Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/shaggy...
[linux-2.6] / fs / gfs2 / glock.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2006 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/completion.h>
14 #include <linux/buffer_head.h>
15 #include <linux/delay.h>
16 #include <linux/sort.h>
17 #include <linux/jhash.h>
18 #include <linux/kallsyms.h>
19 #include <linux/gfs2_ondisk.h>
20 #include <linux/list.h>
21 #include <linux/lm_interface.h>
22 #include <linux/wait.h>
23 #include <linux/rwsem.h>
24 #include <asm/uaccess.h>
25
26 #include "gfs2.h"
27 #include "incore.h"
28 #include "glock.h"
29 #include "glops.h"
30 #include "inode.h"
31 #include "lm.h"
32 #include "lops.h"
33 #include "meta_io.h"
34 #include "quota.h"
35 #include "super.h"
36 #include "util.h"
37
38 struct gfs2_gl_hash_bucket {
39         struct hlist_head hb_list;
40 };
41
42 typedef void (*glock_examiner) (struct gfs2_glock * gl);
43
44 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp);
45 static int dump_glock(struct gfs2_glock *gl);
46 static int dump_inode(struct gfs2_inode *ip);
47 static void gfs2_glock_xmote_th(struct gfs2_holder *gh);
48 static void gfs2_glock_drop_th(struct gfs2_glock *gl);
49 static DECLARE_RWSEM(gfs2_umount_flush_sem);
50
51 #define GFS2_GL_HASH_SHIFT      15
52 #define GFS2_GL_HASH_SIZE       (1 << GFS2_GL_HASH_SHIFT)
53 #define GFS2_GL_HASH_MASK       (GFS2_GL_HASH_SIZE - 1)
54
55 static struct gfs2_gl_hash_bucket gl_hash_table[GFS2_GL_HASH_SIZE];
56
57 /*
58  * Despite what you might think, the numbers below are not arbitrary :-)
59  * They are taken from the ipv4 routing hash code, which is well tested
60  * and thus should be nearly optimal. Later on we might tweek the numbers
61  * but for now this should be fine.
62  *
63  * The reason for putting the locks in a separate array from the list heads
64  * is that we can have fewer locks than list heads and save memory. We use
65  * the same hash function for both, but with a different hash mask.
66  */
67 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) || \
68         defined(CONFIG_PROVE_LOCKING)
69
70 #ifdef CONFIG_LOCKDEP
71 # define GL_HASH_LOCK_SZ        256
72 #else
73 # if NR_CPUS >= 32
74 #  define GL_HASH_LOCK_SZ       4096
75 # elif NR_CPUS >= 16
76 #  define GL_HASH_LOCK_SZ       2048
77 # elif NR_CPUS >= 8
78 #  define GL_HASH_LOCK_SZ       1024
79 # elif NR_CPUS >= 4
80 #  define GL_HASH_LOCK_SZ       512
81 # else
82 #  define GL_HASH_LOCK_SZ       256
83 # endif
84 #endif
85
86 /* We never want more locks than chains */
87 #if GFS2_GL_HASH_SIZE < GL_HASH_LOCK_SZ
88 # undef GL_HASH_LOCK_SZ
89 # define GL_HASH_LOCK_SZ GFS2_GL_HASH_SIZE
90 #endif
91
92 static rwlock_t gl_hash_locks[GL_HASH_LOCK_SZ];
93
94 static inline rwlock_t *gl_lock_addr(unsigned int x)
95 {
96         return &gl_hash_locks[x & (GL_HASH_LOCK_SZ-1)];
97 }
98 #else /* not SMP, so no spinlocks required */
99 static inline rwlock_t *gl_lock_addr(unsigned int x)
100 {
101         return NULL;
102 }
103 #endif
104
105 /**
106  * relaxed_state_ok - is a requested lock compatible with the current lock mode?
107  * @actual: the current state of the lock
108  * @requested: the lock state that was requested by the caller
109  * @flags: the modifier flags passed in by the caller
110  *
111  * Returns: 1 if the locks are compatible, 0 otherwise
112  */
113
114 static inline int relaxed_state_ok(unsigned int actual, unsigned requested,
115                                    int flags)
116 {
117         if (actual == requested)
118                 return 1;
119
120         if (flags & GL_EXACT)
121                 return 0;
122
123         if (actual == LM_ST_EXCLUSIVE && requested == LM_ST_SHARED)
124                 return 1;
125
126         if (actual != LM_ST_UNLOCKED && (flags & LM_FLAG_ANY))
127                 return 1;
128
129         return 0;
130 }
131
132 /**
133  * gl_hash() - Turn glock number into hash bucket number
134  * @lock: The glock number
135  *
136  * Returns: The number of the corresponding hash bucket
137  */
138
139 static unsigned int gl_hash(const struct gfs2_sbd *sdp,
140                             const struct lm_lockname *name)
141 {
142         unsigned int h;
143
144         h = jhash(&name->ln_number, sizeof(u64), 0);
145         h = jhash(&name->ln_type, sizeof(unsigned int), h);
146         h = jhash(&sdp, sizeof(struct gfs2_sbd *), h);
147         h &= GFS2_GL_HASH_MASK;
148
149         return h;
150 }
151
152 /**
153  * glock_free() - Perform a few checks and then release struct gfs2_glock
154  * @gl: The glock to release
155  *
156  * Also calls lock module to release its internal structure for this glock.
157  *
158  */
159
160 static void glock_free(struct gfs2_glock *gl)
161 {
162         struct gfs2_sbd *sdp = gl->gl_sbd;
163         struct inode *aspace = gl->gl_aspace;
164
165         gfs2_lm_put_lock(sdp, gl->gl_lock);
166
167         if (aspace)
168                 gfs2_aspace_put(aspace);
169
170         kmem_cache_free(gfs2_glock_cachep, gl);
171 }
172
173 /**
174  * gfs2_glock_hold() - increment reference count on glock
175  * @gl: The glock to hold
176  *
177  */
178
179 void gfs2_glock_hold(struct gfs2_glock *gl)
180 {
181         atomic_inc(&gl->gl_ref);
182 }
183
184 /**
185  * gfs2_glock_put() - Decrement reference count on glock
186  * @gl: The glock to put
187  *
188  */
189
190 int gfs2_glock_put(struct gfs2_glock *gl)
191 {
192         int rv = 0;
193         struct gfs2_sbd *sdp = gl->gl_sbd;
194
195         write_lock(gl_lock_addr(gl->gl_hash));
196         if (atomic_dec_and_test(&gl->gl_ref)) {
197                 hlist_del(&gl->gl_list);
198                 write_unlock(gl_lock_addr(gl->gl_hash));
199                 BUG_ON(spin_is_locked(&gl->gl_spin));
200                 gfs2_assert(sdp, gl->gl_state == LM_ST_UNLOCKED);
201                 gfs2_assert(sdp, list_empty(&gl->gl_reclaim));
202                 gfs2_assert(sdp, list_empty(&gl->gl_holders));
203                 gfs2_assert(sdp, list_empty(&gl->gl_waiters1));
204                 gfs2_assert(sdp, list_empty(&gl->gl_waiters2));
205                 gfs2_assert(sdp, list_empty(&gl->gl_waiters3));
206                 glock_free(gl);
207                 rv = 1;
208                 goto out;
209         }
210         write_unlock(gl_lock_addr(gl->gl_hash));
211 out:
212         return rv;
213 }
214
215 /**
216  * search_bucket() - Find struct gfs2_glock by lock number
217  * @bucket: the bucket to search
218  * @name: The lock name
219  *
220  * Returns: NULL, or the struct gfs2_glock with the requested number
221  */
222
223 static struct gfs2_glock *search_bucket(unsigned int hash,
224                                         const struct gfs2_sbd *sdp,
225                                         const struct lm_lockname *name)
226 {
227         struct gfs2_glock *gl;
228         struct hlist_node *h;
229
230         hlist_for_each_entry(gl, h, &gl_hash_table[hash].hb_list, gl_list) {
231                 if (!lm_name_equal(&gl->gl_name, name))
232                         continue;
233                 if (gl->gl_sbd != sdp)
234                         continue;
235
236                 atomic_inc(&gl->gl_ref);
237
238                 return gl;
239         }
240
241         return NULL;
242 }
243
244 /**
245  * gfs2_glock_find() - Find glock by lock number
246  * @sdp: The GFS2 superblock
247  * @name: The lock name
248  *
249  * Returns: NULL, or the struct gfs2_glock with the requested number
250  */
251
252 static struct gfs2_glock *gfs2_glock_find(const struct gfs2_sbd *sdp,
253                                           const struct lm_lockname *name)
254 {
255         unsigned int hash = gl_hash(sdp, name);
256         struct gfs2_glock *gl;
257
258         read_lock(gl_lock_addr(hash));
259         gl = search_bucket(hash, sdp, name);
260         read_unlock(gl_lock_addr(hash));
261
262         return gl;
263 }
264
265 /**
266  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
267  * @sdp: The GFS2 superblock
268  * @number: the lock number
269  * @glops: The glock_operations to use
270  * @create: If 0, don't create the glock if it doesn't exist
271  * @glp: the glock is returned here
272  *
273  * This does not lock a glock, just finds/creates structures for one.
274  *
275  * Returns: errno
276  */
277
278 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
279                    const struct gfs2_glock_operations *glops, int create,
280                    struct gfs2_glock **glp)
281 {
282         struct lm_lockname name = { .ln_number = number, .ln_type = glops->go_type };
283         struct gfs2_glock *gl, *tmp;
284         unsigned int hash = gl_hash(sdp, &name);
285         int error;
286
287         read_lock(gl_lock_addr(hash));
288         gl = search_bucket(hash, sdp, &name);
289         read_unlock(gl_lock_addr(hash));
290
291         if (gl || !create) {
292                 *glp = gl;
293                 return 0;
294         }
295
296         gl = kmem_cache_alloc(gfs2_glock_cachep, GFP_KERNEL);
297         if (!gl)
298                 return -ENOMEM;
299
300         gl->gl_flags = 0;
301         gl->gl_name = name;
302         atomic_set(&gl->gl_ref, 1);
303         gl->gl_state = LM_ST_UNLOCKED;
304         gl->gl_hash = hash;
305         gl->gl_owner = NULL;
306         gl->gl_ip = 0;
307         gl->gl_ops = glops;
308         gl->gl_req_gh = NULL;
309         gl->gl_req_bh = NULL;
310         gl->gl_vn = 0;
311         gl->gl_stamp = jiffies;
312         gl->gl_object = NULL;
313         gl->gl_sbd = sdp;
314         gl->gl_aspace = NULL;
315         lops_init_le(&gl->gl_le, &gfs2_glock_lops);
316
317         /* If this glock protects actual on-disk data or metadata blocks,
318            create a VFS inode to manage the pages/buffers holding them. */
319         if (glops == &gfs2_inode_glops || glops == &gfs2_rgrp_glops) {
320                 gl->gl_aspace = gfs2_aspace_get(sdp);
321                 if (!gl->gl_aspace) {
322                         error = -ENOMEM;
323                         goto fail;
324                 }
325         }
326
327         error = gfs2_lm_get_lock(sdp, &name, &gl->gl_lock);
328         if (error)
329                 goto fail_aspace;
330
331         write_lock(gl_lock_addr(hash));
332         tmp = search_bucket(hash, sdp, &name);
333         if (tmp) {
334                 write_unlock(gl_lock_addr(hash));
335                 glock_free(gl);
336                 gl = tmp;
337         } else {
338                 hlist_add_head(&gl->gl_list, &gl_hash_table[hash].hb_list);
339                 write_unlock(gl_lock_addr(hash));
340         }
341
342         *glp = gl;
343
344         return 0;
345
346 fail_aspace:
347         if (gl->gl_aspace)
348                 gfs2_aspace_put(gl->gl_aspace);
349 fail:
350         kmem_cache_free(gfs2_glock_cachep, gl);
351         return error;
352 }
353
354 /**
355  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
356  * @gl: the glock
357  * @state: the state we're requesting
358  * @flags: the modifier flags
359  * @gh: the holder structure
360  *
361  */
362
363 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, unsigned flags,
364                       struct gfs2_holder *gh)
365 {
366         INIT_LIST_HEAD(&gh->gh_list);
367         gh->gh_gl = gl;
368         gh->gh_ip = (unsigned long)__builtin_return_address(0);
369         gh->gh_owner = current;
370         gh->gh_state = state;
371         gh->gh_flags = flags;
372         gh->gh_error = 0;
373         gh->gh_iflags = 0;
374         gfs2_glock_hold(gl);
375 }
376
377 /**
378  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
379  * @state: the state we're requesting
380  * @flags: the modifier flags
381  * @gh: the holder structure
382  *
383  * Don't mess with the glock.
384  *
385  */
386
387 void gfs2_holder_reinit(unsigned int state, unsigned flags, struct gfs2_holder *gh)
388 {
389         gh->gh_state = state;
390         gh->gh_flags = flags;
391         gh->gh_iflags &= 1 << HIF_ALLOCED;
392         gh->gh_ip = (unsigned long)__builtin_return_address(0);
393 }
394
395 /**
396  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
397  * @gh: the holder structure
398  *
399  */
400
401 void gfs2_holder_uninit(struct gfs2_holder *gh)
402 {
403         gfs2_glock_put(gh->gh_gl);
404         gh->gh_gl = NULL;
405         gh->gh_ip = 0;
406 }
407
408 /**
409  * gfs2_holder_get - get a struct gfs2_holder structure
410  * @gl: the glock
411  * @state: the state we're requesting
412  * @flags: the modifier flags
413  * @gfp_flags:
414  *
415  * Figure out how big an impact this function has.  Either:
416  * 1) Replace it with a cache of structures hanging off the struct gfs2_sbd
417  * 2) Leave it like it is
418  *
419  * Returns: the holder structure, NULL on ENOMEM
420  */
421
422 static struct gfs2_holder *gfs2_holder_get(struct gfs2_glock *gl,
423                                            unsigned int state,
424                                            int flags, gfp_t gfp_flags)
425 {
426         struct gfs2_holder *gh;
427
428         gh = kmalloc(sizeof(struct gfs2_holder), gfp_flags);
429         if (!gh)
430                 return NULL;
431
432         gfs2_holder_init(gl, state, flags, gh);
433         set_bit(HIF_ALLOCED, &gh->gh_iflags);
434         gh->gh_ip = (unsigned long)__builtin_return_address(0);
435         return gh;
436 }
437
438 /**
439  * gfs2_holder_put - get rid of a struct gfs2_holder structure
440  * @gh: the holder structure
441  *
442  */
443
444 static void gfs2_holder_put(struct gfs2_holder *gh)
445 {
446         gfs2_holder_uninit(gh);
447         kfree(gh);
448 }
449
450 static void gfs2_holder_dispose_or_wake(struct gfs2_holder *gh)
451 {
452         if (test_bit(HIF_DEALLOC, &gh->gh_iflags)) {
453                 gfs2_holder_put(gh);
454                 return;
455         }
456         clear_bit(HIF_WAIT, &gh->gh_iflags);
457         smp_mb();
458         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
459 }
460
461 static int holder_wait(void *word)
462 {
463         schedule();
464         return 0;
465 }
466
467 static void wait_on_holder(struct gfs2_holder *gh)
468 {
469         might_sleep();
470         wait_on_bit(&gh->gh_iflags, HIF_WAIT, holder_wait, TASK_UNINTERRUPTIBLE);
471 }
472
473 /**
474  * rq_mutex - process a mutex request in the queue
475  * @gh: the glock holder
476  *
477  * Returns: 1 if the queue is blocked
478  */
479
480 static int rq_mutex(struct gfs2_holder *gh)
481 {
482         struct gfs2_glock *gl = gh->gh_gl;
483
484         list_del_init(&gh->gh_list);
485         /*  gh->gh_error never examined.  */
486         set_bit(GLF_LOCK, &gl->gl_flags);
487         clear_bit(HIF_WAIT, &gh->gh_iflags);
488         smp_mb();
489         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
490
491         return 1;
492 }
493
494 /**
495  * rq_promote - process a promote request in the queue
496  * @gh: the glock holder
497  *
498  * Acquire a new inter-node lock, or change a lock state to more restrictive.
499  *
500  * Returns: 1 if the queue is blocked
501  */
502
503 static int rq_promote(struct gfs2_holder *gh)
504 {
505         struct gfs2_glock *gl = gh->gh_gl;
506         struct gfs2_sbd *sdp = gl->gl_sbd;
507
508         if (!relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
509                 if (list_empty(&gl->gl_holders)) {
510                         gl->gl_req_gh = gh;
511                         set_bit(GLF_LOCK, &gl->gl_flags);
512                         spin_unlock(&gl->gl_spin);
513
514                         if (atomic_read(&sdp->sd_reclaim_count) >
515                             gfs2_tune_get(sdp, gt_reclaim_limit) &&
516                             !(gh->gh_flags & LM_FLAG_PRIORITY)) {
517                                 gfs2_reclaim_glock(sdp);
518                                 gfs2_reclaim_glock(sdp);
519                         }
520
521                         gfs2_glock_xmote_th(gh);
522                         spin_lock(&gl->gl_spin);
523                 }
524                 return 1;
525         }
526
527         if (list_empty(&gl->gl_holders)) {
528                 set_bit(HIF_FIRST, &gh->gh_iflags);
529                 set_bit(GLF_LOCK, &gl->gl_flags);
530         } else {
531                 struct gfs2_holder *next_gh;
532                 if (gh->gh_state == LM_ST_EXCLUSIVE)
533                         return 1;
534                 next_gh = list_entry(gl->gl_holders.next, struct gfs2_holder,
535                                      gh_list);
536                 if (next_gh->gh_state == LM_ST_EXCLUSIVE)
537                          return 1;
538         }
539
540         list_move_tail(&gh->gh_list, &gl->gl_holders);
541         gh->gh_error = 0;
542         set_bit(HIF_HOLDER, &gh->gh_iflags);
543
544         gfs2_holder_dispose_or_wake(gh);
545
546         return 0;
547 }
548
549 /**
550  * rq_demote - process a demote request in the queue
551  * @gh: the glock holder
552  *
553  * Returns: 1 if the queue is blocked
554  */
555
556 static int rq_demote(struct gfs2_holder *gh)
557 {
558         struct gfs2_glock *gl = gh->gh_gl;
559
560         if (!list_empty(&gl->gl_holders))
561                 return 1;
562
563         if (gl->gl_state == gh->gh_state || gl->gl_state == LM_ST_UNLOCKED) {
564                 list_del_init(&gh->gh_list);
565                 gh->gh_error = 0;
566                 spin_unlock(&gl->gl_spin);
567                 gfs2_holder_dispose_or_wake(gh);
568                 spin_lock(&gl->gl_spin);
569         } else {
570                 gl->gl_req_gh = gh;
571                 set_bit(GLF_LOCK, &gl->gl_flags);
572                 spin_unlock(&gl->gl_spin);
573
574                 if (gh->gh_state == LM_ST_UNLOCKED ||
575                     gl->gl_state != LM_ST_EXCLUSIVE)
576                         gfs2_glock_drop_th(gl);
577                 else
578                         gfs2_glock_xmote_th(gh);
579
580                 spin_lock(&gl->gl_spin);
581         }
582
583         return 0;
584 }
585
586 /**
587  * run_queue - process holder structures on a glock
588  * @gl: the glock
589  *
590  */
591 static void run_queue(struct gfs2_glock *gl)
592 {
593         struct gfs2_holder *gh;
594         int blocked = 1;
595
596         for (;;) {
597                 if (test_bit(GLF_LOCK, &gl->gl_flags))
598                         break;
599
600                 if (!list_empty(&gl->gl_waiters1)) {
601                         gh = list_entry(gl->gl_waiters1.next,
602                                         struct gfs2_holder, gh_list);
603
604                         if (test_bit(HIF_MUTEX, &gh->gh_iflags))
605                                 blocked = rq_mutex(gh);
606                         else
607                                 gfs2_assert_warn(gl->gl_sbd, 0);
608
609                 } else if (!list_empty(&gl->gl_waiters2) &&
610                            !test_bit(GLF_SKIP_WAITERS2, &gl->gl_flags)) {
611                         gh = list_entry(gl->gl_waiters2.next,
612                                         struct gfs2_holder, gh_list);
613
614                         if (test_bit(HIF_DEMOTE, &gh->gh_iflags))
615                                 blocked = rq_demote(gh);
616                         else
617                                 gfs2_assert_warn(gl->gl_sbd, 0);
618
619                 } else if (!list_empty(&gl->gl_waiters3)) {
620                         gh = list_entry(gl->gl_waiters3.next,
621                                         struct gfs2_holder, gh_list);
622
623                         if (test_bit(HIF_PROMOTE, &gh->gh_iflags))
624                                 blocked = rq_promote(gh);
625                         else
626                                 gfs2_assert_warn(gl->gl_sbd, 0);
627
628                 } else
629                         break;
630
631                 if (blocked)
632                         break;
633         }
634 }
635
636 /**
637  * gfs2_glmutex_lock - acquire a local lock on a glock
638  * @gl: the glock
639  *
640  * Gives caller exclusive access to manipulate a glock structure.
641  */
642
643 static void gfs2_glmutex_lock(struct gfs2_glock *gl)
644 {
645         struct gfs2_holder gh;
646
647         gfs2_holder_init(gl, 0, 0, &gh);
648         set_bit(HIF_MUTEX, &gh.gh_iflags);
649         if (test_and_set_bit(HIF_WAIT, &gh.gh_iflags))
650                 BUG();
651
652         spin_lock(&gl->gl_spin);
653         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
654                 list_add_tail(&gh.gh_list, &gl->gl_waiters1);
655         } else {
656                 gl->gl_owner = current;
657                 gl->gl_ip = (unsigned long)__builtin_return_address(0);
658                 clear_bit(HIF_WAIT, &gh.gh_iflags);
659                 smp_mb();
660                 wake_up_bit(&gh.gh_iflags, HIF_WAIT);
661         }
662         spin_unlock(&gl->gl_spin);
663
664         wait_on_holder(&gh);
665         gfs2_holder_uninit(&gh);
666 }
667
668 /**
669  * gfs2_glmutex_trylock - try to acquire a local lock on a glock
670  * @gl: the glock
671  *
672  * Returns: 1 if the glock is acquired
673  */
674
675 static int gfs2_glmutex_trylock(struct gfs2_glock *gl)
676 {
677         int acquired = 1;
678
679         spin_lock(&gl->gl_spin);
680         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
681                 acquired = 0;
682         } else {
683                 gl->gl_owner = current;
684                 gl->gl_ip = (unsigned long)__builtin_return_address(0);
685         }
686         spin_unlock(&gl->gl_spin);
687
688         return acquired;
689 }
690
691 /**
692  * gfs2_glmutex_unlock - release a local lock on a glock
693  * @gl: the glock
694  *
695  */
696
697 static void gfs2_glmutex_unlock(struct gfs2_glock *gl)
698 {
699         spin_lock(&gl->gl_spin);
700         clear_bit(GLF_LOCK, &gl->gl_flags);
701         gl->gl_owner = NULL;
702         gl->gl_ip = 0;
703         run_queue(gl);
704         BUG_ON(!spin_is_locked(&gl->gl_spin));
705         spin_unlock(&gl->gl_spin);
706 }
707
708 /**
709  * handle_callback - add a demote request to a lock's queue
710  * @gl: the glock
711  * @state: the state the caller wants us to change to
712  *
713  * Note: This may fail sliently if we are out of memory.
714  */
715
716 static void handle_callback(struct gfs2_glock *gl, unsigned int state)
717 {
718         struct gfs2_holder *gh, *new_gh = NULL;
719
720 restart:
721         spin_lock(&gl->gl_spin);
722
723         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
724                 if (test_bit(HIF_DEMOTE, &gh->gh_iflags) &&
725                     gl->gl_req_gh != gh) {
726                         if (gh->gh_state != state)
727                                 gh->gh_state = LM_ST_UNLOCKED;
728                         goto out;
729                 }
730         }
731
732         if (new_gh) {
733                 list_add_tail(&new_gh->gh_list, &gl->gl_waiters2);
734                 new_gh = NULL;
735         } else {
736                 spin_unlock(&gl->gl_spin);
737
738                 new_gh = gfs2_holder_get(gl, state, LM_FLAG_TRY, GFP_NOFS);
739                 if (!new_gh)
740                         return;
741                 set_bit(HIF_DEMOTE, &new_gh->gh_iflags);
742                 set_bit(HIF_DEALLOC, &new_gh->gh_iflags);
743                 set_bit(HIF_WAIT, &new_gh->gh_iflags);
744
745                 goto restart;
746         }
747
748 out:
749         spin_unlock(&gl->gl_spin);
750
751         if (new_gh)
752                 gfs2_holder_put(new_gh);
753 }
754
755 /**
756  * state_change - record that the glock is now in a different state
757  * @gl: the glock
758  * @new_state the new state
759  *
760  */
761
762 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
763 {
764         int held1, held2;
765
766         held1 = (gl->gl_state != LM_ST_UNLOCKED);
767         held2 = (new_state != LM_ST_UNLOCKED);
768
769         if (held1 != held2) {
770                 if (held2)
771                         gfs2_glock_hold(gl);
772                 else
773                         gfs2_glock_put(gl);
774         }
775
776         gl->gl_state = new_state;
777 }
778
779 /**
780  * xmote_bh - Called after the lock module is done acquiring a lock
781  * @gl: The glock in question
782  * @ret: the int returned from the lock module
783  *
784  */
785
786 static void xmote_bh(struct gfs2_glock *gl, unsigned int ret)
787 {
788         struct gfs2_sbd *sdp = gl->gl_sbd;
789         const struct gfs2_glock_operations *glops = gl->gl_ops;
790         struct gfs2_holder *gh = gl->gl_req_gh;
791         int prev_state = gl->gl_state;
792         int op_done = 1;
793
794         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
795         gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
796         gfs2_assert_warn(sdp, !(ret & LM_OUT_ASYNC));
797
798         state_change(gl, ret & LM_OUT_ST_MASK);
799
800         if (prev_state != LM_ST_UNLOCKED && !(ret & LM_OUT_CACHEABLE)) {
801                 if (glops->go_inval)
802                         glops->go_inval(gl, DIO_METADATA);
803         } else if (gl->gl_state == LM_ST_DEFERRED) {
804                 /* We might not want to do this here.
805                    Look at moving to the inode glops. */
806                 if (glops->go_inval)
807                         glops->go_inval(gl, 0);
808         }
809
810         /*  Deal with each possible exit condition  */
811
812         if (!gh)
813                 gl->gl_stamp = jiffies;
814         else if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
815                 spin_lock(&gl->gl_spin);
816                 list_del_init(&gh->gh_list);
817                 gh->gh_error = -EIO;
818                 spin_unlock(&gl->gl_spin);
819         } else if (test_bit(HIF_DEMOTE, &gh->gh_iflags)) {
820                 spin_lock(&gl->gl_spin);
821                 list_del_init(&gh->gh_list);
822                 if (gl->gl_state == gh->gh_state ||
823                     gl->gl_state == LM_ST_UNLOCKED) {
824                         gh->gh_error = 0;
825                 } else {
826                         if (gfs2_assert_warn(sdp, gh->gh_flags &
827                                         (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) == -1)
828                                 fs_warn(sdp, "ret = 0x%.8X\n", ret);
829                         gh->gh_error = GLR_TRYFAILED;
830                 }
831                 spin_unlock(&gl->gl_spin);
832
833                 if (ret & LM_OUT_CANCELED)
834                         handle_callback(gl, LM_ST_UNLOCKED);
835
836         } else if (ret & LM_OUT_CANCELED) {
837                 spin_lock(&gl->gl_spin);
838                 list_del_init(&gh->gh_list);
839                 gh->gh_error = GLR_CANCELED;
840                 spin_unlock(&gl->gl_spin);
841
842         } else if (relaxed_state_ok(gl->gl_state, gh->gh_state, gh->gh_flags)) {
843                 spin_lock(&gl->gl_spin);
844                 list_move_tail(&gh->gh_list, &gl->gl_holders);
845                 gh->gh_error = 0;
846                 set_bit(HIF_HOLDER, &gh->gh_iflags);
847                 spin_unlock(&gl->gl_spin);
848
849                 set_bit(HIF_FIRST, &gh->gh_iflags);
850
851                 op_done = 0;
852
853         } else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
854                 spin_lock(&gl->gl_spin);
855                 list_del_init(&gh->gh_list);
856                 gh->gh_error = GLR_TRYFAILED;
857                 spin_unlock(&gl->gl_spin);
858
859         } else {
860                 if (gfs2_assert_withdraw(sdp, 0) == -1)
861                         fs_err(sdp, "ret = 0x%.8X\n", ret);
862         }
863
864         if (glops->go_xmote_bh)
865                 glops->go_xmote_bh(gl);
866
867         if (op_done) {
868                 spin_lock(&gl->gl_spin);
869                 gl->gl_req_gh = NULL;
870                 gl->gl_req_bh = NULL;
871                 clear_bit(GLF_LOCK, &gl->gl_flags);
872                 run_queue(gl);
873                 spin_unlock(&gl->gl_spin);
874         }
875
876         gfs2_glock_put(gl);
877
878         if (gh)
879                 gfs2_holder_dispose_or_wake(gh);
880 }
881
882 /**
883  * gfs2_glock_xmote_th - Call into the lock module to acquire or change a glock
884  * @gl: The glock in question
885  * @state: the requested state
886  * @flags: modifier flags to the lock call
887  *
888  */
889
890 void gfs2_glock_xmote_th(struct gfs2_holder *gh)
891 {
892         struct gfs2_glock *gl = gh->gh_gl;
893         struct gfs2_sbd *sdp = gl->gl_sbd;
894         int flags = gh->gh_flags;
895         unsigned state = gh->gh_state;
896         const struct gfs2_glock_operations *glops = gl->gl_ops;
897         int lck_flags = flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB |
898                                  LM_FLAG_NOEXP | LM_FLAG_ANY |
899                                  LM_FLAG_PRIORITY);
900         unsigned int lck_ret;
901
902         if (glops->go_xmote_th)
903                 glops->go_xmote_th(gl);
904
905         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
906         gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
907         gfs2_assert_warn(sdp, state != LM_ST_UNLOCKED);
908         gfs2_assert_warn(sdp, state != gl->gl_state);
909
910         gfs2_glock_hold(gl);
911         gl->gl_req_bh = xmote_bh;
912
913         lck_ret = gfs2_lm_lock(sdp, gl->gl_lock, gl->gl_state, state, lck_flags);
914
915         if (gfs2_assert_withdraw(sdp, !(lck_ret & LM_OUT_ERROR)))
916                 return;
917
918         if (lck_ret & LM_OUT_ASYNC)
919                 gfs2_assert_warn(sdp, lck_ret == LM_OUT_ASYNC);
920         else
921                 xmote_bh(gl, lck_ret);
922 }
923
924 /**
925  * drop_bh - Called after a lock module unlock completes
926  * @gl: the glock
927  * @ret: the return status
928  *
929  * Doesn't wake up the process waiting on the struct gfs2_holder (if any)
930  * Doesn't drop the reference on the glock the top half took out
931  *
932  */
933
934 static void drop_bh(struct gfs2_glock *gl, unsigned int ret)
935 {
936         struct gfs2_sbd *sdp = gl->gl_sbd;
937         const struct gfs2_glock_operations *glops = gl->gl_ops;
938         struct gfs2_holder *gh = gl->gl_req_gh;
939
940         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
941         gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
942         gfs2_assert_warn(sdp, !ret);
943
944         state_change(gl, LM_ST_UNLOCKED);
945
946         if (glops->go_inval)
947                 glops->go_inval(gl, DIO_METADATA);
948
949         if (gh) {
950                 spin_lock(&gl->gl_spin);
951                 list_del_init(&gh->gh_list);
952                 gh->gh_error = 0;
953                 spin_unlock(&gl->gl_spin);
954         }
955
956         if (glops->go_drop_bh)
957                 glops->go_drop_bh(gl);
958
959         spin_lock(&gl->gl_spin);
960         gl->gl_req_gh = NULL;
961         gl->gl_req_bh = NULL;
962         clear_bit(GLF_LOCK, &gl->gl_flags);
963         run_queue(gl);
964         spin_unlock(&gl->gl_spin);
965
966         gfs2_glock_put(gl);
967
968         if (gh)
969                 gfs2_holder_dispose_or_wake(gh);
970 }
971
972 /**
973  * gfs2_glock_drop_th - call into the lock module to unlock a lock
974  * @gl: the glock
975  *
976  */
977
978 static void gfs2_glock_drop_th(struct gfs2_glock *gl)
979 {
980         struct gfs2_sbd *sdp = gl->gl_sbd;
981         const struct gfs2_glock_operations *glops = gl->gl_ops;
982         unsigned int ret;
983
984         if (glops->go_drop_th)
985                 glops->go_drop_th(gl);
986
987         gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
988         gfs2_assert_warn(sdp, list_empty(&gl->gl_holders));
989         gfs2_assert_warn(sdp, gl->gl_state != LM_ST_UNLOCKED);
990
991         gfs2_glock_hold(gl);
992         gl->gl_req_bh = drop_bh;
993
994         ret = gfs2_lm_unlock(sdp, gl->gl_lock, gl->gl_state);
995
996         if (gfs2_assert_withdraw(sdp, !(ret & LM_OUT_ERROR)))
997                 return;
998
999         if (!ret)
1000                 drop_bh(gl, ret);
1001         else
1002                 gfs2_assert_warn(sdp, ret == LM_OUT_ASYNC);
1003 }
1004
1005 /**
1006  * do_cancels - cancel requests for locks stuck waiting on an expire flag
1007  * @gh: the LM_FLAG_PRIORITY holder waiting to acquire the lock
1008  *
1009  * Don't cancel GL_NOCANCEL requests.
1010  */
1011
1012 static void do_cancels(struct gfs2_holder *gh)
1013 {
1014         struct gfs2_glock *gl = gh->gh_gl;
1015
1016         spin_lock(&gl->gl_spin);
1017
1018         while (gl->gl_req_gh != gh &&
1019                !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1020                !list_empty(&gh->gh_list)) {
1021                 if (gl->gl_req_bh && !(gl->gl_req_gh &&
1022                                      (gl->gl_req_gh->gh_flags & GL_NOCANCEL))) {
1023                         spin_unlock(&gl->gl_spin);
1024                         gfs2_lm_cancel(gl->gl_sbd, gl->gl_lock);
1025                         msleep(100);
1026                         spin_lock(&gl->gl_spin);
1027                 } else {
1028                         spin_unlock(&gl->gl_spin);
1029                         msleep(100);
1030                         spin_lock(&gl->gl_spin);
1031                 }
1032         }
1033
1034         spin_unlock(&gl->gl_spin);
1035 }
1036
1037 /**
1038  * glock_wait_internal - wait on a glock acquisition
1039  * @gh: the glock holder
1040  *
1041  * Returns: 0 on success
1042  */
1043
1044 static int glock_wait_internal(struct gfs2_holder *gh)
1045 {
1046         struct gfs2_glock *gl = gh->gh_gl;
1047         struct gfs2_sbd *sdp = gl->gl_sbd;
1048         const struct gfs2_glock_operations *glops = gl->gl_ops;
1049
1050         if (test_bit(HIF_ABORTED, &gh->gh_iflags))
1051                 return -EIO;
1052
1053         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1054                 spin_lock(&gl->gl_spin);
1055                 if (gl->gl_req_gh != gh &&
1056                     !test_bit(HIF_HOLDER, &gh->gh_iflags) &&
1057                     !list_empty(&gh->gh_list)) {
1058                         list_del_init(&gh->gh_list);
1059                         gh->gh_error = GLR_TRYFAILED;
1060                         run_queue(gl);
1061                         spin_unlock(&gl->gl_spin);
1062                         return gh->gh_error;
1063                 }
1064                 spin_unlock(&gl->gl_spin);
1065         }
1066
1067         if (gh->gh_flags & LM_FLAG_PRIORITY)
1068                 do_cancels(gh);
1069
1070         wait_on_holder(gh);
1071         if (gh->gh_error)
1072                 return gh->gh_error;
1073
1074         gfs2_assert_withdraw(sdp, test_bit(HIF_HOLDER, &gh->gh_iflags));
1075         gfs2_assert_withdraw(sdp, relaxed_state_ok(gl->gl_state, gh->gh_state,
1076                                                    gh->gh_flags));
1077
1078         if (test_bit(HIF_FIRST, &gh->gh_iflags)) {
1079                 gfs2_assert_warn(sdp, test_bit(GLF_LOCK, &gl->gl_flags));
1080
1081                 if (glops->go_lock) {
1082                         gh->gh_error = glops->go_lock(gh);
1083                         if (gh->gh_error) {
1084                                 spin_lock(&gl->gl_spin);
1085                                 list_del_init(&gh->gh_list);
1086                                 spin_unlock(&gl->gl_spin);
1087                         }
1088                 }
1089
1090                 spin_lock(&gl->gl_spin);
1091                 gl->gl_req_gh = NULL;
1092                 gl->gl_req_bh = NULL;
1093                 clear_bit(GLF_LOCK, &gl->gl_flags);
1094                 run_queue(gl);
1095                 spin_unlock(&gl->gl_spin);
1096         }
1097
1098         return gh->gh_error;
1099 }
1100
1101 static inline struct gfs2_holder *
1102 find_holder_by_owner(struct list_head *head, struct task_struct *owner)
1103 {
1104         struct gfs2_holder *gh;
1105
1106         list_for_each_entry(gh, head, gh_list) {
1107                 if (gh->gh_owner == owner)
1108                         return gh;
1109         }
1110
1111         return NULL;
1112 }
1113
1114 /**
1115  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1116  * @gh: the holder structure to add
1117  *
1118  */
1119
1120 static void add_to_queue(struct gfs2_holder *gh)
1121 {
1122         struct gfs2_glock *gl = gh->gh_gl;
1123         struct gfs2_holder *existing;
1124
1125         BUG_ON(!gh->gh_owner);
1126         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1127                 BUG();
1128
1129         existing = find_holder_by_owner(&gl->gl_holders, gh->gh_owner);
1130         if (existing) {
1131                 print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
1132                 printk(KERN_INFO "pid : %d\n", existing->gh_owner->pid);
1133                 printk(KERN_INFO "lock type : %d lock state : %d\n",
1134                                 existing->gh_gl->gl_name.ln_type, existing->gh_gl->gl_state);
1135                 print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
1136                 printk(KERN_INFO "pid : %d\n", gh->gh_owner->pid);
1137                 printk(KERN_INFO "lock type : %d lock state : %d\n",
1138                                 gl->gl_name.ln_type, gl->gl_state);
1139                 BUG();
1140         }
1141
1142         existing = find_holder_by_owner(&gl->gl_waiters3, gh->gh_owner);
1143         if (existing) {
1144                 print_symbol(KERN_WARNING "original: %s\n", existing->gh_ip);
1145                 print_symbol(KERN_WARNING "new: %s\n", gh->gh_ip);
1146                 BUG();
1147         }
1148
1149         if (gh->gh_flags & LM_FLAG_PRIORITY)
1150                 list_add(&gh->gh_list, &gl->gl_waiters3);
1151         else
1152                 list_add_tail(&gh->gh_list, &gl->gl_waiters3);
1153 }
1154
1155 /**
1156  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1157  * @gh: the holder structure
1158  *
1159  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1160  *
1161  * Returns: 0, GLR_TRYFAILED, or errno on failure
1162  */
1163
1164 int gfs2_glock_nq(struct gfs2_holder *gh)
1165 {
1166         struct gfs2_glock *gl = gh->gh_gl;
1167         struct gfs2_sbd *sdp = gl->gl_sbd;
1168         int error = 0;
1169
1170 restart:
1171         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags))) {
1172                 set_bit(HIF_ABORTED, &gh->gh_iflags);
1173                 return -EIO;
1174         }
1175
1176         set_bit(HIF_PROMOTE, &gh->gh_iflags);
1177
1178         spin_lock(&gl->gl_spin);
1179         add_to_queue(gh);
1180         run_queue(gl);
1181         spin_unlock(&gl->gl_spin);
1182
1183         if (!(gh->gh_flags & GL_ASYNC)) {
1184                 error = glock_wait_internal(gh);
1185                 if (error == GLR_CANCELED) {
1186                         msleep(100);
1187                         goto restart;
1188                 }
1189         }
1190
1191         return error;
1192 }
1193
1194 /**
1195  * gfs2_glock_poll - poll to see if an async request has been completed
1196  * @gh: the holder
1197  *
1198  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1199  */
1200
1201 int gfs2_glock_poll(struct gfs2_holder *gh)
1202 {
1203         struct gfs2_glock *gl = gh->gh_gl;
1204         int ready = 0;
1205
1206         spin_lock(&gl->gl_spin);
1207
1208         if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1209                 ready = 1;
1210         else if (list_empty(&gh->gh_list)) {
1211                 if (gh->gh_error == GLR_CANCELED) {
1212                         spin_unlock(&gl->gl_spin);
1213                         msleep(100);
1214                         if (gfs2_glock_nq(gh))
1215                                 return 1;
1216                         return 0;
1217                 } else
1218                         ready = 1;
1219         }
1220
1221         spin_unlock(&gl->gl_spin);
1222
1223         return ready;
1224 }
1225
1226 /**
1227  * gfs2_glock_wait - wait for a lock acquisition that ended in a GLR_ASYNC
1228  * @gh: the holder structure
1229  *
1230  * Returns: 0, GLR_TRYFAILED, or errno on failure
1231  */
1232
1233 int gfs2_glock_wait(struct gfs2_holder *gh)
1234 {
1235         int error;
1236
1237         error = glock_wait_internal(gh);
1238         if (error == GLR_CANCELED) {
1239                 msleep(100);
1240                 gh->gh_flags &= ~GL_ASYNC;
1241                 error = gfs2_glock_nq(gh);
1242         }
1243
1244         return error;
1245 }
1246
1247 /**
1248  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1249  * @gh: the glock holder
1250  *
1251  */
1252
1253 void gfs2_glock_dq(struct gfs2_holder *gh)
1254 {
1255         struct gfs2_glock *gl = gh->gh_gl;
1256         const struct gfs2_glock_operations *glops = gl->gl_ops;
1257
1258         if (gh->gh_flags & GL_NOCACHE)
1259                 handle_callback(gl, LM_ST_UNLOCKED);
1260
1261         gfs2_glmutex_lock(gl);
1262
1263         spin_lock(&gl->gl_spin);
1264         list_del_init(&gh->gh_list);
1265
1266         if (list_empty(&gl->gl_holders)) {
1267                 spin_unlock(&gl->gl_spin);
1268
1269                 if (glops->go_unlock)
1270                         glops->go_unlock(gh);
1271
1272                 gl->gl_stamp = jiffies;
1273
1274                 spin_lock(&gl->gl_spin);
1275         }
1276
1277         clear_bit(GLF_LOCK, &gl->gl_flags);
1278         run_queue(gl);
1279         spin_unlock(&gl->gl_spin);
1280 }
1281
1282 /**
1283  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1284  * @gh: the holder structure
1285  *
1286  */
1287
1288 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1289 {
1290         gfs2_glock_dq(gh);
1291         gfs2_holder_uninit(gh);
1292 }
1293
1294 /**
1295  * gfs2_glock_nq_num - acquire a glock based on lock number
1296  * @sdp: the filesystem
1297  * @number: the lock number
1298  * @glops: the glock operations for the type of glock
1299  * @state: the state to acquire the glock in
1300  * @flags: modifier flags for the aquisition
1301  * @gh: the struct gfs2_holder
1302  *
1303  * Returns: errno
1304  */
1305
1306 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1307                       const struct gfs2_glock_operations *glops,
1308                       unsigned int state, int flags, struct gfs2_holder *gh)
1309 {
1310         struct gfs2_glock *gl;
1311         int error;
1312
1313         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1314         if (!error) {
1315                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1316                 gfs2_glock_put(gl);
1317         }
1318
1319         return error;
1320 }
1321
1322 /**
1323  * glock_compare - Compare two struct gfs2_glock structures for sorting
1324  * @arg_a: the first structure
1325  * @arg_b: the second structure
1326  *
1327  */
1328
1329 static int glock_compare(const void *arg_a, const void *arg_b)
1330 {
1331         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1332         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1333         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1334         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1335
1336         if (a->ln_number > b->ln_number)
1337                 return 1;
1338         if (a->ln_number < b->ln_number)
1339                 return -1;
1340         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1341         return 0;
1342 }
1343
1344 /**
1345  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1346  * @num_gh: the number of structures
1347  * @ghs: an array of struct gfs2_holder structures
1348  *
1349  * Returns: 0 on success (all glocks acquired),
1350  *          errno on failure (no glocks acquired)
1351  */
1352
1353 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1354                      struct gfs2_holder **p)
1355 {
1356         unsigned int x;
1357         int error = 0;
1358
1359         for (x = 0; x < num_gh; x++)
1360                 p[x] = &ghs[x];
1361
1362         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1363
1364         for (x = 0; x < num_gh; x++) {
1365                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1366
1367                 error = gfs2_glock_nq(p[x]);
1368                 if (error) {
1369                         while (x--)
1370                                 gfs2_glock_dq(p[x]);
1371                         break;
1372                 }
1373         }
1374
1375         return error;
1376 }
1377
1378 /**
1379  * gfs2_glock_nq_m - acquire multiple glocks
1380  * @num_gh: the number of structures
1381  * @ghs: an array of struct gfs2_holder structures
1382  *
1383  * Figure out how big an impact this function has.  Either:
1384  * 1) Replace this code with code that calls gfs2_glock_prefetch()
1385  * 2) Forget async stuff and just call nq_m_sync()
1386  * 3) Leave it like it is
1387  *
1388  * Returns: 0 on success (all glocks acquired),
1389  *          errno on failure (no glocks acquired)
1390  */
1391
1392 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1393 {
1394         int *e;
1395         unsigned int x;
1396         int borked = 0, serious = 0;
1397         int error = 0;
1398
1399         if (!num_gh)
1400                 return 0;
1401
1402         if (num_gh == 1) {
1403                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1404                 return gfs2_glock_nq(ghs);
1405         }
1406
1407         e = kcalloc(num_gh, sizeof(struct gfs2_holder *), GFP_KERNEL);
1408         if (!e)
1409                 return -ENOMEM;
1410
1411         for (x = 0; x < num_gh; x++) {
1412                 ghs[x].gh_flags |= LM_FLAG_TRY | GL_ASYNC;
1413                 error = gfs2_glock_nq(&ghs[x]);
1414                 if (error) {
1415                         borked = 1;
1416                         serious = error;
1417                         num_gh = x;
1418                         break;
1419                 }
1420         }
1421
1422         for (x = 0; x < num_gh; x++) {
1423                 error = e[x] = glock_wait_internal(&ghs[x]);
1424                 if (error) {
1425                         borked = 1;
1426                         if (error != GLR_TRYFAILED && error != GLR_CANCELED)
1427                                 serious = error;
1428                 }
1429         }
1430
1431         if (!borked) {
1432                 kfree(e);
1433                 return 0;
1434         }
1435
1436         for (x = 0; x < num_gh; x++)
1437                 if (!e[x])
1438                         gfs2_glock_dq(&ghs[x]);
1439
1440         if (serious)
1441                 error = serious;
1442         else {
1443                 for (x = 0; x < num_gh; x++)
1444                         gfs2_holder_reinit(ghs[x].gh_state, ghs[x].gh_flags,
1445                                           &ghs[x]);
1446                 error = nq_m_sync(num_gh, ghs, (struct gfs2_holder **)e);
1447         }
1448
1449         kfree(e);
1450
1451         return error;
1452 }
1453
1454 /**
1455  * gfs2_glock_dq_m - release multiple glocks
1456  * @num_gh: the number of structures
1457  * @ghs: an array of struct gfs2_holder structures
1458  *
1459  */
1460
1461 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1462 {
1463         unsigned int x;
1464
1465         for (x = 0; x < num_gh; x++)
1466                 gfs2_glock_dq(&ghs[x]);
1467 }
1468
1469 /**
1470  * gfs2_glock_dq_uninit_m - release multiple glocks
1471  * @num_gh: the number of structures
1472  * @ghs: an array of struct gfs2_holder structures
1473  *
1474  */
1475
1476 void gfs2_glock_dq_uninit_m(unsigned int num_gh, struct gfs2_holder *ghs)
1477 {
1478         unsigned int x;
1479
1480         for (x = 0; x < num_gh; x++)
1481                 gfs2_glock_dq_uninit(&ghs[x]);
1482 }
1483
1484 /**
1485  * gfs2_lvb_hold - attach a LVB from a glock
1486  * @gl: The glock in question
1487  *
1488  */
1489
1490 int gfs2_lvb_hold(struct gfs2_glock *gl)
1491 {
1492         int error;
1493
1494         gfs2_glmutex_lock(gl);
1495
1496         if (!atomic_read(&gl->gl_lvb_count)) {
1497                 error = gfs2_lm_hold_lvb(gl->gl_sbd, gl->gl_lock, &gl->gl_lvb);
1498                 if (error) {
1499                         gfs2_glmutex_unlock(gl);
1500                         return error;
1501                 }
1502                 gfs2_glock_hold(gl);
1503         }
1504         atomic_inc(&gl->gl_lvb_count);
1505
1506         gfs2_glmutex_unlock(gl);
1507
1508         return 0;
1509 }
1510
1511 /**
1512  * gfs2_lvb_unhold - detach a LVB from a glock
1513  * @gl: The glock in question
1514  *
1515  */
1516
1517 void gfs2_lvb_unhold(struct gfs2_glock *gl)
1518 {
1519         gfs2_glock_hold(gl);
1520         gfs2_glmutex_lock(gl);
1521
1522         gfs2_assert(gl->gl_sbd, atomic_read(&gl->gl_lvb_count) > 0);
1523         if (atomic_dec_and_test(&gl->gl_lvb_count)) {
1524                 gfs2_lm_unhold_lvb(gl->gl_sbd, gl->gl_lock, gl->gl_lvb);
1525                 gl->gl_lvb = NULL;
1526                 gfs2_glock_put(gl);
1527         }
1528
1529         gfs2_glmutex_unlock(gl);
1530         gfs2_glock_put(gl);
1531 }
1532
1533 static void blocking_cb(struct gfs2_sbd *sdp, struct lm_lockname *name,
1534                         unsigned int state)
1535 {
1536         struct gfs2_glock *gl;
1537
1538         gl = gfs2_glock_find(sdp, name);
1539         if (!gl)
1540                 return;
1541
1542         handle_callback(gl, state);
1543
1544         spin_lock(&gl->gl_spin);
1545         run_queue(gl);
1546         spin_unlock(&gl->gl_spin);
1547
1548         gfs2_glock_put(gl);
1549 }
1550
1551 /**
1552  * gfs2_glock_cb - Callback used by locking module
1553  * @sdp: Pointer to the superblock
1554  * @type: Type of callback
1555  * @data: Type dependent data pointer
1556  *
1557  * Called by the locking module when it wants to tell us something.
1558  * Either we need to drop a lock, one of our ASYNC requests completed, or
1559  * a journal from another client needs to be recovered.
1560  */
1561
1562 void gfs2_glock_cb(void *cb_data, unsigned int type, void *data)
1563 {
1564         struct gfs2_sbd *sdp = cb_data;
1565
1566         switch (type) {
1567         case LM_CB_NEED_E:
1568                 blocking_cb(sdp, data, LM_ST_UNLOCKED);
1569                 return;
1570
1571         case LM_CB_NEED_D:
1572                 blocking_cb(sdp, data, LM_ST_DEFERRED);
1573                 return;
1574
1575         case LM_CB_NEED_S:
1576                 blocking_cb(sdp, data, LM_ST_SHARED);
1577                 return;
1578
1579         case LM_CB_ASYNC: {
1580                 struct lm_async_cb *async = data;
1581                 struct gfs2_glock *gl;
1582
1583                 down_read(&gfs2_umount_flush_sem);
1584                 gl = gfs2_glock_find(sdp, &async->lc_name);
1585                 if (gfs2_assert_warn(sdp, gl))
1586                         return;
1587                 if (!gfs2_assert_warn(sdp, gl->gl_req_bh))
1588                         gl->gl_req_bh(gl, async->lc_ret);
1589                 gfs2_glock_put(gl);
1590                 up_read(&gfs2_umount_flush_sem);
1591                 return;
1592         }
1593
1594         case LM_CB_NEED_RECOVERY:
1595                 gfs2_jdesc_make_dirty(sdp, *(unsigned int *)data);
1596                 if (sdp->sd_recoverd_process)
1597                         wake_up_process(sdp->sd_recoverd_process);
1598                 return;
1599
1600         case LM_CB_DROPLOCKS:
1601                 gfs2_gl_hash_clear(sdp, NO_WAIT);
1602                 gfs2_quota_scan(sdp);
1603                 return;
1604
1605         default:
1606                 gfs2_assert_warn(sdp, 0);
1607                 return;
1608         }
1609 }
1610
1611 /**
1612  * demote_ok - Check to see if it's ok to unlock a glock
1613  * @gl: the glock
1614  *
1615  * Returns: 1 if it's ok
1616  */
1617
1618 static int demote_ok(struct gfs2_glock *gl)
1619 {
1620         const struct gfs2_glock_operations *glops = gl->gl_ops;
1621         int demote = 1;
1622
1623         if (test_bit(GLF_STICKY, &gl->gl_flags))
1624                 demote = 0;
1625         else if (glops->go_demote_ok)
1626                 demote = glops->go_demote_ok(gl);
1627
1628         return demote;
1629 }
1630
1631 /**
1632  * gfs2_glock_schedule_for_reclaim - Add a glock to the reclaim list
1633  * @gl: the glock
1634  *
1635  */
1636
1637 void gfs2_glock_schedule_for_reclaim(struct gfs2_glock *gl)
1638 {
1639         struct gfs2_sbd *sdp = gl->gl_sbd;
1640
1641         spin_lock(&sdp->sd_reclaim_lock);
1642         if (list_empty(&gl->gl_reclaim)) {
1643                 gfs2_glock_hold(gl);
1644                 list_add(&gl->gl_reclaim, &sdp->sd_reclaim_list);
1645                 atomic_inc(&sdp->sd_reclaim_count);
1646         }
1647         spin_unlock(&sdp->sd_reclaim_lock);
1648
1649         wake_up(&sdp->sd_reclaim_wq);
1650 }
1651
1652 /**
1653  * gfs2_reclaim_glock - process the next glock on the filesystem's reclaim list
1654  * @sdp: the filesystem
1655  *
1656  * Called from gfs2_glockd() glock reclaim daemon, or when promoting a
1657  * different glock and we notice that there are a lot of glocks in the
1658  * reclaim list.
1659  *
1660  */
1661
1662 void gfs2_reclaim_glock(struct gfs2_sbd *sdp)
1663 {
1664         struct gfs2_glock *gl;
1665
1666         spin_lock(&sdp->sd_reclaim_lock);
1667         if (list_empty(&sdp->sd_reclaim_list)) {
1668                 spin_unlock(&sdp->sd_reclaim_lock);
1669                 return;
1670         }
1671         gl = list_entry(sdp->sd_reclaim_list.next,
1672                         struct gfs2_glock, gl_reclaim);
1673         list_del_init(&gl->gl_reclaim);
1674         spin_unlock(&sdp->sd_reclaim_lock);
1675
1676         atomic_dec(&sdp->sd_reclaim_count);
1677         atomic_inc(&sdp->sd_reclaimed);
1678
1679         if (gfs2_glmutex_trylock(gl)) {
1680                 if (list_empty(&gl->gl_holders) &&
1681                     gl->gl_state != LM_ST_UNLOCKED && demote_ok(gl))
1682                         handle_callback(gl, LM_ST_UNLOCKED);
1683                 gfs2_glmutex_unlock(gl);
1684         }
1685
1686         gfs2_glock_put(gl);
1687 }
1688
1689 /**
1690  * examine_bucket - Call a function for glock in a hash bucket
1691  * @examiner: the function
1692  * @sdp: the filesystem
1693  * @bucket: the bucket
1694  *
1695  * Returns: 1 if the bucket has entries
1696  */
1697
1698 static int examine_bucket(glock_examiner examiner, struct gfs2_sbd *sdp,
1699                           unsigned int hash)
1700 {
1701         struct gfs2_glock *gl, *prev = NULL;
1702         int has_entries = 0;
1703         struct hlist_head *head = &gl_hash_table[hash].hb_list;
1704
1705         read_lock(gl_lock_addr(hash));
1706         /* Can't use hlist_for_each_entry - don't want prefetch here */
1707         if (hlist_empty(head))
1708                 goto out;
1709         gl = list_entry(head->first, struct gfs2_glock, gl_list);
1710         while(1) {
1711                 if (gl->gl_sbd == sdp) {
1712                         gfs2_glock_hold(gl);
1713                         read_unlock(gl_lock_addr(hash));
1714                         if (prev)
1715                                 gfs2_glock_put(prev);
1716                         prev = gl;
1717                         examiner(gl);
1718                         has_entries = 1;
1719                         read_lock(gl_lock_addr(hash));
1720                 }
1721                 if (gl->gl_list.next == NULL)
1722                         break;
1723                 gl = list_entry(gl->gl_list.next, struct gfs2_glock, gl_list);
1724         }
1725 out:
1726         read_unlock(gl_lock_addr(hash));
1727         if (prev)
1728                 gfs2_glock_put(prev);
1729         return has_entries;
1730 }
1731
1732 /**
1733  * scan_glock - look at a glock and see if we can reclaim it
1734  * @gl: the glock to look at
1735  *
1736  */
1737
1738 static void scan_glock(struct gfs2_glock *gl)
1739 {
1740         if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object)
1741                 return;
1742
1743         if (gfs2_glmutex_trylock(gl)) {
1744                 if (list_empty(&gl->gl_holders) &&
1745                     gl->gl_state != LM_ST_UNLOCKED && demote_ok(gl))
1746                         goto out_schedule;
1747                 gfs2_glmutex_unlock(gl);
1748         }
1749         return;
1750
1751 out_schedule:
1752         gfs2_glmutex_unlock(gl);
1753         gfs2_glock_schedule_for_reclaim(gl);
1754 }
1755
1756 /**
1757  * gfs2_scand_internal - Look for glocks and inodes to toss from memory
1758  * @sdp: the filesystem
1759  *
1760  */
1761
1762 void gfs2_scand_internal(struct gfs2_sbd *sdp)
1763 {
1764         unsigned int x;
1765
1766         for (x = 0; x < GFS2_GL_HASH_SIZE; x++)
1767                 examine_bucket(scan_glock, sdp, x);
1768 }
1769
1770 /**
1771  * clear_glock - look at a glock and see if we can free it from glock cache
1772  * @gl: the glock to look at
1773  *
1774  */
1775
1776 static void clear_glock(struct gfs2_glock *gl)
1777 {
1778         struct gfs2_sbd *sdp = gl->gl_sbd;
1779         int released;
1780
1781         spin_lock(&sdp->sd_reclaim_lock);
1782         if (!list_empty(&gl->gl_reclaim)) {
1783                 list_del_init(&gl->gl_reclaim);
1784                 atomic_dec(&sdp->sd_reclaim_count);
1785                 spin_unlock(&sdp->sd_reclaim_lock);
1786                 released = gfs2_glock_put(gl);
1787                 gfs2_assert(sdp, !released);
1788         } else {
1789                 spin_unlock(&sdp->sd_reclaim_lock);
1790         }
1791
1792         if (gfs2_glmutex_trylock(gl)) {
1793                 if (list_empty(&gl->gl_holders) &&
1794                     gl->gl_state != LM_ST_UNLOCKED)
1795                         handle_callback(gl, LM_ST_UNLOCKED);
1796                 gfs2_glmutex_unlock(gl);
1797         }
1798 }
1799
1800 /**
1801  * gfs2_gl_hash_clear - Empty out the glock hash table
1802  * @sdp: the filesystem
1803  * @wait: wait until it's all gone
1804  *
1805  * Called when unmounting the filesystem, or when inter-node lock manager
1806  * requests DROPLOCKS because it is running out of capacity.
1807  */
1808
1809 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp, int wait)
1810 {
1811         unsigned long t;
1812         unsigned int x;
1813         int cont;
1814
1815         t = jiffies;
1816
1817         for (;;) {
1818                 cont = 0;
1819                 for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
1820                         if (examine_bucket(clear_glock, sdp, x))
1821                                 cont = 1;
1822                 }
1823
1824                 if (!wait || !cont)
1825                         break;
1826
1827                 if (time_after_eq(jiffies,
1828                                   t + gfs2_tune_get(sdp, gt_stall_secs) * HZ)) {
1829                         fs_warn(sdp, "Unmount seems to be stalled. "
1830                                      "Dumping lock state...\n");
1831                         gfs2_dump_lockstate(sdp);
1832                         t = jiffies;
1833                 }
1834
1835                 down_write(&gfs2_umount_flush_sem);
1836                 invalidate_inodes(sdp->sd_vfs);
1837                 up_write(&gfs2_umount_flush_sem);
1838                 msleep(10);
1839         }
1840 }
1841
1842 /*
1843  *  Diagnostic routines to help debug distributed deadlock
1844  */
1845
1846 /**
1847  * dump_holder - print information about a glock holder
1848  * @str: a string naming the type of holder
1849  * @gh: the glock holder
1850  *
1851  * Returns: 0 on success, -ENOBUFS when we run out of space
1852  */
1853
1854 static int dump_holder(char *str, struct gfs2_holder *gh)
1855 {
1856         unsigned int x;
1857         int error = -ENOBUFS;
1858
1859         printk(KERN_INFO "  %s\n", str);
1860         printk(KERN_INFO "    owner = %ld\n",
1861                    (gh->gh_owner) ? (long)gh->gh_owner->pid : -1);
1862         printk(KERN_INFO "    gh_state = %u\n", gh->gh_state);
1863         printk(KERN_INFO "    gh_flags =");
1864         for (x = 0; x < 32; x++)
1865                 if (gh->gh_flags & (1 << x))
1866                         printk(" %u", x);
1867         printk(" \n");
1868         printk(KERN_INFO "    error = %d\n", gh->gh_error);
1869         printk(KERN_INFO "    gh_iflags =");
1870         for (x = 0; x < 32; x++)
1871                 if (test_bit(x, &gh->gh_iflags))
1872                         printk(" %u", x);
1873         printk(" \n");
1874         print_symbol(KERN_INFO "    initialized at: %s\n", gh->gh_ip);
1875
1876         error = 0;
1877
1878         return error;
1879 }
1880
1881 /**
1882  * dump_inode - print information about an inode
1883  * @ip: the inode
1884  *
1885  * Returns: 0 on success, -ENOBUFS when we run out of space
1886  */
1887
1888 static int dump_inode(struct gfs2_inode *ip)
1889 {
1890         unsigned int x;
1891         int error = -ENOBUFS;
1892
1893         printk(KERN_INFO "  Inode:\n");
1894         printk(KERN_INFO "    num = %llu %llu\n",
1895                     (unsigned long long)ip->i_num.no_formal_ino,
1896                     (unsigned long long)ip->i_num.no_addr);
1897         printk(KERN_INFO "    type = %u\n", IF2DT(ip->i_inode.i_mode));
1898         printk(KERN_INFO "    i_flags =");
1899         for (x = 0; x < 32; x++)
1900                 if (test_bit(x, &ip->i_flags))
1901                         printk(" %u", x);
1902         printk(" \n");
1903
1904         error = 0;
1905
1906         return error;
1907 }
1908
1909 /**
1910  * dump_glock - print information about a glock
1911  * @gl: the glock
1912  * @count: where we are in the buffer
1913  *
1914  * Returns: 0 on success, -ENOBUFS when we run out of space
1915  */
1916
1917 static int dump_glock(struct gfs2_glock *gl)
1918 {
1919         struct gfs2_holder *gh;
1920         unsigned int x;
1921         int error = -ENOBUFS;
1922
1923         spin_lock(&gl->gl_spin);
1924
1925         printk(KERN_INFO "Glock 0x%p (%u, %llu)\n", gl, gl->gl_name.ln_type,
1926                (unsigned long long)gl->gl_name.ln_number);
1927         printk(KERN_INFO "  gl_flags =");
1928         for (x = 0; x < 32; x++) {
1929                 if (test_bit(x, &gl->gl_flags))
1930                         printk(" %u", x);
1931         }
1932         printk(" \n");
1933         printk(KERN_INFO "  gl_ref = %d\n", atomic_read(&gl->gl_ref));
1934         printk(KERN_INFO "  gl_state = %u\n", gl->gl_state);
1935         printk(KERN_INFO "  gl_owner = %s\n", gl->gl_owner->comm);
1936         print_symbol(KERN_INFO "  gl_ip = %s\n", gl->gl_ip);
1937         printk(KERN_INFO "  req_gh = %s\n", (gl->gl_req_gh) ? "yes" : "no");
1938         printk(KERN_INFO "  req_bh = %s\n", (gl->gl_req_bh) ? "yes" : "no");
1939         printk(KERN_INFO "  lvb_count = %d\n", atomic_read(&gl->gl_lvb_count));
1940         printk(KERN_INFO "  object = %s\n", (gl->gl_object) ? "yes" : "no");
1941         printk(KERN_INFO "  le = %s\n",
1942                    (list_empty(&gl->gl_le.le_list)) ? "no" : "yes");
1943         printk(KERN_INFO "  reclaim = %s\n",
1944                     (list_empty(&gl->gl_reclaim)) ? "no" : "yes");
1945         if (gl->gl_aspace)
1946                 printk(KERN_INFO "  aspace = 0x%p nrpages = %lu\n", gl->gl_aspace,
1947                        gl->gl_aspace->i_mapping->nrpages);
1948         else
1949                 printk(KERN_INFO "  aspace = no\n");
1950         printk(KERN_INFO "  ail = %d\n", atomic_read(&gl->gl_ail_count));
1951         if (gl->gl_req_gh) {
1952                 error = dump_holder("Request", gl->gl_req_gh);
1953                 if (error)
1954                         goto out;
1955         }
1956         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1957                 error = dump_holder("Holder", gh);
1958                 if (error)
1959                         goto out;
1960         }
1961         list_for_each_entry(gh, &gl->gl_waiters1, gh_list) {
1962                 error = dump_holder("Waiter1", gh);
1963                 if (error)
1964                         goto out;
1965         }
1966         list_for_each_entry(gh, &gl->gl_waiters2, gh_list) {
1967                 error = dump_holder("Waiter2", gh);
1968                 if (error)
1969                         goto out;
1970         }
1971         list_for_each_entry(gh, &gl->gl_waiters3, gh_list) {
1972                 error = dump_holder("Waiter3", gh);
1973                 if (error)
1974                         goto out;
1975         }
1976         if (gl->gl_ops == &gfs2_inode_glops && gl->gl_object) {
1977                 if (!test_bit(GLF_LOCK, &gl->gl_flags) &&
1978                     list_empty(&gl->gl_holders)) {
1979                         error = dump_inode(gl->gl_object);
1980                         if (error)
1981                                 goto out;
1982                 } else {
1983                         error = -ENOBUFS;
1984                         printk(KERN_INFO "  Inode: busy\n");
1985                 }
1986         }
1987
1988         error = 0;
1989
1990 out:
1991         spin_unlock(&gl->gl_spin);
1992         return error;
1993 }
1994
1995 /**
1996  * gfs2_dump_lockstate - print out the current lockstate
1997  * @sdp: the filesystem
1998  * @ub: the buffer to copy the information into
1999  *
2000  * If @ub is NULL, dump the lockstate to the console.
2001  *
2002  */
2003
2004 static int gfs2_dump_lockstate(struct gfs2_sbd *sdp)
2005 {
2006         struct gfs2_glock *gl;
2007         struct hlist_node *h;
2008         unsigned int x;
2009         int error = 0;
2010
2011         for (x = 0; x < GFS2_GL_HASH_SIZE; x++) {
2012
2013                 read_lock(gl_lock_addr(x));
2014
2015                 hlist_for_each_entry(gl, h, &gl_hash_table[x].hb_list, gl_list) {
2016                         if (gl->gl_sbd != sdp)
2017                                 continue;
2018
2019                         error = dump_glock(gl);
2020                         if (error)
2021                                 break;
2022                 }
2023
2024                 read_unlock(gl_lock_addr(x));
2025
2026                 if (error)
2027                         break;
2028         }
2029
2030
2031         return error;
2032 }
2033
2034 int __init gfs2_glock_init(void)
2035 {
2036         unsigned i;
2037         for(i = 0; i < GFS2_GL_HASH_SIZE; i++) {
2038                 INIT_HLIST_HEAD(&gl_hash_table[i].hb_list);
2039         }
2040 #ifdef GL_HASH_LOCK_SZ
2041         for(i = 0; i < GL_HASH_LOCK_SZ; i++) {
2042                 rwlock_init(&gl_hash_locks[i]);
2043         }
2044 #endif
2045         return 0;
2046 }
2047