GNU Linux-libre 4.9.308-gnu1
[releases.git] / fs / ocfs2 / dlmglue.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * dlmglue.c
5  *
6  * Code which implements an OCFS2 specific interface to our DLM.
7  *
8  * Copyright (C) 2003, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/types.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
29 #include <linux/mm.h>
30 #include <linux/kthread.h>
31 #include <linux/pagemap.h>
32 #include <linux/debugfs.h>
33 #include <linux/seq_file.h>
34 #include <linux/time.h>
35 #include <linux/quotaops.h>
36
37 #define MLOG_MASK_PREFIX ML_DLM_GLUE
38 #include <cluster/masklog.h>
39
40 #include "ocfs2.h"
41 #include "ocfs2_lockingver.h"
42
43 #include "alloc.h"
44 #include "dcache.h"
45 #include "dlmglue.h"
46 #include "extent_map.h"
47 #include "file.h"
48 #include "heartbeat.h"
49 #include "inode.h"
50 #include "journal.h"
51 #include "stackglue.h"
52 #include "slot_map.h"
53 #include "super.h"
54 #include "uptodate.h"
55 #include "quota.h"
56 #include "refcounttree.h"
57 #include "acl.h"
58
59 #include "buffer_head_io.h"
60
61 struct ocfs2_mask_waiter {
62         struct list_head        mw_item;
63         int                     mw_status;
64         struct completion       mw_complete;
65         unsigned long           mw_mask;
66         unsigned long           mw_goal;
67 #ifdef CONFIG_OCFS2_FS_STATS
68         ktime_t                 mw_lock_start;
69 #endif
70 };
71
72 static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres);
73 static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres);
74 static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres);
75 static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres);
76
77 /*
78  * Return value from ->downconvert_worker functions.
79  *
80  * These control the precise actions of ocfs2_unblock_lock()
81  * and ocfs2_process_blocked_lock()
82  *
83  */
84 enum ocfs2_unblock_action {
85         UNBLOCK_CONTINUE        = 0, /* Continue downconvert */
86         UNBLOCK_CONTINUE_POST   = 1, /* Continue downconvert, fire
87                                       * ->post_unlock callback */
88         UNBLOCK_STOP_POST       = 2, /* Do not downconvert, fire
89                                       * ->post_unlock() callback. */
90 };
91
92 struct ocfs2_unblock_ctl {
93         int requeue;
94         enum ocfs2_unblock_action unblock_action;
95 };
96
97 /* Lockdep class keys */
98 struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES];
99
100 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
101                                         int new_level);
102 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
103
104 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
105                                      int blocking);
106
107 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
108                                        int blocking);
109
110 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
111                                      struct ocfs2_lock_res *lockres);
112
113 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres);
114
115 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
116                                             int new_level);
117 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
118                                          int blocking);
119
120 #define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
121
122 /* This aids in debugging situations where a bad LVB might be involved. */
123 static void ocfs2_dump_meta_lvb_info(u64 level,
124                                      const char *function,
125                                      unsigned int line,
126                                      struct ocfs2_lock_res *lockres)
127 {
128         struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
129
130         mlog(level, "LVB information for %s (called from %s:%u):\n",
131              lockres->l_name, function, line);
132         mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
133              lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
134              be32_to_cpu(lvb->lvb_igeneration));
135         mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
136              (unsigned long long)be64_to_cpu(lvb->lvb_isize),
137              be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
138              be16_to_cpu(lvb->lvb_imode));
139         mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
140              "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
141              (long long)be64_to_cpu(lvb->lvb_iatime_packed),
142              (long long)be64_to_cpu(lvb->lvb_ictime_packed),
143              (long long)be64_to_cpu(lvb->lvb_imtime_packed),
144              be32_to_cpu(lvb->lvb_iattr));
145 }
146
147
148 /*
149  * OCFS2 Lock Resource Operations
150  *
151  * These fine tune the behavior of the generic dlmglue locking infrastructure.
152  *
153  * The most basic of lock types can point ->l_priv to their respective
154  * struct ocfs2_super and allow the default actions to manage things.
155  *
156  * Right now, each lock type also needs to implement an init function,
157  * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
158  * should be called when the lock is no longer needed (i.e., object
159  * destruction time).
160  */
161 struct ocfs2_lock_res_ops {
162         /*
163          * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
164          * this callback if ->l_priv is not an ocfs2_super pointer
165          */
166         struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
167
168         /*
169          * Optionally called in the downconvert thread after a
170          * successful downconvert. The lockres will not be referenced
171          * after this callback is called, so it is safe to free
172          * memory, etc.
173          *
174          * The exact semantics of when this is called are controlled
175          * by ->downconvert_worker()
176          */
177         void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
178
179         /*
180          * Allow a lock type to add checks to determine whether it is
181          * safe to downconvert a lock. Return 0 to re-queue the
182          * downconvert at a later time, nonzero to continue.
183          *
184          * For most locks, the default checks that there are no
185          * incompatible holders are sufficient.
186          *
187          * Called with the lockres spinlock held.
188          */
189         int (*check_downconvert)(struct ocfs2_lock_res *, int);
190
191         /*
192          * Allows a lock type to populate the lock value block. This
193          * is called on downconvert, and when we drop a lock.
194          *
195          * Locks that want to use this should set LOCK_TYPE_USES_LVB
196          * in the flags field.
197          *
198          * Called with the lockres spinlock held.
199          */
200         void (*set_lvb)(struct ocfs2_lock_res *);
201
202         /*
203          * Called from the downconvert thread when it is determined
204          * that a lock will be downconverted. This is called without
205          * any locks held so the function can do work that might
206          * schedule (syncing out data, etc).
207          *
208          * This should return any one of the ocfs2_unblock_action
209          * values, depending on what it wants the thread to do.
210          */
211         int (*downconvert_worker)(struct ocfs2_lock_res *, int);
212
213         /*
214          * LOCK_TYPE_* flags which describe the specific requirements
215          * of a lock type. Descriptions of each individual flag follow.
216          */
217         int flags;
218 };
219
220 /*
221  * Some locks want to "refresh" potentially stale data when a
222  * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
223  * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
224  * individual lockres l_flags member from the ast function. It is
225  * expected that the locking wrapper will clear the
226  * OCFS2_LOCK_NEEDS_REFRESH flag when done.
227  */
228 #define LOCK_TYPE_REQUIRES_REFRESH 0x1
229
230 /*
231  * Indicate that a lock type makes use of the lock value block. The
232  * ->set_lvb lock type callback must be defined.
233  */
234 #define LOCK_TYPE_USES_LVB              0x2
235
236 static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
237         .get_osb        = ocfs2_get_inode_osb,
238         .flags          = 0,
239 };
240
241 static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = {
242         .get_osb        = ocfs2_get_inode_osb,
243         .check_downconvert = ocfs2_check_meta_downconvert,
244         .set_lvb        = ocfs2_set_meta_lvb,
245         .downconvert_worker = ocfs2_data_convert_worker,
246         .flags          = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
247 };
248
249 static struct ocfs2_lock_res_ops ocfs2_super_lops = {
250         .flags          = LOCK_TYPE_REQUIRES_REFRESH,
251 };
252
253 static struct ocfs2_lock_res_ops ocfs2_rename_lops = {
254         .flags          = 0,
255 };
256
257 static struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = {
258         .flags          = 0,
259 };
260
261 static struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = {
262         .flags          = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
263 };
264
265 static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
266         .get_osb        = ocfs2_get_dentry_osb,
267         .post_unlock    = ocfs2_dentry_post_unlock,
268         .downconvert_worker = ocfs2_dentry_convert_worker,
269         .flags          = 0,
270 };
271
272 static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
273         .get_osb        = ocfs2_get_inode_osb,
274         .flags          = 0,
275 };
276
277 static struct ocfs2_lock_res_ops ocfs2_flock_lops = {
278         .get_osb        = ocfs2_get_file_osb,
279         .flags          = 0,
280 };
281
282 static struct ocfs2_lock_res_ops ocfs2_qinfo_lops = {
283         .set_lvb        = ocfs2_set_qinfo_lvb,
284         .get_osb        = ocfs2_get_qinfo_osb,
285         .flags          = LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB,
286 };
287
288 static struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = {
289         .check_downconvert = ocfs2_check_refcount_downconvert,
290         .downconvert_worker = ocfs2_refcount_convert_worker,
291         .flags          = 0,
292 };
293
294 static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
295 {
296         return lockres->l_type == OCFS2_LOCK_TYPE_META ||
297                 lockres->l_type == OCFS2_LOCK_TYPE_RW ||
298                 lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
299 }
300
301 static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
302 {
303         return container_of(lksb, struct ocfs2_lock_res, l_lksb);
304 }
305
306 static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
307 {
308         BUG_ON(!ocfs2_is_inode_lock(lockres));
309
310         return (struct inode *) lockres->l_priv;
311 }
312
313 static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
314 {
315         BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
316
317         return (struct ocfs2_dentry_lock *)lockres->l_priv;
318 }
319
320 static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres)
321 {
322         BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO);
323
324         return (struct ocfs2_mem_dqinfo *)lockres->l_priv;
325 }
326
327 static inline struct ocfs2_refcount_tree *
328 ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res)
329 {
330         return container_of(res, struct ocfs2_refcount_tree, rf_lockres);
331 }
332
333 static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
334 {
335         if (lockres->l_ops->get_osb)
336                 return lockres->l_ops->get_osb(lockres);
337
338         return (struct ocfs2_super *)lockres->l_priv;
339 }
340
341 static int ocfs2_lock_create(struct ocfs2_super *osb,
342                              struct ocfs2_lock_res *lockres,
343                              int level,
344                              u32 dlm_flags);
345 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
346                                                      int wanted);
347 static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
348                                    struct ocfs2_lock_res *lockres,
349                                    int level, unsigned long caller_ip);
350 static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb,
351                                         struct ocfs2_lock_res *lockres,
352                                         int level)
353 {
354         __ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_);
355 }
356
357 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
358 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
359 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
360 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
361 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
362                                         struct ocfs2_lock_res *lockres);
363 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
364                                                 int convert);
365 #define ocfs2_log_dlm_error(_func, _err, _lockres) do {                                 \
366         if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY)                               \
367                 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n",        \
368                      _err, _func, _lockres->l_name);                                    \
369         else                                                                            \
370                 mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n",  \
371                      _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name,  \
372                      (unsigned int)ocfs2_get_dentry_lock_ino(_lockres));                \
373 } while (0)
374 static int ocfs2_downconvert_thread(void *arg);
375 static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
376                                         struct ocfs2_lock_res *lockres);
377 static int ocfs2_inode_lock_update(struct inode *inode,
378                                   struct buffer_head **bh);
379 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
380 static inline int ocfs2_highest_compat_lock_level(int level);
381 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
382                                               int new_level);
383 static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
384                                   struct ocfs2_lock_res *lockres,
385                                   int new_level,
386                                   int lvb,
387                                   unsigned int generation);
388 static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
389                                         struct ocfs2_lock_res *lockres);
390 static int ocfs2_cancel_convert(struct ocfs2_super *osb,
391                                 struct ocfs2_lock_res *lockres);
392
393
394 static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
395                                   u64 blkno,
396                                   u32 generation,
397                                   char *name)
398 {
399         int len;
400
401         BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
402
403         len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
404                        ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
405                        (long long)blkno, generation);
406
407         BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
408
409         mlog(0, "built lock resource with name: %s\n", name);
410 }
411
412 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
413
414 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
415                                        struct ocfs2_dlm_debug *dlm_debug)
416 {
417         mlog(0, "Add tracking for lockres %s\n", res->l_name);
418
419         spin_lock(&ocfs2_dlm_tracking_lock);
420         list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
421         spin_unlock(&ocfs2_dlm_tracking_lock);
422 }
423
424 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
425 {
426         spin_lock(&ocfs2_dlm_tracking_lock);
427         if (!list_empty(&res->l_debug_list))
428                 list_del_init(&res->l_debug_list);
429         spin_unlock(&ocfs2_dlm_tracking_lock);
430 }
431
432 #ifdef CONFIG_OCFS2_FS_STATS
433 static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
434 {
435         res->l_lock_refresh = 0;
436         memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
437         memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
438 }
439
440 static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
441                                     struct ocfs2_mask_waiter *mw, int ret)
442 {
443         u32 usec;
444         ktime_t kt;
445         struct ocfs2_lock_stats *stats;
446
447         if (level == LKM_PRMODE)
448                 stats = &res->l_lock_prmode;
449         else if (level == LKM_EXMODE)
450                 stats = &res->l_lock_exmode;
451         else
452                 return;
453
454         kt = ktime_sub(ktime_get(), mw->mw_lock_start);
455         usec = ktime_to_us(kt);
456
457         stats->ls_gets++;
458         stats->ls_total += ktime_to_ns(kt);
459         /* overflow */
460         if (unlikely(stats->ls_gets == 0)) {
461                 stats->ls_gets++;
462                 stats->ls_total = ktime_to_ns(kt);
463         }
464
465         if (stats->ls_max < usec)
466                 stats->ls_max = usec;
467
468         if (ret)
469                 stats->ls_fail++;
470 }
471
472 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
473 {
474         lockres->l_lock_refresh++;
475 }
476
477 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
478 {
479         mw->mw_lock_start = ktime_get();
480 }
481 #else
482 static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
483 {
484 }
485 static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
486                            int level, struct ocfs2_mask_waiter *mw, int ret)
487 {
488 }
489 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
490 {
491 }
492 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
493 {
494 }
495 #endif
496
497 static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
498                                        struct ocfs2_lock_res *res,
499                                        enum ocfs2_lock_type type,
500                                        struct ocfs2_lock_res_ops *ops,
501                                        void *priv)
502 {
503         res->l_type          = type;
504         res->l_ops           = ops;
505         res->l_priv          = priv;
506
507         res->l_level         = DLM_LOCK_IV;
508         res->l_requested     = DLM_LOCK_IV;
509         res->l_blocking      = DLM_LOCK_IV;
510         res->l_action        = OCFS2_AST_INVALID;
511         res->l_unlock_action = OCFS2_UNLOCK_INVALID;
512
513         res->l_flags         = OCFS2_LOCK_INITIALIZED;
514
515         ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
516
517         ocfs2_init_lock_stats(res);
518 #ifdef CONFIG_DEBUG_LOCK_ALLOC
519         if (type != OCFS2_LOCK_TYPE_OPEN)
520                 lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
521                                  &lockdep_keys[type], 0);
522         else
523                 res->l_lockdep_map.key = NULL;
524 #endif
525 }
526
527 void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
528 {
529         /* This also clears out the lock status block */
530         memset(res, 0, sizeof(struct ocfs2_lock_res));
531         spin_lock_init(&res->l_lock);
532         init_waitqueue_head(&res->l_event);
533         INIT_LIST_HEAD(&res->l_blocked_list);
534         INIT_LIST_HEAD(&res->l_mask_waiters);
535         INIT_LIST_HEAD(&res->l_holders);
536 }
537
538 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
539                                enum ocfs2_lock_type type,
540                                unsigned int generation,
541                                struct inode *inode)
542 {
543         struct ocfs2_lock_res_ops *ops;
544
545         switch(type) {
546                 case OCFS2_LOCK_TYPE_RW:
547                         ops = &ocfs2_inode_rw_lops;
548                         break;
549                 case OCFS2_LOCK_TYPE_META:
550                         ops = &ocfs2_inode_inode_lops;
551                         break;
552                 case OCFS2_LOCK_TYPE_OPEN:
553                         ops = &ocfs2_inode_open_lops;
554                         break;
555                 default:
556                         mlog_bug_on_msg(1, "type: %d\n", type);
557                         ops = NULL; /* thanks, gcc */
558                         break;
559         };
560
561         ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
562                               generation, res->l_name);
563         ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
564 }
565
566 static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
567 {
568         struct inode *inode = ocfs2_lock_res_inode(lockres);
569
570         return OCFS2_SB(inode->i_sb);
571 }
572
573 static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
574 {
575         struct ocfs2_mem_dqinfo *info = lockres->l_priv;
576
577         return OCFS2_SB(info->dqi_gi.dqi_sb);
578 }
579
580 static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
581 {
582         struct ocfs2_file_private *fp = lockres->l_priv;
583
584         return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
585 }
586
587 static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
588 {
589         __be64 inode_blkno_be;
590
591         memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
592                sizeof(__be64));
593
594         return be64_to_cpu(inode_blkno_be);
595 }
596
597 static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
598 {
599         struct ocfs2_dentry_lock *dl = lockres->l_priv;
600
601         return OCFS2_SB(dl->dl_inode->i_sb);
602 }
603
604 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
605                                 u64 parent, struct inode *inode)
606 {
607         int len;
608         u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
609         __be64 inode_blkno_be = cpu_to_be64(inode_blkno);
610         struct ocfs2_lock_res *lockres = &dl->dl_lockres;
611
612         ocfs2_lock_res_init_once(lockres);
613
614         /*
615          * Unfortunately, the standard lock naming scheme won't work
616          * here because we have two 16 byte values to use. Instead,
617          * we'll stuff the inode number as a binary value. We still
618          * want error prints to show something without garbling the
619          * display, so drop a null byte in there before the inode
620          * number. A future version of OCFS2 will likely use all
621          * binary lock names. The stringified names have been a
622          * tremendous aid in debugging, but now that the debugfs
623          * interface exists, we can mangle things there if need be.
624          *
625          * NOTE: We also drop the standard "pad" value (the total lock
626          * name size stays the same though - the last part is all
627          * zeros due to the memset in ocfs2_lock_res_init_once()
628          */
629         len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
630                        "%c%016llx",
631                        ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
632                        (long long)parent);
633
634         BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
635
636         memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
637                sizeof(__be64));
638
639         ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
640                                    OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
641                                    dl);
642 }
643
644 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
645                                       struct ocfs2_super *osb)
646 {
647         /* Superblock lockres doesn't come from a slab so we call init
648          * once on it manually.  */
649         ocfs2_lock_res_init_once(res);
650         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
651                               0, res->l_name);
652         ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
653                                    &ocfs2_super_lops, osb);
654 }
655
656 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
657                                        struct ocfs2_super *osb)
658 {
659         /* Rename lockres doesn't come from a slab so we call init
660          * once on it manually.  */
661         ocfs2_lock_res_init_once(res);
662         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
663         ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
664                                    &ocfs2_rename_lops, osb);
665 }
666
667 static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
668                                          struct ocfs2_super *osb)
669 {
670         /* nfs_sync lockres doesn't come from a slab so we call init
671          * once on it manually.  */
672         ocfs2_lock_res_init_once(res);
673         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
674         ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
675                                    &ocfs2_nfs_sync_lops, osb);
676 }
677
678 static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
679                                             struct ocfs2_super *osb)
680 {
681         ocfs2_lock_res_init_once(res);
682         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
683         ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
684                                    &ocfs2_orphan_scan_lops, osb);
685 }
686
687 void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
688                               struct ocfs2_file_private *fp)
689 {
690         struct inode *inode = fp->fp_file->f_mapping->host;
691         struct ocfs2_inode_info *oi = OCFS2_I(inode);
692
693         ocfs2_lock_res_init_once(lockres);
694         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
695                               inode->i_generation, lockres->l_name);
696         ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
697                                    OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
698                                    fp);
699         lockres->l_flags |= OCFS2_LOCK_NOCACHE;
700 }
701
702 void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
703                                struct ocfs2_mem_dqinfo *info)
704 {
705         ocfs2_lock_res_init_once(lockres);
706         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
707                               0, lockres->l_name);
708         ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
709                                    OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
710                                    info);
711 }
712
713 void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
714                                   struct ocfs2_super *osb, u64 ref_blkno,
715                                   unsigned int generation)
716 {
717         ocfs2_lock_res_init_once(lockres);
718         ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
719                               generation, lockres->l_name);
720         ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
721                                    &ocfs2_refcount_block_lops, osb);
722 }
723
724 void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
725 {
726         if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
727                 return;
728
729         ocfs2_remove_lockres_tracking(res);
730
731         mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
732                         "Lockres %s is on the blocked list\n",
733                         res->l_name);
734         mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
735                         "Lockres %s has mask waiters pending\n",
736                         res->l_name);
737         mlog_bug_on_msg(spin_is_locked(&res->l_lock),
738                         "Lockres %s is locked\n",
739                         res->l_name);
740         mlog_bug_on_msg(res->l_ro_holders,
741                         "Lockres %s has %u ro holders\n",
742                         res->l_name, res->l_ro_holders);
743         mlog_bug_on_msg(res->l_ex_holders,
744                         "Lockres %s has %u ex holders\n",
745                         res->l_name, res->l_ex_holders);
746
747         /* Need to clear out the lock status block for the dlm */
748         memset(&res->l_lksb, 0, sizeof(res->l_lksb));
749
750         res->l_flags = 0UL;
751 }
752
753 /*
754  * Keep a list of processes who have interest in a lockres.
755  * Note: this is now only uesed for check recursive cluster locking.
756  */
757 static inline void ocfs2_add_holder(struct ocfs2_lock_res *lockres,
758                                    struct ocfs2_lock_holder *oh)
759 {
760         INIT_LIST_HEAD(&oh->oh_list);
761         oh->oh_owner_pid = get_pid(task_pid(current));
762
763         spin_lock(&lockres->l_lock);
764         list_add_tail(&oh->oh_list, &lockres->l_holders);
765         spin_unlock(&lockres->l_lock);
766 }
767
768 static inline void ocfs2_remove_holder(struct ocfs2_lock_res *lockres,
769                                        struct ocfs2_lock_holder *oh)
770 {
771         spin_lock(&lockres->l_lock);
772         list_del(&oh->oh_list);
773         spin_unlock(&lockres->l_lock);
774
775         put_pid(oh->oh_owner_pid);
776 }
777
778 static inline int ocfs2_is_locked_by_me(struct ocfs2_lock_res *lockres)
779 {
780         struct ocfs2_lock_holder *oh;
781         struct pid *pid;
782
783         /* look in the list of holders for one with the current task as owner */
784         spin_lock(&lockres->l_lock);
785         pid = task_pid(current);
786         list_for_each_entry(oh, &lockres->l_holders, oh_list) {
787                 if (oh->oh_owner_pid == pid) {
788                         spin_unlock(&lockres->l_lock);
789                         return 1;
790                 }
791         }
792         spin_unlock(&lockres->l_lock);
793
794         return 0;
795 }
796
797 static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
798                                      int level)
799 {
800         BUG_ON(!lockres);
801
802         switch(level) {
803         case DLM_LOCK_EX:
804                 lockres->l_ex_holders++;
805                 break;
806         case DLM_LOCK_PR:
807                 lockres->l_ro_holders++;
808                 break;
809         default:
810                 BUG();
811         }
812 }
813
814 static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
815                                      int level)
816 {
817         BUG_ON(!lockres);
818
819         switch(level) {
820         case DLM_LOCK_EX:
821                 BUG_ON(!lockres->l_ex_holders);
822                 lockres->l_ex_holders--;
823                 break;
824         case DLM_LOCK_PR:
825                 BUG_ON(!lockres->l_ro_holders);
826                 lockres->l_ro_holders--;
827                 break;
828         default:
829                 BUG();
830         }
831 }
832
833 /* WARNING: This function lives in a world where the only three lock
834  * levels are EX, PR, and NL. It *will* have to be adjusted when more
835  * lock types are added. */
836 static inline int ocfs2_highest_compat_lock_level(int level)
837 {
838         int new_level = DLM_LOCK_EX;
839
840         if (level == DLM_LOCK_EX)
841                 new_level = DLM_LOCK_NL;
842         else if (level == DLM_LOCK_PR)
843                 new_level = DLM_LOCK_PR;
844         return new_level;
845 }
846
847 static void lockres_set_flags(struct ocfs2_lock_res *lockres,
848                               unsigned long newflags)
849 {
850         struct ocfs2_mask_waiter *mw, *tmp;
851
852         assert_spin_locked(&lockres->l_lock);
853
854         lockres->l_flags = newflags;
855
856         list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
857                 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
858                         continue;
859
860                 list_del_init(&mw->mw_item);
861                 mw->mw_status = 0;
862                 complete(&mw->mw_complete);
863         }
864 }
865 static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
866 {
867         lockres_set_flags(lockres, lockres->l_flags | or);
868 }
869 static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
870                                 unsigned long clear)
871 {
872         lockres_set_flags(lockres, lockres->l_flags & ~clear);
873 }
874
875 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
876 {
877         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
878         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
879         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
880         BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
881
882         lockres->l_level = lockres->l_requested;
883         if (lockres->l_level <=
884             ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
885                 lockres->l_blocking = DLM_LOCK_NL;
886                 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
887         }
888         lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
889 }
890
891 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
892 {
893         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
894         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
895
896         /* Convert from RO to EX doesn't really need anything as our
897          * information is already up to data. Convert from NL to
898          * *anything* however should mark ourselves as needing an
899          * update */
900         if (lockres->l_level == DLM_LOCK_NL &&
901             lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
902                 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
903
904         lockres->l_level = lockres->l_requested;
905
906         /*
907          * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
908          * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
909          * downconverting the lock before the upconvert has fully completed.
910          * Do not prevent the dc thread from downconverting if NONBLOCK lock
911          * had already returned.
912          */
913         if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED))
914                 lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
915         else
916                 lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED);
917
918         lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
919 }
920
921 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
922 {
923         BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
924         BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
925
926         if (lockres->l_requested > DLM_LOCK_NL &&
927             !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
928             lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
929                 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
930
931         lockres->l_level = lockres->l_requested;
932         lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
933         lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
934 }
935
936 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
937                                      int level)
938 {
939         int needs_downconvert = 0;
940
941         assert_spin_locked(&lockres->l_lock);
942
943         if (level > lockres->l_blocking) {
944                 /* only schedule a downconvert if we haven't already scheduled
945                  * one that goes low enough to satisfy the level we're
946                  * blocking.  this also catches the case where we get
947                  * duplicate BASTs */
948                 if (ocfs2_highest_compat_lock_level(level) <
949                     ocfs2_highest_compat_lock_level(lockres->l_blocking))
950                         needs_downconvert = 1;
951
952                 lockres->l_blocking = level;
953         }
954
955         mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
956              lockres->l_name, level, lockres->l_level, lockres->l_blocking,
957              needs_downconvert);
958
959         if (needs_downconvert)
960                 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
961         mlog(0, "needs_downconvert = %d\n", needs_downconvert);
962         return needs_downconvert;
963 }
964
965 /*
966  * OCFS2_LOCK_PENDING and l_pending_gen.
967  *
968  * Why does OCFS2_LOCK_PENDING exist?  To close a race between setting
969  * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock().  See ocfs2_unblock_lock()
970  * for more details on the race.
971  *
972  * OCFS2_LOCK_PENDING closes the race quite nicely.  However, it introduces
973  * a race on itself.  In o2dlm, we can get the ast before ocfs2_dlm_lock()
974  * returns.  The ast clears OCFS2_LOCK_BUSY, and must therefore clear
975  * OCFS2_LOCK_PENDING at the same time.  When ocfs2_dlm_lock() returns,
976  * the caller is going to try to clear PENDING again.  If nothing else is
977  * happening, __lockres_clear_pending() sees PENDING is unset and does
978  * nothing.
979  *
980  * But what if another path (eg downconvert thread) has just started a
981  * new locking action?  The other path has re-set PENDING.  Our path
982  * cannot clear PENDING, because that will re-open the original race
983  * window.
984  *
985  * [Example]
986  *
987  * ocfs2_meta_lock()
988  *  ocfs2_cluster_lock()
989  *   set BUSY
990  *   set PENDING
991  *   drop l_lock
992  *   ocfs2_dlm_lock()
993  *    ocfs2_locking_ast()               ocfs2_downconvert_thread()
994  *     clear PENDING                     ocfs2_unblock_lock()
995  *                                        take_l_lock
996  *                                        !BUSY
997  *                                        ocfs2_prepare_downconvert()
998  *                                         set BUSY
999  *                                         set PENDING
1000  *                                        drop l_lock
1001  *   take l_lock
1002  *   clear PENDING
1003  *   drop l_lock
1004  *                      <window>
1005  *                                        ocfs2_dlm_lock()
1006  *
1007  * So as you can see, we now have a window where l_lock is not held,
1008  * PENDING is not set, and ocfs2_dlm_lock() has not been called.
1009  *
1010  * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
1011  * set by ocfs2_prepare_downconvert().  That wasn't nice.
1012  *
1013  * To solve this we introduce l_pending_gen.  A call to
1014  * lockres_clear_pending() will only do so when it is passed a generation
1015  * number that matches the lockres.  lockres_set_pending() will return the
1016  * current generation number.  When ocfs2_cluster_lock() goes to clear
1017  * PENDING, it passes the generation it got from set_pending().  In our
1018  * example above, the generation numbers will *not* match.  Thus,
1019  * ocfs2_cluster_lock() will not clear the PENDING set by
1020  * ocfs2_prepare_downconvert().
1021  */
1022
1023 /* Unlocked version for ocfs2_locking_ast() */
1024 static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
1025                                     unsigned int generation,
1026                                     struct ocfs2_super *osb)
1027 {
1028         assert_spin_locked(&lockres->l_lock);
1029
1030         /*
1031          * The ast and locking functions can race us here.  The winner
1032          * will clear pending, the loser will not.
1033          */
1034         if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
1035             (lockres->l_pending_gen != generation))
1036                 return;
1037
1038         lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
1039         lockres->l_pending_gen++;
1040
1041         /*
1042          * The downconvert thread may have skipped us because we
1043          * were PENDING.  Wake it up.
1044          */
1045         if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1046                 ocfs2_wake_downconvert_thread(osb);
1047 }
1048
1049 /* Locked version for callers of ocfs2_dlm_lock() */
1050 static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1051                                   unsigned int generation,
1052                                   struct ocfs2_super *osb)
1053 {
1054         unsigned long flags;
1055
1056         spin_lock_irqsave(&lockres->l_lock, flags);
1057         __lockres_clear_pending(lockres, generation, osb);
1058         spin_unlock_irqrestore(&lockres->l_lock, flags);
1059 }
1060
1061 static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1062 {
1063         assert_spin_locked(&lockres->l_lock);
1064         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1065
1066         lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1067
1068         return lockres->l_pending_gen;
1069 }
1070
1071 static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1072 {
1073         struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1074         struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1075         int needs_downconvert;
1076         unsigned long flags;
1077
1078         BUG_ON(level <= DLM_LOCK_NL);
1079
1080         mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, "
1081              "type %s\n", lockres->l_name, level, lockres->l_level,
1082              ocfs2_lock_type_string(lockres->l_type));
1083
1084         /*
1085          * We can skip the bast for locks which don't enable caching -
1086          * they'll be dropped at the earliest possible time anyway.
1087          */
1088         if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1089                 return;
1090
1091         spin_lock_irqsave(&lockres->l_lock, flags);
1092         needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1093         if (needs_downconvert)
1094                 ocfs2_schedule_blocked_lock(osb, lockres);
1095         spin_unlock_irqrestore(&lockres->l_lock, flags);
1096
1097         wake_up(&lockres->l_event);
1098
1099         ocfs2_wake_downconvert_thread(osb);
1100 }
1101
1102 static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1103 {
1104         struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1105         struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1106         unsigned long flags;
1107         int status;
1108
1109         spin_lock_irqsave(&lockres->l_lock, flags);
1110
1111         status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1112
1113         if (status == -EAGAIN) {
1114                 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1115                 goto out;
1116         }
1117
1118         if (status) {
1119                 mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1120                      lockres->l_name, status);
1121                 spin_unlock_irqrestore(&lockres->l_lock, flags);
1122                 return;
1123         }
1124
1125         mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, "
1126              "level %d => %d\n", lockres->l_name, lockres->l_action,
1127              lockres->l_unlock_action, lockres->l_level, lockres->l_requested);
1128
1129         switch(lockres->l_action) {
1130         case OCFS2_AST_ATTACH:
1131                 ocfs2_generic_handle_attach_action(lockres);
1132                 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1133                 break;
1134         case OCFS2_AST_CONVERT:
1135                 ocfs2_generic_handle_convert_action(lockres);
1136                 break;
1137         case OCFS2_AST_DOWNCONVERT:
1138                 ocfs2_generic_handle_downconvert_action(lockres);
1139                 break;
1140         default:
1141                 mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, "
1142                      "flags 0x%lx, unlock: %u\n",
1143                      lockres->l_name, lockres->l_action, lockres->l_flags,
1144                      lockres->l_unlock_action);
1145                 BUG();
1146         }
1147 out:
1148         /* set it to something invalid so if we get called again we
1149          * can catch it. */
1150         lockres->l_action = OCFS2_AST_INVALID;
1151
1152         /* Did we try to cancel this lock?  Clear that state */
1153         if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1154                 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1155
1156         /*
1157          * We may have beaten the locking functions here.  We certainly
1158          * know that dlm_lock() has been called :-)
1159          * Because we can't have two lock calls in flight at once, we
1160          * can use lockres->l_pending_gen.
1161          */
1162         __lockres_clear_pending(lockres, lockres->l_pending_gen,  osb);
1163
1164         wake_up(&lockres->l_event);
1165         spin_unlock_irqrestore(&lockres->l_lock, flags);
1166 }
1167
1168 static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
1169 {
1170         struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1171         unsigned long flags;
1172
1173         mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n",
1174              lockres->l_name, lockres->l_unlock_action);
1175
1176         spin_lock_irqsave(&lockres->l_lock, flags);
1177         if (error) {
1178                 mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
1179                      "unlock_action %d\n", error, lockres->l_name,
1180                      lockres->l_unlock_action);
1181                 spin_unlock_irqrestore(&lockres->l_lock, flags);
1182                 return;
1183         }
1184
1185         switch(lockres->l_unlock_action) {
1186         case OCFS2_UNLOCK_CANCEL_CONVERT:
1187                 mlog(0, "Cancel convert success for %s\n", lockres->l_name);
1188                 lockres->l_action = OCFS2_AST_INVALID;
1189                 /* Downconvert thread may have requeued this lock, we
1190                  * need to wake it. */
1191                 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1192                         ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
1193                 break;
1194         case OCFS2_UNLOCK_DROP_LOCK:
1195                 lockres->l_level = DLM_LOCK_IV;
1196                 break;
1197         default:
1198                 BUG();
1199         }
1200
1201         lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1202         lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1203         wake_up(&lockres->l_event);
1204         spin_unlock_irqrestore(&lockres->l_lock, flags);
1205 }
1206
1207 /*
1208  * This is the filesystem locking protocol.  It provides the lock handling
1209  * hooks for the underlying DLM.  It has a maximum version number.
1210  * The version number allows interoperability with systems running at
1211  * the same major number and an equal or smaller minor number.
1212  *
1213  * Whenever the filesystem does new things with locks (adds or removes a
1214  * lock, orders them differently, does different things underneath a lock),
1215  * the version must be changed.  The protocol is negotiated when joining
1216  * the dlm domain.  A node may join the domain if its major version is
1217  * identical to all other nodes and its minor version is greater than
1218  * or equal to all other nodes.  When its minor version is greater than
1219  * the other nodes, it will run at the minor version specified by the
1220  * other nodes.
1221  *
1222  * If a locking change is made that will not be compatible with older
1223  * versions, the major number must be increased and the minor version set
1224  * to zero.  If a change merely adds a behavior that can be disabled when
1225  * speaking to older versions, the minor version must be increased.  If a
1226  * change adds a fully backwards compatible change (eg, LVB changes that
1227  * are just ignored by older versions), the version does not need to be
1228  * updated.
1229  */
1230 static struct ocfs2_locking_protocol lproto = {
1231         .lp_max_version = {
1232                 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
1233                 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
1234         },
1235         .lp_lock_ast            = ocfs2_locking_ast,
1236         .lp_blocking_ast        = ocfs2_blocking_ast,
1237         .lp_unlock_ast          = ocfs2_unlock_ast,
1238 };
1239
1240 void ocfs2_set_locking_protocol(void)
1241 {
1242         ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version);
1243 }
1244
1245 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1246                                                 int convert)
1247 {
1248         unsigned long flags;
1249
1250         spin_lock_irqsave(&lockres->l_lock, flags);
1251         lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1252         lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1253         if (convert)
1254                 lockres->l_action = OCFS2_AST_INVALID;
1255         else
1256                 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1257         spin_unlock_irqrestore(&lockres->l_lock, flags);
1258
1259         wake_up(&lockres->l_event);
1260 }
1261
1262 /* Note: If we detect another process working on the lock (i.e.,
1263  * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1264  * to do the right thing in that case.
1265  */
1266 static int ocfs2_lock_create(struct ocfs2_super *osb,
1267                              struct ocfs2_lock_res *lockres,
1268                              int level,
1269                              u32 dlm_flags)
1270 {
1271         int ret = 0;
1272         unsigned long flags;
1273         unsigned int gen;
1274
1275         mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1276              dlm_flags);
1277
1278         spin_lock_irqsave(&lockres->l_lock, flags);
1279         if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1280             (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1281                 spin_unlock_irqrestore(&lockres->l_lock, flags);
1282                 goto bail;
1283         }
1284
1285         lockres->l_action = OCFS2_AST_ATTACH;
1286         lockres->l_requested = level;
1287         lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1288         gen = lockres_set_pending(lockres);
1289         spin_unlock_irqrestore(&lockres->l_lock, flags);
1290
1291         ret = ocfs2_dlm_lock(osb->cconn,
1292                              level,
1293                              &lockres->l_lksb,
1294                              dlm_flags,
1295                              lockres->l_name,
1296                              OCFS2_LOCK_ID_MAX_LEN - 1);
1297         lockres_clear_pending(lockres, gen, osb);
1298         if (ret) {
1299                 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1300                 ocfs2_recover_from_dlm_error(lockres, 1);
1301         }
1302
1303         mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1304
1305 bail:
1306         return ret;
1307 }
1308
1309 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1310                                         int flag)
1311 {
1312         unsigned long flags;
1313         int ret;
1314
1315         spin_lock_irqsave(&lockres->l_lock, flags);
1316         ret = lockres->l_flags & flag;
1317         spin_unlock_irqrestore(&lockres->l_lock, flags);
1318
1319         return ret;
1320 }
1321
1322 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1323
1324 {
1325         wait_event(lockres->l_event,
1326                    !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1327 }
1328
1329 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1330
1331 {
1332         wait_event(lockres->l_event,
1333                    !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1334 }
1335
1336 /* predict what lock level we'll be dropping down to on behalf
1337  * of another node, and return true if the currently wanted
1338  * level will be compatible with it. */
1339 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1340                                                      int wanted)
1341 {
1342         BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1343
1344         return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1345 }
1346
1347 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1348 {
1349         INIT_LIST_HEAD(&mw->mw_item);
1350         init_completion(&mw->mw_complete);
1351         ocfs2_init_start_time(mw);
1352 }
1353
1354 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1355 {
1356         wait_for_completion(&mw->mw_complete);
1357         /* Re-arm the completion in case we want to wait on it again */
1358         reinit_completion(&mw->mw_complete);
1359         return mw->mw_status;
1360 }
1361
1362 static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1363                                     struct ocfs2_mask_waiter *mw,
1364                                     unsigned long mask,
1365                                     unsigned long goal)
1366 {
1367         BUG_ON(!list_empty(&mw->mw_item));
1368
1369         assert_spin_locked(&lockres->l_lock);
1370
1371         list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1372         mw->mw_mask = mask;
1373         mw->mw_goal = goal;
1374 }
1375
1376 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
1377  * if the mask still hadn't reached its goal */
1378 static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1379                                       struct ocfs2_mask_waiter *mw)
1380 {
1381         int ret = 0;
1382
1383         assert_spin_locked(&lockres->l_lock);
1384         if (!list_empty(&mw->mw_item)) {
1385                 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1386                         ret = -EBUSY;
1387
1388                 list_del_init(&mw->mw_item);
1389                 init_completion(&mw->mw_complete);
1390         }
1391
1392         return ret;
1393 }
1394
1395 static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1396                                       struct ocfs2_mask_waiter *mw)
1397 {
1398         unsigned long flags;
1399         int ret = 0;
1400
1401         spin_lock_irqsave(&lockres->l_lock, flags);
1402         ret = __lockres_remove_mask_waiter(lockres, mw);
1403         spin_unlock_irqrestore(&lockres->l_lock, flags);
1404
1405         return ret;
1406
1407 }
1408
1409 static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1410                                              struct ocfs2_lock_res *lockres)
1411 {
1412         int ret;
1413
1414         ret = wait_for_completion_interruptible(&mw->mw_complete);
1415         if (ret)
1416                 lockres_remove_mask_waiter(lockres, mw);
1417         else
1418                 ret = mw->mw_status;
1419         /* Re-arm the completion in case we want to wait on it again */
1420         reinit_completion(&mw->mw_complete);
1421         return ret;
1422 }
1423
1424 static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1425                                 struct ocfs2_lock_res *lockres,
1426                                 int level,
1427                                 u32 lkm_flags,
1428                                 int arg_flags,
1429                                 int l_subclass,
1430                                 unsigned long caller_ip)
1431 {
1432         struct ocfs2_mask_waiter mw;
1433         int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1434         int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1435         unsigned long flags;
1436         unsigned int gen;
1437         int noqueue_attempted = 0;
1438         int dlm_locked = 0;
1439         int kick_dc = 0;
1440
1441         if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) {
1442                 mlog_errno(-EINVAL);
1443                 return -EINVAL;
1444         }
1445
1446         ocfs2_init_mask_waiter(&mw);
1447
1448         if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1449                 lkm_flags |= DLM_LKF_VALBLK;
1450
1451 again:
1452         wait = 0;
1453
1454         spin_lock_irqsave(&lockres->l_lock, flags);
1455
1456         if (catch_signals && signal_pending(current)) {
1457                 ret = -ERESTARTSYS;
1458                 goto unlock;
1459         }
1460
1461         mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1462                         "Cluster lock called on freeing lockres %s! flags "
1463                         "0x%lx\n", lockres->l_name, lockres->l_flags);
1464
1465         /* We only compare against the currently granted level
1466          * here. If the lock is blocked waiting on a downconvert,
1467          * we'll get caught below. */
1468         if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1469             level > lockres->l_level) {
1470                 /* is someone sitting in dlm_lock? If so, wait on
1471                  * them. */
1472                 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1473                 wait = 1;
1474                 goto unlock;
1475         }
1476
1477         if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1478                 /*
1479                  * We've upconverted. If the lock now has a level we can
1480                  * work with, we take it. If, however, the lock is not at the
1481                  * required level, we go thru the full cycle. One way this could
1482                  * happen is if a process requesting an upconvert to PR is
1483                  * closely followed by another requesting upconvert to an EX.
1484                  * If the process requesting EX lands here, we want it to
1485                  * continue attempting to upconvert and let the process
1486                  * requesting PR take the lock.
1487                  * If multiple processes request upconvert to PR, the first one
1488                  * here will take the lock. The others will have to go thru the
1489                  * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1490                  * downconvert request.
1491                  */
1492                 if (level <= lockres->l_level)
1493                         goto update_holders;
1494         }
1495
1496         if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1497             !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1498                 /* is the lock is currently blocked on behalf of
1499                  * another node */
1500                 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1501                 wait = 1;
1502                 goto unlock;
1503         }
1504
1505         if (level > lockres->l_level) {
1506                 if (noqueue_attempted > 0) {
1507                         ret = -EAGAIN;
1508                         goto unlock;
1509                 }
1510                 if (lkm_flags & DLM_LKF_NOQUEUE)
1511                         noqueue_attempted = 1;
1512
1513                 if (lockres->l_action != OCFS2_AST_INVALID)
1514                         mlog(ML_ERROR, "lockres %s has action %u pending\n",
1515                              lockres->l_name, lockres->l_action);
1516
1517                 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1518                         lockres->l_action = OCFS2_AST_ATTACH;
1519                         lkm_flags &= ~DLM_LKF_CONVERT;
1520                 } else {
1521                         lockres->l_action = OCFS2_AST_CONVERT;
1522                         lkm_flags |= DLM_LKF_CONVERT;
1523                 }
1524
1525                 lockres->l_requested = level;
1526                 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1527                 gen = lockres_set_pending(lockres);
1528                 spin_unlock_irqrestore(&lockres->l_lock, flags);
1529
1530                 BUG_ON(level == DLM_LOCK_IV);
1531                 BUG_ON(level == DLM_LOCK_NL);
1532
1533                 mlog(ML_BASTS, "lockres %s, convert from %d to %d\n",
1534                      lockres->l_name, lockres->l_level, level);
1535
1536                 /* call dlm_lock to upgrade lock now */
1537                 ret = ocfs2_dlm_lock(osb->cconn,
1538                                      level,
1539                                      &lockres->l_lksb,
1540                                      lkm_flags,
1541                                      lockres->l_name,
1542                                      OCFS2_LOCK_ID_MAX_LEN - 1);
1543                 lockres_clear_pending(lockres, gen, osb);
1544                 if (ret) {
1545                         if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1546                             (ret != -EAGAIN)) {
1547                                 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1548                                                     ret, lockres);
1549                         }
1550                         ocfs2_recover_from_dlm_error(lockres, 1);
1551                         goto out;
1552                 }
1553                 dlm_locked = 1;
1554
1555                 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1556                      lockres->l_name);
1557
1558                 /* At this point we've gone inside the dlm and need to
1559                  * complete our work regardless. */
1560                 catch_signals = 0;
1561
1562                 /* wait for busy to clear and carry on */
1563                 goto again;
1564         }
1565
1566 update_holders:
1567         /* Ok, if we get here then we're good to go. */
1568         ocfs2_inc_holders(lockres, level);
1569
1570         ret = 0;
1571 unlock:
1572         lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1573
1574         /* ocfs2_unblock_lock reques on seeing OCFS2_LOCK_UPCONVERT_FINISHING */
1575         kick_dc = (lockres->l_flags & OCFS2_LOCK_BLOCKED);
1576
1577         spin_unlock_irqrestore(&lockres->l_lock, flags);
1578         if (kick_dc)
1579                 ocfs2_wake_downconvert_thread(osb);
1580 out:
1581         /*
1582          * This is helping work around a lock inversion between the page lock
1583          * and dlm locks.  One path holds the page lock while calling aops
1584          * which block acquiring dlm locks.  The voting thread holds dlm
1585          * locks while acquiring page locks while down converting data locks.
1586          * This block is helping an aop path notice the inversion and back
1587          * off to unlock its page lock before trying the dlm lock again.
1588          */
1589         if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1590             mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1591                 wait = 0;
1592                 spin_lock_irqsave(&lockres->l_lock, flags);
1593                 if (__lockres_remove_mask_waiter(lockres, &mw)) {
1594                         if (dlm_locked)
1595                                 lockres_or_flags(lockres,
1596                                         OCFS2_LOCK_NONBLOCK_FINISHED);
1597                         spin_unlock_irqrestore(&lockres->l_lock, flags);
1598                         ret = -EAGAIN;
1599                 } else {
1600                         spin_unlock_irqrestore(&lockres->l_lock, flags);
1601                         goto again;
1602                 }
1603         }
1604         if (wait) {
1605                 ret = ocfs2_wait_for_mask(&mw);
1606                 if (ret == 0)
1607                         goto again;
1608                 mlog_errno(ret);
1609         }
1610         ocfs2_update_lock_stats(lockres, level, &mw, ret);
1611
1612 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1613         if (!ret && lockres->l_lockdep_map.key != NULL) {
1614                 if (level == DLM_LOCK_PR)
1615                         rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1616                                 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1617                                 caller_ip);
1618                 else
1619                         rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1620                                 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1621                                 caller_ip);
1622         }
1623 #endif
1624         return ret;
1625 }
1626
1627 static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1628                                      struct ocfs2_lock_res *lockres,
1629                                      int level,
1630                                      u32 lkm_flags,
1631                                      int arg_flags)
1632 {
1633         return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1634                                     0, _RET_IP_);
1635 }
1636
1637
1638 static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1639                                    struct ocfs2_lock_res *lockres,
1640                                    int level,
1641                                    unsigned long caller_ip)
1642 {
1643         unsigned long flags;
1644
1645         spin_lock_irqsave(&lockres->l_lock, flags);
1646         ocfs2_dec_holders(lockres, level);
1647         ocfs2_downconvert_on_unlock(osb, lockres);
1648         spin_unlock_irqrestore(&lockres->l_lock, flags);
1649 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1650         if (lockres->l_lockdep_map.key != NULL)
1651                 rwsem_release(&lockres->l_lockdep_map, 1, caller_ip);
1652 #endif
1653 }
1654
1655 static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1656                                  struct ocfs2_lock_res *lockres,
1657                                  int ex,
1658                                  int local)
1659 {
1660         int level =  ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1661         unsigned long flags;
1662         u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1663
1664         spin_lock_irqsave(&lockres->l_lock, flags);
1665         BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1666         lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1667         spin_unlock_irqrestore(&lockres->l_lock, flags);
1668
1669         return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1670 }
1671
1672 /* Grants us an EX lock on the data and metadata resources, skipping
1673  * the normal cluster directory lookup. Use this ONLY on newly created
1674  * inodes which other nodes can't possibly see, and which haven't been
1675  * hashed in the inode hash yet. This can give us a good performance
1676  * increase as it'll skip the network broadcast normally associated
1677  * with creating a new lock resource. */
1678 int ocfs2_create_new_inode_locks(struct inode *inode)
1679 {
1680         int ret;
1681         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1682
1683         BUG_ON(!ocfs2_inode_is_new(inode));
1684
1685         mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1686
1687         /* NOTE: That we don't increment any of the holder counts, nor
1688          * do we add anything to a journal handle. Since this is
1689          * supposed to be a new inode which the cluster doesn't know
1690          * about yet, there is no need to.  As far as the LVB handling
1691          * is concerned, this is basically like acquiring an EX lock
1692          * on a resource which has an invalid one -- we'll set it
1693          * valid when we release the EX. */
1694
1695         ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1696         if (ret) {
1697                 mlog_errno(ret);
1698                 goto bail;
1699         }
1700
1701         /*
1702          * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1703          * don't use a generation in their lock names.
1704          */
1705         ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1706         if (ret) {
1707                 mlog_errno(ret);
1708                 goto bail;
1709         }
1710
1711         ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1712         if (ret)
1713                 mlog_errno(ret);
1714
1715 bail:
1716         return ret;
1717 }
1718
1719 int ocfs2_rw_lock(struct inode *inode, int write)
1720 {
1721         int status, level;
1722         struct ocfs2_lock_res *lockres;
1723         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1724
1725         mlog(0, "inode %llu take %s RW lock\n",
1726              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1727              write ? "EXMODE" : "PRMODE");
1728
1729         if (ocfs2_mount_local(osb))
1730                 return 0;
1731
1732         lockres = &OCFS2_I(inode)->ip_rw_lockres;
1733
1734         level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1735
1736         status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level, 0,
1737                                     0);
1738         if (status < 0)
1739                 mlog_errno(status);
1740
1741         return status;
1742 }
1743
1744 void ocfs2_rw_unlock(struct inode *inode, int write)
1745 {
1746         int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1747         struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1748         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1749
1750         mlog(0, "inode %llu drop %s RW lock\n",
1751              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1752              write ? "EXMODE" : "PRMODE");
1753
1754         if (!ocfs2_mount_local(osb))
1755                 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1756 }
1757
1758 /*
1759  * ocfs2_open_lock always get PR mode lock.
1760  */
1761 int ocfs2_open_lock(struct inode *inode)
1762 {
1763         int status = 0;
1764         struct ocfs2_lock_res *lockres;
1765         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1766
1767         mlog(0, "inode %llu take PRMODE open lock\n",
1768              (unsigned long long)OCFS2_I(inode)->ip_blkno);
1769
1770         if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1771                 goto out;
1772
1773         lockres = &OCFS2_I(inode)->ip_open_lockres;
1774
1775         status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1776                                     DLM_LOCK_PR, 0, 0);
1777         if (status < 0)
1778                 mlog_errno(status);
1779
1780 out:
1781         return status;
1782 }
1783
1784 int ocfs2_try_open_lock(struct inode *inode, int write)
1785 {
1786         int status = 0, level;
1787         struct ocfs2_lock_res *lockres;
1788         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1789
1790         mlog(0, "inode %llu try to take %s open lock\n",
1791              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1792              write ? "EXMODE" : "PRMODE");
1793
1794         if (ocfs2_is_hard_readonly(osb)) {
1795                 if (write)
1796                         status = -EROFS;
1797                 goto out;
1798         }
1799
1800         if (ocfs2_mount_local(osb))
1801                 goto out;
1802
1803         lockres = &OCFS2_I(inode)->ip_open_lockres;
1804
1805         level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1806
1807         /*
1808          * The file system may already holding a PRMODE/EXMODE open lock.
1809          * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1810          * other nodes and the -EAGAIN will indicate to the caller that
1811          * this inode is still in use.
1812          */
1813         status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1814                                     level, DLM_LKF_NOQUEUE, 0);
1815
1816 out:
1817         return status;
1818 }
1819
1820 /*
1821  * ocfs2_open_unlock unlock PR and EX mode open locks.
1822  */
1823 void ocfs2_open_unlock(struct inode *inode)
1824 {
1825         struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1826         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1827
1828         mlog(0, "inode %llu drop open lock\n",
1829              (unsigned long long)OCFS2_I(inode)->ip_blkno);
1830
1831         if (ocfs2_mount_local(osb))
1832                 goto out;
1833
1834         if(lockres->l_ro_holders)
1835                 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1836                                      DLM_LOCK_PR);
1837         if(lockres->l_ex_holders)
1838                 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1839                                      DLM_LOCK_EX);
1840
1841 out:
1842         return;
1843 }
1844
1845 static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1846                                      int level)
1847 {
1848         int ret;
1849         struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1850         unsigned long flags;
1851         struct ocfs2_mask_waiter mw;
1852
1853         ocfs2_init_mask_waiter(&mw);
1854
1855 retry_cancel:
1856         spin_lock_irqsave(&lockres->l_lock, flags);
1857         if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1858                 ret = ocfs2_prepare_cancel_convert(osb, lockres);
1859                 if (ret) {
1860                         spin_unlock_irqrestore(&lockres->l_lock, flags);
1861                         ret = ocfs2_cancel_convert(osb, lockres);
1862                         if (ret < 0) {
1863                                 mlog_errno(ret);
1864                                 goto out;
1865                         }
1866                         goto retry_cancel;
1867                 }
1868                 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1869                 spin_unlock_irqrestore(&lockres->l_lock, flags);
1870
1871                 ocfs2_wait_for_mask(&mw);
1872                 goto retry_cancel;
1873         }
1874
1875         ret = -ERESTARTSYS;
1876         /*
1877          * We may still have gotten the lock, in which case there's no
1878          * point to restarting the syscall.
1879          */
1880         if (lockres->l_level == level)
1881                 ret = 0;
1882
1883         mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1884              lockres->l_flags, lockres->l_level, lockres->l_action);
1885
1886         spin_unlock_irqrestore(&lockres->l_lock, flags);
1887
1888 out:
1889         return ret;
1890 }
1891
1892 /*
1893  * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1894  * flock() calls. The locking approach this requires is sufficiently
1895  * different from all other cluster lock types that we implement a
1896  * separate path to the "low-level" dlm calls. In particular:
1897  *
1898  * - No optimization of lock levels is done - we take at exactly
1899  *   what's been requested.
1900  *
1901  * - No lock caching is employed. We immediately downconvert to
1902  *   no-lock at unlock time. This also means flock locks never go on
1903  *   the blocking list).
1904  *
1905  * - Since userspace can trivially deadlock itself with flock, we make
1906  *   sure to allow cancellation of a misbehaving applications flock()
1907  *   request.
1908  *
1909  * - Access to any flock lockres doesn't require concurrency, so we
1910  *   can simplify the code by requiring the caller to guarantee
1911  *   serialization of dlmglue flock calls.
1912  */
1913 int ocfs2_file_lock(struct file *file, int ex, int trylock)
1914 {
1915         int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1916         unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1917         unsigned long flags;
1918         struct ocfs2_file_private *fp = file->private_data;
1919         struct ocfs2_lock_res *lockres = &fp->fp_flock;
1920         struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1921         struct ocfs2_mask_waiter mw;
1922
1923         ocfs2_init_mask_waiter(&mw);
1924
1925         if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1926             (lockres->l_level > DLM_LOCK_NL)) {
1927                 mlog(ML_ERROR,
1928                      "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1929                      "level: %u\n", lockres->l_name, lockres->l_flags,
1930                      lockres->l_level);
1931                 return -EINVAL;
1932         }
1933
1934         spin_lock_irqsave(&lockres->l_lock, flags);
1935         if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1936                 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1937                 spin_unlock_irqrestore(&lockres->l_lock, flags);
1938
1939                 /*
1940                  * Get the lock at NLMODE to start - that way we
1941                  * can cancel the upconvert request if need be.
1942                  */
1943                 ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
1944                 if (ret < 0) {
1945                         mlog_errno(ret);
1946                         goto out;
1947                 }
1948
1949                 ret = ocfs2_wait_for_mask(&mw);
1950                 if (ret) {
1951                         mlog_errno(ret);
1952                         goto out;
1953                 }
1954                 spin_lock_irqsave(&lockres->l_lock, flags);
1955         }
1956
1957         lockres->l_action = OCFS2_AST_CONVERT;
1958         lkm_flags |= DLM_LKF_CONVERT;
1959         lockres->l_requested = level;
1960         lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1961
1962         lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1963         spin_unlock_irqrestore(&lockres->l_lock, flags);
1964
1965         ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
1966                              lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
1967         if (ret) {
1968                 if (!trylock || (ret != -EAGAIN)) {
1969                         ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1970                         ret = -EINVAL;
1971                 }
1972
1973                 ocfs2_recover_from_dlm_error(lockres, 1);
1974                 lockres_remove_mask_waiter(lockres, &mw);
1975                 goto out;
1976         }
1977
1978         ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
1979         if (ret == -ERESTARTSYS) {
1980                 /*
1981                  * Userspace can cause deadlock itself with
1982                  * flock(). Current behavior locally is to allow the
1983                  * deadlock, but abort the system call if a signal is
1984                  * received. We follow this example, otherwise a
1985                  * poorly written program could sit in kernel until
1986                  * reboot.
1987                  *
1988                  * Handling this is a bit more complicated for Ocfs2
1989                  * though. We can't exit this function with an
1990                  * outstanding lock request, so a cancel convert is
1991                  * required. We intentionally overwrite 'ret' - if the
1992                  * cancel fails and the lock was granted, it's easier
1993                  * to just bubble success back up to the user.
1994                  */
1995                 ret = ocfs2_flock_handle_signal(lockres, level);
1996         } else if (!ret && (level > lockres->l_level)) {
1997                 /* Trylock failed asynchronously */
1998                 BUG_ON(!trylock);
1999                 ret = -EAGAIN;
2000         }
2001
2002 out:
2003
2004         mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
2005              lockres->l_name, ex, trylock, ret);
2006         return ret;
2007 }
2008
2009 void ocfs2_file_unlock(struct file *file)
2010 {
2011         int ret;
2012         unsigned int gen;
2013         unsigned long flags;
2014         struct ocfs2_file_private *fp = file->private_data;
2015         struct ocfs2_lock_res *lockres = &fp->fp_flock;
2016         struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
2017         struct ocfs2_mask_waiter mw;
2018
2019         ocfs2_init_mask_waiter(&mw);
2020
2021         if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
2022                 return;
2023
2024         if (lockres->l_level == DLM_LOCK_NL)
2025                 return;
2026
2027         mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
2028              lockres->l_name, lockres->l_flags, lockres->l_level,
2029              lockres->l_action);
2030
2031         spin_lock_irqsave(&lockres->l_lock, flags);
2032         /*
2033          * Fake a blocking ast for the downconvert code.
2034          */
2035         lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
2036         lockres->l_blocking = DLM_LOCK_EX;
2037
2038         gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
2039         lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
2040         spin_unlock_irqrestore(&lockres->l_lock, flags);
2041
2042         ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
2043         if (ret) {
2044                 mlog_errno(ret);
2045                 return;
2046         }
2047
2048         ret = ocfs2_wait_for_mask(&mw);
2049         if (ret)
2050                 mlog_errno(ret);
2051 }
2052
2053 static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
2054                                         struct ocfs2_lock_res *lockres)
2055 {
2056         int kick = 0;
2057
2058         /* If we know that another node is waiting on our lock, kick
2059          * the downconvert thread * pre-emptively when we reach a release
2060          * condition. */
2061         if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
2062                 switch(lockres->l_blocking) {
2063                 case DLM_LOCK_EX:
2064                         if (!lockres->l_ex_holders && !lockres->l_ro_holders)
2065                                 kick = 1;
2066                         break;
2067                 case DLM_LOCK_PR:
2068                         if (!lockres->l_ex_holders)
2069                                 kick = 1;
2070                         break;
2071                 default:
2072                         BUG();
2073                 }
2074         }
2075
2076         if (kick)
2077                 ocfs2_wake_downconvert_thread(osb);
2078 }
2079
2080 #define OCFS2_SEC_BITS   34
2081 #define OCFS2_SEC_SHIFT  (64 - 34)
2082 #define OCFS2_NSEC_MASK  ((1ULL << OCFS2_SEC_SHIFT) - 1)
2083
2084 /* LVB only has room for 64 bits of time here so we pack it for
2085  * now. */
2086 static u64 ocfs2_pack_timespec(struct timespec *spec)
2087 {
2088         u64 res;
2089         u64 sec = spec->tv_sec;
2090         u32 nsec = spec->tv_nsec;
2091
2092         res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2093
2094         return res;
2095 }
2096
2097 /* Call this with the lockres locked. I am reasonably sure we don't
2098  * need ip_lock in this function as anyone who would be changing those
2099  * values is supposed to be blocked in ocfs2_inode_lock right now. */
2100 static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2101 {
2102         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2103         struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2104         struct ocfs2_meta_lvb *lvb;
2105
2106         lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2107
2108         /*
2109          * Invalidate the LVB of a deleted inode - this way other
2110          * nodes are forced to go to disk and discover the new inode
2111          * status.
2112          */
2113         if (oi->ip_flags & OCFS2_INODE_DELETED) {
2114                 lvb->lvb_version = 0;
2115                 goto out;
2116         }
2117
2118         lvb->lvb_version   = OCFS2_LVB_VERSION;
2119         lvb->lvb_isize     = cpu_to_be64(i_size_read(inode));
2120         lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2121         lvb->lvb_iuid      = cpu_to_be32(i_uid_read(inode));
2122         lvb->lvb_igid      = cpu_to_be32(i_gid_read(inode));
2123         lvb->lvb_imode     = cpu_to_be16(inode->i_mode);
2124         lvb->lvb_inlink    = cpu_to_be16(inode->i_nlink);
2125         lvb->lvb_iatime_packed  =
2126                 cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
2127         lvb->lvb_ictime_packed =
2128                 cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
2129         lvb->lvb_imtime_packed =
2130                 cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
2131         lvb->lvb_iattr    = cpu_to_be32(oi->ip_attr);
2132         lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2133         lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2134
2135 out:
2136         mlog_meta_lvb(0, lockres);
2137 }
2138
2139 static void ocfs2_unpack_timespec(struct timespec *spec,
2140                                   u64 packed_time)
2141 {
2142         spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2143         spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2144 }
2145
2146 static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
2147 {
2148         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2149         struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2150         struct ocfs2_meta_lvb *lvb;
2151
2152         mlog_meta_lvb(0, lockres);
2153
2154         lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2155
2156         /* We're safe here without the lockres lock... */
2157         spin_lock(&oi->ip_lock);
2158         oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2159         i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2160
2161         oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2162         oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2163         ocfs2_set_inode_flags(inode);
2164
2165         /* fast-symlinks are a special case */
2166         if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2167                 inode->i_blocks = 0;
2168         else
2169                 inode->i_blocks = ocfs2_inode_sector_count(inode);
2170
2171         i_uid_write(inode, be32_to_cpu(lvb->lvb_iuid));
2172         i_gid_write(inode, be32_to_cpu(lvb->lvb_igid));
2173         inode->i_mode    = be16_to_cpu(lvb->lvb_imode);
2174         set_nlink(inode, be16_to_cpu(lvb->lvb_inlink));
2175         ocfs2_unpack_timespec(&inode->i_atime,
2176                               be64_to_cpu(lvb->lvb_iatime_packed));
2177         ocfs2_unpack_timespec(&inode->i_mtime,
2178                               be64_to_cpu(lvb->lvb_imtime_packed));
2179         ocfs2_unpack_timespec(&inode->i_ctime,
2180                               be64_to_cpu(lvb->lvb_ictime_packed));
2181         spin_unlock(&oi->ip_lock);
2182 }
2183
2184 static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2185                                               struct ocfs2_lock_res *lockres)
2186 {
2187         struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2188
2189         if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2190             && lvb->lvb_version == OCFS2_LVB_VERSION
2191             && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2192                 return 1;
2193         return 0;
2194 }
2195
2196 /* Determine whether a lock resource needs to be refreshed, and
2197  * arbitrate who gets to refresh it.
2198  *
2199  *   0 means no refresh needed.
2200  *
2201  *   > 0 means you need to refresh this and you MUST call
2202  *   ocfs2_complete_lock_res_refresh afterwards. */
2203 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2204 {
2205         unsigned long flags;
2206         int status = 0;
2207
2208 refresh_check:
2209         spin_lock_irqsave(&lockres->l_lock, flags);
2210         if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2211                 spin_unlock_irqrestore(&lockres->l_lock, flags);
2212                 goto bail;
2213         }
2214
2215         if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2216                 spin_unlock_irqrestore(&lockres->l_lock, flags);
2217
2218                 ocfs2_wait_on_refreshing_lock(lockres);
2219                 goto refresh_check;
2220         }
2221
2222         /* Ok, I'll be the one to refresh this lock. */
2223         lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2224         spin_unlock_irqrestore(&lockres->l_lock, flags);
2225
2226         status = 1;
2227 bail:
2228         mlog(0, "status %d\n", status);
2229         return status;
2230 }
2231
2232 /* If status is non zero, I'll mark it as not being in refresh
2233  * anymroe, but i won't clear the needs refresh flag. */
2234 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2235                                                    int status)
2236 {
2237         unsigned long flags;
2238
2239         spin_lock_irqsave(&lockres->l_lock, flags);
2240         lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2241         if (!status)
2242                 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2243         spin_unlock_irqrestore(&lockres->l_lock, flags);
2244
2245         wake_up(&lockres->l_event);
2246 }
2247
2248 /* may or may not return a bh if it went to disk. */
2249 static int ocfs2_inode_lock_update(struct inode *inode,
2250                                   struct buffer_head **bh)
2251 {
2252         int status = 0;
2253         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2254         struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2255         struct ocfs2_dinode *fe;
2256         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2257
2258         if (ocfs2_mount_local(osb))
2259                 goto bail;
2260
2261         spin_lock(&oi->ip_lock);
2262         if (oi->ip_flags & OCFS2_INODE_DELETED) {
2263                 mlog(0, "Orphaned inode %llu was deleted while we "
2264                      "were waiting on a lock. ip_flags = 0x%x\n",
2265                      (unsigned long long)oi->ip_blkno, oi->ip_flags);
2266                 spin_unlock(&oi->ip_lock);
2267                 status = -ENOENT;
2268                 goto bail;
2269         }
2270         spin_unlock(&oi->ip_lock);
2271
2272         if (!ocfs2_should_refresh_lock_res(lockres))
2273                 goto bail;
2274
2275         /* This will discard any caching information we might have had
2276          * for the inode metadata. */
2277         ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2278
2279         ocfs2_extent_map_trunc(inode, 0);
2280
2281         if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2282                 mlog(0, "Trusting LVB on inode %llu\n",
2283                      (unsigned long long)oi->ip_blkno);
2284                 ocfs2_refresh_inode_from_lvb(inode);
2285         } else {
2286                 /* Boo, we have to go to disk. */
2287                 /* read bh, cast, ocfs2_refresh_inode */
2288                 status = ocfs2_read_inode_block(inode, bh);
2289                 if (status < 0) {
2290                         mlog_errno(status);
2291                         goto bail_refresh;
2292                 }
2293                 fe = (struct ocfs2_dinode *) (*bh)->b_data;
2294
2295                 /* This is a good chance to make sure we're not
2296                  * locking an invalid object.  ocfs2_read_inode_block()
2297                  * already checked that the inode block is sane.
2298                  *
2299                  * We bug on a stale inode here because we checked
2300                  * above whether it was wiped from disk. The wiping
2301                  * node provides a guarantee that we receive that
2302                  * message and can mark the inode before dropping any
2303                  * locks associated with it. */
2304                 mlog_bug_on_msg(inode->i_generation !=
2305                                 le32_to_cpu(fe->i_generation),
2306                                 "Invalid dinode %llu disk generation: %u "
2307                                 "inode->i_generation: %u\n",
2308                                 (unsigned long long)oi->ip_blkno,
2309                                 le32_to_cpu(fe->i_generation),
2310                                 inode->i_generation);
2311                 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2312                                 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2313                                 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
2314                                 (unsigned long long)oi->ip_blkno,
2315                                 (unsigned long long)le64_to_cpu(fe->i_dtime),
2316                                 le32_to_cpu(fe->i_flags));
2317
2318                 ocfs2_refresh_inode(inode, fe);
2319                 ocfs2_track_lock_refresh(lockres);
2320         }
2321
2322         status = 0;
2323 bail_refresh:
2324         ocfs2_complete_lock_res_refresh(lockres, status);
2325 bail:
2326         return status;
2327 }
2328
2329 static int ocfs2_assign_bh(struct inode *inode,
2330                            struct buffer_head **ret_bh,
2331                            struct buffer_head *passed_bh)
2332 {
2333         int status;
2334
2335         if (passed_bh) {
2336                 /* Ok, the update went to disk for us, use the
2337                  * returned bh. */
2338                 *ret_bh = passed_bh;
2339                 get_bh(*ret_bh);
2340
2341                 return 0;
2342         }
2343
2344         status = ocfs2_read_inode_block(inode, ret_bh);
2345         if (status < 0)
2346                 mlog_errno(status);
2347
2348         return status;
2349 }
2350
2351 /*
2352  * returns < 0 error if the callback will never be called, otherwise
2353  * the result of the lock will be communicated via the callback.
2354  */
2355 int ocfs2_inode_lock_full_nested(struct inode *inode,
2356                                  struct buffer_head **ret_bh,
2357                                  int ex,
2358                                  int arg_flags,
2359                                  int subclass)
2360 {
2361         int status, level, acquired;
2362         u32 dlm_flags;
2363         struct ocfs2_lock_res *lockres = NULL;
2364         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2365         struct buffer_head *local_bh = NULL;
2366
2367         mlog(0, "inode %llu, take %s META lock\n",
2368              (unsigned long long)OCFS2_I(inode)->ip_blkno,
2369              ex ? "EXMODE" : "PRMODE");
2370
2371         status = 0;
2372         acquired = 0;
2373         /* We'll allow faking a readonly metadata lock for
2374          * rodevices. */
2375         if (ocfs2_is_hard_readonly(osb)) {
2376                 if (ex)
2377                         status = -EROFS;
2378                 goto getbh;
2379         }
2380
2381         if ((arg_flags & OCFS2_META_LOCK_GETBH) ||
2382             ocfs2_mount_local(osb))
2383                 goto update;
2384
2385         if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2386                 ocfs2_wait_for_recovery(osb);
2387
2388         lockres = &OCFS2_I(inode)->ip_inode_lockres;
2389         level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2390         dlm_flags = 0;
2391         if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2392                 dlm_flags |= DLM_LKF_NOQUEUE;
2393
2394         status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2395                                       arg_flags, subclass, _RET_IP_);
2396         if (status < 0) {
2397                 if (status != -EAGAIN)
2398                         mlog_errno(status);
2399                 goto bail;
2400         }
2401
2402         /* Notify the error cleanup path to drop the cluster lock. */
2403         acquired = 1;
2404
2405         /* We wait twice because a node may have died while we were in
2406          * the lower dlm layers. The second time though, we've
2407          * committed to owning this lock so we don't allow signals to
2408          * abort the operation. */
2409         if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2410                 ocfs2_wait_for_recovery(osb);
2411
2412 update:
2413         /*
2414          * We only see this flag if we're being called from
2415          * ocfs2_read_locked_inode(). It means we're locking an inode
2416          * which hasn't been populated yet, so clear the refresh flag
2417          * and let the caller handle it.
2418          */
2419         if (inode->i_state & I_NEW) {
2420                 status = 0;
2421                 if (lockres)
2422                         ocfs2_complete_lock_res_refresh(lockres, 0);
2423                 goto bail;
2424         }
2425
2426         /* This is fun. The caller may want a bh back, or it may
2427          * not. ocfs2_inode_lock_update definitely wants one in, but
2428          * may or may not read one, depending on what's in the
2429          * LVB. The result of all of this is that we've *only* gone to
2430          * disk if we have to, so the complexity is worthwhile. */
2431         status = ocfs2_inode_lock_update(inode, &local_bh);
2432         if (status < 0) {
2433                 if (status != -ENOENT)
2434                         mlog_errno(status);
2435                 goto bail;
2436         }
2437 getbh:
2438         if (ret_bh) {
2439                 status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2440                 if (status < 0) {
2441                         mlog_errno(status);
2442                         goto bail;
2443                 }
2444         }
2445
2446 bail:
2447         if (status < 0) {
2448                 if (ret_bh && (*ret_bh)) {
2449                         brelse(*ret_bh);
2450                         *ret_bh = NULL;
2451                 }
2452                 if (acquired)
2453                         ocfs2_inode_unlock(inode, ex);
2454         }
2455
2456         if (local_bh)
2457                 brelse(local_bh);
2458
2459         return status;
2460 }
2461
2462 /*
2463  * This is working around a lock inversion between tasks acquiring DLM
2464  * locks while holding a page lock and the downconvert thread which
2465  * blocks dlm lock acquiry while acquiring page locks.
2466  *
2467  * ** These _with_page variantes are only intended to be called from aop
2468  * methods that hold page locks and return a very specific *positive* error
2469  * code that aop methods pass up to the VFS -- test for errors with != 0. **
2470  *
2471  * The DLM is called such that it returns -EAGAIN if it would have
2472  * blocked waiting for the downconvert thread.  In that case we unlock
2473  * our page so the downconvert thread can make progress.  Once we've
2474  * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2475  * that called us can bubble that back up into the VFS who will then
2476  * immediately retry the aop call.
2477  */
2478 int ocfs2_inode_lock_with_page(struct inode *inode,
2479                               struct buffer_head **ret_bh,
2480                               int ex,
2481                               struct page *page)
2482 {
2483         int ret;
2484
2485         ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2486         if (ret == -EAGAIN) {
2487                 unlock_page(page);
2488                 /*
2489                  * If we can't get inode lock immediately, we should not return
2490                  * directly here, since this will lead to a softlockup problem.
2491                  * The method is to get a blocking lock and immediately unlock
2492                  * before returning, this can avoid CPU resource waste due to
2493                  * lots of retries, and benefits fairness in getting lock.
2494                  */
2495                 if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
2496                         ocfs2_inode_unlock(inode, ex);
2497                 ret = AOP_TRUNCATED_PAGE;
2498         }
2499
2500         return ret;
2501 }
2502
2503 int ocfs2_inode_lock_atime(struct inode *inode,
2504                           struct vfsmount *vfsmnt,
2505                           int *level)
2506 {
2507         int ret;
2508
2509         ret = ocfs2_inode_lock(inode, NULL, 0);
2510         if (ret < 0) {
2511                 mlog_errno(ret);
2512                 return ret;
2513         }
2514
2515         /*
2516          * If we should update atime, we will get EX lock,
2517          * otherwise we just get PR lock.
2518          */
2519         if (ocfs2_should_update_atime(inode, vfsmnt)) {
2520                 struct buffer_head *bh = NULL;
2521
2522                 ocfs2_inode_unlock(inode, 0);
2523                 ret = ocfs2_inode_lock(inode, &bh, 1);
2524                 if (ret < 0) {
2525                         mlog_errno(ret);
2526                         return ret;
2527                 }
2528                 *level = 1;
2529                 if (ocfs2_should_update_atime(inode, vfsmnt))
2530                         ocfs2_update_inode_atime(inode, bh);
2531                 if (bh)
2532                         brelse(bh);
2533         } else
2534                 *level = 0;
2535
2536         return ret;
2537 }
2538
2539 void ocfs2_inode_unlock(struct inode *inode,
2540                        int ex)
2541 {
2542         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2543         struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2544         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2545
2546         mlog(0, "inode %llu drop %s META lock\n",
2547              (unsigned long long)OCFS2_I(inode)->ip_blkno,
2548              ex ? "EXMODE" : "PRMODE");
2549
2550         if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
2551             !ocfs2_mount_local(osb))
2552                 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
2553 }
2554
2555 /*
2556  * This _tracker variantes are introduced to deal with the recursive cluster
2557  * locking issue. The idea is to keep track of a lock holder on the stack of
2558  * the current process. If there's a lock holder on the stack, we know the
2559  * task context is already protected by cluster locking. Currently, they're
2560  * used in some VFS entry routines.
2561  *
2562  * return < 0 on error, return == 0 if there's no lock holder on the stack
2563  * before this call, return == 1 if this call would be a recursive locking.
2564  */
2565 int ocfs2_inode_lock_tracker(struct inode *inode,
2566                              struct buffer_head **ret_bh,
2567                              int ex,
2568                              struct ocfs2_lock_holder *oh)
2569 {
2570         int status;
2571         int arg_flags = 0, has_locked;
2572         struct ocfs2_lock_res *lockres;
2573
2574         lockres = &OCFS2_I(inode)->ip_inode_lockres;
2575         has_locked = ocfs2_is_locked_by_me(lockres);
2576         /* Just get buffer head if the cluster lock has been taken */
2577         if (has_locked)
2578                 arg_flags = OCFS2_META_LOCK_GETBH;
2579
2580         if (likely(!has_locked || ret_bh)) {
2581                 status = ocfs2_inode_lock_full(inode, ret_bh, ex, arg_flags);
2582                 if (status < 0) {
2583                         if (status != -ENOENT)
2584                                 mlog_errno(status);
2585                         return status;
2586                 }
2587         }
2588         if (!has_locked)
2589                 ocfs2_add_holder(lockres, oh);
2590
2591         return has_locked;
2592 }
2593
2594 void ocfs2_inode_unlock_tracker(struct inode *inode,
2595                                 int ex,
2596                                 struct ocfs2_lock_holder *oh,
2597                                 int had_lock)
2598 {
2599         struct ocfs2_lock_res *lockres;
2600
2601         lockres = &OCFS2_I(inode)->ip_inode_lockres;
2602         /* had_lock means that the currect process already takes the cluster
2603          * lock previously. If had_lock is 1, we have nothing to do here, and
2604          * it will get unlocked where we got the lock.
2605          */
2606         if (!had_lock) {
2607                 ocfs2_remove_holder(lockres, oh);
2608                 ocfs2_inode_unlock(inode, ex);
2609         }
2610 }
2611
2612 int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2613 {
2614         struct ocfs2_lock_res *lockres;
2615         struct ocfs2_orphan_scan_lvb *lvb;
2616         int status = 0;
2617
2618         if (ocfs2_is_hard_readonly(osb))
2619                 return -EROFS;
2620
2621         if (ocfs2_mount_local(osb))
2622                 return 0;
2623
2624         lockres = &osb->osb_orphan_scan.os_lockres;
2625         status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2626         if (status < 0)
2627                 return status;
2628
2629         lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2630         if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2631             lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2632                 *seqno = be32_to_cpu(lvb->lvb_os_seqno);
2633         else
2634                 *seqno = osb->osb_orphan_scan.os_seqno + 1;
2635
2636         return status;
2637 }
2638
2639 void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2640 {
2641         struct ocfs2_lock_res *lockres;
2642         struct ocfs2_orphan_scan_lvb *lvb;
2643
2644         if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2645                 lockres = &osb->osb_orphan_scan.os_lockres;
2646                 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2647                 lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2648                 lvb->lvb_os_seqno = cpu_to_be32(seqno);
2649                 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2650         }
2651 }
2652
2653 int ocfs2_super_lock(struct ocfs2_super *osb,
2654                      int ex)
2655 {
2656         int status = 0;
2657         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2658         struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2659
2660         if (ocfs2_is_hard_readonly(osb))
2661                 return -EROFS;
2662
2663         if (ocfs2_mount_local(osb))
2664                 goto bail;
2665
2666         status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2667         if (status < 0) {
2668                 mlog_errno(status);
2669                 goto bail;
2670         }
2671
2672         /* The super block lock path is really in the best position to
2673          * know when resources covered by the lock need to be
2674          * refreshed, so we do it here. Of course, making sense of
2675          * everything is up to the caller :) */
2676         status = ocfs2_should_refresh_lock_res(lockres);
2677         if (status) {
2678                 status = ocfs2_refresh_slot_info(osb);
2679
2680                 ocfs2_complete_lock_res_refresh(lockres, status);
2681
2682                 if (status < 0) {
2683                         ocfs2_cluster_unlock(osb, lockres, level);
2684                         mlog_errno(status);
2685                 }
2686                 ocfs2_track_lock_refresh(lockres);
2687         }
2688 bail:
2689         return status;
2690 }
2691
2692 void ocfs2_super_unlock(struct ocfs2_super *osb,
2693                         int ex)
2694 {
2695         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2696         struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2697
2698         if (!ocfs2_mount_local(osb))
2699                 ocfs2_cluster_unlock(osb, lockres, level);
2700 }
2701
2702 int ocfs2_rename_lock(struct ocfs2_super *osb)
2703 {
2704         int status;
2705         struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2706
2707         if (ocfs2_is_hard_readonly(osb))
2708                 return -EROFS;
2709
2710         if (ocfs2_mount_local(osb))
2711                 return 0;
2712
2713         status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2714         if (status < 0)
2715                 mlog_errno(status);
2716
2717         return status;
2718 }
2719
2720 void ocfs2_rename_unlock(struct ocfs2_super *osb)
2721 {
2722         struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2723
2724         if (!ocfs2_mount_local(osb))
2725                 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2726 }
2727
2728 int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2729 {
2730         int status;
2731         struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2732
2733         if (ocfs2_is_hard_readonly(osb))
2734                 return -EROFS;
2735
2736         if (ocfs2_mount_local(osb))
2737                 return 0;
2738
2739         status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2740                                     0, 0);
2741         if (status < 0)
2742                 mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2743
2744         return status;
2745 }
2746
2747 void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2748 {
2749         struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2750
2751         if (!ocfs2_mount_local(osb))
2752                 ocfs2_cluster_unlock(osb, lockres,
2753                                      ex ? LKM_EXMODE : LKM_PRMODE);
2754 }
2755
2756 int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2757 {
2758         int ret;
2759         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2760         struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2761         struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2762
2763         BUG_ON(!dl);
2764
2765         if (ocfs2_is_hard_readonly(osb)) {
2766                 if (ex)
2767                         return -EROFS;
2768                 return 0;
2769         }
2770
2771         if (ocfs2_mount_local(osb))
2772                 return 0;
2773
2774         ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2775         if (ret < 0)
2776                 mlog_errno(ret);
2777
2778         return ret;
2779 }
2780
2781 void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2782 {
2783         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2784         struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2785         struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2786
2787         if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
2788                 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2789 }
2790
2791 /* Reference counting of the dlm debug structure. We want this because
2792  * open references on the debug inodes can live on after a mount, so
2793  * we can't rely on the ocfs2_super to always exist. */
2794 static void ocfs2_dlm_debug_free(struct kref *kref)
2795 {
2796         struct ocfs2_dlm_debug *dlm_debug;
2797
2798         dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2799
2800         kfree(dlm_debug);
2801 }
2802
2803 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2804 {
2805         if (dlm_debug)
2806                 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2807 }
2808
2809 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2810 {
2811         kref_get(&debug->d_refcnt);
2812 }
2813
2814 struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2815 {
2816         struct ocfs2_dlm_debug *dlm_debug;
2817
2818         dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2819         if (!dlm_debug) {
2820                 mlog_errno(-ENOMEM);
2821                 goto out;
2822         }
2823
2824         kref_init(&dlm_debug->d_refcnt);
2825         INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2826         dlm_debug->d_locking_state = NULL;
2827 out:
2828         return dlm_debug;
2829 }
2830
2831 /* Access to this is arbitrated for us via seq_file->sem. */
2832 struct ocfs2_dlm_seq_priv {
2833         struct ocfs2_dlm_debug *p_dlm_debug;
2834         struct ocfs2_lock_res p_iter_res;
2835         struct ocfs2_lock_res p_tmp_res;
2836 };
2837
2838 static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2839                                                  struct ocfs2_dlm_seq_priv *priv)
2840 {
2841         struct ocfs2_lock_res *iter, *ret = NULL;
2842         struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2843
2844         assert_spin_locked(&ocfs2_dlm_tracking_lock);
2845
2846         list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2847                 /* discover the head of the list */
2848                 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2849                         mlog(0, "End of list found, %p\n", ret);
2850                         break;
2851                 }
2852
2853                 /* We track our "dummy" iteration lockres' by a NULL
2854                  * l_ops field. */
2855                 if (iter->l_ops != NULL) {
2856                         ret = iter;
2857                         break;
2858                 }
2859         }
2860
2861         return ret;
2862 }
2863
2864 static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2865 {
2866         struct ocfs2_dlm_seq_priv *priv = m->private;
2867         struct ocfs2_lock_res *iter;
2868
2869         spin_lock(&ocfs2_dlm_tracking_lock);
2870         iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2871         if (iter) {
2872                 /* Since lockres' have the lifetime of their container
2873                  * (which can be inodes, ocfs2_supers, etc) we want to
2874                  * copy this out to a temporary lockres while still
2875                  * under the spinlock. Obviously after this we can't
2876                  * trust any pointers on the copy returned, but that's
2877                  * ok as the information we want isn't typically held
2878                  * in them. */
2879                 priv->p_tmp_res = *iter;
2880                 iter = &priv->p_tmp_res;
2881         }
2882         spin_unlock(&ocfs2_dlm_tracking_lock);
2883
2884         return iter;
2885 }
2886
2887 static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
2888 {
2889 }
2890
2891 static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
2892 {
2893         struct ocfs2_dlm_seq_priv *priv = m->private;
2894         struct ocfs2_lock_res *iter = v;
2895         struct ocfs2_lock_res *dummy = &priv->p_iter_res;
2896
2897         spin_lock(&ocfs2_dlm_tracking_lock);
2898         iter = ocfs2_dlm_next_res(iter, priv);
2899         list_del_init(&dummy->l_debug_list);
2900         if (iter) {
2901                 list_add(&dummy->l_debug_list, &iter->l_debug_list);
2902                 priv->p_tmp_res = *iter;
2903                 iter = &priv->p_tmp_res;
2904         }
2905         spin_unlock(&ocfs2_dlm_tracking_lock);
2906
2907         return iter;
2908 }
2909
2910 /*
2911  * Version is used by debugfs.ocfs2 to determine the format being used
2912  *
2913  * New in version 2
2914  *      - Lock stats printed
2915  * New in version 3
2916  *      - Max time in lock stats is in usecs (instead of nsecs)
2917  */
2918 #define OCFS2_DLM_DEBUG_STR_VERSION 3
2919 static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
2920 {
2921         int i;
2922         char *lvb;
2923         struct ocfs2_lock_res *lockres = v;
2924
2925         if (!lockres)
2926                 return -EINVAL;
2927
2928         seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
2929
2930         if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
2931                 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
2932                            lockres->l_name,
2933                            (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
2934         else
2935                 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
2936
2937         seq_printf(m, "%d\t"
2938                    "0x%lx\t"
2939                    "0x%x\t"
2940                    "0x%x\t"
2941                    "%u\t"
2942                    "%u\t"
2943                    "%d\t"
2944                    "%d\t",
2945                    lockres->l_level,
2946                    lockres->l_flags,
2947                    lockres->l_action,
2948                    lockres->l_unlock_action,
2949                    lockres->l_ro_holders,
2950                    lockres->l_ex_holders,
2951                    lockres->l_requested,
2952                    lockres->l_blocking);
2953
2954         /* Dump the raw LVB */
2955         lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2956         for(i = 0; i < DLM_LVB_LEN; i++)
2957                 seq_printf(m, "0x%x\t", lvb[i]);
2958
2959 #ifdef CONFIG_OCFS2_FS_STATS
2960 # define lock_num_prmode(_l)            ((_l)->l_lock_prmode.ls_gets)
2961 # define lock_num_exmode(_l)            ((_l)->l_lock_exmode.ls_gets)
2962 # define lock_num_prmode_failed(_l)     ((_l)->l_lock_prmode.ls_fail)
2963 # define lock_num_exmode_failed(_l)     ((_l)->l_lock_exmode.ls_fail)
2964 # define lock_total_prmode(_l)          ((_l)->l_lock_prmode.ls_total)
2965 # define lock_total_exmode(_l)          ((_l)->l_lock_exmode.ls_total)
2966 # define lock_max_prmode(_l)            ((_l)->l_lock_prmode.ls_max)
2967 # define lock_max_exmode(_l)            ((_l)->l_lock_exmode.ls_max)
2968 # define lock_refresh(_l)               ((_l)->l_lock_refresh)
2969 #else
2970 # define lock_num_prmode(_l)            (0)
2971 # define lock_num_exmode(_l)            (0)
2972 # define lock_num_prmode_failed(_l)     (0)
2973 # define lock_num_exmode_failed(_l)     (0)
2974 # define lock_total_prmode(_l)          (0ULL)
2975 # define lock_total_exmode(_l)          (0ULL)
2976 # define lock_max_prmode(_l)            (0)
2977 # define lock_max_exmode(_l)            (0)
2978 # define lock_refresh(_l)               (0)
2979 #endif
2980         /* The following seq_print was added in version 2 of this output */
2981         seq_printf(m, "%u\t"
2982                    "%u\t"
2983                    "%u\t"
2984                    "%u\t"
2985                    "%llu\t"
2986                    "%llu\t"
2987                    "%u\t"
2988                    "%u\t"
2989                    "%u\t",
2990                    lock_num_prmode(lockres),
2991                    lock_num_exmode(lockres),
2992                    lock_num_prmode_failed(lockres),
2993                    lock_num_exmode_failed(lockres),
2994                    lock_total_prmode(lockres),
2995                    lock_total_exmode(lockres),
2996                    lock_max_prmode(lockres),
2997                    lock_max_exmode(lockres),
2998                    lock_refresh(lockres));
2999
3000         /* End the line */
3001         seq_printf(m, "\n");
3002         return 0;
3003 }
3004
3005 static const struct seq_operations ocfs2_dlm_seq_ops = {
3006         .start =        ocfs2_dlm_seq_start,
3007         .stop =         ocfs2_dlm_seq_stop,
3008         .next =         ocfs2_dlm_seq_next,
3009         .show =         ocfs2_dlm_seq_show,
3010 };
3011
3012 static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
3013 {
3014         struct seq_file *seq = file->private_data;
3015         struct ocfs2_dlm_seq_priv *priv = seq->private;
3016         struct ocfs2_lock_res *res = &priv->p_iter_res;
3017
3018         ocfs2_remove_lockres_tracking(res);
3019         ocfs2_put_dlm_debug(priv->p_dlm_debug);
3020         return seq_release_private(inode, file);
3021 }
3022
3023 static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
3024 {
3025         struct ocfs2_dlm_seq_priv *priv;
3026         struct ocfs2_super *osb;
3027
3028         priv = __seq_open_private(file, &ocfs2_dlm_seq_ops, sizeof(*priv));
3029         if (!priv) {
3030                 mlog_errno(-ENOMEM);
3031                 return -ENOMEM;
3032         }
3033
3034         osb = inode->i_private;
3035         ocfs2_get_dlm_debug(osb->osb_dlm_debug);
3036         priv->p_dlm_debug = osb->osb_dlm_debug;
3037         INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
3038
3039         ocfs2_add_lockres_tracking(&priv->p_iter_res,
3040                                    priv->p_dlm_debug);
3041
3042         return 0;
3043 }
3044
3045 static const struct file_operations ocfs2_dlm_debug_fops = {
3046         .open =         ocfs2_dlm_debug_open,
3047         .release =      ocfs2_dlm_debug_release,
3048         .read =         seq_read,
3049         .llseek =       seq_lseek,
3050 };
3051
3052 static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
3053 {
3054         int ret = 0;
3055         struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
3056
3057         dlm_debug->d_locking_state = debugfs_create_file("locking_state",
3058                                                          S_IFREG|S_IRUSR,
3059                                                          osb->osb_debug_root,
3060                                                          osb,
3061                                                          &ocfs2_dlm_debug_fops);
3062         if (!dlm_debug->d_locking_state) {
3063                 ret = -EINVAL;
3064                 mlog(ML_ERROR,
3065                      "Unable to create locking state debugfs file.\n");
3066                 goto out;
3067         }
3068
3069         ocfs2_get_dlm_debug(dlm_debug);
3070 out:
3071         return ret;
3072 }
3073
3074 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
3075 {
3076         struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
3077
3078         if (dlm_debug) {
3079                 debugfs_remove(dlm_debug->d_locking_state);
3080                 ocfs2_put_dlm_debug(dlm_debug);
3081         }
3082 }
3083
3084 int ocfs2_dlm_init(struct ocfs2_super *osb)
3085 {
3086         int status = 0;
3087         struct ocfs2_cluster_connection *conn = NULL;
3088
3089         if (ocfs2_mount_local(osb)) {
3090                 osb->node_num = 0;
3091                 goto local;
3092         }
3093
3094         status = ocfs2_dlm_init_debug(osb);
3095         if (status < 0) {
3096                 mlog_errno(status);
3097                 goto bail;
3098         }
3099
3100         /* launch downconvert thread */
3101         osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc-%s",
3102                         osb->uuid_str);
3103         if (IS_ERR(osb->dc_task)) {
3104                 status = PTR_ERR(osb->dc_task);
3105                 osb->dc_task = NULL;
3106                 mlog_errno(status);
3107                 goto bail;
3108         }
3109
3110         /* for now, uuid == domain */
3111         status = ocfs2_cluster_connect(osb->osb_cluster_stack,
3112                                        osb->osb_cluster_name,
3113                                        strlen(osb->osb_cluster_name),
3114                                        osb->uuid_str,
3115                                        strlen(osb->uuid_str),
3116                                        &lproto, ocfs2_do_node_down, osb,
3117                                        &conn);
3118         if (status) {
3119                 mlog_errno(status);
3120                 goto bail;
3121         }
3122
3123         status = ocfs2_cluster_this_node(conn, &osb->node_num);
3124         if (status < 0) {
3125                 mlog_errno(status);
3126                 mlog(ML_ERROR,
3127                      "could not find this host's node number\n");
3128                 ocfs2_cluster_disconnect(conn, 0);
3129                 goto bail;
3130         }
3131
3132 local:
3133         ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3134         ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3135         ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
3136         ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3137
3138         osb->cconn = conn;
3139 bail:
3140         if (status < 0) {
3141                 ocfs2_dlm_shutdown_debug(osb);
3142                 if (osb->dc_task)
3143                         kthread_stop(osb->dc_task);
3144         }
3145
3146         return status;
3147 }
3148
3149 void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3150                         int hangup_pending)
3151 {
3152         ocfs2_drop_osb_locks(osb);
3153
3154         /*
3155          * Now that we have dropped all locks and ocfs2_dismount_volume()
3156          * has disabled recovery, the DLM won't be talking to us.  It's
3157          * safe to tear things down before disconnecting the cluster.
3158          */
3159
3160         if (osb->dc_task) {
3161                 kthread_stop(osb->dc_task);
3162                 osb->dc_task = NULL;
3163         }
3164
3165         ocfs2_lock_res_free(&osb->osb_super_lockres);
3166         ocfs2_lock_res_free(&osb->osb_rename_lockres);
3167         ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3168         ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3169
3170         ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3171         osb->cconn = NULL;
3172
3173         ocfs2_dlm_shutdown_debug(osb);
3174 }
3175
3176 static int ocfs2_drop_lock(struct ocfs2_super *osb,
3177                            struct ocfs2_lock_res *lockres)
3178 {
3179         int ret;
3180         unsigned long flags;
3181         u32 lkm_flags = 0;
3182
3183         /* We didn't get anywhere near actually using this lockres. */
3184         if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3185                 goto out;
3186
3187         if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3188                 lkm_flags |= DLM_LKF_VALBLK;
3189
3190         spin_lock_irqsave(&lockres->l_lock, flags);
3191
3192         mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3193                         "lockres %s, flags 0x%lx\n",
3194                         lockres->l_name, lockres->l_flags);
3195
3196         while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3197                 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3198                      "%u, unlock_action = %u\n",
3199                      lockres->l_name, lockres->l_flags, lockres->l_action,
3200                      lockres->l_unlock_action);
3201
3202                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3203
3204                 /* XXX: Today we just wait on any busy
3205                  * locks... Perhaps we need to cancel converts in the
3206                  * future? */
3207                 ocfs2_wait_on_busy_lock(lockres);
3208
3209                 spin_lock_irqsave(&lockres->l_lock, flags);
3210         }
3211
3212         if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3213                 if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3214                     lockres->l_level == DLM_LOCK_EX &&
3215                     !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3216                         lockres->l_ops->set_lvb(lockres);
3217         }
3218
3219         if (lockres->l_flags & OCFS2_LOCK_BUSY)
3220                 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3221                      lockres->l_name);
3222         if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3223                 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3224
3225         if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3226                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3227                 goto out;
3228         }
3229
3230         lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3231
3232         /* make sure we never get here while waiting for an ast to
3233          * fire. */
3234         BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3235
3236         /* is this necessary? */
3237         lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3238         lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3239         spin_unlock_irqrestore(&lockres->l_lock, flags);
3240
3241         mlog(0, "lock %s\n", lockres->l_name);
3242
3243         ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3244         if (ret) {
3245                 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3246                 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3247                 ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3248                 BUG();
3249         }
3250         mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3251              lockres->l_name);
3252
3253         ocfs2_wait_on_busy_lock(lockres);
3254 out:
3255         return 0;
3256 }
3257
3258 static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3259                                        struct ocfs2_lock_res *lockres);
3260
3261 /* Mark the lockres as being dropped. It will no longer be
3262  * queued if blocking, but we still may have to wait on it
3263  * being dequeued from the downconvert thread before we can consider
3264  * it safe to drop.
3265  *
3266  * You can *not* attempt to call cluster_lock on this lockres anymore. */
3267 void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
3268                                 struct ocfs2_lock_res *lockres)
3269 {
3270         int status;
3271         struct ocfs2_mask_waiter mw;
3272         unsigned long flags, flags2;
3273
3274         ocfs2_init_mask_waiter(&mw);
3275
3276         spin_lock_irqsave(&lockres->l_lock, flags);
3277         lockres->l_flags |= OCFS2_LOCK_FREEING;
3278         if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) {
3279                 /*
3280                  * We know the downconvert is queued but not in progress
3281                  * because we are the downconvert thread and processing
3282                  * different lock. So we can just remove the lock from the
3283                  * queue. This is not only an optimization but also a way
3284                  * to avoid the following deadlock:
3285                  *   ocfs2_dentry_post_unlock()
3286                  *     ocfs2_dentry_lock_put()
3287                  *       ocfs2_drop_dentry_lock()
3288                  *         iput()
3289                  *           ocfs2_evict_inode()
3290                  *             ocfs2_clear_inode()
3291                  *               ocfs2_mark_lockres_freeing()
3292                  *                 ... blocks waiting for OCFS2_LOCK_QUEUED
3293                  *                 since we are the downconvert thread which
3294                  *                 should clear the flag.
3295                  */
3296                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3297                 spin_lock_irqsave(&osb->dc_task_lock, flags2);
3298                 list_del_init(&lockres->l_blocked_list);
3299                 osb->blocked_lock_count--;
3300                 spin_unlock_irqrestore(&osb->dc_task_lock, flags2);
3301                 /*
3302                  * Warn if we recurse into another post_unlock call.  Strictly
3303                  * speaking it isn't a problem but we need to be careful if
3304                  * that happens (stack overflow, deadlocks, ...) so warn if
3305                  * ocfs2 grows a path for which this can happen.
3306                  */
3307                 WARN_ON_ONCE(lockres->l_ops->post_unlock);
3308                 /* Since the lock is freeing we don't do much in the fn below */
3309                 ocfs2_process_blocked_lock(osb, lockres);
3310                 return;
3311         }
3312         while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3313                 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3314                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3315
3316                 mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3317
3318                 status = ocfs2_wait_for_mask(&mw);
3319                 if (status)
3320                         mlog_errno(status);
3321
3322                 spin_lock_irqsave(&lockres->l_lock, flags);
3323         }
3324         spin_unlock_irqrestore(&lockres->l_lock, flags);
3325 }
3326
3327 void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3328                                struct ocfs2_lock_res *lockres)
3329 {
3330         int ret;
3331
3332         ocfs2_mark_lockres_freeing(osb, lockres);
3333         ret = ocfs2_drop_lock(osb, lockres);
3334         if (ret)
3335                 mlog_errno(ret);
3336 }
3337
3338 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3339 {
3340         ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3341         ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3342         ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3343         ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3344 }
3345
3346 int ocfs2_drop_inode_locks(struct inode *inode)
3347 {
3348         int status, err;
3349
3350         /* No need to call ocfs2_mark_lockres_freeing here -
3351          * ocfs2_clear_inode has done it for us. */
3352
3353         err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3354                               &OCFS2_I(inode)->ip_open_lockres);
3355         if (err < 0)
3356                 mlog_errno(err);
3357
3358         status = err;
3359
3360         err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3361                               &OCFS2_I(inode)->ip_inode_lockres);
3362         if (err < 0)
3363                 mlog_errno(err);
3364         if (err < 0 && !status)
3365                 status = err;
3366
3367         err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3368                               &OCFS2_I(inode)->ip_rw_lockres);
3369         if (err < 0)
3370                 mlog_errno(err);
3371         if (err < 0 && !status)
3372                 status = err;
3373
3374         return status;
3375 }
3376
3377 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3378                                               int new_level)
3379 {
3380         assert_spin_locked(&lockres->l_lock);
3381
3382         BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3383
3384         if (lockres->l_level <= new_level) {
3385                 mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3386                      "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3387                      "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
3388                      new_level, list_empty(&lockres->l_blocked_list),
3389                      list_empty(&lockres->l_mask_waiters), lockres->l_type,
3390                      lockres->l_flags, lockres->l_ro_holders,
3391                      lockres->l_ex_holders, lockres->l_action,
3392                      lockres->l_unlock_action, lockres->l_requested,
3393                      lockres->l_blocking, lockres->l_pending_gen);
3394                 BUG();
3395         }
3396
3397         mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n",
3398              lockres->l_name, lockres->l_level, new_level, lockres->l_blocking);
3399
3400         lockres->l_action = OCFS2_AST_DOWNCONVERT;
3401         lockres->l_requested = new_level;
3402         lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3403         return lockres_set_pending(lockres);
3404 }
3405
3406 static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3407                                   struct ocfs2_lock_res *lockres,
3408                                   int new_level,
3409                                   int lvb,
3410                                   unsigned int generation)
3411 {
3412         int ret;
3413         u32 dlm_flags = DLM_LKF_CONVERT;
3414
3415         mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name,
3416              lockres->l_level, new_level);
3417
3418         /*
3419          * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always
3420          * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that
3421          * we can recover correctly from node failure. Otherwise, we may get
3422          * invalid LVB in LKB, but without DLM_SBF_VALNOTVALID being set.
3423          */
3424         if (ocfs2_userspace_stack(osb) &&
3425             lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3426                 lvb = 1;
3427
3428         if (lvb)
3429                 dlm_flags |= DLM_LKF_VALBLK;
3430
3431         ret = ocfs2_dlm_lock(osb->cconn,
3432                              new_level,
3433                              &lockres->l_lksb,
3434                              dlm_flags,
3435                              lockres->l_name,
3436                              OCFS2_LOCK_ID_MAX_LEN - 1);
3437         lockres_clear_pending(lockres, generation, osb);
3438         if (ret) {
3439                 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3440                 ocfs2_recover_from_dlm_error(lockres, 1);
3441                 goto bail;
3442         }
3443
3444         ret = 0;
3445 bail:
3446         return ret;
3447 }
3448
3449 /* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3450 static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3451                                         struct ocfs2_lock_res *lockres)
3452 {
3453         assert_spin_locked(&lockres->l_lock);
3454
3455         if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3456                 /* If we're already trying to cancel a lock conversion
3457                  * then just drop the spinlock and allow the caller to
3458                  * requeue this lock. */
3459                 mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name);
3460                 return 0;
3461         }
3462
3463         /* were we in a convert when we got the bast fire? */
3464         BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3465                lockres->l_action != OCFS2_AST_DOWNCONVERT);
3466         /* set things up for the unlockast to know to just
3467          * clear out the ast_action and unset busy, etc. */
3468         lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3469
3470         mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3471                         "lock %s, invalid flags: 0x%lx\n",
3472                         lockres->l_name, lockres->l_flags);
3473
3474         mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3475
3476         return 1;
3477 }
3478
3479 static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3480                                 struct ocfs2_lock_res *lockres)
3481 {
3482         int ret;
3483
3484         ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3485                                DLM_LKF_CANCEL);
3486         if (ret) {
3487                 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3488                 ocfs2_recover_from_dlm_error(lockres, 0);
3489         }
3490
3491         mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3492
3493         return ret;
3494 }
3495
3496 static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3497                               struct ocfs2_lock_res *lockres,
3498                               struct ocfs2_unblock_ctl *ctl)
3499 {
3500         unsigned long flags;
3501         int blocking;
3502         int new_level;
3503         int level;
3504         int ret = 0;
3505         int set_lvb = 0;
3506         unsigned int gen;
3507
3508         spin_lock_irqsave(&lockres->l_lock, flags);
3509
3510 recheck:
3511         /*
3512          * Is it still blocking? If not, we have no more work to do.
3513          */
3514         if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3515                 BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3516                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3517                 ret = 0;
3518                 goto leave;
3519         }
3520
3521         if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3522                 /* XXX
3523                  * This is a *big* race.  The OCFS2_LOCK_PENDING flag
3524                  * exists entirely for one reason - another thread has set
3525                  * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3526                  *
3527                  * If we do ocfs2_cancel_convert() before the other thread
3528                  * calls dlm_lock(), our cancel will do nothing.  We will
3529                  * get no ast, and we will have no way of knowing the
3530                  * cancel failed.  Meanwhile, the other thread will call
3531                  * into dlm_lock() and wait...forever.
3532                  *
3533                  * Why forever?  Because another node has asked for the
3534                  * lock first; that's why we're here in unblock_lock().
3535                  *
3536                  * The solution is OCFS2_LOCK_PENDING.  When PENDING is
3537                  * set, we just requeue the unblock.  Only when the other
3538                  * thread has called dlm_lock() and cleared PENDING will
3539                  * we then cancel their request.
3540                  *
3541                  * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3542                  * at the same time they set OCFS2_DLM_BUSY.  They must
3543                  * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3544                  */
3545                 if (lockres->l_flags & OCFS2_LOCK_PENDING) {
3546                         mlog(ML_BASTS, "lockres %s, ReQ: Pending\n",
3547                              lockres->l_name);
3548                         goto leave_requeue;
3549                 }
3550
3551                 ctl->requeue = 1;
3552                 ret = ocfs2_prepare_cancel_convert(osb, lockres);
3553                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3554                 if (ret) {
3555                         ret = ocfs2_cancel_convert(osb, lockres);
3556                         if (ret < 0)
3557                                 mlog_errno(ret);
3558                 }
3559                 goto leave;
3560         }
3561
3562         /*
3563          * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3564          * set when the ast is received for an upconvert just before the
3565          * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3566          * on the heels of the ast, we want to delay the downconvert just
3567          * enough to allow the up requestor to do its task. Because this
3568          * lock is in the blocked queue, the lock will be downconverted
3569          * as soon as the requestor is done with the lock.
3570          */
3571         if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3572                 goto leave_requeue;
3573
3574         /*
3575          * How can we block and yet be at NL?  We were trying to upconvert
3576          * from NL and got canceled.  The code comes back here, and now
3577          * we notice and clear BLOCKING.
3578          */
3579         if (lockres->l_level == DLM_LOCK_NL) {
3580                 BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3581                 mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name);
3582                 lockres->l_blocking = DLM_LOCK_NL;
3583                 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3584                 spin_unlock_irqrestore(&lockres->l_lock, flags);
3585                 goto leave;
3586         }
3587
3588         /* if we're blocking an exclusive and we have *any* holders,
3589          * then requeue. */
3590         if ((lockres->l_blocking == DLM_LOCK_EX)
3591             && (lockres->l_ex_holders || lockres->l_ro_holders)) {
3592                 mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3593                      lockres->l_name, lockres->l_ex_holders,
3594                      lockres->l_ro_holders);
3595                 goto leave_requeue;
3596         }
3597
3598         /* If it's a PR we're blocking, then only
3599          * requeue if we've got any EX holders */
3600         if (lockres->l_blocking == DLM_LOCK_PR &&
3601             lockres->l_ex_holders) {
3602                 mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n",
3603                      lockres->l_name, lockres->l_ex_holders);
3604                 goto leave_requeue;
3605         }
3606
3607         /*
3608          * Can we get a lock in this state if the holder counts are
3609          * zero? The meta data unblock code used to check this.
3610          */
3611         if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3612             && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) {
3613                 mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n",
3614                      lockres->l_name);
3615                 goto leave_requeue;
3616         }
3617
3618         new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3619
3620         if (lockres->l_ops->check_downconvert
3621             && !lockres->l_ops->check_downconvert(lockres, new_level)) {
3622                 mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n",
3623                      lockres->l_name);
3624                 goto leave_requeue;
3625         }
3626
3627         /* If we get here, then we know that there are no more
3628          * incompatible holders (and anyone asking for an incompatible
3629          * lock is blocked). We can now downconvert the lock */
3630         if (!lockres->l_ops->downconvert_worker)
3631                 goto downconvert;
3632
3633         /* Some lockres types want to do a bit of work before
3634          * downconverting a lock. Allow that here. The worker function
3635          * may sleep, so we save off a copy of what we're blocking as
3636          * it may change while we're not holding the spin lock. */
3637         blocking = lockres->l_blocking;
3638         level = lockres->l_level;
3639         spin_unlock_irqrestore(&lockres->l_lock, flags);
3640
3641         ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3642
3643         if (ctl->unblock_action == UNBLOCK_STOP_POST) {
3644                 mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n",
3645                      lockres->l_name);
3646                 goto leave;
3647         }
3648
3649         spin_lock_irqsave(&lockres->l_lock, flags);
3650         if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3651                 /* If this changed underneath us, then we can't drop
3652                  * it just yet. */
3653                 mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, "
3654                      "Recheck\n", lockres->l_name, blocking,
3655                      lockres->l_blocking, level, lockres->l_level);
3656                 goto recheck;
3657         }
3658
3659 downconvert:
3660         ctl->requeue = 0;
3661
3662         if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3663                 if (lockres->l_level == DLM_LOCK_EX)
3664                         set_lvb = 1;
3665
3666                 /*
3667                  * We only set the lvb if the lock has been fully
3668                  * refreshed - otherwise we risk setting stale
3669                  * data. Otherwise, there's no need to actually clear
3670                  * out the lvb here as it's value is still valid.
3671                  */
3672                 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3673                         lockres->l_ops->set_lvb(lockres);
3674         }
3675
3676         gen = ocfs2_prepare_downconvert(lockres, new_level);
3677         spin_unlock_irqrestore(&lockres->l_lock, flags);
3678         ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3679                                      gen);
3680
3681 leave:
3682         if (ret)
3683                 mlog_errno(ret);
3684         return ret;
3685
3686 leave_requeue:
3687         spin_unlock_irqrestore(&lockres->l_lock, flags);
3688         ctl->requeue = 1;
3689
3690         return 0;
3691 }
3692
3693 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3694                                      int blocking)
3695 {
3696         struct inode *inode;
3697         struct address_space *mapping;
3698         struct ocfs2_inode_info *oi;
3699
3700         inode = ocfs2_lock_res_inode(lockres);
3701         mapping = inode->i_mapping;
3702
3703         if (S_ISDIR(inode->i_mode)) {
3704                 oi = OCFS2_I(inode);
3705                 oi->ip_dir_lock_gen++;
3706                 mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
3707                 goto out_forget;
3708         }
3709
3710         if (!S_ISREG(inode->i_mode))
3711                 goto out;
3712
3713         /*
3714          * We need this before the filemap_fdatawrite() so that it can
3715          * transfer the dirty bit from the PTE to the
3716          * page. Unfortunately this means that even for EX->PR
3717          * downconverts, we'll lose our mappings and have to build
3718          * them up again.
3719          */
3720         unmap_mapping_range(mapping, 0, 0, 0);
3721
3722         if (filemap_fdatawrite(mapping)) {
3723                 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3724                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
3725         }
3726         sync_mapping_buffers(mapping);
3727         if (blocking == DLM_LOCK_EX) {
3728                 truncate_inode_pages(mapping, 0);
3729         } else {
3730                 /* We only need to wait on the I/O if we're not also
3731                  * truncating pages because truncate_inode_pages waits
3732                  * for us above. We don't truncate pages if we're
3733                  * blocking anything < EXMODE because we want to keep
3734                  * them around in that case. */
3735                 filemap_fdatawait(mapping);
3736         }
3737
3738 out_forget:
3739         forget_all_cached_acls(inode);
3740
3741 out:
3742         return UNBLOCK_CONTINUE;
3743 }
3744
3745 static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3746                                  struct ocfs2_lock_res *lockres,
3747                                  int new_level)
3748 {
3749         int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3750
3751         BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
3752         BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
3753
3754         if (checkpointed)
3755                 return 1;
3756
3757         ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
3758         return 0;
3759 }
3760
3761 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
3762                                         int new_level)
3763 {
3764         struct inode *inode = ocfs2_lock_res_inode(lockres);
3765
3766         return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
3767 }
3768
3769 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
3770 {
3771         struct inode *inode = ocfs2_lock_res_inode(lockres);
3772
3773         __ocfs2_stuff_meta_lvb(inode);
3774 }
3775
3776 /*
3777  * Does the final reference drop on our dentry lock. Right now this
3778  * happens in the downconvert thread, but we could choose to simplify the
3779  * dlmglue API and push these off to the ocfs2_wq in the future.
3780  */
3781 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
3782                                      struct ocfs2_lock_res *lockres)
3783 {
3784         struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3785         ocfs2_dentry_lock_put(osb, dl);
3786 }
3787
3788 /*
3789  * d_delete() matching dentries before the lock downconvert.
3790  *
3791  * At this point, any process waiting to destroy the
3792  * dentry_lock due to last ref count is stopped by the
3793  * OCFS2_LOCK_QUEUED flag.
3794  *
3795  * We have two potential problems
3796  *
3797  * 1) If we do the last reference drop on our dentry_lock (via dput)
3798  *    we'll wind up in ocfs2_release_dentry_lock(), waiting on
3799  *    the downconvert to finish. Instead we take an elevated
3800  *    reference and push the drop until after we've completed our
3801  *    unblock processing.
3802  *
3803  * 2) There might be another process with a final reference,
3804  *    waiting on us to finish processing. If this is the case, we
3805  *    detect it and exit out - there's no more dentries anyway.
3806  */
3807 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
3808                                        int blocking)
3809 {
3810         struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3811         struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
3812         struct dentry *dentry;
3813         unsigned long flags;
3814         int extra_ref = 0;
3815
3816         /*
3817          * This node is blocking another node from getting a read
3818          * lock. This happens when we've renamed within a
3819          * directory. We've forced the other nodes to d_delete(), but
3820          * we never actually dropped our lock because it's still
3821          * valid. The downconvert code will retain a PR for this node,
3822          * so there's no further work to do.
3823          */
3824         if (blocking == DLM_LOCK_PR)
3825                 return UNBLOCK_CONTINUE;
3826
3827         /*
3828          * Mark this inode as potentially orphaned. The code in
3829          * ocfs2_delete_inode() will figure out whether it actually
3830          * needs to be freed or not.
3831          */
3832         spin_lock(&oi->ip_lock);
3833         oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
3834         spin_unlock(&oi->ip_lock);
3835
3836         /*
3837          * Yuck. We need to make sure however that the check of
3838          * OCFS2_LOCK_FREEING and the extra reference are atomic with
3839          * respect to a reference decrement or the setting of that
3840          * flag.
3841          */
3842         spin_lock_irqsave(&lockres->l_lock, flags);
3843         spin_lock(&dentry_attach_lock);
3844         if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
3845             && dl->dl_count) {
3846                 dl->dl_count++;
3847                 extra_ref = 1;
3848         }
3849         spin_unlock(&dentry_attach_lock);
3850         spin_unlock_irqrestore(&lockres->l_lock, flags);
3851
3852         mlog(0, "extra_ref = %d\n", extra_ref);
3853
3854         /*
3855          * We have a process waiting on us in ocfs2_dentry_iput(),
3856          * which means we can't have any more outstanding
3857          * aliases. There's no need to do any more work.
3858          */
3859         if (!extra_ref)
3860                 return UNBLOCK_CONTINUE;
3861
3862         spin_lock(&dentry_attach_lock);
3863         while (1) {
3864                 dentry = ocfs2_find_local_alias(dl->dl_inode,
3865                                                 dl->dl_parent_blkno, 1);
3866                 if (!dentry)
3867                         break;
3868                 spin_unlock(&dentry_attach_lock);
3869
3870                 if (S_ISDIR(dl->dl_inode->i_mode))
3871                         shrink_dcache_parent(dentry);
3872
3873                 mlog(0, "d_delete(%pd);\n", dentry);
3874
3875                 /*
3876                  * The following dcache calls may do an
3877                  * iput(). Normally we don't want that from the
3878                  * downconverting thread, but in this case it's ok
3879                  * because the requesting node already has an
3880                  * exclusive lock on the inode, so it can't be queued
3881                  * for a downconvert.
3882                  */
3883                 d_delete(dentry);
3884                 dput(dentry);
3885
3886                 spin_lock(&dentry_attach_lock);
3887         }
3888         spin_unlock(&dentry_attach_lock);
3889
3890         /*
3891          * If we are the last holder of this dentry lock, there is no
3892          * reason to downconvert so skip straight to the unlock.
3893          */
3894         if (dl->dl_count == 1)
3895                 return UNBLOCK_STOP_POST;
3896
3897         return UNBLOCK_CONTINUE_POST;
3898 }
3899
3900 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
3901                                             int new_level)
3902 {
3903         struct ocfs2_refcount_tree *tree =
3904                                 ocfs2_lock_res_refcount_tree(lockres);
3905
3906         return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
3907 }
3908
3909 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
3910                                          int blocking)
3911 {
3912         struct ocfs2_refcount_tree *tree =
3913                                 ocfs2_lock_res_refcount_tree(lockres);
3914
3915         ocfs2_metadata_cache_purge(&tree->rf_ci);
3916
3917         return UNBLOCK_CONTINUE;
3918 }
3919
3920 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
3921 {
3922         struct ocfs2_qinfo_lvb *lvb;
3923         struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
3924         struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3925                                             oinfo->dqi_gi.dqi_type);
3926
3927         lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3928         lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
3929         lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
3930         lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
3931         lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
3932         lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
3933         lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
3934         lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
3935 }
3936
3937 void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3938 {
3939         struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3940         struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3941         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3942
3943         if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
3944                 ocfs2_cluster_unlock(osb, lockres, level);
3945 }
3946
3947 static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
3948 {
3949         struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3950                                             oinfo->dqi_gi.dqi_type);
3951         struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3952         struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3953         struct buffer_head *bh = NULL;
3954         struct ocfs2_global_disk_dqinfo *gdinfo;
3955         int status = 0;
3956
3957         if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
3958             lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
3959                 info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
3960                 info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
3961                 oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
3962                 oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
3963                 oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
3964                 oinfo->dqi_gi.dqi_free_entry =
3965                                         be32_to_cpu(lvb->lvb_free_entry);
3966         } else {
3967                 status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode,
3968                                                      oinfo->dqi_giblk, &bh);
3969                 if (status) {
3970                         mlog_errno(status);
3971                         goto bail;
3972                 }
3973                 gdinfo = (struct ocfs2_global_disk_dqinfo *)
3974                                         (bh->b_data + OCFS2_GLOBAL_INFO_OFF);
3975                 info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
3976                 info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
3977                 oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
3978                 oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
3979                 oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
3980                 oinfo->dqi_gi.dqi_free_entry =
3981                                         le32_to_cpu(gdinfo->dqi_free_entry);
3982                 brelse(bh);
3983                 ocfs2_track_lock_refresh(lockres);
3984         }
3985
3986 bail:
3987         return status;
3988 }
3989
3990 /* Lock quota info, this function expects at least shared lock on the quota file
3991  * so that we can safely refresh quota info from disk. */
3992 int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3993 {
3994         struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3995         struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3996         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3997         int status = 0;
3998
3999         /* On RO devices, locking really isn't needed... */
4000         if (ocfs2_is_hard_readonly(osb)) {
4001                 if (ex)
4002                         status = -EROFS;
4003                 goto bail;
4004         }
4005         if (ocfs2_mount_local(osb))
4006                 goto bail;
4007
4008         status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
4009         if (status < 0) {
4010                 mlog_errno(status);
4011                 goto bail;
4012         }
4013         if (!ocfs2_should_refresh_lock_res(lockres))
4014                 goto bail;
4015         /* OK, we have the lock but we need to refresh the quota info */
4016         status = ocfs2_refresh_qinfo(oinfo);
4017         if (status)
4018                 ocfs2_qinfo_unlock(oinfo, ex);
4019         ocfs2_complete_lock_res_refresh(lockres, status);
4020 bail:
4021         return status;
4022 }
4023
4024 int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
4025 {
4026         int status;
4027         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4028         struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
4029         struct ocfs2_super *osb = lockres->l_priv;
4030
4031
4032         if (ocfs2_is_hard_readonly(osb))
4033                 return -EROFS;
4034
4035         if (ocfs2_mount_local(osb))
4036                 return 0;
4037
4038         status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
4039         if (status < 0)
4040                 mlog_errno(status);
4041
4042         return status;
4043 }
4044
4045 void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
4046 {
4047         int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4048         struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
4049         struct ocfs2_super *osb = lockres->l_priv;
4050
4051         if (!ocfs2_mount_local(osb))
4052                 ocfs2_cluster_unlock(osb, lockres, level);
4053 }
4054
4055 static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
4056                                        struct ocfs2_lock_res *lockres)
4057 {
4058         int status;
4059         struct ocfs2_unblock_ctl ctl = {0, 0,};
4060         unsigned long flags;
4061
4062         /* Our reference to the lockres in this function can be
4063          * considered valid until we remove the OCFS2_LOCK_QUEUED
4064          * flag. */
4065
4066         BUG_ON(!lockres);
4067         BUG_ON(!lockres->l_ops);
4068
4069         mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name);
4070
4071         /* Detect whether a lock has been marked as going away while
4072          * the downconvert thread was processing other things. A lock can
4073          * still be marked with OCFS2_LOCK_FREEING after this check,
4074          * but short circuiting here will still save us some
4075          * performance. */
4076         spin_lock_irqsave(&lockres->l_lock, flags);
4077         if (lockres->l_flags & OCFS2_LOCK_FREEING)
4078                 goto unqueue;
4079         spin_unlock_irqrestore(&lockres->l_lock, flags);
4080
4081         status = ocfs2_unblock_lock(osb, lockres, &ctl);
4082         if (status < 0)
4083                 mlog_errno(status);
4084
4085         spin_lock_irqsave(&lockres->l_lock, flags);
4086 unqueue:
4087         if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
4088                 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
4089         } else
4090                 ocfs2_schedule_blocked_lock(osb, lockres);
4091
4092         mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name,
4093              ctl.requeue ? "yes" : "no");
4094         spin_unlock_irqrestore(&lockres->l_lock, flags);
4095
4096         if (ctl.unblock_action != UNBLOCK_CONTINUE
4097             && lockres->l_ops->post_unlock)
4098                 lockres->l_ops->post_unlock(osb, lockres);
4099 }
4100
4101 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
4102                                         struct ocfs2_lock_res *lockres)
4103 {
4104         unsigned long flags;
4105
4106         assert_spin_locked(&lockres->l_lock);
4107
4108         if (lockres->l_flags & OCFS2_LOCK_FREEING) {
4109                 /* Do not schedule a lock for downconvert when it's on
4110                  * the way to destruction - any nodes wanting access
4111                  * to the resource will get it soon. */
4112                 mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n",
4113                      lockres->l_name, lockres->l_flags);
4114                 return;
4115         }
4116
4117         lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
4118
4119         spin_lock_irqsave(&osb->dc_task_lock, flags);
4120         if (list_empty(&lockres->l_blocked_list)) {
4121                 list_add_tail(&lockres->l_blocked_list,
4122                               &osb->blocked_lock_list);
4123                 osb->blocked_lock_count++;
4124         }
4125         spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4126 }
4127
4128 static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
4129 {
4130         unsigned long processed;
4131         unsigned long flags;
4132         struct ocfs2_lock_res *lockres;
4133
4134         spin_lock_irqsave(&osb->dc_task_lock, flags);
4135         /* grab this early so we know to try again if a state change and
4136          * wake happens part-way through our work  */
4137         osb->dc_work_sequence = osb->dc_wake_sequence;
4138
4139         processed = osb->blocked_lock_count;
4140         /*
4141          * blocked lock processing in this loop might call iput which can
4142          * remove items off osb->blocked_lock_list. Downconvert up to
4143          * 'processed' number of locks, but stop short if we had some
4144          * removed in ocfs2_mark_lockres_freeing when downconverting.
4145          */
4146         while (processed && !list_empty(&osb->blocked_lock_list)) {
4147                 lockres = list_entry(osb->blocked_lock_list.next,
4148                                      struct ocfs2_lock_res, l_blocked_list);
4149                 list_del_init(&lockres->l_blocked_list);
4150                 osb->blocked_lock_count--;
4151                 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4152
4153                 BUG_ON(!processed);
4154                 processed--;
4155
4156                 ocfs2_process_blocked_lock(osb, lockres);
4157
4158                 spin_lock_irqsave(&osb->dc_task_lock, flags);
4159         }
4160         spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4161 }
4162
4163 static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
4164 {
4165         int empty = 0;
4166         unsigned long flags;
4167
4168         spin_lock_irqsave(&osb->dc_task_lock, flags);
4169         if (list_empty(&osb->blocked_lock_list))
4170                 empty = 1;
4171
4172         spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4173         return empty;
4174 }
4175
4176 static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
4177 {
4178         int should_wake = 0;
4179         unsigned long flags;
4180
4181         spin_lock_irqsave(&osb->dc_task_lock, flags);
4182         if (osb->dc_work_sequence != osb->dc_wake_sequence)
4183                 should_wake = 1;
4184         spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4185
4186         return should_wake;
4187 }
4188
4189 static int ocfs2_downconvert_thread(void *arg)
4190 {
4191         int status = 0;
4192         struct ocfs2_super *osb = arg;
4193
4194         /* only quit once we've been asked to stop and there is no more
4195          * work available */
4196         while (!(kthread_should_stop() &&
4197                 ocfs2_downconvert_thread_lists_empty(osb))) {
4198
4199                 wait_event_interruptible(osb->dc_event,
4200                                          ocfs2_downconvert_thread_should_wake(osb) ||
4201                                          kthread_should_stop());
4202
4203                 mlog(0, "downconvert_thread: awoken\n");
4204
4205                 ocfs2_downconvert_thread_do_work(osb);
4206         }
4207
4208         osb->dc_task = NULL;
4209         return status;
4210 }
4211
4212 void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4213 {
4214         unsigned long flags;
4215
4216         spin_lock_irqsave(&osb->dc_task_lock, flags);
4217         /* make sure the voting thread gets a swipe at whatever changes
4218          * the caller may have made to the voting state */
4219         osb->dc_wake_sequence++;
4220         spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4221         wake_up(&osb->dc_event);
4222 }