994         struct gfs2_glock_iter *gi = seq->private;
1995
1996         (*pos)++;
1997         gi->last_pos = *pos;
1998         gfs2_glock_iter_next(gi, 1);
1999         return gi->gl;
2000 }
2001
2002 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
2003         __releases(RCU)
2004 {
2005         struct gfs2_glock_iter *gi = seq->private;
2006
2007         rhashtable_walk_stop(&gi->hti);
2008 }
2009
2010 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2011 {
2012         dump_glock(seq, iter_ptr);
2013         return 0;
2014 }
2015
2016 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2017 {
2018         preempt_disable();
2019         if (*pos >= GFS2_NR_SBSTATS)
2020                 return NULL;
2021         return pos;
2022 }
2023
2024 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2025                                    loff_t *pos)
2026 {
2027         (*pos)++;
2028         if (*pos >= GFS2_NR_SBSTATS)
2029                 return NULL;
2030         return pos;
2031 }
2032
2033 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2034 {
2035         preempt_enable();
2036 }
2037
2038 static const struct seq_operations gfs2_glock_seq_ops = {
2039         .start = gfs2_glock_seq_start,
2040         .next  = gfs2_glock_seq_next,
2041         .stop  = gfs2_glock_seq_stop,
2042         .show  = gfs2_glock_seq_show,
2043 };
2044
2045 static const struct seq_operations gfs2_glstats_seq_ops = {
2046         .start = gfs2_glock_seq_start,
2047         .next  = gfs2_glock_seq_next,
2048         .stop  = gfs2_glock_seq_stop,
2049         .show  = gfs2_glstats_seq_show,
2050 };
2051
2052 static const struct seq_operations gfs2_sbstats_seq_ops = {
2053         .start = gfs2_sbstats_seq_start,
2054         .next  = gfs2_sbstats_seq_next,
2055         .stop  = gfs2_sbstats_seq_stop,
2056         .show  = gfs2_sbstats_seq_show,
2057 };
2058
2059 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2060
2061 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2062                               const struct seq_operations *ops)
2063 {
2064         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2065         if (ret == 0) {
2066                 struct seq_file *seq = file->private_data;
2067                 struct gfs2_glock_iter *gi = seq->private;
2068
2069                 gi->sdp = inode->i_private;
2070                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2071                 if (seq->buf)
2072                         seq->size = GFS2_SEQ_GOODSIZE;
2073                 /*
2074                  * Initially, we are "before" the first hash table entry; the
2075                  * first call to rhashtable_walk_next gets us the first entry.
2076                  */
2077                 gi->last_pos = -1;
2078                 gi->gl = NULL;
2079                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2080         }
2081         return ret;
2082 }
2083
2084 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2085 {
2086         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2087 }
2088
2089 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2090 {
2091         struct seq_file *seq = file->private_data;
2092         struct gfs2_glock_iter *gi = seq->private;
2093
2094         if (gi->gl)
2095                 gfs2_glock_put(gi->gl);
2096         rhashtable_walk_exit(&gi->hti);
2097         return seq_release_private(inode, file);
2098 }
2099
2100 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2101 {
2102         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2103 }
2104
2105 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2106 {
2107         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2108         if (ret == 0) {
2109                 struct seq_file *seq = file->private_data;
2110                 seq->private = inode->i_private;  /* sdp */
2111         }
2112         return ret;
2113 }
2114
2115 static const struct file_operations gfs2_glocks_fops = {
2116         .owner   = THIS_MODULE,
2117         .open    = gfs2_glocks_open,
2118         .read    = seq_read,
2119         .llseek  = seq_lseek,
2120         .release = gfs2_glocks_release,
2121 };
2122
2123 static const struct file_operations gfs2_glstats_fops = {
2124         .owner   = THIS_MODULE,
2125         .open    = gfs2_glstats_open,
2126         .read    = seq_read,
2127         .llseek  = seq_lseek,
2128         .release = gfs2_glocks_release,
2129 };
2130
2131 static const struct file_operations gfs2_sbstats_fops = {
2132         .owner   = THIS_MODULE,
2133         .open    = gfs2_sbstats_open,
2134         .read    = seq_read,
2135         .llseek  = seq_lseek,
2136         .release = seq_release,
2137 };
2138
2139 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2140 {
2141         struct dentry *dent;
2142
2143         dent = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2144         if (IS_ERR_OR_NULL(dent))
2145                 goto fail;
2146         sdp->debugfs_dir = dent;
2147
2148         dent = debugfs_create_file("glocks",
2149                                    S_IFREG | S_IRUGO,
2150                                    sdp->debugfs_dir, sdp,
2151                                    &gfs2_glocks_fops);
2152         if (IS_ERR_OR_NULL(dent))
2153                 goto fail;
2154         sdp->debugfs_dentry_glocks = dent;
2155
2156         dent = debugfs_create_file("glstats",
2157                                    S_IFREG | S_IRUGO,
2158                                    sdp->debugfs_dir, sdp,
2159                                    &gfs2_glstats_fops);
2160         if (IS_ERR_OR_NULL(dent))
2161                 goto fail;
2162         sdp->debugfs_dentry_glstats = dent;
2163
2164         dent = debugfs_create_file("sbstats",
2165                                    S_IFREG | S_IRUGO,
2166                                    sdp->debugfs_dir, sdp,
2167                                    &gfs2_sbstats_fops);
2168         if (IS_ERR_OR_NULL(dent))
2169                 goto fail;
2170         sdp->debugfs_dentry_sbstats = dent;
2171
2172         return 0;
2173 fail:
2174         gfs2_delete_debugfs_file(sdp);
2175         return dent ? PTR_ERR(dent) : -ENOMEM;
2176 }
2177
2178 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2179 {
2180         if (sdp->debugfs_dir) {
2181                 if (sdp->debugfs_dentry_glocks) {
2182                         debugfs_remove(sdp->debugfs_dentry_glocks);
2183                         sdp->debugfs_dentry_glocks = NULL;
2184                 }
2185                 if (sdp->debugfs_dentry_glstats) {
2186                         debugfs_remove(sdp->debugfs_dentry_glstats);
2187                         sdp->debugfs_dentry_glstats = NULL;
2188                 }
2189                 if (sdp->debugfs_dentry_sbstats) {
2190                         debugfs_remove(sdp->debugfs_dentry_sbstats);
2191                         sdp->debugfs_dentry_sbstats = NULL;
2192                 }
2193                 debugfs_remove(sdp->debugfs_dir);
2194                 sdp->debugfs_dir = NULL;
2195         }
2196 }
2197
2198 int gfs2_register_debugfs(void)
2199 {
2200         gfs2_root = debugfs_create_dir("gfs2", NULL);
2201         if (IS_ERR(gfs2_root))
2202                 return PTR_ERR(gfs2_root);
2203         return gfs2_root ? 0 : -ENOMEM;
2204 }
2205
2206 void gfs2_unregister_debugfs(void)
2207 {
2208         debugfs_remove(gfs2_root);
2209         gfs2_root = NULL;
2210 }